Thèse en co-tutelle. présentée en vue de l'obtention du titre de

Transcription

1 Thèse en co-tutelle présentée en vue de l'obtention du titre de Docteur de l'université Paul Sabatier Toulouse III (France) et De Pusan National University (Korea) Discipline : ECOLOGIE & EVOLUTION Spécialité : HYDROBIOLOGIE Ecological quality assessment of stream ecosystems using benthic macroinvertebrates par Mi-Young SONG Soutenu le 30 october 2007, devant le jury composé de M. KEI TOKITA University Osaka, Japan M. TAE-SOO CHON University Pusan, Korea M. GAE-JAE JOO University Pusan, Korea M. SOVAN LEK University Toulouse III, France M. YOUNG-SEUK PARK University Kyung Hee, Korea Mm. SITHAN LEK-ANG University Toulouse III, France

2 CONTENTS LIST OF FIGURES.. i LIST OF TABLES.. vi GENERAL INTRODUCTION I : CHARACTERIZATION OF BENTHIC MACROINVERTEBRATE COMMUNITIES IN A RESTORED STREAM BY USING SELF-ORGANIZING MAP INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION. 22 II : AGRICULTURAL LAND USE EFFECTS ON MACROINVERTEBRATES IN STREAMS OF THE GARONNE RIVER CATCHMENT (SW FRANCE) INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION 37 III : SELF-ORGANIZING MAPPING OF BENTHIC MACROINVERTEBRATE COMMUNITIES IMPLEMENTED TO COMMUNITY ASSESSMENT AND WATER QUALITY EVALUATION INTRODUCTION

3 MATERIALS AND METHODS RESULTS DISCUSSION. 52 IV : COMMUNITY PATTERNS OF BENTHIC MACROINVERTEBRATES COLLECTED ON THE NATIONAL SCALE IN KOREA INTRODUCTION MATERIALS AND METHODS RESULTS DISCUSSION. 67 GENERAL CONCLUSION.. 70 REFERANCES.. 73 APPENDICES SUMMARY (in France). 94 SUMMARY (in English) ACKNOWLEDGEMENTS... 98

4 LIST OF FIGURES Fig The sample sites located in the Yangjae Stream in the Han River, Korea. Fig Classification of the sample sites by SOM based on environmental variables measured in the Yangjae Stream in a small scale from April 1996 to March 1998: a) sample sites, and b) cluster analysis with the Ward s linkage method. Acronyms in units stand for the samples: the letter represents the name of sample sites (see Fig. 1), while the numbers indicate the year and month of collection (e.g., U9604; site U collected in April 1996, A9711; site A collected in November 1997; and F9803; site F collected in March 1998). Fig Mean abundance of the selected taxa in different clusters corresponding to SOM (Fig. 2). The height of the bar represents the mean, while the whisker indicates the confidential interval (mean±0.95). The same letters located on top of the whisker indicate no significant difference (p>0.05) between clusters based on the nonparametric Kruskall Wallis test with the unequal number of samples. Among the sampled communities, the taxa with statistical differences were only presented in the figure. Fig Patterning of temporal changes in macroinvertebrate communities collected at site A from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m2); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters. Fig Patterning of temporal changes in macroinvertebrate communities collected at site H from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian i

5 distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m2); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters (Fig. 2). Fig Patterning of temporal changes in macroinvertebrate communities collected at site E from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m2); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters. Fig Comparison of substrate volumes in different particle size: a) proportion (%) of substrates among different sample sites and b) nonparametric comparison of volumes between different sample sites in different substrate sizes. Fig Location of the sampling sites in tributary streams of the Garonne River catchment, France. Black circles represent the different sampling sites. Fig Classification of sampling sites on the self-organizing map (SOM). a) The patterned SOM map showing the classification of sample sites according to their macroinvertebrates composition. b) Hierarchical classification of the cells of SOM map. c) Mean and SE of EPTC species richness in different clusters defined in the SOM. The Mann-Whitney test was significant for all three clusters (p<0.001). d) Comparison of the Shannon-Weaver diversity index (mean and SE) in different clusters defined with SOM (Fig. 2). The same letters indicate no significant difference. Fig Distribution patterns of some species on the SOM. Scale bars indicate the weight vector of each species in corresponding SOM units. ii

6 Fig Box plots showing occurrence probability (%) of each species in different clusters. Values were obtained from weight vectors of the trained SOM. median, 25-75%, non-outlier range. Different colors are used simply to aid the reader in the identification of different species across the 3 clusters. Fig Environmental characteristics of different clusters. Error bars indicate standard error. The same letters indicate no significant difference. Fig Location of the sample sites. a) Korean Peninsula, b) the Suyong River and c) the Nakdong River. Fig The map trained by SOM for pattering benthic macroinvertebrates reported from different streams in South Korea from 1984 to 2000: (a) sample sites; (b) mean and S.E. of EPT richness in different clusters defined in the SOM (n = 31 (I), n = 21 (II), n = 63 (III), n = 64 (IV)); (c) mean and S.E. of BMWP scores in different clusters defined in the SOM (n = 31 (I), n=21 (II), n = 63 (III), n = 64 (IV)). The different alphabets indicate significant difference in the Mann-Whitney test (p < 0.001). Fig Profile of abundance of the prevalent taxa matched to clusters based on the trained SOM. The values in the vertical bar indicate densities (individuals/square meter). Fig Monitoring of benthic macroinvertebrate communities collected at YCK in the Suyong stream from November 1992 to April 1995 according to the trained SOM (the sample was not collected in December 1994). (a) Recognition of the samples: November 1992.November 1993 (dots); January 1994.January 1995 (solid), and (b) mean and S.E. of biological and physicochemical indices in different clusters defined in the SOM (n = 4 (I), n = 12 (III), n = 12 (IV)). The different alphabets indicate significant difference in the Mann-Whitney test (p < 0.001). Fig (a) Monitoring of benthic macroinvertebrate communities collected at THP in the Suyong River from March 1992 to April 1995 according to the trained SOM (n = 35). (b) Mean iii

7 and S.E. of biological and physico-chemical indices. Fig Monitoring of benthic macroinvertebrate communities collected at LTER sites according to the trained SOM: (a) recognition of the samples (the number in the sample name indicates the month of collection) and (b) mean and S.E. of biological and physico-chemical indices (n = 2 (I), n = 8 (IV)). Fig Relations between number of individuals in log scale and species richness in the 1970 samples used in this study. Each point indicates each sampling site. (a) All samples, (b i) the samples separately grouped in clusters 1, 2, 3,..., 8, respectively. Fig Log-normal model of community structure by grouping species into abundance categories (octaves, i.e. power of 2). Fig Classification of the samples according to the trained SOM. (a) The SOM units were grouped to eight clusters, (b) the dendrogram according to the Ward linkage method based on Euclidean distance, (c) geographical location of the sampling sites matching to clusters according to the SOM (Fig. 3a). Cluster 6 is not indicated on the geographical map because it did not show any specific geographic area. Fig Community characterization in different clusters according to the SOM (Fig. 3a). (a) Species richness (number of species), (b) abundance (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable). Fig Environmental variables in different clusters according to the SOM (Fig. 3a). (a) Altitude, (b) depth, (c) conductivity, (d) velocity (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable. Conductivity was not available at the samples in cluster 1). iv

8 Fig Variation in biological indices in different clusters according to the SOM (Fig. 3a). (a) EPT richness, (b) EPT abundance, (c) Shannon diversity index, (d) Biological Monitoring Working Party (BMWP) score, (e) evenness (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable). v

9 LIST OF TABLES Table 1-1. Summary of environmental variables in averages (min-max) and list of abundant taxa in different clusters defined by SOM (Fig. 2). Table 1-2. Comparison of SR and BMWP in different clusters defined by SOM (Figs. 4 6). Table 4-1. Correlation coefficients among community parameters and biological indices used in the datasets. Table 4-2. Community parameters, indicator species and environmental descriptions in different clusters. vi

10 GENERAL INTRODUCTION Sustainable management of aquatic ecosystems has been one of the most urgent concerns in environmental issues due to water resource shortage and its contamination. In order to achieve successful management of aquatic ecosystems, the objective assessment of water quality is a prerequisite for execution of appropriate management polices. Such management requires the understanding of how these ecosystems function, and thus how communities are related to the environment (Lek et al. 2005). The assessments made and the predictions forecast are hoped to lead to improvements in the physical and chemical characteristics of freshwater ecosystems. Among biological communities, benthic macroinvertebrates have been widely used for ecological assessment of water quality. Macroinvertebrates are sedentary and have intermediate life span (from months to a few years). Additionally, benthic macroinvertebrates play a key role in food web dynamics, linking producers and top carnivores. The different kinds of species had different levels of tolerance, so the community structures had higher relationship with the disturbance. It is easy to collect and identified. Consequently, macroinvertebrates have been suitable for reflecting ecological water quality (Barbour et al., 1996; Butcher et al., 2003; Davies et al., 2000; Hawkes, 1979; Hellawell, 1986; Resh et al., 1995; Reynoldson et al., 1997; Richard et al., 1997; Rosenberg and Resh, 1993; Wright et al., 1993; Wright et al., 2000). Data for community dynamics are complex and difficult to analyze, since communities consist of many species varying in a non-linear fashion in spatial and temporal domains. There have been numerous accounts of multi-variate statistical analyses regarding characterization of community data in ecology (e.g., Bunn et al., 1986; Legendre and Legendre, 1998; Ludwig and Reynolds, 1988; Quinn et al., 1991). Data for community classification and ordination have 1

11 been available by measuring degree of association among the sampled communities and taxa (Legendre and Legendre, 1998; Ludwig and Reynolds, 1988). However, conventional multivariate methods are generally limited in the sense that they are mainly applicable to linear data and have less flexibility in representing ecological data, for instance, handling noise and data management (Chon et al., 1996; Lek and Guégan, 1999, 2000; Recknagel, 2003). The Self-Organizing Map (SOM) is an efficient tool for mining non-linear data and has been extensively used for patterning community data since 1990s (e.g., Chon et al., 1996, 2000, 2002; Kwak et al., 2000; Levine et al., 1996; Park et al., 2001, 2003a,b, 2004). Chon et al. (1996) classified benthic macroinvertebrate communities in polluted streams with the SOM and elucidated community patterning according to anthropogenic disturbances and locality of the sample sites. The species richness and distribution are the main fact of the ecology research for the theoretical and conversational aims (Krebs, 1994). Along with analysis in species richness, investigation of species abundance patterns has been regarded as an important topic in elucidating patterns of communities responding to the disturbances. Preston s canonical lognormal distribution has been the most widely accepted formalization of the relative commonness and rarity of species (Preston, 1962; Brown, 1981). Regarding that the species are often vulnerable to various environmental disturbances, the existence of rare species is a key issue in community ecology in relation to risk assessment. This type of complex relationships in community changes and environment disturbances would be accordingly addressed by studying community compositions in relation to abundance patterns per taxa, i.e. relations between species richness and abundance. This thesis documents bioassessment using benthic macroinvertebrates in different scale and pollution level in streams by using ecological informatics. Chapter I, we intend to reveal 2

12 changes in macroinvertebrate communities intensively collected within a limited area in a midstream reach of a polluted stream after restoration project. We investigated spatial heterogeneity by selecting the sample sites short distances (5-10m) apart, and correspondingly characterized abundance patterns of benthic macroinvertberates in different habitats. This study, in chapter II, were (i) to assess how agricultural land use disturbs EPTC assemblages at the spatial scale of reach and (ii) to employ community analyses in revealing the effect of agricultural land use. We accordingly selected reference (i.e., streamsides composed of woody and grass vegetation) and agriculture-impacted (i.e., streamsides composed of croplands) sites. Chapter III, we further elaborated to show the trained SOM as a means of providing a comprehensive view on ecological states of the communities and to use the SOM as a map for assessing biological water quality. In Chapter IV, we apply the SOM to mining the large-scale community data and to further relating the community patterns to variation caused by geographic distribution and different degrees of disturbances. 3

13 I. Characterization of benthic macroinvertebrate communities in a restored stream by using self-organizing map Introduction With the advantages of taxonomic diversity, sedentary in behaviors and long life cycles, benthic macroinvertebrates characteristically respond to anthropogenic disturbances in an integrated and continuous manner, and consequently have been widely used for assessing the water quality and ecological status of aquatic systems (Resh and Rosenberg, 1984; Hellawell, 1986; Rosenberg and Resh, 1993). There have been numerous studies on community characterization and water quality evaluation over a broad scope from clean to severely polluted states (Hellawell, 1986; Rosenberg and Resh, 1993; Barbour et al., 1996; Reynoldson et al., 1997; Richards et al., 1997; Davies et al., 2000; Wright et al., 2000; Chon et al., 2002; Butcher et al., 2003). In most cases, the surveys have been carried out in a large scale in the order of km between the sample sites. Species traits are usually distinct and community compositions are easy to characterize between different sites in these cases (Townsend and Hildrew, 1994; Death, 1995; Resh et al., 1994; Poff, 1997; Rabeni et al., 2002; Lamouroux et al., 2004). Spatial heterogeneity, however, still exists in a small scale (e.g., less than one kilometer) in lotic conditions according to hydro-morphological characteristics of streams (e.g., pool, riffle, etc.) (Minshall, 1988; Poff and Ward, 1990; Townsend and Hildrew, 1994; Copper et al., 1997; Palmer et al., 1997; Palmer and Poff, 1997; Poff, 1997; Poole, 2002). Accordingly, the 4

14 community patterns are variable in different habitats in a small scale (Lancaster and Hildrew, 1993; Resh et al., 1994; Townsend and Hildrew, 1994; Armitage and Cannan, 2000; Brown, 2003; Roy et al., 2003; Gebler, 2004), although overall composition of species may not be widely variable as shown in a large scale. Benthic macroinvertebrates have been investigated in a relatively small scale of hundred-meter distances (e.g., Cummins and Lauff, 1969; Hildrew et al., 1980; Lancaster and Hildrew, 1993; Brown, 2003; Roy et al., 2003). There have been a limited number of studies that reveal the relationships between hydraulic variables and distribution of macroinvertebrates at micro-habitats (Lancaster and Hildrew, 1993; Lamouroux et al., 2004; Brooks et al., 2005). Downes et al. (1993) described important hydro-morphological factors influencing the spatial distribution of invertebrates after investigating small scale patchness. Brooks et al. (2005) recently demonstrated the importance of small-scale differences in hydraulic conditions characterized by water velocity, depth and substrate roughness in determining the spatial distribution of macroinvertebrate assemblages in riffle habitats. The characteristics of small scale habitats are important factors for the success of stream restoration activities. Monitoring in a small scale habitat, heterogeneity provides a measure of ecosystem restoration (Pik et al., 2002). For example, Purcell et al. (2002) evaluated the effects of restoration of a small stream using benthic macroinvertebrate communities. In this study, we intend to reveal changes in macroinvertebrate communities intensively collected within a limited area in a midstream reach of a polluted stream after a restoration project. We investigated spatial heterogeneity by selecting the sample sites short distances (5 10 m) apart, and correspondingly characterized abundance patterns of benthic macroinvertebrates in different habitats. 5

15 Materials and methods Study sites The field survey was carried out within a 200 m reach (Hakyeoul) in the Yangjae Stream, a tributary of the Han River in the south of Seoul, Korea ( N, E) (Fig. 1-1). It flows through the metropolitan and agricultural areas in the city, and has been mainly polluted with organic matter (KICT, 1997). The stream has a year-round flow of m in width and cm in depth. Water discharge rapidly increases in the period of summer flooding and decreases in the dry winter season. Recently, a campaign for water recovery has been carried out by the local government. Benthic macroinvertebrates were sampled at 11 sampling sites in the study area based on location, hydromorphological characters and a 5 10 m distance between sites (Fig. 1-1). According to topographical conditions the sample area was partitioned into three zones: 1) the upstream zone covered the sample sites located in the upper part of the sample area, and was characterized by large substrates and relatively high water velocity (sites, U, A, B, C and D), 2) the downstream zone included the sample sites with lower velocity and higher sedimentation at the lower part of the sample area (sites H, I, and J) (Fig. 1-1), and 3) the pool zone was located at the curved down stream area and were characterized with low values in water depth and velocity, and high levels of sedimentation (sites E, F, and G). Except sites U and J which were located at the upper part and the lower part of the study sites, respectively, longitudinal or curvilinear sampling was carried out to reveal local topographic characters of the sample sites. 6

16 Fig The sample sites located in the Yangjae Stream in the Han River, Korea. The survey area has been partially manipulated for a restoration project of the stream. Vegetation channel revetment technique was carried out in the riparian zone close to site A (Fig. 1-1) (KICT, 1997). The stream bed was artificially planted with large-sized cobbles approximately 5 cm in diameter. The upstream islet, near where site F was located, was artificially constructed for management of siltation as well as for providing habitats for birds and other animals, while the downstream islet, around where G was sampled, was naturally formed. Restoration was also carried out on the riparian zone close to site D, but the stream bed 7

17 was not affected by the restoration project in this case. Site E was located close to the edge area at the pool zone, and silts were highly accumulated around this site. We selected four environmental variables to reveal spatial heterogeneity in the sample area. Firstly, the water depth and velocity were used to represent hydromorphological characters of the sample sites. Secondly, considering that topographic conditions of streams affect stream beds, substrate roughness was recorded to represent spatial heterogeneity of the sample sites (Statzner et al., 1988; Poff and Ward, 1990). Lastly, the percentage of silt was measured to represent a fine level of substrate compositions, according to Statzner et al. (1988). The substrate composition in each site was measured in different diameters (D): coarse cobbles (mean D sizes 100 mm), fine cobbles (50mm D<100mm), pebbles (30mm D<50mm), fine pebbles (16 mm D<30mm), coarse gravel (8 mm D<16mm), and the smaller substrates (4 mm D<8 mm, 2 mm D<4 mm, 1 mm D<2 mm, 0.5 mm D<1 mm, 0.25 mm D<0.5 mm, 0.125mm D<0.25mmand D<0.125mm) (Cummins and Lauff, 1969). The volumes of larger substrates ( 8 mm) were determined by the volumetric bucket in the field, while substrates smaller than 8 mm in diameter were separately sampled in plastic containers (50 ml) in triplications. Substrate roughness was expressed as K=(5C1+3C2+C3) /9, where the subscripts 1, 2 and 3 represent the 1st, 2nd and 3rdmost dominant substratum type, respectively (Statzner et al., 1988). Coarseness value C was correspondingly assigned to the size of the dominant substrates: 1, 2, 3 and 4 if the size, k, is in the range of k<0.125mm, 0.5 mm k<4 mm, 8 mm k<30 mm, k 30 mm, respectively. The coarseness classes have been slightly modified from those given by Statzner et al. (1988) to the size used in this study. Community data At each sample site, three to four benthic samples were collected with the Surber sampler 8

18 (30 cm 30 cm, 500 µm mesh; APHA et al., 1985) approximately 10 cm in depth at monthly intervals for two years starting in April The collected macroinvertebrates were preserved in 7% Formalin solution. In the laboratory, the invertebrate specimens were sorted, identified to genus level and counted for the number of specimens under microscopes. Identification was based on Yun (1988), Brighnam et al. (1982), Merritt and Cummins (1984), Pennak (1978), and Quigley (1977). Chironomidae was separately identified based on Wiederholm (1983), while Oligochaeta was checked with Brinkhurst and Jamieson (1971) and Brinkhust (1986). In the datasets, 24 genera were identified, showing that only a few taxa were highly abundant at the polluted sample sites. Chironomidae, abundant with Chironumus sp., and Oligochaeta, mostly consisting of Limnodrilus hoffmeisteri (Tubificidae), were the dominant taxa. In order to represent water quality of the sample sites, species richness (SR) and BMWP (Biological Monitoring Working Party, Walley and Hawkes, 1997), two conventionally used biotic indices, were estimated from the sampled community data. Modeling procedure First we defined hydro-morphological patterns of the sample sites based on four environmental variables (depth, velocity, substrate roughness and the percentage of silt) by the learning process of SOM. Subsequently, we further revealed temporal changes in macroinvertebrate communities at the selected sites by SOM. Both environmental variables and community data were scaled between 0 and 1 in the range of the minimum and maximum values within each variable. In order to reduce high variation, abundance data used for training with SOM was log-transformed before the analysis of each taxon. SOM is an adaptive unsupervised learning algorithm and approximates the probability density function of the input data (Kohonen, 2001). SOM consists of input and output layers connected with computational 9

19 weights (connection intensities). The array of input nodes (i.e., computational units) operates as a flow-through layer for the input vectors, whereas the output layer consists of a twodimensional network of nodes arranged in a hexagonal lattice. In the learning process of SOM, initially the input data (data matrix for either environmental variables or taxa abundance in this study) were subjected to the network. Each raw input vector consisting of the values for different environmental variables (or abundance data in different taxa) was provided sequentially as input data. In this case the number of the input node was equal to the number of variables (or number of taxa), while the output layer consisted of N output nodes (i.e., computational units) which usually constitute a 2D grid for better visualization. Subsequently, the weights of the network were trained for a given dataset. Weights were initially generated as small random numbers. Each node of the output layer computes the summed distance between weight vector and input vector. The output nodes are considered as virtual units to represent typical patterns of the input dataset assigned to their units after the learning process. Among all virtual units, the best matching unit (BMU), which has the minimum distance between weight and input vectors, becomes the winner. For the BMU and its neighborhood units, the new weight vectors are updated by the SOM learning rule. This results in training the network to classify the input vectors by the weight vectors they are closest to. The detailed algorithm of SOM for ecological applications can be found in Chon et al. (1996) and Park et al. (2003a). After training, the Ward s linkage method based on the Euclidian distance (Ward, 1963) was applied to the weights of the nodes in SOM for further clustering (Jain and Dubes, 1988; Park et al., 2003b). After preliminary training, we used N=80 (10 8) of SOM output units for patterning samples with environmental data, and N=20 (5 4) units for patterning temporal variation of community at the selected sites. We used the functions provided in the SOM toolbox (Alhoniemi et al., 2000) in Matlab (The Mathworks, 2001). 10

20 Statistical analysis To test the null hypotheses of no significant differences in environmental variables, taxa abundance and biotic indices in different clusters, we carried out nonparametric multiple comparisons after the Kruskall Wallis test with the unequal number of samples (Zar, 1999), considering wide variations in community data and the different number of the samples in clusters patterned by SOM. Results Classification of sample sites When the sample sites were trained with SOM (80=10 8 nodes) based on environmental variables, two large clusters (I and II) were formed according to the dendrogram of Ward s linkage method (Fig. 1-2a, b). SOM output units were further subclustered into four groups (Ia, Ib, IIa, and IIb). At the highest level, SOM units were vertically divided: cluster I for the upper area and cluster II for the bottom area in the map (Fig. 1-2a). This grouping coincided with the locations of the sample sites. In cluster I, the samples were collected in both pool and downstream zones, including sites E to J (Fig. 1-1), where water velocity was relatively slower and small size substrates were more abundantly present. Additionally, some samples collected at sites C and D were grouped in cluster I. In contrast to cluster I, cluster II accommodated the sample sites (U, A, B, etc.) belonging to the upstream zone with the large size substrates where water velocity was higher. Subclustering was further obtained based on spatial variation. Within cluster I, cluster Ia was mainly represented by the pool zone such as sites E, F and G, while cluster Ib was more 11

21 associated with the sample sites located at the downstream zone, such as H and J (Fig. 1-2a). Within cluster II, subclustering was formed based on the location of the sample sites and environmental factors. Samples located in the upstream zone were mainly grouped in cluster IIa (e.g., sample sites U, A, B, etc.), while cluster IIb accommodated various sample sites with high levels of water depth and velocity. A majority of the samples collected in summer were grouped in cluster IIb, and the flooding effect appeared more clearly at the sites belonging to the downstream zone. Environmental variables varied in different clusters of the sample sites (Table 1-1). Cluster II was characterized by a higher velocity, higher substrate roughness and lower percentage of silt, and vice versa for cluster I. Within cluster II, cluster IIa was differentiated from cluster IIb by higher levels of substrate roughness. Cluster Ia was mainly different from cluster Ib by a higher percentage of silt. Overall, environmental factors in different clusters revealed heterogeneity of the sample sites. The percentage of silt was higher at the pool zone of sites E and G, whereas lower at the upstream zone of sites U and A. Environmental variables in each cluster were made distinct by using the nonparametric Kruskall Wallis test (Table 1-1). Differences in depth, velocity and substrate roughness were statistically significant between all clusters. The variables were generally higher in cluster II than in cluster I. Depth and velocity were in the lowest range in cluster Ia, while in the highest range in cluster IIb. Substrate roughness was lower in cluster I and higher in cluster II, showing the lowest value in cluster Ib and the highest value in cluster IIa. The percentage of silt was the same level between clusters IIa and IIb, while the percentage of silt showed the highest value in cluster Ia. The statistical results revealed differences of environmental variables in different clusters. 12

22 Fig Classification of the sample sites by SOM based on environmental variables measured in the Yangjae Stream in a small scale from April 1996 to March 1998: a) sample sites, and b) cluster analysis with the Ward s linkage method. Acronyms in units stand for the samples: the letter represents the name of sample sites (see Fig. 1-1), while the numbers indicate the year and month of collection (e.g., U9604; site U collected in April 1996, A9711; site A collected in November 1997; and F9803; site F collected in March 1998). Abundances of the selected taxa in benthic macroinvertebrates varied with different clusters (Fig. 1-3). A majority of taxa were abundant in subcluster IIa. Orthocladius sp. and Cricotopus sp., which are known to be present in recovering water (Ferringto and Crisp, 1989), were highly collected in subcluster IIa. Additionally, Tanypus sp. and other invertebrates including Erpobdella sp. and Glossiphora sp. showed the highest level of abundance in this subcluster. However, the tolerant species such as Chironomus sp. and L. hoffmeisteri showed different patterns, being most abundant in cluster Ia (Table 1-1, Fig. 1-3). 13

23 Fig Mean abundance of the selected taxa in different clusters corresponding to SOM (Fig. 1-2). The height of the bar represents the mean, while the whisker indicates the confidential interval (mean±0.95). The same letters located on top of the whisker indicate no significant difference (p>0.05) between clusters based on the nonparametric Kruskall Wallis test with the unequal number of samples. Among the sampled communities, the taxa with statistical differences were only presented in the figure. Overall, environmental variables in different clusters were associated with different patterns of taxa abundance. Cluster IIa was characterized with the highest value of substrate roughness and the lowest value of the percentage of silt (Table 1-1). Abundance of various taxa was associated with this subcluster including Chironomus sp., Orthocladius sp., Cricotopus sp., Tanypus sp., Erpobdella sp., Glossiphora sp., and Physa sp. In comparison with cluster IIa, cluster IIb was represented by the highest level of water depth and velocity due to high precipitation in summer. In this subcluster, Tanypus sp., L. hoffmeisteri and Glossiphora sp. were abundantly present. Subcluster Ia was further characterized with the lowest level of water depth and velocity, 14

24 and with the highest level of percentage of silt (Table 1-1). Subcluster Ia covered the sites E, F and G at the pool zone of the survey area (Fig. 1-2). The tolerant taxa L. hoffmeisteri and Chironomus sp. were most abundant in this subcluster. L. hoffmeisteri and Chironomus sp. belong to Tubificidae and Chironomidae, respectively, and both taxa are burrowing or tube building types, being commonly associated with soft, depositing area (Hellawell, 1986). In this subcluster, the percent of silt was at the highest level (Table 1-1). Cluster Ia was additionally associated with Cricotopus sp., Glossiphora sp. etc. In comparison with subcluster Ia, subcluster Ib was relatively higher in water depth and velocity, and was relatively lower in the amount of silt. This indicated a smaller effect of sedimentation on the sample sites (e.g., H and I) in the downstream zone. In this subcluster, the associated taxa were not as diverse as shown in cluster Ia. Orthocladius sp., Erpobdella sp. and Glossiphora sp. were each abundant in cluster Ib. Table 1-1. Summary of environmental variables in averages (min-max) and list of abundant taxa in different clusters defined by SOM (Fig. 1-2). The same alphabets listed in superscript of environmental variables indicate no significant difference (p > 0.05) between clusters based on the nonparametric Kruskall-Wallis test with unequal number of samples. 15

25 Temporal variation of communities After patterning the samples based on the four environmental variables, we further chose the sample sites E, H and A, which would typically represent different clusters Ia, Ib, and IIa based on the differences of environmental factors, respectively (Fig. 1-2a), to reveal temporal variation of communities at different habitat types. The site representing cluster IIb was not chosen since the cluster was mixed with various sample sites, and a majority of the samples were observed in the temporally unstable period of summer in this subcluster (Fig. 1-2a). Community abundance data were subsequently trained with SOM (Figs. 1-4~1-6) for each site. Temporal patterns of macroinvertebrates were identified in the map. At site A representing cluster IIa, the sample sites were further divided to two main clusters with each cluster producing two more subclusters (Fig. 1-4a, b). Clusters A1 and A2 were separated according to different sampling periods. Subclusters showed further temporal changes sequentially from A1a (April 1996 January 1997), A1b (February May 1997), A2a (June August 1997) to A2b (September November 1997) in the order of sampling time (Fig. 1-4a, b). In the last phase, however, the samples collected in this period belonged again to A2a (January March 1998) after A2b (September November 1997). Community abundance patterns were correspondingly different in various subclusters. Different taxa appeared in a diverse manner as time progressed (Fig. 1-4c). In subcluster A1a in the earliest period, abundant taxa were not observed. The following subcluster A1b, mainly observed in March 1997, was selectively matched to Enchytraeus sp. Both tolerant genus (e.g., Chironomus sp.) and intolerant genus (e.g., Orthocladius sp., Cricotopus sp.) were also commonly grouped in this cluster. Two genera, however, were further associated with subcluster A2a in the next phase. The appearance of recovering species such as Orthocladius sp. and Cricotopus sp., indicated the recovery of water quality in the periods corresponding to 16

26 A1b A2a (February 1997 March 1998). Fig Patterning of temporal changes in macroinvertebrate communities collected at site A from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters. A large number of taxa were additionally present in the following subcluster A2a. Especially, Baetis, a well known genus appearing in recovering water (Hellawell, 1986), was observed in this subcluster. The appearance of Baetis sp. confirmed the phase of water recovery in the period matching to cluster A2a (Fig. 1-4c). In the following subcluster A2b, even a larger number of taxa were associated (10 genus) including various aquatic insects and other invertebrates (e.g., Erpobdella sp. and Physa sp.). L. hoffmeisteri, which was the dominant 17

27 species in the survey area, showed different patterns of occurrence compared with other species. The species was highly associated with subcluster A2b, but was also abundant over a broad area of the map (Fig. 1-4c). Biotic indices such as SR and BMWP changed accordingly with time (Fig. 1-4d). The clusters listed below x-axis (month) in Fig. 1-4d indicate the sample communities patterned by SOM (Fig. 1-4a). Subclusters appeared along with changes in biotic indices as time progressed. SR and BMWP increased gradually, peaking in October 1997 (Fig. 1-4d). The changes in indices represented the trend of water quality improvement after the restoration project. Main clusters A1 and A2 were separated according to the sampling time, June The samples grouped in cluster A1weremainly collected in the early sampling period up to May 1997, whereas the samples in cluster A2 were in the later sampling period starting from June Fig. 5 shows temporal changes in macroinvertebrate communities collected at the sample site H, which represents the downstream zone of the sampling area defined in cluster Ib (Fig. 1-2a). Similar to site A, the samples were grouped to two main clusters according to the sampling periods (Fig. 1-5a, b). Cluster H1 mainly represented the earlier periods of the survey, whereas the samples collected at the later periods were grouped in cluster H2. Each cluster was further divided to two subclusters. However, the degree of temporal patterning at the level of subclusters was not as strong as shown in the subclustering at site A (Fig. 1-4). Different taxa were associated with different clusters at site A (Fig. 1-5c). At the early phase in cluster H1, however, not many taxa were abundant. Subcluster H1a was loosely associated with Baetis sp., while subcluster H1b was grouped with 2 species of Ceratopsche sp. and Tubifex tubifex (Fig. 1-5c). 18

28 Fig Patterning of temporal changes in macroinvertebrate communities collected at site H from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters (Fig. 1-2). Subclusters in H2, however, were associated with diverse taxa. Both tolerant (e.g., Chironomus sp.) and intolerant (e.g., Cricotopus sp.) species were grouped in subcluster H2a. Occurrences of Cricotopus sp. along with the tolerant species, Chironomus sp., indicated recovery of water at this phase. Subcluster H2b was also diversely related to various species but were different from species composition observed in cluster H2a. Species in Chironomidae such as Orthocladius sp. and Tanypus sp. were more associated with cluster H2b. Similar to site A, the tolerant species were abundant and showed different patterns in abundance. Chironomus sp. was grouped in subclusters H2a and H2b, while L. hoffmeisteri was broadly present, covering 19

29 H1b and H2b on the map (Fig. 1-5c). Biotic indices also varied according to different clusters at site H (Fig. 1-5d). However, the temporal patterns were not as distinct as shown at site A (Fig. 1-4d). While SR and BMWP were in the ranges of 0 15 and 0 40, respectively at site A, the indices were in narrower ranges, 0 10 and 0 30 at site H, respectively. The samples were observed in different temporal periods at the level of main clusters H1 and H2 (Fig. 1-5d). This type of grouping was revealed at the level of subclusters at site A (Fig. 1-4d). Separation of the sampling time was observed in September 1997 at site H. Fig. 6 shows temporal changes in macroinvertebrate communities collected at site E representing cluster Ia (Fig. 1-2a). Similar to the other sites, the samples were divided into two main clusters with subdivisions (Fig. 1-6a, b). Grouping patterns, however, appeared differently. Most samples were associated with cluster E1b, while only a few samples were sparsely located in other clusters. Cluster E1 was mainly sampled in the early periods of the survey, whereas cluster E2 was collected in the later periods (Fig. 1-6a). Subclustering was mostly based on differences in community abundance. Subclusters in cluster E1 were associated with the selected taxa (Fig. 1-6c). Cluster E1a was related to Tanypus (sp.) and Carabidae (sp.). Although a large number of sample sites were grouped (Fig. 1-6a), cluster E1b was associated with a limited number of taxa such as Lumbriculus (sp.) and Enchytraeus (sp.). Communities collected at the later periods in cluster E2a, however, were diverse, being grouped with Erpobdella sp., Orthocladius sp., Glossiphora sp., etc. Other taxa such as Physa (sp.), Cricotopus (sp.), and Chironomus (sp.) also appeared in cluster E2a, but they were also associated with E2b. Baetis (sp.), however, was randomly present in cluster E2b (Fig. 1-6c). Tolerant species, L. hoffmeisteri and T. tubifex showed the patterns similar to site A. They were broadly abundant over clusters E1b and E2b. Biotic indices were also variable at site E in response to temporal changes (Fig. 1-6d). 20

30 Fig Patterning of temporal changes in macroinvertebrate communities collected at site E from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters. SR and BMWP showed similar patterns as those of site A (Fig. 1-4d). However, ranges of biotic indices were not as wide as at site A. The sample sites were mixed at the level of subclusters, but were differentiated by the sampling periods at the level of the main clusters. Samples in cluster E1 were collected mostly at the early sampling period up to July 1997, whereas samples in cluster E2 were present at the later sampling periods beginning in August At sites A, E, and H, both biotic indices SR and BMWP were significantly different between the early sampling periods and the later sampling periods (Table 1-2), representing 21

31 recovery of water quality at the sampling sites after the restoration project. Table 1-2. Comparison of SR and BMWP in different clusters defined by SOM (Figs. 1-4~1-6) Index Site 1 st cluster (A1, H1, E1) 2 nd cluster (A2, H2, E2) P b SR A 5.0 (1.6) a 9.9 (2.4) H 3.1 (1.3) 5.6 (2.2) E 3.2 (1.3) 6.4 (1.5) BMWP A 8.8 (2.7) 20.0 (8.4) H 4.2 (2.4) 12.2 (6.6) E 4.6 (2.3) 15.0 (6.9) a Mean (SD), b probability based on the nonparametric Kruskall-Wallis test Discussion The different patterns of the sample sites intensively collected in a polluted stream in the small scale were elucidated through SOM. Based on four input variables (depth, velocity, substrate roughness and silt (%)), the clusters were defined in a hierarchical manner depending on the impact of hydromorphological factors. The results confirmed the variation of the community abundance depending upon spatial heterogeneity in the small scale (Table 1-1, Fig. 1-3) (Brown, 2003; Roy et al., 2003; Brooks et al., 2005), and in biotic indices along with different time periods (Figs. 1-4d, 1-5d, 1-6d, Table 1-2). This study demonstrated that improvement of water quality could be differently monitored at the sample sites in a small scale. In general, water quality changes were more clearly addressed at site A than at sites E and H (Figs. 1-4~1-6).Communities varied depending upon the hydro-morphological condition of the 22

32 sample sites even though the locations are about 5 10 m apart. As stated above, large substrates (>30 mm) were artificially planted at site A (Fig. 1-7). At site H, which was adjacent to A, but not planted with large substrates, substrate in smaller size ranging 0.5mm 2mm were accumulated more abundantly (Fig. 1-7). Meanwhile substrate at site E was mainly composed of smaller particles less than 0.5 mm in diameter. Fig Comparison of substrate volumes in different particle size: a) proportion (%) of substrates among different sample sites and b) nonparametric comparison of volumes between different sample sites in different substrate sizes. These differences in substrates were important in characterizing the spatial heterogeneity and were therefore, crucial in determining community abundance patterns at micro-habitats. Considering that large substrates were artificially planted for the restoration project at site A, it could be stated that higher amounts of large-size substrates would accommodate diverse community composition and would reveal sensitive changes in communities, especially in the recovery of water. Consequently, large substrates used in the restoration project would be useful for monitoring community changes, confirming the previous reports that maximizing heterogeneity in ecological restoration projects may promote diverse communities and may be useful for the management of aquatic communities (Brown, 2003). There were a few samples sites unexpectedly mixed with other clusters. The sites C and D belonged to the upstream zone 23

33 (Fig. 1-1), but a majority of the samples at these sites were grouped in cluster Ia for the pool zone along with the sites E, F and G with SOM (Fig. 1-2). Hydro-morphological characters of the sites C and D, obliquely located across the sites A and B (Fig. 1-1), however, were closer to the sites in the pool zone, and were consequently characterized by slower water velocity and the increased amount of silt. Considering that biotic indices may be different according to spatial heterogeneity, consistency in selecting habitats for evaluation of biotic indices may be required. The community patterns appeared differently in each habitat (Table 1-1, Fig. 1-3), and in different temporal periods within the sample sites (Figs. 1-4~1-6). This type of sampling consistency could be further discussed in the future along with the problem of habitat suitability and patchness in spatial distribution. Community development could be observed in regressive succession for recovery (Hawkes, 1979; Sládeèek, 1979; Hellawell, 1986) in the observed data. In recovering water, community compositions become more diverse, and these consequences were observed in community development at the sample sites in the period of water recovery (April 1996 October 1997) (Figs. 1-4d, 1-5d, 1-6d). The genus indicating recovery of water quality such as Orthocladius (sp.), Cricotopus (sp.) and Tanypus (sp.) were collected from February 1997 to March Densities of Baetis (sp.) were high from July to September 1997, while they were not collected in the corresponding periods in 1996.At the phase of higher biotic indices (July November 1997), species in Coleoptera, Ephemeroptera and Odonata were diversely collected at low densities (Figs. 1-4c d, 1-5c d, 1-6c d). The occurrence patterns of recovering species, however, were variable according to the location of the sample sites. The indicator genus such as Baetis (sp.), Orthocladius (sp.), Cricotopus (sp.) and Tanypus (sp.) were collected over a longer period at site A starting from February At site E, the indicator species were also selectively present. Certain species appeared in a longer period. Orthocladius (sp.), for instance, 24

34 was present from February to October 1997 at site E. These differences indicate that studies on community development in recovery should be also related to spatial heterogeneity. Especially in the situation of disturbance with organic pollution, spatial condition would be greatly affected by the changes in composition of substrates. This type of study should be closely checked with the restoration process and water recovery. However, the topic on the restoration project is beyond the scope of the current study and could be further investigated in the future. 25

35 II. Agricultural land use effects on macroinvertebrates in streams of the Garonne River catchment (SW France) Introduction Five interacting categories of human-induced perturbations have been reported to threaten freshwater biodiversity in a broad scope: degradation and destruction of habitats, water pollution, over-exploitation, flow modifications, and invasion of alien species (see review by Dudgeon et al., 2006). In particular, the global transition from undisturbed areas to humandominated landscapes has strongly impacted both physical and biological features of lotic ecosystems (Allan and Flecker, 1993; Harding et al., 1998; Townsend et al., 2003; Allan, 2004). Stream ecosystems are especially vulnerable regarding spatially nested hierarchies residing in the habitats. Larger scale features (e.g., landscape, basin, segment), consequently, constrain smaller habitat units such as reaches, riffle-pool sequences and micro-habitats (Frissel et al., 1986). A great number of studies have been carried out to reveal changes in physical patterns and their influence on biological components of streams in various spatial and temporal scales, especially regarding assessment of land use impact (see review by Allan, 2004). Agricultural land use anthropogenically modifies stream ecosystems by (i) increasing inputs of sediments, nutrients, organic matter and contaminants (i.e.; pesticides), (ii) clearing riparian vegetation and opening canopy, and (iii) altering flows and reducing habitat heterogeneity (Allan, 2004). These effects have been identified as a major cause of loss of biodiversity in agriculture-impacted catchments in streams in various studies (Lenat and Crawford, 1994; Delong and Brusven, 1998; Schulz and Liess, 1999; Hall et al., 2001; Allan, 26

36 2004). The Garonne River basin has a long history of agricultural development (Fortuné, 1988). Forest areas were gradually replaced by agriculture lands in the Garonne River valley along with its main tributaries in the eighteenth and nineteenth centuries (see Chauvet and Descamps, 1989). Due to severe agricultural practices, the areas of the riparian forest surrounding streams were reduced to a few meters or even disappeared, while riparian forest corridors were known to enhance stream biodiversity by the flows of materials and the movement of organisms (Petersen et al., 2004; Baxter et al., 2005). Consequently, community compositions have been substantially affected by agricultural practices being carried out around the Garonne River. Changes in community composition responding to environmental disturbances have been frequently indicated by the group of benthic insects among aquatic organisms since benthic insects have been regarded as a sensitive indicator of long-term environmental changes in water and habitat quality (Johnson et al., 1993). This group has been widely used to assess biotic integrity of streams and rivers. In particular, Ephemeroptera, Trichoptera, and Plecoptera (EPT) taxa are well documented as good biological indicators (Rosenberg and Resh, 1993). Additionally, species in Coleoptera (e.g., Elmidae), have also been reported to indicate water quality (Compin and Céréghino, 2003). Taxa richness in both groups, EPT and Coleoptera, has been used in combination to assess water quality and habitat suitability (e.g., Barbour et al., 1996; Wallace et al., 1996). We used the assemblages of both groups in elucidating the impact of agricultural disturbances in benthic macroinvertebrate communities. Most of the studies dealing with aquatic insect assemblages of the Garonne River basin have been carried out in a larger regional scale (e.g. Céréghino et al., 2001, 2003; Compin and Céréghino, 2003). However, it is necessary to evaluate changes in communities more specifically at the local level to elucidate community-environment relationships caused by 27

37 disturbances. Small scale research has shown that spatial heterogeneity at the local level plays an important role in specifically characterizing community patterns in a polluted stream (Song et al., 2006). The aims of this study were (i) to assess how agricultural land use disturbs EPTC assemblages at the spatial scale of reach and (ii) to employ community analyses in revealing the effect of agricultural land use. We accordingly selected reference (i.e., streamsides composed of woody and grass vegetation) and agriculture-impacted (i.e., streamsides composed of croplands) sites. Materials and Methods Study sites The survey area is located in the Garonne River basin in South France (between the Pyrenees and the Garonne valley near Toulouse) (Fig. 2-1), which is the third of the large French Atlantic Rivers. We investigated three tributaries of the Garonne river that mainly drain through agriculture lands: the Save stream (length (L), 143 km; catchment area (CA), 1150 km²), the Touch stream (L, 60 km; CA, 515 km²) and the Sousson stream (L, 53 km; CA, 120 km²). Four reference sites were selected in forest environment (G, T1, SA1 and R), while the other sample sites (see Fig. 2-1) were surrounded by croplands. All the sites are located between 137 and 361 m in altitude above the sea level. 28

38 Fig Location of the sampling sites in tributary streams of the Garonne River catchment, France. Black circles represent the different sampling sites. Macroinvertebrate sampling Benthic macroinvertebrates were sampled three times in each site from February to July Two samples were collected in each site. One sample was collected in the riffle zone using the Surber sampler (30 cm 30 cm, 200 µm) while the other sampling was carried out at the sediment in the pool zone with a plastic bottle (2 cm 10 cm 20 cm). Samples were preserved in 70% ethanol. Ephemerotpera, Plecoptera, Trichoptera and Coleoptera (EPTC) were identified at the species level. Total EPTC richness and Shannon-Weaver diversity index were determined accordingly for the sampled communities. 29

39 Environmental variables For each site, 8 environmental variables were measured along with the collection of benthic macroinvertebrates. Water samples were additionally collected in the riffle for measurement of Chemical Oxygen Demand (COD), Nitrate (NO3) and Total Dissolved Solid (TDS) according to the Standard Method. Dissolved Oxygen (DO) was additionally obtained using a portable meter on site. Depth, velocity and width were measured at the sampling point of benthic macroinvertebrates. The sites were also classified semi-quantitatively by the degree of riparian forest development: 1 (no riparian forest), 2 (0-10 m width of riparian forest) and 3 (forest area; > 10 m width of riparian forest). Data analysis SOM was utilized to cluster overall communities sampled in the survey area and to visualize environmental factors and abundance patterns of communities corresponding to the clusters. Non-linearity is generally embedded in ecological data resulting from complex interactions between environmental variables and communities (Legendre and Legendre, 1998; Lek and Guégan, 2000). The SOM (Kohonen, 1989, 2001) is an unsupervised neural network and has been implemented in various ecological researches (Chon et al., 1996;, Lek and Guégan, 2000; Park et al., 2001, 2003ab; Recknagel, 2003; Lek et al., 2005). The SOM provides an alternative to traditional ordination and classification methods (Lek and Guégan, 2000). The SOM consists of two layers which are composed of neurons in the form of computational units: input and output layers connected with weight vectors (i.e. connection intensity). When an input vector (abundance of each species) x is sent to the input layer of the network, each neuron k of the network computes the distance between the weight vector w and the input vector x. The output layer consists of D output neurons (i.e., computational units, 24 = 30

40 6 4 in this study), which are arranged into a two dimensional grid. The best arrangement for the output layer is a hexagonal lattice, because it does not favor horizontal and vertical directions as much as the rectangular array (Kohonen, 2001). Among all D output neurons, the best matching unit (BMU), which has minimum distance between weight and input vectors, is the winner. For the BMU and its neighborhood neurons, the weight vectors w are updated by the SOM learning rule. The map size (number of output units) of the SOM is critical for accommodating hierarchical levels in community classification (Park et al., 2004). We trained the SOM with different map sizes, and chose the optimum map size based on low topographic errors (Kohonen, 2001). We looked at whether relevant groups of samples characterized distinct EPTC assemblages by performing a hierarchical cluster analysis (Ward s linkage and Euclidean distance). To do so, we used a new matrix (24 76, output neurons x species) of the connection intensity values (i.e., vector weights) estimated by the SOM. As a preprocessing procedure before SOM application, an abundance of 76 EPTC species were transformed with natural logarithm (i.e., log (x+1)). The transformed values were also rescaled in the range of the minimum and maximum values (i.e., 0-1) before training the model. We used the functions implemented in the SOM toolbox (Alhoniemi et al., 2000) for Matlab (The Mathworks, 2001) developed by the Laboratory of Information and Computer Science at the Helsinki University of Technology. The detailed algorithm of the SOM can be found in Kohonen (1989, 2001) for theoretical considerations, and in Chon et al. (1996) and Park et al. (2003a) for ecological applications. 31

41 Results Characterization of EPTC assemblages Abundance data for EPTC species provided to the SOM and the sample sites were patterned on the 2D lattice map through learning processes (Fig. 2-2a). The samples were mainly grouped according to location of the sample sites. Based on the hierarchical cluster analysis, the samples were classified into three main clusters (Fig. 2-2b). The samples from Géze (G) belonging to the reference sites were grouped in the lower-right areas of the SOM (cluster 3), while the samples (SA1-SA3) in the catchment area from the Save stream were located in the lower-left areas of the SOM (cluster 2). The samples from the other streams (e.g., Touch (T1-T4), Sousson (SO1-SO3), Marguestaud (Ma) and Rieutord (R)) were grouped widely in the upper area of the map (cluster 1). Most of the samples from the Touch and Sousson streams belonged to this group. Community parameters could be accordingly utilized based on the SOM grouping. Species richness and Shannon-Weaver diversity indices were distinctively different among the clusters on the SOM (Fig. 2-2c). The parameters were accordingly higher in clusters 2 and 3. The Mann- Whitney test revealed that Species richness and the Shannon-Weaver diversity index were significantly lower in cluster 1 (Fig. 2-2c,d; p<0.001). This decrease in species richness and diversity suggested that the samples in cluster 1 were affected by disturbances. 32

42 Fig Classification of sampling sites on the self-organizing map (SOM). a) The patterned SOM map showing the classification of sample sites according to their macroinvertebrates composition. b) Hierarchical classification of the cells of SOM map. c) Mean and SE of EPTC species richness in different clusters defined in the SOM. The Mann-Whitney test was significant for all three clusters (p<0.001). d) Comparison of the Shannon-Weaver diversity index (mean and SE) in different clusters defined with SOM (Fig. 2-2). The same letters indicate no significant difference. Figure 2-3 displays the estimated profiles of species projected on the SOM units on a grey scale. Groups of species were characteristically visualized on the SOM. Baetis vernus, Amphinemura sp. and Limnephilus spp. appeared at cluster 1. Cluster 2 was characterized by high values of Baetis fuscatus, Oligoneuriella rhenena and Heptagenia sulphurea. Hydropsyche siltalai, Brachyptera risi and Capnia bifrons appeared in cluster 3 as indicators in the reference sites. Baetis rhodani and Serratella ignita were widely distributed on the map. Caenis luctuosa and Amphinemura standfussi appeared at the right and left area on the map. The SOM clearly visualized different occurrences at the reference and polluted sites. Even the species in the same genus demonstrated contrasting tolerance to agricultural disturbances. Occurrence of Baetis vernus characteristically matched the area of cluster 1, while Baetis fuscatus mainly appeared in 33

43 the area of cluster 2 (Figs. 2-2 and 2-3). According to the SOM, Baetis vernus in the same genus could be identified as an indicator species in areas with agricultural disturbances. Fig Distribution patterns of some species on the SOM. Scale bars indicate the weight vector of each species in corresponding SOM units. Figure 2-4 shows the list of species observed in the order of abundance in cluster 3 of the reference sites. Abundance of the same taxa was correspondingly shown in clusters 2 and 3 in the figure. The gradient in the abundance pattern was accordingly observed in cluster 3 (Fig. 4, Table 1). Baetis rhodani, Elmis spp. and Hydropsyche siltalai were in the range of the highest occurrence in cluster 3, while Caenis luctuosa and Cheumatopsyche lepida were in the range of the lowest occurrence. Abundance patterns appeared differently in clusters 1 and 2 compared with cluster 3. In cluster 1, the overall level of abundance was generally lower than the abundance shown in clusters 2 and 3. Communities in the disturbed area were characteristically collected in cluster 1. 34

44 The species abundantly collected in cluster 2 (e.g., Hydropsyche siltalai and Dupophilus brevis) were not present in cluster 1. Also, a subsequent number of species did not occur in cluster 1 (e.g., Lectura fusica, Brachyptera risi, Heptagenia sulphurea and Baetis fuscatus). Baetis rhodani was most abundant. It is notable that a wide range in abundance was observed in genus Baetis, while the level of abundance of Baetis vernus and Baetis fuscatus were in the minimum range. The gradient shown in cluster 3 was not observed in cluster 2, although some of the abundant species (e.g., Baetis rhotani and Elmis sp.) in cluster 3 also appeared to be abundant in cluster 2. Abundance patterns in cluster 2 were found to be in the intermediate range. While the gradient shown in cluster 3 was not observed, species were more diverse compared with cluster 1. Esolus parallelepipedus, Serratella ignita and Habrophlebia sp. were more abundant, while Baetis vernus, Capnia bifrons and Psychomyiia pusilla were in the range of least abundance in cluster 2. Species occurring in low abundance in cluster 3 were mostly collected in higher abundance in cluster 2. Some selected species abundant in cluster 3 (e.g., Capnia bifrons, Habrophlebia sp. and Brachyptera risi) were collected in the lowest range in cluster 2. Figure 2-4 provides an overall view on the patterns of occurrence in the reference and disturbed sites. Species such as Baetis rhodani and Serratella ignita were commonly observed in three clusters. Elmis spp. and Esolus parallelepipedus, however, were highly distributed in clusters 2 and 3, while they were observed to be in a low-minimal range in cluster 1. In some species, the gradients were observed across the clusters. Hydropsyche siltalai, Dupophilus brevis, and Brachyptera sp. appeared to decrease gradually from cluster 3 of the reference sites to cluster 1 of the disturbed sites. Comparison of community abundance in different clusters showed that communities in the reference sites (cluster 3) were affected by agricultural disturbances (clusters 1 and 2). 35

45 Fig Box plots showing occurrence probability (%) of each species in different clusters. Values were obtained from weight vectors of the trained SOM. median, 25-75%, non-outlier range. Different colors are used simply to aid the reader in the identification of different species across the 3 clusters. Environmental factors were accordingly different in each cluster. All of the environmental variables significantly differed among clusters (Fig. 2-5). Agricultural impact was accordingly revealed in cluster 1. Nitrate (0.01~1.11 meq/l), COD (1.28~10.88 mg/l) and TDS (160.9~ mg/l) were in the highest range in cluster 1, while these variables were in the lowest range in cluster 3 (Nitrate; 0.04~0.19 meq/l, COD; 0.98~2.58 mg/l, TDS; 81.5~121.4 mg/l). Similarly, the percentage of forest showed its highest level in cluster 3, in contrast to cluster 1. Depth (2~45 m), velocity (0~1.7 m/s) and width (0~12.5 m) were in the highest range in cluster 2, while DO (0.84mg/L ~12.84mg/L) was not significantly different among the clusters. 36

46 Fig Environmental characteristics of different clusters. Error bars indicate standard error. The same letters indicate no significant difference. Discussion The impact of agricultural land use on water quality and aquatic communities has been extensively documented (e.g., Lenat and Crawford, 1994; Hall et al., 2001; Genito et al., 2002; Riley et al., 2003; Allan, 2004; Probst et al., 2005). Overall, our results were consistent with results from the previous studies. The disturbed areas due to agricultural practices were accordingly associated with occurrences of the selected tolerable species (i.e., low EPTC species richness and diversity) and the high levels of organic matter and nutrients (Figs. 2-2 ~ 2-5). According to the SOM analysis, some of the agriculture-impacted sites were classified in cluster 2. Cluster 2 was characterized by samples with diverse EPTC species along with levels of environmental factors such as high depth, width and current velocity (Figs. 2-2 ~ 2-5). In particular, Baetis fuscatus, Oligoneuriella rhenana and Heptagenia sulphurea are sensitive species and prefer high current velocity (Moog et al., 1997; Beketov, 2004). By considering 37

47 mixture of the sample sites from the reference sites and some of polluted sites, we speculate that the impact of agricultural land use on these sites was masked by natural hydrologic variability and high flows that make instream habitats more heterogeneous and then more diverse in macroinvertebrates (Allan, 2004). This confirms that the covariation between natural landscape features and anthropogenic factors makes it difficult to assess the biotic integrity of stream ecosystems (e.g., Richards et al., 1997; Fitzpatrick et al., 2001). The longitudinal nature of rivers and streams has long been recognized as a major force in structuring lotic communities (see Hynes, 1970). Our analysis based on the SOM (Fig. 2-2a), however, indicated that the between-sites differences in EPTC composition cannot be related to the upstream-downstream gradient as well as to seasonal variations. Both upstream and downstream sites were classified in the same clusters of the SOM (e.g., the sites of the Touch stream) (Fig. 2-2a). Absence of the longitudinal gradient has also been reported in catchments dominated by agricultural land use (e.g., Delong and Brusven, 1998). These results suggest that homogeneous assemblages of benthic macroinvertebrates are capable of tolerating agricultural non-point sources of pollution without being overly compounded by natural factors such as longitudinal location and season (Allan, 2004). In this study, identification of specimens was carried out at the species level. The specieslevel identification has been recommended to assess small scale community composition between-sites and between-date differences (Lenat and Resh, 2001). Here, community parameters such as diversity index and richness can be strongly under-evaluated at the family and genus level, especially in the non-impacted streams (Guerold, 2000). Identification of aquatic insects at the species level allowed for a clearer characterization of the sample sites since the tolerance levels for environmental factors could be specifically provided for each species. Species data are useful for provision of specific information for the management of 38

48 communities, especially including conservation of rare species (e.g., endemic species). As stated before, the most tolerant species dominated the agriculture-impacted sites (i.e., those belonging to cluster 1), especially Baetis vernus. However, densities and frequency of occurrence were relatively low in this study (Table 1). This tolerant species (Moog et al., 1997; Beketov, 2004) is known to exist mostly at low current habitats (Armitage, 1976). In contrast, Baetis rhodani dominated the reference sites of cluster 3. These sites (e.g., Géze) were also inhabited by the sensitive taxa such as Hydropsyche siltalai and Dupophilus brevis that preferred high current velocity and low temperature (Thomas and Berthélemy, 1991; Stuijfzand et al., 1999; Stubauer and Moog, 2000; Urbanič et al., 2005). We also observed that numerous sensitive species absent in the agriculture-impacted sites were historically common from the Pyrenees to the Garonne River. Specifically, Brachyptera risi and Capnia bifrons (Plecoptera) were only sampled in the references area of cluster 3, whereas they were widely distributed historically in the study area according to previous records (Berthélemy, 1966). Loss of sensitive species in deforested areas due to the alteration of habitats and water quality (Quinn, 2000; Allan, 2004) was also reported in many studies (e.g., Lenat and Crawford, 1994; Collier, 1995; Wang et al., 1997). The reference sites of our study were also characterized by rare EPTC species that were historically widespread. Considerations are required to be given to some of the sample sites in the reference areas. The samples from the reference sites T1 and R were grouped with the agriculture-impacted sites in cluster 1, whereas the samples were localized in the forest area. Indeed, the spring area in the upstream T1 was exposed to agricultural land use along an area 1 km in size, which may explain the high level of organic matter and nutrients in this reference site. This underlines that the whole catchment feature is also important for explaining the structure of local macroinvertebrate communities (Allan et al., 1997; Vondracek et al., 2005). In addition, the site R was characterized by low 39

49 flow conditions and was dominated by small pool habitats during the sampling period compared to the other reference sites. Harsh environmental factors such as stream drying required high tolerance or specific adaptations of aquatic insects (Williams, 1996; Meyer et al., 2003). Regarding the existence of disturbed sites in the reference area (Figs. 1-2a), we recommend that conservation efforts be undertaken to maintain the regional biodiversity of the Garonne River basin. Finally, in this paper we have only dealt with the impact of agricultural land use at the reach scale. Instructive and complementary results could be further obtained by comparing different land use practices in the catchment. This could permit researchers to test whether landscape indicators at multiple spatial scales (i.e., reach, catchment) are integrative and efficient when assessing the impact of anthropogenic activities on the riverine ecosystems (see review by Gergel et al., 2002). 40

50 III. Self-Organizing Mapping of Benthic Macroinvertebrate Communities Implemented to Community Assessment and Water Quality Evaluation Introduction Sustainable management of aquatic ecosystems has been one of the most urgent concerns in environmental issues due to water resource shortage and its contamination. In order to achieve successful management of aquatic ecosystems, the objective assessment of water quality is a prerequisite for execution of appropriate management polices. Biological organisms residing in the habitats in aquatic ecosystems would convey the integrative and continuous characters of water quality (Allan, 1995; Hawkes, 1979; Hellawell, 1986; Rosenberg and Resh, 1993; Sladecek, 1979; Tittizer and Kothe, 1979). Among biological communities, benthic macroinvertebrates have been widely used for ecological assessment of water quality. Macroinvertebrates are sedentary and have intermediate life span (from months to a few years). Additionally, benthic macroinvertebrates play a key role in food web dynamics, linking producers and top carnivores. Consequently, macroinvertebrates have been suitable for reflecting ecological water quality (Barbour et al., 1996; Butcher et al., 2003; Davies et al., 2000; Hawkes, 1979; Hellawell, 1986; Resh et al., 1995; Reynoldson et al., 1997; Richard et al., 1997; Rosenberg and Resh, 1993; Wright et al., 1993; Wright et al., 2000). However, it is a difficult task to develop an objective and quantifiable indicator system from community data. Especially data for benthic macroinvertebrates are complex since the communities consist of multi-variables (i.e., diverse taxa) in a complex manner. There have 41

51 been numerous accounts of multi-variate statistical analyses regarding characterization of community data in ecology (e.g., Bunn et al., 1986; Legendre and Legendre, 1998; Ludwig and Reynolds, 1988; Quinn et al., 1991). Community classification and information on component factors have been available by measuring degree of association among the sampled communities and taxa (Legendre and Legendre, 1998; Ludwig and Reynolds, 1988). However, conventional multivariate methods are generally limited in the sense that they are mainly applicable to linear data and have less flexibility in representing ecological data, for instance, handling non-linearity and data management (Chon et al., 1996; Lek and Guegan 1999, 2000; Recknagel, 2003). The Self-Organizing Map (SOM) is an efficient tool for mining non-linear data and has been extensively used for patterning community data since 1990s (e.g. Chon et al., 1996; Chon et al., 2000, 2002; Kwak et al., 2000; Levine et al., 1996; Park et al., 2001, 2003a,b, 2004). Chon et al. (1996) classified benthic macroinvertebrate communities in polluted streams with the SOM and characterized the source of variation of communities according to anthropogenic disturbances and locality of the sample sites. Chon et al. (2002) further implemented the SOM to a large scale data collected in different river systems in the Korean Peninsula for 16 years. The large scale data were accordingly arranged to reveal the impact of environmental disturbances. In this study we further elaborated clustering of the community samples as a means of comprehensive characterization of ecological status and evaluation of biological water quality in the national scale. Materials and Methods Self-Organizing Map (SOM) The SOM based on the Kohonen network (Kohonen, 1989) efficiently mines complex data 42

52 without templates (or teachers) in an unsupervised manner and has been reported as a reliable classifier of ecological data (Chon et al., 1996, 2002; Kwak et al., 2000; Park et al., 2004). In the SOM a linear array of M 2 artificial neurons (i.e., computation nodes), with each neuron being represented as j, is arranged in two dimensions for convenience of visual understanding. Suppose a community data containing N species, and the density of species, i, is expressed as a vector x i. The vector x i is considered to be an input layer to the SOM. In the network each neuron, j, is supposed to be connected to each node, i, of the input layer. The connectivities are represented as weights, w ij (t), adaptively changing at each iteration of calculation, t. Initially the weights are randomly assigned in small values. When the input vector is sent through the network, each neuron of the network computes the summed distance between the weight and input as shown below: N 1 i= 0 d j ( t ) = ( xi wij( t )) 2 The input values with greatly different numerical values in densities are avoided for training. The data were transformed by natural logarithm in order to emphasize the differences in low densities. Subsequently the transformed data were proportionally normalized between 0.01 and 0.99 in the range of the maximum and minimum density for each species collected during the survey period. The neuron responding maximally to a given input vector is chosen to be the winning neuron, the weight vector of which has the shortest distance to the input vector. The winning neuron and possibly its neighboring neurons are allowed to learn by changing the weights in the manner to further reduce the distance between the weight and the input vector as shown below: 43

53 w ( t+1) = w ( t ) +η( t )( x w ( t )) Z ij ij i ij j where Z j is assigned 1 for winning (and its neighboring) neuron(s) while it is assigned 0 for the rest neurons, and η(t) (e.g., ) denotes the fractional increment of the correction. The radius defining neighborhood is usually set to a larger value early in the training process, and is gradually reduced as convergence is reached. Detailed algorithm could be referred to Kohonen (1989), Zurada (1992), and Chon et al. (1996). After training, the Ward s linkage method (Ward, 1963) was applied to the weights of the SOM for further clustering of the patterned nodes. After preliminary training, we used N=80 (10 8) of the SOM output units for patterning community data. Community data The community data collected from the twenty-five published papers in Korea from 1984 to 2000 were used for evaluating water quality in streams (Chon et al., 2002). The communities were in the streams mainly collected with the Surber net, and the community data to the Family level were used as input to the SOM. We further extended the input data (Chon et al, 2002) by adding additional sample sites in different degrees of pollution: Piago valley in the Somjin River, Tokchon stream in the Nam River, the Miryang River in the Nakdong River system. We used 179 cases of the data consisting of 109 families for training. New community data additionally collected in the natural and urban areas on the regular basis were used for testing feasibility of the SOM through recognition (Fig. 3-1). The sample sites in the Suyong River were recognized by the trained SOM (Fig. 3-1a, b). The sample sites in the Suyong River showed various levels of pollution in the urban area, and the sample sites have been used for the long term monitoring of the urban streams (Kwon and Chon, 1993; Yoon and Chon, 1996). The sample sites YCK (28 44

54 cases) in the Suyong stream is located in the agricultural area, while THP (35 cases) in the Soktae stream is highly polluted with domestic sewage. We additionally surveyed the sample sites in the Baenae stream (2 cases) and Nakdong River (10 cases) which have been recently selected for LTER (National Long-Term Ecological Research Project in Korea) since 2005 (Fig. 1a, 1c). Fig Location of the sample sites. a) Korean Peninsula, b) the Suyong River and c) the Nakdong River. Water quality indices were obtained based on community data clustered by the SOM. EPT richness (the number of species in Ephemeroptera, Plecoptera and Trichoptera; see Resh et al., 45

55 1995) and BMWP Index (Biological Monitoring Working Party; Hawkes, 1997; National Water Council, 1981; Walley and Hawkes, 1996, 1997) were measured from the samples of benthic macroinvertebrates. The indices have been reported to correspondingly decrease with the increase in organic pollution (Hellawell, 1986; Rosenberg and Resh, 1993). Results Clustering of benthic macroinvertebrate communities Community abundance data collected from the Southern Peninsula of Korea (Fig. 3-1) were provided for training (Fig. 3-2a). The sample sites were firstly grouped according to the impact of pollution, and secondly to the location of streams and rivers. The clusters were vertically arranged according to degree of pollution: the sample sites with severe pollution appeared in the lower areas of the map (e.g., Cluster IV), while the clean sites were grouped in the upper right areas (e.g., Clusters I II). For instance, the groups of the sample sites from the natural areas (e.g., Somjin River, and Tamjin River), and the sample sites from the clean area of the Han River were located in Clusters I II, while communities from the polluted sites such as the Han, Nakdong, Masan and Suyoung Rivers were found in Cluster IV. Cluster III was placed in the middle area, being occupied by less polluted sites. The gradient shown in the mapping indicated that the SOM could serve as a map in revealing different impact of environmental disturbances. Within the clusters communities were further grouped according to location of the streams and rivers. Samples from the same locations tended to be grouped together. Biological water quality was measured based on clustering by the SOM. Figures 3-2b and 3-2c show the levels of EPT richness and BMWP corresponding to the clusters shown in the SOM (Fig. 3-2a). The levels of EPT richness and BMWP accordingly decreased in the clusters 46

56 at the lower areas of the map (Figs. 3-2b, 3-2c). The values of the indices were statistically significant among different clusters based on the nonparametric multiple comparison test (Mann-Whitney test, Zar, 1999)(Figs. 3-2b, 3-2c). BMWP was differentiated in all clusters, while the levels of EPT richness were partially in the same range between clusters II and III. This indicated that community clustering by the SOM properly reveals differences in biological water quality. Fig The map trained by SOM for pattering benthic macroinvertebrates reported from different streams in South Korea from 1984 to 2000: (a) sample sites; (b) mean and S.E. of EPT richness in different clusters defined in the SOM (n = 31 (I), n = 21 (II), n = 63 (III), n = 64 (IV)); (c) mean and S.E. of BMWP scores in different clusters defined in the SOM (n = 31 (I), n=21 (II), n = 63 (III), n = 64 (IV)). The different alphabets indicate significant difference in the Mann-Whitney test (p < 0.001). Profiles of Family in benthic macroinvertebrates could be obtained from the trained SOM. Abundance of the Families is visualized in the corresponding clusters (Fig. 3-3). In the clean zone, cluster I was mainly characterized by Perlidae and Lepidostomatidae while cluster II was 47

57 dominated by Gomphidae and Corydalidae. In the less polluted zone, cluster III was associated with Hydropsychidae and Baetidae. Cluster IV in the polluted zone was strongly occupied by Tubificidae and Chironomidae. The association of indicator taxa and clusters based on the SOM were generally in accord with filed observation. For instance, Tubificidae has been reported to be tolerant in polluted water, while Perlidae is frequently collected in clean water (Hellawell 1986, Rosenberg and Resh 1993). In the intermidate zone in cluster III, Hydropsychidae and Baetidae appeared as the indicator taxa for slightly polluted condition (Hellawell, 1986, Rosenberg and Resh 1993). Fig Profile of abundance of the prevalent taxa matched to clusters based on the trained SOM. The values in the vertical bar indicate densities (individuals/square meter). 48

58 Monitoring community patterns with the SOM We tested the trained network (Fig. 3-2a) with new data sets and monitored changes in community states as time progressed. The communities collected monthly at YCK in the Suyong River from November 1992 to April 1995 were recognized in a sequence by the SOM (Fig. 3-4a). In the early period (November 1992~November 1993), communities were mostly located in clusters III and IV, frequently crossing over the two clusters. Regarding that values of biological indices in clusters III and IV were low (Figs. 3-2b and 3-2c), water quality appeared to be poor at this stage. The fact that the sample communities frequently crossed the border of the clusters III and IV indicated that community patterns were somewhat variable although the sampled communities belonged to the polluted state overall. It was also notable that some communities tend to be located close together (e.g., group of nodes (3 (x), 2 (y)), (2, 3) and (3,3)). These nodes were characteristic in community composition: Chironomidae and Tubificidae were abundant in this group (Fig. 3-3). As the time progressed, communities moved to cluster I in the clean state in the later period of survey (January~April 1994; Fig. 3-4a), indicating recovery of water quality in winter of 1994 and 1995 briefly. However, the state of communities returned to the polluted state in cluster IV in the later period of survey (Fig. 3-4a). The overall tracks recorded on the map demonstrated that the states of communities collected on the regular basis could be monitored by the trained the SOM. Fig. 3-4b shows differences in biological and physico-chemical indices obtained from the newly recognized sample sites according to the clusters shown in Fig. 3-2a. Biological indices such as EPT richness and BMWP were clearly different with statistical significance among different clusters. However, physico-chemical indices such as BOD and turbidity were not statistically different, although there was a trend to be higher in BOD and turbidity in averages in clusters III and IV. 49

59 Fig Monitoring of benthic macroinvertebrate communities collected at YCK in the Suyong stream from November 1992 to April 1995 according to the trained SOM (the sample was not collected in December 1994). (a) Recognition of the samples: November 1992.November 1993 (dots); January 1994.January 1995 (solid), and (b) mean and S.E. of biological and physicochemical indices in different clusters defined in the SOM (n = 4 (I), n = 12 (III), n = 12 (IV)). The different alphabets indicate significant difference in the Mann-Whitney test (p < 0.001). Additionally, we tested community data from the severely polluted site, THP in the Soktae stream (Figs. 3-1b, 3-5). The monthly community data collected from March 1993 to April 1995 were invariably recognized at one node (5, 1) only in cluster IV all through the survey period (Fig. 3-5a). At this site, the taxa tolerant to organic pollution such as Oligochaetes and Chironomids were highly dominant. In addition, Psychodidae and Physidae were characteristically observed in THP (Fig. 3-5). Water quality data were also in the lowest range (Fig. 3-5b). Biological indices such as EPT richness and BMWP were invariably low. Especially EPT richness appeared to be zero all through the sampling period. Physico-chemical indices such as BOD (16.0 ± 17.4 mg/l) and turbidity (10.0 NTU ± 15.3 NTU) were also accordingly low but showed a higher degree of variability compared with the biological indices. 50

60 Fig (a) Monitoring of benthic macroinvertebrate communities collected at THP in the Suyong River from March 1992 to April 1995 according to the trained SOM (n = 35). (b) Mean and S.E. of biological and physico-chemical indices. We also monitored the sample sites used for LTER (Fig. 3-1c) on the trained SOM. The sample sites were accordingly recognized in two clusters on the map (Fig. 3-6a). The sites in the Baenae stream, BN, in the area were located in Cluster I, while the samples collected in the Nakdong River, NSJ, NKJ, NJP and NMK, were all placed in Cluster IV (Fig. 3-6a). The sample sites in the Nakdong River, however, were divided to two groups: a group of nodes around the node (3, 4) for the samples collected at NSJ and NKJ, etc, and the single node (6, 1) for NMK-6 and NJP-6. While the dominant taxa, Tubificidae and Chironomidae were abundant in both groups, Psychodidae and Physidae were only observed in the node (6, 1) as shown in THP (Fig. 3-5a). Water quality data was different according to the clusters. Biological indices such as EPT richness and BMWP were invariably low in cluster IV (Fig. 3-6b). Physico-chemical indices such as BOD (16.0 mg/l ± 6.7 mg/l) and turbidity (10.0 NTU ± 4.9 mg/l) were also low but showed relatively higher degree of variability. In cluster I where the site BN was recognized, the biological indices were in the highest range while BOD and turbidity were in the lowest range 51

61 (Fig. 3-6b). Fig Monitoring of benthic macroinvertebrate communities collected at LTER sites according to the trained SOM: (a) recognition of the samples (the number in the sample name indicates the month of collection) and (b) mean and S.E. of biological and physico-chemical indices (n = 2 (I), n = 8 (IV)). Discussion This study illustrated that clustering by the SOM would be useful for showing the gradient of water quality and providing a comprehensive view on overall states of community changes in response to environmental disturbances. The trained SOM readily accommodated diverse scope of community states across different degrees of pollution (Fig. 3-2a). Consequently, the SOM could serve as an alternative tool for monitoring community data to elucidate temporal changes in community states in response to disturbances. In the training data, physico-chemical data were not available for each sample site. As an alternative, we checked the differences in BOD and turbidity from the tested samples (Figs. 3-52

62 4~3-6). Although the biological indices (BMWP and EPT richness) were clearly differentiated, physico-chemical factors were not statistically different among the clusters. This indicated that biological indices appeared to be more sensitive in representing water quality based on the clustered community data. However, this would not necessarily mean that biological water quality indices are invariably superior to physico-chemical indices, regarding that biological indices were obtained from the same biological community data in this study. The sensitivity study between biological and physico-chemical factors should be more investigated with the sufficient data. In this study, the samples have not been sufficiently accumulated to evaluate differences in physicochemical factors in the testing data. The sample collection is still on going from the long-term monitoring sites (Fig. 3-1). Cross-evaluation of biological indices and physico-chemical indices should be carried out with accumulation of the sufficient data in the future. Due to limit of taxonomic information at the species level in the available data, training was carried out at the Family level in this study. The problem of information limit in taxonomy is critical in the Family of Chironomidae: Chironomidae includes diverse taxa but classification at the species level (especially, larvae) is extremely difficult. In the samples collected in the Baenae stream for testing, diverse species were newly identified in Chironomidae. The majority of species in this case were the specimens appearing in clean water. Diverse species in Chironomidae have been reported to be collected in cleanwater (Hellawell, 1986; Rosenberg and Resh, 1993). In this study, however, the tolerant species in Chironomidae such as Chironomus were mainly collected and were used for training. Consequently Chironomidae was presented as an indicator of polluted water along with Tubificidae in Oligochaetes in the training data. Due to this problem of training in Chironomidae, the samples newly collected with clean water species in the Baenae stream could not be included for testing at the present 53

63 time. Further information on taxonomic identification at the species level in benthic macroinvertebrates is required to be updated for training data in the future in order to obtain a higher predictability in water quality assessment through community patterning as the long-term field survey continues. 54

64 IV. Community patterns of benthic macroinvertebrates collected on the national scale in Korea Introduction The community composition depends on stability of their habitats which provide the resources for development of the residing populations (Cummins, 1979; Ward and Stanford, 1979; Malmqvist and Otto, 1987). The structures of community assemblages are potentially determined by various environmental factors acting on the habitats in different spatial and time scales (Stevenson, 1997; Snyder et al., 2002). Regarding that different species show different spatial temporal dynamics in responding to the impact of environments, understanding community abundance patterns is a fundamental step to achieve the sustainable management of aquatic ecosystems. With the advantages of taxonomic diversity, sedentariness in behavior and long life cycles, benthic macroinvertebrates respond to environmental disturbances in an integrated and continuous manner. Consequently, macroinvertebrates have been widely used for assessing ecological water quality in aquatic ecosystems. There have been a numerous accounts of benthic macroinvertebrates used for indicators of the shortand long-term environmental changes in running waters (Hellawell, 1978; Lenat, 1988; Smith et al., 1999; Hawkins et al., 2000). In natural conditions, species richness (i.e. the number of species occurring in a given area) has been commonly used as an integrative descriptor of the community (Lenat, 1988), since species richness is influenced by a large number of environmental factors in a continuous period with regards to environmental stability (Cummins, 1979; Ward and Stanford, 1979), ecosystem 55

65 productivity (Lavandier and Décamps, 1984) and heterogeneity (Malmqvist and Otto, 1987), and other biological factors (MacArthur, 1965; Feminella and Resh, 1990). The interactions of the environmental factors can consequently determine the gradients in species richness in streams (Vannote et al., 1980; Minshall et al., 1985). Additionally, species richness is useful for revealing the impact of environmental factors in disturbed conditions. The species richness of aquatic invertebrates has been influenced by natural and/or anthropogenic disturbances (Rosenberg and Resh, 1993), which may lead to spatial discontinuities of predictable gradients (Ward and Stanford, 1979, 1983) and loss of taxa (Brittain and Saltveit, 1989). Along with analysis in species richness, investigation of species abundance patterns has been regarded as an important topic in elucidating patterns of communities responding to the disturbances. Preston s canonical log-normal distribution has been the most widely accepted formalization of the relative commonness and rarity of species (Preston, 1962; Brown, 1981). Regarding that the species are often vulnerable to various environmental disturbances, the existence of rare species is a key issue in community ecology in relation to risk assessment. This type of complex relationships in community changes and environment disturbances would be accordingly addressed by studying community compositions in relation to abundance patterns per taxa, i.e. relations between species richness and abundance. Considering that the relationships between species and abundance are closely associated with overall ecological conditions, and communities are influenced by the given environmental factors limited by the geographic regions, the species abundance patterns would be reflected by ecoregions. The effect of ecoregions would appear especially on a large scale. Many studies have been carried out to classify ecoregions based on the distribution patterns of aquatic organisms on regional or national scale, showing the importance of geographic differences in biotic and abiotic characteristics of streams (Tison et al., 2005). 56

66 Similarly, overall variation in community composition would be reflected on a map characterizing the community patterns. When processing the large-scale data, the first step is usually to derive a classification of sites or systems by analysing the degree of association among community data. In this study, we apply the technique in ecological informatics, Self-Organizing Map, to mining the large-scale community data and to further relating the community patterns to variation caused by geographic distribution and different degrees of disturbances. We aim (1) to classify benthic macroinvertebrate communities on the national scale in Korea, (2) to accordingly reveal the relationships between community grouping and environmental factors, and (3) to elucidate association of community clusters with the species abundance patterns. Materials and methods Ecological data Benthic macroinvertebrates were collected in relatively clean to intermediately polluted areas in South Korea, being initiated by the National Natural Environment Monitoring Project, Ministry of Environment of Korea. Benthic macroinvertebrates were sampled from the 1970 sampling sites in total in the major rivers in Korea from 1997 to The samples were collected mostly at the 1st 3rd stream order using a Surber sampler (30cm 30 cm). The number of collected individuals was estimated in the unit area (1m 2 ). The general techniques for sampling benthic macroinvertebrates and measuring environmental variables have been conducted according to the National Natural Environment Monitoring protocol (Ministry of Environment, 1997). We used four variables to describe environmental conditions for the sampling site: altitude, depth, current velocity and conductivity. In the dataset, 687 species was 57

67 recorded in total. Diptera (including Chironomidae), Ephemeroptera, Trichoptera, Oligochaeta and Gastropod were mostly dominant. In addition to species richness and abundance, community and biological indices were estimated to evaluate ecological status of the sampling sites: Shannon diversity and evenness indices, and Biological Monitoring Working Party (BMWP) score (National Water Council, 1981; Walley and Hawkes, 1997). We also checked abundance of three major insect orders, i.e. Ephemeroptera, Plecoptera, and Trichoptera (EPT), which are commonly identified at the species level in freshwater studies. Subsequently, the EPT richness (i.e. the number of species occurring in a given area) and abundance were recorded at each sampling site. The differences of variables between the different groups identified through the models were evaluated by the Unequal N HSD multiple comparison test. Statistical analyses were conducted with the software package, Statistica (StatSoft, 2004). Abundance of each species was proportionally normalized between 0 and 1 (in the range of the minimum and maximum values) as input data for modelling. Modelling procedure First we extracted information on the relationships between species richness and abundance using scatter plots and log-normal distribution plots. Then we utilized the Self- Organizing Map (SOM) (Kohonen, 1982) for patterning macroinvertebrate communities. The SOM approximates the probability density function of the input data, and has been frequently used for clustering, visualization, and abstraction of ecological data (i.e. the idea of which is to show the dataset in another, more usable, representation form) (Kohonen, 2001; Park et al., 2003). The SOM consists of two layers: input and output layers connected by connection intensities (weights). Input layer gets information from data matrix, while output layer visualizes the computational results. When an input vector x (abundance of species) is sent 58

68 through the network, each neuron k of the network computes the distance between the weight vector w and the input vector x. The output layer consists of D output neurons, which are usually arranged into a two-dimensional grid for better visualization. There are no strict rules regarding the choice of the number of output neurons. To choose a suitable map size, we trained the SOM with different map sizes. Based on the recommendations of field ecologists in analyzing the data and the results from different map sizes, we chose 300 (D=20 15) neurons as the number of output neurons in this study. Each output neuron is a computational unit in the learning process. The best arrangement for the output layer is a hexagonal lattice, as it does not favor horizontal and vertical directions as much as a rectangular array (Kohonen, 2001). Among all the output neurons, the best matching unit (BMU) with minimum distance between the weight and input vectors is the winner. For the BMU and its neighborhood neurons, the weight vectors w are updated using the SOM learning rule. A detailed description of the SOM algorithm could be provided by Kohonen (2001) for computation and by Park et al. (2003) for ecological application. The learning process of the SOM was carried out using the SOM Toolbox (Alhoniemi et al., 2000) developed by the Laboratory of Information and Computer Science in the Helsinki University of Technology ( in Matlab environments (The Mathworks, 2001), and we adopted the initialization and training methods suggested by the authors of the SOM Toolbox that allow the algorithm to be optimized (Vesanto et al., 1999). 59

69 Results Relations between species richness and abundance Species richness in the samples was in the broad range from 1 to 54. It showed the maximum value around the abundance of 12 of log 2 (individuals) (Fig. 4-1a). While species richness showed the widest range in the middle point of abundance level, the range of species richness became narrower at either lower or higher values of abundance a) b) c) Number of species d) e) f) g) h) i) log 2 (abundance) log 2 (abundance) log 2 (abundance) Fig Relations between number of individuals in log scale and species richness in the 1970 samples used in this study. Each point indicates each sampling site. (a) All samples, (b~i) the samples separately grouped in clusters 1, 2, 3,..., 8, respectively. Fig. 4-2 shows the log-normal model of community structure by grouping species into abundance categories (octaves, i.e. power of 2). Although the distribution approached to a 60

70 normal distribution, the distribution was not statistically significant for the normal distribution (p < 0.01 for the Lilliefors normality test). 80 Number of species log 2 (abundance) Fig Log-normal model of community structure by grouping species into abundance categories (octaves, i.e. power of 2). Patterning benthic communities Benthic macroinvertebrate communities were patterned according to the similarity of community compositions through training with the SOM. The classified samples were accordingly visualized on the map (Fig. 4-3). The node includes another hexagon within itself in order to indicate the size of the patterned samples (Fig. 4-3a). As the number of patterned samples increase in a SOM output unit, the size of hexagon correspondingly increases. Simultaneously, the gray level becomes darker with the increasing number of the patterned samples in each output unit. The number of the patterned samples in the different output units ranged from 0 to 37. Overall, the SOM units were classified to two main clusters based on the dendrogram of the cluster analysis with the Ward s linkage method (Fig. 4-3b). The clusters were further divided to eight subclusters at different levels of the Euclidean distance. The clusters accordingly reflected geographical differences among the sampling sites. Fig. 4-3c 61

71 shows the representative location for each cluster in the geographical map of Korea. Cluster 1 was characterized by the samples collected mainly from the four regions Goyang, Paju, Gongju, Imsil and Jinhae, whereas cluster 2 was from Uljin in the southern area of Korea. However, cluster 6 did not show any specific geographic areas, being scattered widely over Korea. Therefore, cluster 6 was not indicated in the map in Fig. 4-3c. According to the clusters shown in Fig. 3a, the patterns of species richness and abundance were correspondingly differentiated in different clusters (Fig. 4-1b~i). Fig Classification of the samples according to the trained SOM. (a) The SOM units were grouped to eight clusters, (b) the dendrogram according to the Ward linkage method based on Euclidean distance, (c) geographical location of the sampling sites matching to clusters according to the SOM (Fig. 4-3a). Cluster 6 is not indicated on the geographical map because it did not show any specific geographic area. Species richness and abundance were statistically significant in different clusters (Fig. 4-4). Abundance was extremely low in cluster 1, while species richness in this cluster was in the intermediate range (Fig. 4-1b). The samples in cluster 1 were mainly collected at the pool zones with higher values of depth and lower values of velocity at the low altitude areas (Fig. 4-5). In contrast, the samples in cluster 2 (Fig. 4-1c) displayed the high level of abundance and the intermediate level of species richness, and were mainly collected in the southeastern areas in 62

72 Korea. The sample sites in these areas were characterized with high electric conductivity (Fig. 4-5). This was due to the wide distribution of the limestone in this area. Overall, the clusters projected on the trained SOM were in accordance with the different geographical areas in Korea. While species richness was in the highest range in cluster 3 in the mountainous area (Fig. 4-1d), the highest range of abundance was observed in cluster 7 in the urbanized area (Fig. 4-1h). The samples in cluster 7 were mainly collected from the slightly polluted streams in urbanizing area in the suburbs of Seoul. It was notable that the bottom areas of the SOM map showed high values of both species richness and abundance, including cluster 3 (Fig. 4-1d), cluster 4 (Fig. 4-1e), and cluster 8 (Fig. 4-1i) (Fig. 4-4). This was clearly distinguished at the level of the main clusters obtained by the Ward linkage method (Fig. 4-3b). Fig Community characterization in different clusters according to the SOM (Fig. 4-3a). (a) Species richness (number of species), (b) abundance (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable). 63

73 Fig Environmental variables in different clusters according to the SOM (Fig. 4-3a). (a) Altitude, (b) depth, (c) conductivity, (d) velocity (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable. Conductivity was not available at the samples in cluster 1). The samples of clusters 3, 4 and 8 were collected mostly from the mountain streams showing relatively high altitude (Fig. 4-5a). Generally, clusters showing high species richness displayed high abundance in this study. The sampling sites of clusters 4 and 5 were closely located each other in the geographical map (Fig. 4-3c). However, their community compositions were greatly different, separately belonging to the two main groups of the dendrogram (Fig. 4-3b). While the samples in cluster 4 (Fig. 4-1e) showed higher species richness in a broad range, the samples in cluster 5 (Fig. 4-1f) were in the lower ranges for both species richness and abundance. It was notable that cluster 8 showed unique patterns with high species richness and low abundance (Fig. 4-1i). The samples in cluster 8 were collected mainly in the Jiri Mountain 64

74 (Fig. 4-3c), indicating clean water with low conductivity (Fig. 4-5). Cluster 7 (Fig. 4-1h) was characteristic with regards that communities belonging to this cluster showed the negative relationships with communities grouped in cluster 8 in between species richness and abundance. Community patterns in cluster 6 were characteristic with the lowest range in both species richness and abundance (Fig. 4-1g). The samples in cluster 6 were identified as the communities collected from the widely distributed disturbed areas in the country due to the construction and rehabilitation programs carried out in the streams during the survey period. We evaluated each cluster with biological indices BMWP score, EPT richness, EPT abundance, Shannon diversity, and evenness (Fig. 4-6). In general, biological indices matched with the gradients of environmental disturbances shown by the trained SOM (Fig. 4-3a). EPT richness ranged from 0 to 38, showing lower values in clusters 1 and 6, and higher values in clusters 3, 4 and 8 (Fig. 4-6a). Clusters 2, 5, and 7 showed intermediate values. EPT richness accordingly revealed the degree of disturbances according to the clustered communities. EPT abundance ranged from 0 to 10,756, showing low values in clusters 1 and 6 and high values in clusters 2, 3 and 7 (Fig. 4-6b). The differences in EPT abundance among clusters were in general similar to the patterns of overall abundance, although EPT abundance was in the highest range in clusters 2 and 7, while overall abundance was in the highest range only in cluster 7 (Fig. 4-4b). BMWP score was between 0 and 146, showing the highest value in cluster 3 and the lowest value in cluster 6 (Fig. 4-6c). Community parameters also accordingly revealed the group characters of the clusters based on the trained SOM (Fig. 4-3a). As expected, Shannon diversity index was in the highest range in cluster 3, and was in the lowest range in cluster 6 (Fig. 4-6d). Evenness appeared to be lower in cluster 6 along with low values of species richness and abundance, while the index showed the highest value in cluster 8 with high species richness and low abundance (Fig. 4-6e). The patterns on the SOM well reflected the degree of 65

75 association among water quality indices. Correlation coefficient was between overall species richness and EPT richness were higher, between overall species richness and BMWP score, and was between EPT richness and BMWP score (Table 4-1). Shannon diversity index showed the higher correlation coefficients with species richness, EPT richness, and evenness. Each cluster was also characteristically associated with the indicator species in the corresponding clusters. In Table 4-2, indicator species (or taxa) characterizing each cluster were presented along with summary of species richness and abundance. Fig Variation in biological indices in different clusters according to the SOM (Fig. 4-3a). (a) EPT richness, (b) EPT abundance, (c) Shannon diversity index, (d) Biological Monitoring Working Party (BMWP) score, (e) evenness (different alphabets indicate significant differences between the clusters based on the Unequal N HSD multiple comparison test (p = 0.05). Error bars indicate mean and standard error of each variable). 66

76 Table 4-1. Correlation coefficients among community parameters and biological indices used in the datasets Species richness Abundance EPT richness EPT abundance BMWP Shannon index Abundance EPT richness EPT abundance BMWP score Shannon index Evenness Table 4-2. Community parameters, indicator species and environmental descriptions in different clusters Cluster Species richness Abundance Environments Indicator species 1 Intermediate Low Lowland, pool zone Physa acuta, Pantala flavescens 2 Intermediate High High conductivity Baetis nla, Ephemera strigata 3 High High High altitude Ecdyonurus kibunensis, Epeorus latifolium 4 High Intermediate High altitude Ecdyonurus kibunensis, Epeorus curvatulus 5 Intermediate Low Intermediate altitude Uracanthella rufa, Ecdyonurus levis 6 Intermediate Low Intermediate conductivity Baetis ursinus, Simulium uchidai 7 Intermediate High Low depth Limnodrilus socialis, Uracanthella rufa 8 High Low High altitude Semisulcospira libertine, Epeorus latifolium Discussion In this study, we classified benthic macroinvertebrate communities collected at the relatively less polluted areas (e.g., mountains, suburbs, etc.) for the purpose of conservation and diversity recording. The SOM accordingly classified the sample sites to eight groups based on species compositions. The classification matched with geographical distribution of the sampling sites, and showed that spatial variation was the main factor for characterizing benthic macroinvertebrate communities collected in Korea on a large scale (Fig. 4-3). Geographic location was effectively identified with the clusters according to the trained SOM. Moreover, 67

77 we demonstrated that the geographically identified communities matched with the species abundance patterns (Fig. 4-1b~i): lowland area (cluster 1, Fig. 4-1b), the limestone area (cluster 2, Fig. 4-1c), mountain area (clusters 3 (Fig. 4-1d), 4 (Fig. 4-1e), and 8 (Fig. 4-1i)), and slightly organic polluted area (cluster Fig. 4-1h). The SOM demonstrated feasibility in mapping of the clusters regarding provision of information on geographic distribution and species abundance patterns at the same time. Regional or national surveys of the stream ecosystems provide large volumes of the sitespecific data, which may carry some valuable information to derive certain spatial patterns of biological communities on different scales (i.e. from a local to a regional area). It has been a key issue in defining community patterns on a large scale for sustainable ecosystem management. There have been many attempts to classify streams based on the distribution patterns of aquatic organisms (Huet, 1954; Holmes et al., 1998; Wright et al., 1998, 2000; Cowx et al., 2004).Wasson et al. (2002) defined hydro-ecoregions in surface water bodies in France based on geology, relief, and climate, and validated the classification with benthic macroinvertebrate fauna in streams. Tison et al. (2005) also validated the hydro-ecoregions in France with diatoms in streams. Through the learning process of the SOM, we demonstrated that the characteristics of the samples on a large scale were distinctively identified in the clusters. The samples from mountain streams were grouped in cluster 8, showing high species richness and low abundance (Fig. 4-4). Samples from the disturbed areas due to the rehabilitation and construction on streams were grouped together in cluster 6, showing extremely low species richness and low abundance. This is commonly observed in the physically disturbed areas with poor condition of colonization by aquatic organisms. Efficiency of mapping was further demonstrated in the clusters located closely. The sampling sites of clusters 4 and 5, for instance, were close each other in the geographical map (Fig. 4-3c). However, their community compositions were greatly 68

78 different: while the samples in cluster 4 (Fig. 4-1e) showed higher species richness in a broad range, the samples in cluster 5 (Fig. 4-1f) were in the lower range for both species richness and abundance. This feasibility in identifying critical characteristics by the SOM could be efficiently used for characterizing typology of the river systems. However, to clearly define typology based on the benthic macroinvertebrates, we need the sufficient number of reference sites along with records of physical and chemical environmental factors and sufficient geographical information. Further studies are required to define typologies in Korea on the national scale. 69

79 GENERAL CONCLUSION In this study, we applied the SOM for ecological quality assessment using benthic macroinvertebrates in aquatic ecosystem. The chapters were divided to different scale and disturbances in streams. It was revealed the gradient of pollution and regional using benthic macroinvertebrates. First, Chapter I, SOM was utilized to extract information from complex data of environmental variables and benthic macroinvertebrate communities residing in different micro-habitats. Although the sampling was carried out in a limited area, the patterns of environmental variables revealed spatial heterogeneity. The clustering of benthic macroinvertebrate communities in the trained SOM was efficient in showing temporal variation and evaluating water quality according to the conditions of different micro-habitats. Consequently, local spatial heterogeneity is important in revealing dynamics of community abundance and biotic indices, especially regarding restoration processes in polluted streams. Chapter II, the samples were grouped into three main clusters corresponding to distinct EPTC assemblages in the tributary streams of the Garonne River catchment, southern France. Lower richness and diversity of macroinvertebrates were observed in the areas affected by agricultural land use, being associated with high Total Dissolved Solids (TDS), Nitrate (NO3) and Chemical Oxygen Demand (COD). Tolerant EPTC species were identified as controlling parameters for the changes in the assemblages collected at the agriculture-impacted sites. These findings confirmed that agricultural land use has disturbing effects on instream conditions of the riparian forest corridor that naturally serves as a buffer zone. In chapter III, the trained SOM showed the diverse scope of communities and were efficient in tracking temporal changes in community states. Biological water quality indices were correspondingly different in different clusters by the trained SOM. The SOM could be an 70

80 alternative map for covering diverse scope of community states in water quality monitoring for the large-scale, long-term data, thus would broaden the scope of ecological informatics implemented to ecosystem assessment. In chapter IV, the SOM was efficient in patterning and visualizing community characteristics and accordingly classified the samples corresponding to the geographical regions and to the degree of disturbances. The patterned groups accordingly revealed the impact of pollution: the species abundance and richness accordingly reflected geographical conditions and anthropogenic disturbances, while some tolerant species were selectively collected in the disturbed areas. The communities clustered by the trained network were also characterized with the species-abundant patterns. The SOM was efficient in linking various information on community composition, geographic distribution, and species abundance patterns in an integrative manner. The SOM could serve as an efficient ecological map for specifying ecoregions and for providing comprehensive view on ecological status of the communities sampled on a large scale. DePauw & Vanhooren (1983) suggested that as water quality can changes extremely rapidly, any one sample of water chemistry will only offer a snap-shot of true environmental conditions. Whereas abiotic measurements can only provide a view of environmental conditions at any one point in time (DePauw & Vanhooren, 1983), organisms are continually exposed to, and therefore reflect, the long-term average and extreme conditions of the environment in which they reside (Metacalfe-Smith, 1994). Especially, benthic macroinvertebrates well represented the ecological quality assessment in aquatic ecosystems. The different kinds of species had different levels of tolerance, so the community structures had higher relationship with the disturbance. The gradient of water quality appeared corresponding to the impact of pollution, and 71

81 biological water quality indices were correspondingly different according to the clusters by the trained SOM. The trained SOM showed the diverse scope of communities and were efficient in tracking temporal changes in community states. Based on the results, the SOM could be an alternative map for determining ecological status of water quality for the long term data, thus would broaden the scope of ecological informatics implemented to ecosystem assessment. This study was carried out two countries in South Korea and France. Diverse species were appeared in each country (Appendices 1-2). However, there are no same species. It is difficult to compare the community data in two countries directly. It will be useful to use the bio-indicator such as diversity index and tolerant index. The species-level identification has been recommended to assess small scale community composition between-sites and between-date differences (Lenat and Resh, 2001). Community parameters such as diversity index and richness can be strongly under-evaluated at the family and genus level, especially in the non-impacted streams (Guerold, 2000). Identification of aquatic insects at the species level allowed for a clearer characterization of the sample sites since the tolerance levels for environmental factors could be specifically provided for each species. Species data are useful for provision of specific information for the management of communities, especially including conservation of rare species (e.g., endemic species). 72

82 REFERANCES Alhoniemi, E., Himberg, J., Parhankangas, J. and Vesanto, J., SOM Toolbox. [online] Allan, J.D., Stream Ecology Structure and function of running waters. Chapman & hall. Allan, J.D. and A.S. Flecker, Biodiversity conservation in running waters. Bioscience 43: Allan, J.D., Landscapes and riverscapes: The influence of land use on stream ecosystems. Annual Review of Ecology and Systematics 35: Allan, J.D., D.L. Erickson and J. Fay, The influence of catchment land use on stream integrity across multiple spatial scales. Freshwater Biology 37: APHA, AWWA and WPCF., Standard methods for the examination of water and waste (16th ed.). Washington D.C., 1134 pp. Armitage, P.O., A quantitative study of the invertebrates of the River Tees below Cow Green. Reservoir. Freshwater Biology 6: Armitage, P.D. and Cannan, C.E., Annual changes in summer patterns of mesohabitat distribution and associated macorinvertebrate assemblages. Hydrol. Process., 14, Barbour, M.T., J. Gerritsen, G.E. Griffith, R. Frydenborg, E. McCarron, J.S. White and M.L. Bastian, A framework for biological criteria for Florida streams using benthic macroinvertebrates. Journal of the North American Benthological Society 15: Baxter, C. V., K. D. Fausch and W. C. Saunders, Tangled webs: reciprocal flows of invertebrate prey link streams and riparian zones. Freshwater Biology 50: Beketov, M.A., Different sensitivity of mayflies (Insecta, Ephemerotera) to ammonia, nitrite and nitrate: linkage between experimental and observational data. Hydrobiologia 528: 73

83 Berthélemy, C., Recherches écologiques et biogéographiques sur les plécoptéres et coléoptéres d eau courante (Hydraena et Elminthidae) des Pyrénées. Annales de limnologie 2: Brighnam, A.R., Brighnam, W.U. and Gnika, A., Aquatic insects and Oligochaetea of North and South Carolina. Midwest Aquatic Enterprise. Brinkhurst, R.O. and Jamieson, B.G.M., Aquatic Oligochaeta of the World. OliverandBody Edinburgh. Brinkhust, R. O., Guide to the Freshwater Aquatic Microdrile Oligachaetes of North America. Canadian Special Publication of Fidheries and Aquatic Sciences 84, 259 pp. Brittain, J.E., Saltveit, S.J., A review of the effect of river regulation on mayflies (Ephemeroptera). Reg. Riv.: Res. Manage. 3, Brooks, A.J., Haeusler, T., Reinfelds, I. and Williams, S., Hydraulic microhabitats and the distribution of macroinvertebrate assemblages in riffles. Freshwater Biol. 50, Brown, B.L., Spatial heterogeneity reduces temporal variability in stream insect communities. Ecol. Lett. 6, Brown, J.H., Two decades of homage to Santa Rosalia: toward a general theory of diversity. Am. Zool. 21, Bunn, S.E., Edward, D.H., Loneragan, N.R., Spatial and temporal variation in the macroinvertebrate fauna of streams of the northern jarrah forest, Western Australia: community structure. Freshwater Biology, 16, Butcher, J.T., Stewart, P.M., Simon, T.P., A benthic community index for streams in the northern lakes and forests ecoregion. Ecological indicators, 3, Céréghino, R., J.L. Giraudel and A. Compin, Spatial analysis of stream invertebrates 74

84 distribution in the Adour-Garonne drainage basin (France), using Kohonen Self Organizing Maps. Ecological Modelling 146: Céréghino, R., Y.S. Park, A. Compin and S. Lek, Predicting species richness of aquatic insects in streams using a limited number of environmental variables. Journal of the North American Benthological Society 22(3): Chauvet, E. and H. Décamps, Lateral interactions in a fluvial landscape: the River Garonne, France. Journal of the North American Benthological Society 8(1): Chon, T.-S., Park, Y.S., Moon, K.H., Cha, E.Y., Patternizing communities by using an artificial neural network. Ecol. Model. 90, Chon, T.-S., Park, Y.S., Park, J.H., Determining temporal pattern of community dynamics by using unsupervised learning algorithms. Ecol. Model. 132, Chon, T-S., Kwak, I-S., Song, M-Y., Park, Y-S., Cho, H.D, Kim, M.J., Cha, E.Y. and Lek, S., Benthic Macro Invertebrates in Streams of South Korea in Different Levels of Pollution and Patterning of Communities by Implementing the Self-Organizing Mapping. Pages in Ecology of Korea. Lee, D. (Editor). Bumwoo Publishing Company, Seoul. Collier, K. J., Environmental factors affecting the taxonomic composition of aquatic macroinvertebrate communities in lowland waterways of Northland, New Zealand. New Zealand Journal of Marine and Freshwater Research 29: Compin, A. and R. Céréghino, Sensitivity of aquatic insect species richness to disturbance in the Adour Garonne stream system (France). Ecological Indicators 3: Copper, S.D., Barmuta, L., Sarnelle, O., Kratz, K. and Diehl, S., Quantifying spatial heterogeneity in streams. J. N. Am. Benthol. Soc. 16, Cowx, I.G., Noble, R.A., Nunn, A.D., Harvey, J.P., Flow and level criteria for coarse fish and conservation species. Environment Agency R&D Report, W Bristol. 75

85 Cummins, K.W., The natural stream ecosystem. In: Ward, J.V., Stanford, J.A. (Eds.), The Ecology of Regulated Streams. Plenum Press, New York, pp Cummins, K.W. and Lauff, G.H., The influence of substrate particle size on the microditribution of stream macrobenthos. Hydrobiologia 34, Davies, N.M., Norris, R.H., Thomas, M., Prediction and assessment of local stream habitat features using large-scale catchment characteristics. Freshwater Biology, 45, Death, R.G., Spatial patterns in benthic invertebrate community structure: products of habitat stability or are they habitat specific? Freshwater Biol. 33, Delong, M.D. and M.A. Brusven, Macroinvertebrate community structure along the longithudinal gradient of an agriculturally impacted stream. Environmental Management 22: DePauw, N. and Vanhooren, G., Method for biological quality assessment of watercourses in Belgium. Hydrobiologia 100: Dolédec, S. and D. Chessel, Co-inertia analysis: an alternative method for studying species-environmental relationships. Freshwater Biology 31: Downes, B.J. Lake, P.S. and Schreiber, E.S.G., Spatial variation in the distribution of stream invertebrates: implications of patchiness for models of community organization. Freshwater Biol Dray, S., D. Chessel and J. Thioulouse, Co-inertia analysis and the linking of ecological data tables. Ecology 84: Dudgeon, D., A.H. Arthington, M.O. Gessner, Z. Kawabata, D.J. Knowler, C. Leveque, R.J. Naiman, A. Prieur-Richard, H.D. Soto, M.L.J. Stiassny and C.A. Sullivan, Freshwater biodiversity: importance, threats, status and conservation challenges. Biological Reviews (in press). 76

86 Feminella, J.W., Resh, V.H., Hydrologic influences, disturbance, and intra specific competition in a stream caddisfly population. Ecology 71, Ferringto, L.C. and Crisp, N.H., Water chemistry characteristics of receiving streams and the occurrence of Chironomus riparius and other chironomidae in Kensas. Acta Biol. Debr. Oecol. Hung. 3, Fitzpatrick, F.A., B.C. Scudder, B.N. Lenz and D.J. Sullivan, Effects of multi-scale environmental characteristics on agricultural stream biota in eastern Wisconsin. Journal of the American Water Resources Association 37(6): Fortuné, M., Usages passées et écologie de la Garonne. Thesis, Toulouse. Frissell, C.A., W.J. Liss, C.E. Warren and M.D. Hurley, A hierarchical framework for stream habitat classification: viewing streams in a watershed context. Environmental Management 12: Gebler, J.B., Mesoscale spatial variability of selected aquatic invertebrate community metrics from a minimally impaired stream segment. J. N. Am. Benthol. Soc. 23, Genito, D., W.J. Gburek and A.N. Sharpley, Response of stream macro invertebrates to agricultural land cover in a small watershed. Journal of Freshwater Ecology 17: Gergel, S.E., M.G. Turner, J.R. Miller, J.M. Melack and E.H. Stanley, Lanscape indicators of human impacts to riverine systems. Aquatic Sciences 64: Gittins, R., Canonical Analysis, a Review with Applications in Ecology. Springer-Verlag, Berlin, 351 pp. Guerold, F., Influence of taxonomic determination level on several community indices. Water Research 34: Hall, M.J., G.P. Closs and R.H. Riley, Relationships between land use and stream invertebrate community structure in a South Island, New Zealand, coastal stream catchment. 77

87 New Zealand Journal of Marine and Freshwater Research 35: Harding, J.S., E.F. Benfield, P.V. Bolstad, G.S. Helfman and E.B.D. III Jones, Stream biodiversity: The ghost of land use past. Proceedings of the National Academy of Sciences USA 95: Hynes, H.B.N., The Ecology of Running Waters. University of Toronto Press, Toronto, 555 pp. Hawkes, H.A., Invertebrates as indicators of river water quality. Chapter 2 in Biological indicators of water quality. James A. and Evision L. (Editors). John Wiley and Sons, Chishester, Great Britain. Hawkes, H.A., Origin and development of the biological monitoring working party score system. Water Res. 32, Hawkins, C.P., Norris, R.H., Gerritsen, J., Hughes, R.M., Jackson, S.K., Johnson, R.K., Stevenson, R.J., Evaluation of the use of landscape classifications for the prediction of freshwater biota: synthesis and recommendations. J. N. Am. Benthol. Soc. 19, Hellawell, J.M., Biological Surveillance of Rivers. Water Research Center, Stevenage Laboratory, England. Hellawell, J.M., Biological Indicators of Freshwater Pollution and Environmental Management. Elsevier, London, p 546. Hildrew, A.G., Townsend, C.R. and Henderson, J., Interactions between larval size, microdistribution and substrate in the stoneflies of an iron-rich stream. Oikos 35, Holmes, N.T.H., Boon, P.J., Rowell, T.A., A revised classification system for British rivers based on their aquatic plant communities. Aquat. Conserv. -Mar. Freshwater Ecosyst. 8 (4), Huet, M., Biologie, profils en long et en travers des eaux courantes. Bull. Fran. Pisci. 175, 78

88 Jain, A.K. and Dubes, R.C., Algorithms for clustering data. Prentice-Hall, Englewood Hills, NJ. Johnson, R.K., T. Wiederholm and D.M. Rosenberg, Freshwater biomonitoring using individual organisms, populations, and species assemblages of benthic macroinvertebrates. In: D.M. Rosenberg and V.H. Resh (eds). Freshwater biomonitoring and benthic macroinvertrbrates. Chapman & Hall, London, UK. KICT (Korea Institute of Construction Technology), Development of close-to-nature river improvement techniques (CTNRIT) adapted to the Korean streams. Report 2. Kohonen, T., Self-organized formation of topologically correct feature maps. Biol. Cybern. 43, Kohonen, T., Self-organization and Associative Memory. Springer-Verlag. Berlin, p 312. Kohonen, T., Self-organizing maps. 3 rd ed.. Springer, Berlin. Krebs, C.J., Ecology: The experimental analysis of distribution and abundance. Harper & Row, New York. Kwak, I.S., Liu, G.C., Chon, T.-S, Park, Y.S., Community Patterning of Benthic Macroinvertebrates in Streams of South Korea by Utilizing an Artificial Neural Network. Korean Journal of Limnology. 33, Kwon, T-S., Chon, T-S., Ecological studies on benthic macroinvertebrates in the Suyong River. III. Water quality estimations using chemical and biological indices. Kor J Lim. 26, Lamouroux, N., Doledec, S. and Gayraud, S., Biological traits of stream macroinvertebrate communities : effects of microhabitat, reach, and basin filters. J. N. Am. Benthol. Soc. 23,

89 Lancaster, J. and Hildrew, A.G., Flow refugia and the microistribution of lotic macroinvertebrates. J. N. Am. Benthol. Soc. 12, Lavandier, P., Décamps, H., Estaragne. In: Whitton, B.A. (Ed.), Ecology of European Rivers. Blackwell, London, pp Legendre, P., Legendre, L., Numerical Ecology. Elsevier Science, Netherlands. Lek, S., Guégan, J-F., Artificial Neuronal Network as a tool in ecological modelling: an introduction. Ecol. Model. 120, Lek, S., Guégan, J-F., Artificial Neuronal Networks: Application to Ecology and Evolution. Springer, Berlin. Lek, S., M. Scardi, P.F.M. Verdonschot, J.P. Descy and Y.-S. Park, Modelling Community Structure in Freshwater Ecosystems. Springer-Verlag, Berlin Heidlberg, 518 pp. Lenat, D.R., Water quality assessment of streams using a qualitative collection method for benthic macroinvertebrates. J. N. Am. Benth. Soc. 7, Lenat, D.R. and J.K. Crawford, Effects of land use on water quality and aquatic biota of three North Carolina Piedmont streams. Hydrobiologia 294: Lenat, D.R. and V.H. Resh, Taxonomy and stream ecology-the benefits of genus- and species-level identification. Journal of the North American Benthological Society 20: Levine, E.R., Kimes, D.S., Sigillito, V.G., Classifying soil structure using neural networks. Ecol. Model. 92, Ludwig, J.A., Reynolds, J.F., Statistical Ecology: A Primer of Methods and Computing. John Wiley and Sons, New York. MacArthur, R.H., Patterns of species diversity. Biol. Rev. 40, Malmqvist, B., Otto, C., The influence of substrate stability on the composition of stream benthos: an experimental study. Oikos 48,

90 Merritt, R.W. and Cummins, K.W., An Introduction to the Aquatic Insects of North America. Hunt Publishing Company, Dubugue. 722 pp. Metcalfe-Smith, J.L., Biological water-quality assessment of rivers: use of macroinvertebrates. In: The Rivers Handbook-Hydrological and Ecological Principles. Blackwell Scientific Publications: Meyer, A., E.I. Meyer and C. Meyer, Lotic communities of two small temporary karstic stream systems (East Westphalia, Germany) along a longitudinal gradient of hydrological intermittency. Limnologica 33: Ministry of Environment, The National Natural Environment Monitoring Protocol. Ministry of Environment, Seoul. Minshall, G.W., Stream ecosystem theory: a global perspective. J. N. Am. Benthol. Soc. 7, Minshall, G.W., Petersen, R.C., Nimz, C.F., Species richness in streams of different size from the same drainage basin. Am. Nat. 125, Moog, O, E. Bauernfeind and P. Weichselbaumer, The use of Ephemeroptera as saprobic indicators in Austria. In P. Landolt and M. Sartori (eds). Ephemeroptera & Plecoptera: Biology- Ecology-Systematics. Fribourg, pp National Water Council, River Quality: The 1980 Survey and Future Out-look. London. Palmer, M.A., Hakenkamp, C.C. and Nelson-Baker, K., Ecological heterogeneity in streams: why variance matters. J. N. Am. Benthol. Soc. 16, Palmer M.A. and Poff N.L., The influence of environmental heterogeneity on patterns and processes in streams. J. N. Am. Benthol. Soc. 16, Park, Y-S., Chon, T-S., Kwak, IS., Kim, J-K., Jørgensen, SE., Implementation of artificial neural networks in patterning and prediction of exergy in response to temporal dynamics of 81

91 benthic macroinvertebrate communities in streams. Ecol Model. 146, Park, Y-S., Céréghino, R., Compin, A. and Lek, S., 2003a. Applications of artificial neural networks for patterning and predicting aquatic insect species richness in running waters. Ecol Model. 160, Park, Y-S., Chang, J., Lek, S., Cao, W. and Brosse, S., 2003b. Conservation strategies for endemic fish species threatened by the Three Gorges Dam. Conserv. Biol. 17, Park.,Y-S., Chon, T-S., Kwak, IS., Lek, S., Hierarchical community classification and assessment of aquatic ecosystems using artificial neural networks. Science of the Total Environment, 327, Pennak, R.W., Fresh-water Invertebrates of the United States. John Wieley and Sons, Inc., New York. 803 pp. Petersen, I., E. Masters, A.G. Hildrew and S.J. Ormerod, Dispersal of adult aquatic insects in catchments of differing land use. Journal of Applied Ecology 41: Pik, A., Dangerfield, J.M., Bramble, R.A., Angus, C. and Nipperess, D.A., The use of invertebrates to detect small-scale habitat heterogeneity and its application to restoration practices. Environ. Monit. Assess. 75, Poff, N.L., Landscape filters and species traits: towards mechanistic understanding and prediction in stream ecology. J. N. Am. Benthol. Soc. 16, Poff, N.L. and Ward, J.V., Physical habitat template of lotic systems: recovery in the context of historical pattern of spatiotemporal heterogeneity. Environ. Manage., 14, Poole, G.C., Fluvial landscape ecology: addressing uniqueness within the river discontinuum. Freshwater Biol. 47, Probst, M., N. Berenzen, N. Lentzen-Godding, R. Schulz and M. Liess, Linking land use variables and invertebrate taxon richness in small and medium-sized agricultural streams on a 82

92 landscape level. Ecotoxicology and Environmental Safety 60: Purcell, A.H., Friedrich, C. and Resh, V.H., Restoration Project in Northern California. Restor. Ecol. 10, Quigley, M., Invertebrates of streams and rivers. Edward A. (publishers) Ltd., Colchester, 84 pp. Quinn, M.A., Halbert, S.E., Williams, III L., Spatial and temporal changes in aphid (Homoptera: Aphididae) species assemblages collected with suction traps in Idaho. J. Econ. Entomol. 84, Quinn, J.M., Effects of pastoral development. In K.J. Collier and M.J. Winterbourn (eds). In New Zealand Stream Invertebrates: Ecology and Implications for Management. Christchurch, NZ: Caxton, pp Rabeni, C., Doisy, K.E. and Galat, D.L., Testing the biological basis of a stream habitat classification using benthic invertebrates. Ecol. Appl. 12, Recknagel, F., Ecological Informatics: Understanding ecology by biologically-inspired computation. Springer, Berlin, p 398. Resh, V.H. and Rosenberg, DM., The ecology of aquatic insects. Praeger Publishers, New York, 625 pp. Resh, V.H., Hildrew, A.G., Statzner, B. and Townsend, C.R., Theoretical habitat templets, species traits and species richness: A synthesis of long-term ecological research on the Upper Rhone River in the context of concurrently developed ecological theory. Freshwater Biol. 31, Resh, V.H., Norris, R.H., Barbour, M.T., Design and implementation of rapid assessment approaches for water resource monitoring using benthic macroinvertebrates. Australian Journal of Ecology. 20,

93 Reynoldson, T.B., Norris, R.H., Resh, V.H., Day, K.E. and Rosenberg, D.M., The reference condition: a comparison of multimetric and multivariate approaches to assess waterquality impairment using benthic macroinvertebrates. J.N. Benthol. Soc. 16, Richards, C., Haro, R.J., Johnson, L.B. and Host, G.E., Catchment and reach-scale properties as indicators of macroinvertebrate species traits. Freshwater Biol. 37, Riley, R.H., C.R. Townsend, D.K. Niyogi, C.A. Arbuckle and K.A. Peacock, Headwater stream response to grassland agricultural development in New Zealand. New Zealand Journal of Marine and Freshwater Research 37: Rosenberg, D.M. and Resh, V.H., Freshwater Biomonitoring in Benthic Macroinvertebrates. Chapman and Hall, New York, 488 pp. Roy, A.H., Rosemond, A.D., Leigh, D.S., Paul, M.J. and Wallace, B., Habitat-specific responses of stream insects to land cover disturbance: biological consequences and monitoring implications. J. N. Am. Benthol. Soc. 22, Schulz, R. and M. Liess, A field study of the effects of agriculturally derived insecticide input on stream macroinvertebrate dynamics. Aquatic Toxicology 46: Sladecek, V., Continental systems for the assessment of river water quality. In: James, A., Evison, L. (Eds.), Biological Indicators of Water Quality. John Wiley & Sons, Chichester. Smith, M.J., Kay, W.R., Edward, D.H.D., Papas, P.J., Richardson, K.St.J., Simpson, J.C., Pinder, A.M., Cale, D.J., Horwitz, P.H.J., Davis, J.A., Yung, F.H., Norris, R.H., Halse, S.A., AusRivAS: using macroinvertebrates to assess ecological condition of rivers in Western Australia. Freshwater Biol. 41, Snyder, E.B., Robinson, C.T., Minshall, G.W., Rushforth, S.R., Regional patterns in periphyton accrual and diatom assemblages structure in a heterogeneous nutrient landscape. Can. J. Fish. Aquat. Sci. 59,

94 Song, M.-Y., Y.-S. Park, I.-S. Kwak, H. Woo and T.-S. Chon, Characterization of benthic macroinvertebrate communities in a restored stream by using self-organizing map. Ecological Informatics 1: StatSoft, Inc., STATISTICA (data analysis software system), version 7, Statzner, B., Gore, J.A. and Resh, V.H., Hydraulic stream ecology: observed patterns and potential applications. J. N. Am. Benthol. Soc. 7, Stevenson, R.J., Scale-dependent determinants and consequences of benthic algal heterogeneity. J. N. Am. Benth. Soc. 16, Stubauer, I. and O. Moog, Taxonomic sufficiency versus need for information-comments on Austrian experience in biological water quality monitoring. Verh. Int. Verein. Limnol. 27: Stuijfzand, S.C., S. Engels, E. Ammelrooy and M. Jonker, Caddisflies (Trichoptera Hydropsychidae) used for evaluating water quality of large European Rivers. Archives of Environmental Contamination and Toxicology 36: The Mathworks, lnc., MATLAB Version 6.1, Massachusetts. Thiere, G. and R. Schulz, Runoff-related agricultural impact in relation to macroinvertebrate communities of the Lourens River, South Africa. Water Research 38: Thomas, A. and C. Berthélemy, Préférences et limites écologiques des Elmidae (Coleoptera) dans le Sud-Ouest de la France. Bull. Soc. Hist. Nat., Toulouse, 127: Thioulouse, J., S. Dolédec, D. Chessel and J.M. Olivier, ADE-4: a multivariate analysis and graphical display software. Statistics and Computing 7: Tison, J., Park, Y.-S., Coste, M., Wasson, J.G., Ector, L., Rimet, F., Delmas, F., Typology 85

95 of diatom communities and the influence of hydro-ecoregions: a study at the French hydrosystem scale. Water Res. 39, Tittizer, T.T., Koth, P., Possibilities and limitations of biological methods of water analysis. In: James, A., Evison, L. (Eds.), Biological Indicators of Water Quality. John Wiley and Sons, Chichester. Townsend, C.R., S. Dolédec, R. Norris, K. Peacock and C. Arbuckle, The influence of scale and geography on relationships between stream community composition and landscape variables: description and prediction. Freshwater Biology 48: Townsend, C.R. and Hildrew, A.G., Species traits in relation to a habitat templet for river systems. Freshwater Biol. 31, Urbanič, G., M.J. Toman and C. Krušnik, Microhabitat type selection of caddisfly larvae (Insecta: Trichoptera) in a shallow lowland stream. Hydrobiologia 541: Vannote, R.L., Minshall, G.W., Cummins, K.W., Sedell, J.R., Cushing, C.E., The river continuum concept. Can. J. Fish. Aquat. Sci. 37, Vesanto, J., Himberg, J., Alhoniemi, E., Parhankangas, J., Self-organizing map in Matlab: the SOM Toolbox. In: Proceedings of the MATLAB Digital Signal Processing Conference, Espoo, Finland, pp Vondracek, B., K.L. Blann, C.B. Cox, J.A. Frost Nerbonne, K.G. Mumford, B.A. Nerbonne, L.A. Sovell and J.K.H. Zimmmerman, Land use, spatial scale and stream systems: Lessons from an agricultural region. Environmental Management 36: Wallace, J. B., J. W. Grubaugh and M. R. Whiles, Biotic indices and stream ecosystem processes: results from an experimental study. Ecological Applications 6: Walley, WJ., Hawkes, HA., A computer-based reappraisal of Biological Monitoring Working Party scores using data from the 1990 River Quality Survey of England and Wales. 86

96 Water Res. 30, Walley, WJ., Hawkes, HA., A computer-based development of the Biological Monitoring Working Party score system incorporating abundance rating, biotope type and indicator value. Water Res. 31, Wang, L., J. Lyons, P. Kanehl, and R. Gatti, Influences of watershed land use on habitat quality and biotic integrity in Wisconsin streams. Fisheries 22: Ward, J.H., Hierarchical grouping to optimize an objective function. J. Amer. Statist. Assoc. 58, Ward, J.V., Stanford, J.A., Ecological factors controlling stream zoobenthos with emphasis on thermal modification of regulated streams. In: Ward, J.V., Stanford, J.A. (Eds.), The Ecology of Regulated Streams. Plenum Press, New York, pp Ward, J.V., Stanford, J.A., The intermediate disturbance hypothesis: an explanation for biotic diversity patterns in lotic systems. In: Fontaine, T.D., Bartell, S.M. (Eds.), Dynamics of Lotic Ecosystems. Ann Arbor Sciences, Ann Arbor, Michigan, pp Wasson, J.G., Chandesris, A., Pella, H., Blanc, L., Typology and reference conditions for surface water bodies in France the hydro-ecoregion approach. In: Proceeding of Typology and Ecological Classification of Lakes and Rivers, Finnish Environment Institute (SYKE), Helsinki, Finland, October 25 26, 2002 (published in TemaNord). Wiederholm, T Chironomidae of the Holactic Region. Keys and Diagnoses. (Part 1. Larvae). Ent. Scand. Suppl pp. Williams, D.D., Environmental constraints in temporary fresh waters and their consequences for the insect fauna. Journal of the North American Benthological Society 15: Wright, J.F., Sutcliffe, D.W. and Furse, M.T., Assessing the biological quality of fresh 87

97 waters: RIVPACS and other techniques. Freshwater Biological Association, Ambleside, UK, 400 pp. Wright, JF., Furse, MT., Armitage, PD., RIVPACS: A technique for evaluating the biological quality of rivers in the UK. European Water Pollution Control. 3, Wright, J.F., Furse, M.T., Moss, D., River classification using invertebrates: RIVPACS applications. Aquat. Conserv. -Mar. Freshwater Ecosyst. 8 (4), Wright, J.F., Sutcliffe, D.W., Furse, M.T. (Eds.), Assessing the Biological Quality of Fresh Waters: RIVPACS and Other Techniques. Freshwater Biological Association, Ambleside. Yun, I-B., Illustrated Encyclopedia of Fauna and Flora of Korea Vol. 30. (Aquatic insects), Ministry of Education, Seoul. 840 pp. Yoon, B.J., Chon, T.S., Community analysis in chironomids and biological assessment of water qualities in the Suyong and Soktae streams of the Suyong River. Kor. J. Lim. 29, Zar J.H Biostatistical Analysis. 4 th ed.. Prentice-Hall Inc.. New Jersey. p.698. Zurada, J.M., Introduction to Artificial Neural Systems. West Publishing Company. New York, p

98 APPENDICES APPENDIX 1. List of EPTC species in streams of the Garonne River catchment (SW France) Order Family Genus Species Siphlonuridae Siphlonurus S. aestivalis B. rhodani Baetidae Baetis Procloeon Cloeon Centroptilum B. vernus B cf lutheri B. fuscatus P. sp P. bifidum C. cognatum C. luteolum Oligoneuridae Oligoneuriella O. rhenana E. dispar Ecdyonurus E. venosus E. insignis Ephemeroptera Heptageniidae Electrogena El. sp. Epeorus E. torrentium Heptagenia H. sulphurea Rhithrogena R. gr semicolorata Potamanthidae Potamanthus P. luteus Ephemeridae Ephemera E. danica E. lineata Polymitarcyidae Ephoron E. virgo Ephemerellidae Serratella S. ignita Caenidae Leptophlebiidae Caenis Paraleptophlebia Habroleptoides Habrophlebia C. luctuosa C. pusilla P. submarginata H. confusa Hbia sp 89

99 Order Family Genus Species Taeniopterygidae Brachyptera B. risi Protonemura P. intricata A. sp. Nemouridae Amphinemura A. standfussi N. sp. Plecoptera Nemoura N. cinerea Capniidae Capnia C. bifrons Leuctridae Leuctra L. sp.(fusca) Euleuctra E. geniculata Perlodidae Rhyacophilidae Isoperla Rhyacophila I. gr grammatica R. dorsalis R. fasciata R. obliterata R. sp Polycentropodidae Polycentropodus P. kingi Psychomyiidae Psychomyiia P. pusilla H. contubernalis Hydropsychidae Hydropsyche H. siltalai H. cf exocellata Cheumatopsyche C. lepida Chimarra C. marginata Philopotamidae Philopotamus P. montanus Trichoptera Hydroptilidae Hydroptila H. sp Limnephilidae Limnephilus L. spp Odontoceridae Odontocerum O. sp G. pilosa Goeridae Goera G. sp A. bilineatus Leptoceridae Athripodes A. sp Mystacides M. azurea Lepidostomatidae Genus sp. ND Brachycentridae Brachycentrus B. subnubilus Glossomatidae Agapetus A. laniger Sericostomatidae Sericostoma S. personatum 90

100 Order Family Genus Species Gyrinidae Orectochilus O. villosus Dytiscidae ND ND Hydrophilidae ND ND Hydraenidae ND ND Dryopidae Dryops D. sp. Dupophilus D. brevis Elmis E. spp. Coleoptera Esolus E. parallelepipedus Oulimnius O. troglodytes Elmidae L. sp Limnius L. opacus Stenelmis S. canaliculata Riolus R. cupreus Scirtidae Helodes H. sp. Helophoridae ND ND 91

101 APPENDIX 2. List of EPTC species in Daechon stream of Busan (Korea) Order Family Genus Species Ameletidae Ameletus A. costalis Siphlonuridae Siphlonurus S. chankae B. japonica Baetidae Baetiella Baetis Ecdyonurus B. tuberculata B. pseudothermicus B. sp1 B. sp2 E. kibunensis E. levis E. sp E. bajikovae Heptageniidae H. kifada Ephemeroptera Heptagenia H. sp E. curvatulus Epeorus E. latifolium Leptophlebiidae Paraleptophlebia P. chocorata Choroterpes C. altioculus E. strigata Ephemeridae Ephemera E. separigata Ephemerella E. keijoensis Uracanthella U. rufa Ephemerellidae Acerella A. longicaudata Serratella S. setigera Cincticostella C. castanea 92

102 APPENDIX 2. (continued) Order Family Genus Species Leuctridae Rhopalopsole R. mahunkai Kiotina K. decorata Plecoptera Neoperla N. quadrata Perlidae Paragnetina P. flavotincta Pteronarcidae Pternarcys P. sachalina H. sp1 H. sp2 Hydropsychidae Hydropsyche H. sp3 H. sp4 H. sp5 Cheumatopsche C. brevilineata Trichoptera Rhyacophilidae Rhyacophila R. sp R. shikotsuensis R. sibirica Leptoceridae Ceraclea C. sp Polycentropodidae Plectrocnemia P. sp Glossosomatidae Glossosoma G. sp G. sp1 Lepidostomatidae Goerodes G. sp2 Limnephilidae ND ND Elmidae ND ND Stenelmis S. vulgaris Coleoptera Chrysomelidae Galerucella G. sp Hydrophilidae Helochares H. striatus Sternolophus S. rufipes 93

103 Evaluation de la qualité écologique des écosystèmes lotiques utilisant des macroinvertébrés Résumé Dans cette étude, nous avons appliqué le modèle SOM (Self-Organizing Map) pour une évaluation de la qualité écologique de l écosystème aquatique en se basant sur les macroinvertébrés benthiques. D abord, dans le chapitre I, le SOM a été utilisé pour extraire les informations à partir de grandes matrices de données complexes de variables environnementales et des macroinvertébrés benthiques des différents microhabitats. Bien que les échantillonnages aient été réalisés dans une zone limitée, le patron des variables environnementales révèle une hétérogénéité spatiale. Les classes de macroinvertébrés dans le modèle SOM ont montré une variation temporelle et permettent l évaluation de la qualité de l eau en accord avec les conditions des différents microhabitats. En conséquence, l hétérogénéité spatiale locale est importante en révélant les dynamiques des communautés et les indices biotiques, particulièrement en rapport avec le processus de restauration des rivières polluées. Dans le chapitre II, les échantillons ont été classés en 3 groupes principaux par le modèle SOM pour distinguer les assemblages EPTC (Ephéméroptères, Plécoptères, Trichoptères, Coléoptères) dans les cours d eaux du Coteau de Gascogne, tributaires de la Garonne (sud-ouest France). Des faibles richesses et diversités ont été observées dans cette zone affectée par une pratique agricole intensive, caractérisée par de très fortes valeurs de TDS (Total Dissolved Solids), nitrate (NO3) et COD (Chemical Oxygen Demand). Les espèces EPTC tolérantes ont été utilisées comme paramètres de contrôle pour les changements d assemblage de communautés collectées dans des sites impactés par la pratique agricole. Ces constatations confirment que la pratique agricole a perturbé les conditions des cours d eau, avec le corridor de ripisylves qui sert 94

104 de zone tampon naturelle. Dans le chapitre III, le modèle SOM a montré divers groupes de communautés et a été efficace pour révéler l état de changement temporel de communautés. Les indices biologiques de la qualité de l eau correspondent bien aux différents groupes du modèle SOM. Le SOM pourrait être ainsi une méthode alternative d étude des communautés pour évaluer la qualité de l eau à large échelle, des longues séries de données, élargissant ainsi la place de l écologie informatique pour implémenter les évaluations des écosystèmes. Dans le chapitre IV, le SOM a été efficace pour identifier le patron des communautés en visualisant les caractéristiques des communautés, classant ainsi les échantillons correspondant à des régions géographiques et des degrés de perturbations. Les groupements obtenus révèlent les impacts de pollution : la richesse et l abondance des espèces reflètent en conséquence les conditions géographiques et anthropogéniques de perturbations, alors que certaines espèces tolérantes ont été sélectivement collectées dans des zones perturbées. Les communautés classées par le modèle SOM montrent un patron cohérent avec l abondance et la richesse. Le SOM est ainsi une méthode efficace pour relier les informations diverses sur la composition des communautés, la distribution géographique et le patron de l abondance des espèces d une manière intégrative. Le SOM pourrait ainsi servir comme méthode pour déterminer la carte écologique des écorégions et fournir une vue compréhensible de l état écologique de communautés échantillonnées à large échelle. Mots clés : macroinvertébrés benthiques, écosystème aquatique, évaluation, carte autoorganisatrice 95

105 Ecological quality assessment of stream ecosystems using benthic macroinvertebrates Abstract In this study, we applied the SOM for ecological assessment using benthic macroinvertebrates in aquatic ecosystem. First, Chapter I, SOM was utilized to extract information from complex data of environmental variables and benthic macroinvertebrate communities residing in different micro-habitats. Although the sampling was carried out in a limited area, the patterns of environmental variables revealed spatial heterogeneity. The clustering of benthic macroinvertebrate communities in the trained SOM was efficient in showing temporal variation and evaluating water quality according to the conditions of different micro-habitats. Consequently, local spatial heterogeneity is important in revealing dynamics of community abundance and biotic indices, especially regarding restoration processes in polluted streams. Chapter II, the samples were grouped into three main clusters corresponding to distinct EPTC assemblages in the tributary streams of the Garonne River catchment, southern France. Lower richness and diversity of macroinvertebrates were observed in the areas affected by agricultural land use, being associated with high Total Dissolved Solids (TDS), Nitrate (NO3) and Chemical Oxygen Demand (COD). Tolerant EPTC species were identified as controlling parameters for the changes in the assemblages collected at the agriculture-impacted sites. These findings confirmed that agricultural land use has disturbing effects on instream conditions of the riparian forest corridor that naturally serves as a buffer zone. In chapter III, the trained SOM showed the diverse scope of communities and were efficient in tracking temporal changes in community states. Biological water quality indices were correspondingly different in different clusters by the trained SOM. The SOM could be an alternative map for covering diverse scope 96

106 of community states in water quality monitoring for the large-scale, long-term data, thus would broaden the scope of ecological informatics implemented to ecosystem assessment. In chapter IV, the SOM was efficient in patterning and visualizing community characteristics and accordingly classified the samples corresponding to the geographical regions and to the degree of disturbances. The patterned groups accordingly revealed the impact of pollution: the species abundance and richness accordingly reflected geographical conditions and anthropogenic disturbances, while some tolerant species were selectively collected in the disturbed areas. The communities clustered by the trained network were also characterized with the species-abundant patterns. The SOM was efficient in linking various information on community composition, geographic distribution, and species abundance patterns in an integrative manner. The SOM could serve as an efficient ecological map for specifying ecoregions and for providing comprehensive view on ecological status of the communities sampled on a large scale. Key words : benthic macroinvertebrates, aquatic ecosystem, assessment, self-organizing map 97

107 ACKNOWLEDGEMENTS I am very grateful to Drs. Sovan Lek and Sithan Lek, my supervisors, in the Dept. of ECOLOGIE & EVOLUTION, de l'université Paul Sabatier Toulouse III, to finish my course fruitfully in France. And I am greatly appreciated to Dr. Tae-Soo Chon, my advisor and mentor, in the Dept. of Biology, Pusan National Univ. They gave a chance for the joint supervision of thesis between Pusan National Univ. and l'université Paul Sabatier Toulouse III. I m appreciating all of the committee members of the thesis evaluation, Dr. Michele Scardi, Proffesor in University Tor Vergata (Rome, Italy), Dr. Kei Tokita, Professor in University Osaka (Japan) and Dr. Gea-Jae Joo, Professor in Pusan National Univ.. All the professors in the Dept. of Biology I appreciate, Professors Tae-Hyung Ryu, Won- Ho Lee, Jung-Hee Hong, Won-Chul Choi, Young-Sang Kim, Byung-Ki Kim, Geun-Seob Lee, Eun-Sang Choi and Chang-Geun Kang. They beholded me so long time from the undergrate to the Ph. D. course. They gave invaluable lectures on the biology, and whenever I feel need on any information, they were enthusiastically solved my curiosity. I appreciate, Dr. Young-Seuk Park in the Kyunghee Univ., Dr. Inn-Sil Kwak in the Chonnam National Univ., supported valuable comments to me. In my laboratory in PNU, there were many of lab members and they gave me so much helps on my Ph. D. works. I appreciate Chang Woo Ji, Dr. Xiaodong Qu, Seung Tae Kim, Hyun Ju Hwang, Ji Eun Pakr, Kyung Im Jung, Ju Young Chung, Ae Jung Seo and Su Jin Song for their support on field trips and the comments on my manuscripts. Also Dr. Joo-Mi Lee, Voki Woo, Ju-Young Sohn and all other friends in the department including people in the department office I would like to thank. Although it was only one year (6 months in 2004 and 6 months in 2005) to stay in Toulouse, I got help from many people. I am appreciating all of the members in LADYBIO, Dr. 98

108 Alain Tomas, Dr. Sébastien Brosse, Dr. Géraldine Loot, Dr. Muriel Gevrey, Wafa Bouzid, Fabien Leprieur and all other friends. On this paper, I would like to express my deepest regret to my family. 99

109 ANNEX some published papers 100

110 ECOLOGICAL INFORMATICS 1 (2006) available at Characterization of benthic macroinvertebrate communities in a restored stream by using self-organizing map Mi-Young Song a, Young-Seuk Park b, Inn-Sil Kwak c, Hyoseop Woo d, Tae-Soo Chon a, a Divison of Biological Sciences, Pusan National University, Busan , Republic of Korea b Department of Biology, Kyung Hee University, Seoul , Republic of Korea c Bioresources Utilization Program, College of Fisheries and Ocean Sciences, Chonnam National University, Yeosu , Republic of Korea d River and Environments Research Group, Korea Institute of Construction Technology, Kyonggi-Do , Republic of Korea ARTICLE INFO Article history: Received 12 March 2005 Received in revised form 18 December 2005 Accepted 27 December 2005 Keywords: Macroinvertebrates Spatial heterogeneity Biotic indices Artificial neural network Polluted stream Restoration ABSTRACT The Self-Organizing Map (SOM) was used for revealing the ecological states of streams in recovery through patterning of benthic macroinvertebrate communities. SOM was capable of showing different clusters of the sample sites in a small scale according to changes in environmental variables such as water velocity, depth, substrate roughness and the amount of silt. Community abundance correspondingly varied in different clusters of the sample sites. Within each cluster, data for community abundance were further grouped according to temporal changes in water quality. The patterns of benthic macroinvertebrate communities in the trained SOM were efficient in assessing recovery processes in the polluted sample sites, revealing the effects of river restoration projects in stream ecosystems. The study showed that spatial heterogeneity at the local level plays an important role in characterizing community patterns and consequently biological water quality assessment Elsevier B.V. All rights reserved. 1. Introduction With the advantages of taxonomic diversity, sedentary in behaviors and long life cycles, benthic macroinvertebrates characteristically respond to anthropogenic disturbances in an integrated and continuous manner, and consequently have been widely used for assessing the water quality and ecological status of aquatic systems (Resh and Rosenberg, 1984; Hellawell, 1986; Rosenberg and Resh, 1993). There have been numerous studies on community characterization and water quality evaluation over a broad scope from clean to severely polluted states (Hellawell, 1986; Rosenberg and Resh, 1993; Barbour et al., 1996; Reynoldson et al., 1997; Richards et al., 1997; Davies et al., 2000; Wright et al., 2000; Chon et al., 2002; Butcher et al., 2003). In most cases, the surveys have been carried out in a large scale in the order of km between the sample sites. Species traits are usually distinct and community compositions are easy to characterize between different sites in these cases (Townsend and Hildrew, 1994; Death, 1995; Resh et al., 1994; Poff, 1997; Rabeni et al., 2002; Lamouroux et al., 2004). Spatial heterogeneity, however, still exists in a small scale (e.g., less than one kilometer) in lotic conditions according to hydro-morphological characteristics of streams (e.g., pool, riffle, etc.) (Minshall, 1988; Poff and Ward, 1990; Townsend and Hildrew, 1994; Copper et al., 1997; Palmer et al., 1997; Palmer and Poff, 1997; Poff, 1997; Poole, 2002). Accordingly, the community patterns are variable in different habitats in a small scale (Lancaster and Hildrew, 1993; Resh et al., 1994; Townsend and Hildrew, 1994; Armitage and Cannan, 2000; Brown, 2003; Roy et al., 2003; Gebler, 2004), although overall composition of species may not be widely variable as shown in a large scale. Benthic macroinvertebrates have been Corresponding author. Fax: address: [email protected] (T.-S. Chon) /$ - see front matter 2006 Elsevier B.V. All rights reserved. doi: /j.ecoinf

111 296 ECOLOGICAL INFORMATICS 1 (2006) investigated in a relatively small scale of hundred-meter distances (e.g., Cummins and Lauff, 1969; Hildrew et al., 1980; Lancaster and Hildrew, 1993; Brown, 2003; Roy et al., 2003). There have been a limited number of studies that reveal the relationships between hydraulic variables and distribution of macroinvertebrates at micro-habitats (Lancaster and Hildrew, 1993; Lamouroux et al., 2004; Brooks et al., 2005). Downes et al. (1993) described important hydro-morphological factors influencing the spatial distribution of invertebrates after investigating small scale patchness. Brooks et al. (2005) recently demonstrated the importance of small-scale differences in hydraulic conditions characterized by water velocity, depth and substrate roughness in determining the spatial distribution of macroinvertebrate assemblages in riffle habitats. The characteristics of small scale habitats are important factors for the success of stream restoration activities. Monitoring in a small scale habitat, heterogeneity provides a measure of ecosystem restoration (Pik et al., 2002). For example, Purcell et al. (2002) evaluated the effects of restoration of a small stream using benthic macroinvertebrate communities. In this study, we intend to reveal changes in macroinvertebrate communities intensively collected within a limited area in a midstream reach of a polluted stream after a restoration project. We investigated spatial heterogeneity by selecting the sample sites short distances (5 10 m) apart, and correspondingly characterized abundance patterns of benthic macroinvertebrates in different habitats. 2. Materials and methods 2.1. Study sites The field survey was carried out within a 200 m reach (Hakyeoul) in the Yangjae Stream, a tributary of the Han River in the south of Seoul, Korea ( N, E) (Fig. 1). It flows through the metropolitan and agricultural areas in the city, and has been mainly polluted with organic matter (KICT, 1997). The stream has a year-round flow of m in width and cm in depth. Water discharge rapidly increases in the period of summer flooding and decreases in the dry winter season. Recently, a campaign for water recovery has been carried out by the local government. Benthic macroinvertebrates were sampled at 11 sampling sites in the study area based on location, hydromorphological characters and a 5 10 m distance between sites (Fig. 1). According to topographical conditions the sample area was partitioned into three zones: 1) the upstream zone covered the sample sites located in the upper part of the sample area, and was characterized by large substrates and relatively high water velocity (sites, U, A, B, C and D), 2) the downstream zone included the sample sites with lower velocity and higher sedimentation at the lower part of the sample area (sites H, I, and J) (Fig. 1), and 3) the pool zone was located at the curved down stream area and were Fig. 1 The sample sites located in the Yangjae Stream in the Han River, Korea.

112 ECOLOGICAL INFORMATICS 1 (2006) characterized with low values in water depth and velocity, and high levels of sedimentation (sites E, F, and G). Except sites U and J which were located at the upper part and the lower part of the study sites, respectively, longitudinal or curvilinear sampling was carried out to reveal local topographic characters of the sample sites. The survey area has been partially manipulated for a restoration project of the stream. Vegetation channel revetment technique was carried out in the riparian zone close to site A (Fig. 1)(KICT, 1997). The stream bed was artificially planted with large-sized cobbles approximately 5 cm in diameter. The upstream islet, near where site F was located, was artificially constructed for management of siltation as well as for providing habitats for birds and other animals, while the downstream islet, around where G was sampled, was naturally formed. Restoration was also carried out on the riparian zone close to site D, but the stream bed was not affected by the restoration project in this case. Site E was located close to the edge area at the pool zone, and silts were highly accumulated around this site. We selected four environmental variables to reveal spatial heterogeneity in the sample area. Firstly, the water depth and velocity were used to represent hydromorphological characters of the sample sites. Secondly, considering that topographic conditions of streams affect stream beds, substrate roughness was recorded to represent spatial heterogeneity of the sample sites (Statzner et al., 1988; Poff and Ward, 1990). Lastly, the percentage of silt was measured to represent a fine level of substrate compositions, according to Statzner et al. (1988). The substrate composition in each site was measured in different diameters (D): coarse cobbles (mean D sizes 100 mm), fine cobbles (50 mm D<100 mm), pebbles (30 mm D<50mm), fine pebbles (16 mm D<30 mm), coarse gravel (8 mm D<16 mm), and the smaller substrates (4 mm D<8 mm, 2 mm D<4 mm, 1 mm D<2 mm, 0.5 mm D<1 mm, 0.25 mm D<0.5 mm, mm D<0.25 mm and D<0.125 mm)(cummins and Lauff, 1969). The volumes of larger substrates ( 8 mm) were determined by the volumetric bucket in the field, while substrates smaller than 8 mm in diameter were separately sampled in plastic containers (50 ml) in triplications. Substrate roughness was expressed as K=(5C 1 +3C 2 +C 3 )/9, where the subscripts 1, 2 and 3 represent the 1st, 2nd and 3rd most dominant substratum type, respectively (Statzner et al., 1988). Coarseness value C was correspondingly assigned to the size of the dominant substrates: 1, 2, 3 and 4 if the size, k, is in the range of k<0.125 mm, 0.5 mm k<4 mm, 8 mm k<30 mm, k 30 mm, respectively. The coarseness classes have been slightly modified from those given by Statzner et al. (1988) to the size used in this study Community data At each sample site, three to four benthic samples were collected with the Surber sampler (30 cm 30 cm, 500 μm mesh; APHA et al., 1985) approximately 10 cm in depth at monthly intervals for two years starting in April The collected macroinvertebrates were preserved in 7% Formalin solution. In the laboratory, the invertebrate specimens were sorted, identified to genus level and counted for the number of specimens under microscopes. Identification was based on Yun (1988), Brighnam et al. (1982), Merritt and Cummins (1984), Pennak (1978), and Quigley (1977). Chironomidae was separately identified based on Wiederholm (1983), while Oligochaeta was checked with Brinkhurst and Jamieson (1971) and Brinkhust (1986). In the datasets, 24 genera were identified, showing that only a few taxa were highly abundant at the polluted sample sites. Chironomidae, abundant with Chironumus sp., and Oligochaeta, mostly consisting of Limnodrilus hoffmeisteri (Tubificidae), were the dominant taxa. In order to represent water quality of the sample sites, species richness (SR) and BMWP (Biological Monitoring Working Party, Walley and Hawkes, 1997), two conventionally used biotic indices, were estimated from the sampled community data Modeling procedure First we defined hydro-morphological patterns of the sample sites based on four environmental variables (depth, velocity, substrate roughness and the percentage of silt) by the learning process of SOM. Subsequently, we further revealed temporal changes in macroinvertebrate communities at the selected sites by SOM. Both environmental variables and community data were scaled between 0 and 1 in the range of the minimum and maximum values within each variable. In order to reduce high variation, abundance data used for training with SOM was log-transformed before the analysis of each taxon. SOM is an adaptive unsupervised learning algorithm and approximates the probability density function of the input data (Kohonen, 2001). SOM consists of input and output layers connected with computational weights (connection intensities). The array of input nodes (i.e., computational units) operates as a flow-through layer for the input vectors, whereas the output layer consists of a two-dimensional network of nodes arranged in a hexagonal lattice. In the learning process of SOM, initially the input data (data matrix for either environmental variables or taxa abundance in this study) were subjected to the network. Each raw input vector consisting of the values for different environmental variables (or abundance data in different taxa) was provided sequentially as input data. In this case the number of the input node was equal to the number of variables (or number of taxa), while the output layer consisted of N output nodes (i.e., computational units) which usually constitute a 2D grid for better visualization. Subsequently, the weights of the network were trained for a given dataset. Weights were initially generated as small random numbers. Each node of the output layer computes the summed distance between weight vector and input vector. The output nodes are considered as virtual units to represent typical patterns of the input dataset assigned to their units after the learning process. Among all virtual units, the best matching unit (BMU), which has the minimum distance between weight and input vectors, becomes the winner. For the BMU and its neighborhood units, the new weight vectors are updated by the SOM learning rule. This results in training the network to classify the input vectors by the weight vectors they are closest to. The detailed algorithm of SOM for ecological applications can be found in Chon et al. (1996) and Park et al. (2003a). After training, the Ward s linkage method based on the Euclidian distance (Ward, 1963) was applied to the weights of the nodes in SOM for further clustering (Jain and Dubes, 1988; Park et al., 2003b). After preliminary training, we used N=80

113 298 ECOLOGICAL INFORMATICS 1 (2006) (10 8) of SOM output units for patterning samples with environmental data, and N=20 (5 4) units for patterning temporal variation of community at the selected sites. We used the functions provided in the SOM toolbox (Alhoniemi et al., 2000) in Matlab (The Mathworks, 2001) Statistical analysis To test the null hypotheses of no significant differences in environmental variables, taxa abundance and biotic indices in different clusters, we carried out nonparametric multiple comparisons after the Kruskall Wallis test with the unequal number of samples (Zar, 1999), considering wide variations in community data and the different number of the samples in clusters patterned by SOM. 3. Results 3.1. Classification of sample sites When the sample sites were trained with SOM (80=10 8 nodes) based on environmental variables, two large clusters (I and II) were formed according to the dendrogram of Ward s linkage method (Fig. 2a, b). SOM output units were further subclustered into four groups (Ia, Ib, IIa, and IIb). At the highest level, SOM units were vertically divided: cluster I for the upper area and cluster II for the bottom area in the map (Fig. 2a). This grouping coincided with the locations of the sample sites. In cluster I, the samples were collected in both pool and downstream zones, including sites E to J (Fig. 1), where water velocity was relatively slower and small size substrates were more abundantly present. Additionally, some samples collected at sites C and D were grouped in cluster I. In contrast to cluster I, cluster II accommodated the sample sites (U, A, B, etc.) belonging to the upstream zone with the large size substrates where water velocity was higher. Subclustering was further obtained based on spatial variation. Within cluster I, cluster Ia was mainly represented by the pool zone such as sites E, F and G, while cluster Ib was more associated with the sample sites located at the downstream zone, such as H and J (Fig. 2a). Within cluster II, subclustering was formed based on the location of the sample sites and environmental factors. Samples located in the upstream zone were mainly grouped in cluster IIa (e.g., sample sites U, A, B, etc.), while cluster IIb accommodated various sample sites with high levels of water depth and velocity. A majority of the samples collected in summer were grouped in cluster IIb, and the flooding effect appeared more clearly at the sites belonging to the downstream zone. Environmental variables varied in different clusters of the sample sites (Table 1). Cluster II was characterized by a higher velocity, higher substrate roughness and lower percentage of silt, and vice versa for cluster I. Within cluster II, cluster IIa was differentiated from cluster IIb by higher levels of substrate Fig. 2 Classification of the sample sites by SOM based on environmental variables measured in the Yangjae Stream in a small scale from April 1996 to March 1998: a) sample sites, and b) cluster analysis with the Ward s linkage method. Acronyms in units stand for the samples: the letter represents the name of sample sites (see Fig. 1), while the numbers indicate the year and month of collection (e.g., U9604; site U collected in April 1996, A9711; site A collected in November 1997; and F9803; site F collected in March 1998).

114 ECOLOGICAL INFORMATICS 1 (2006) Table 1 Summary of environmental variables in averages (min max) and list of abundant taxa in different clusters defined by SOM (Fig. 2) Clusters Ia Ib IIa IIb Number of samples Zones Pool Downstream Upstream Downstream Environmental variables Depth 7.08 ( ) d ( ) b ( ) c ( ) a Velocity ( ) d ( ) c ( ) b ( ) a Roughness 1.54 ( ) c 1.51 ( ) d 2.78 ( ) a 1.88 ( ) b Silt (%) 6.01 ( ) a 3.83 ( ) b 1.61 ( ) c 2.2 ( ) c Representative species Chironomus (sp.) Orthocladius (sp.) Chironomus (sp.) Tanypus (sp.) Cricotopus (sp.) Tanypus (sp.) Orthocladius (sp.) Limnodrilus hoffmeisteri Tanypus (sp.) Erpobdella (sp.) Cricotopus (sp.) Glossiphora (sp.) Limnodrilus hoffmeisteri Glossiphora (sp.) Tanypus (sp). Glossiphora (sp.) Erpobdella (sp.) Physa (sp.) Glossiphora (sp.) Physa (sp.) The same letters listed in superscript of environmental variables indicate no significant difference (p>0.05) between clusters based on the nonparametric Kruskall Wallis test with unequal number of samples. roughness. Cluster Ia was mainly different from cluster Ib by a higher percentage of silt. Overall, environmental factors in different clusters revealed heterogeneity of the sample sites. The percentage of silt was higher at the pool zone of sites E and G, whereas lower at the upstream zone of sites U and A. Environmental variables in each cluster were made distinct by using the nonparametric Kruskall Wallis test (Table 1). Differences in depth, velocity and substrate roughness were statistically significant between all clusters. The variables were generally higher in cluster II than in cluster I. Depth and velocity were in the lowest range in cluster Ia, while in the highest range in cluster IIb. Substrate roughness was lower in cluster I and higher in cluster II, showing the lowest value in cluster Ib and the highest value in cluster IIa. The percentage of silt was the same level between clusters IIa and IIb, while the percentage of silt showed the highest value in cluster Ia. The statistical results revealed differences of environmental variables in different clusters. Fig. 3 Mean abundance of the selected taxa in different clusters corresponding to SOM (Fig. 2). The height of the bar represents the mean, while the whisker indicates the confidential interval (mean±0.95). The same letters located on top of the whisker indicate no significant difference ( p>0.05) between clusters based on the nonparametric Kruskall Wallis test with the unequal number of samples. Among the sampled communities, the taxa with statistical differences were only presented in the figure.

115 300 ECOLOGICAL INFORMATICS 1 (2006) Abundances of the selected taxa in benthic macroinvertebrates varied with different clusters (Fig. 3). A majority of taxa were abundant in subcluster IIa. Orthocladius sp. and Cricotopus sp., which are known to be present in recovering water (Ferringto and Crisp, 1989), were highly collected in subcluster IIa. Additionally, Tanypus sp. and other invertebrates including Erpobdella sp. and Glossiphora sp. showed the highest level of abundance in this subcluster. However, the tolerant species such as Chironomus sp. and L. hoffmeisteri showed different patterns, being most abundant in cluster Ia (Table 1, Fig. 3). Overall, environmental variables in different clusters were associated with different patterns of taxa abundance. Cluster IIa was characterized with the highest value of substrate roughness and the lowest value of the percentage of silt (Table 1). Abundance of various taxa was associated with this subcluster including Chironomus sp., Orthocladius sp., Cricotopus sp., Tanypus sp., Erpobdella sp., Glossiphora sp., and Physa sp. In comparison with cluster IIa, cluster IIb was represented by the highest level of water depth and velocity due to high precipitation in summer. In this subcluster, Tanypus sp., L. hoffmeisteri and Glossiphora sp. were abundantly present. Subcluster Ia was further characterized with the lowest level of water depth and velocity, and with the highest level of percentage of silt (Table 1). Subcluster Ia covered the sites E, F and G at the pool zone of the survey area (Fig. 2). The tolerant taxa L. hoffmeisteri and Chironomus sp. were most abundant in this subcluster. L. hoffmeisteri and Chironomus sp. belong to Tubificidae and Chironomidae, respectively, and both taxa are burrowing or tube building types, being commonly associated with soft, depositing area (Hellawell, 1986). In this subcluster, the percent of silt was at the highest level (Table 1). Cluster Ia was additionally associated with Cricotopus sp., Glossiphora sp. etc. In comparison with subcluster Ia, subcluster Ib was relatively higher in water depth and velocity, and was relatively lower in the amount of silt. This indicated a smaller effect of sedimentation on the sample sites (e.g., H and I) in the downstream zone. In this subcluster, the associated taxa were not as diverse as shown in cluster Ia. Orthocladius sp., Erpobdella sp. and Glossiphora sp. were each abundant in cluster Ib Temporal variation of communities After patterning the samples based on the four environmental variables, we further chose the sample sites E, H and A, which would typically represent different clusters Ia, Ib, and IIa based on the differences of environmental factors, respectively (Fig. 2a), to reveal temporal variation of communities at Fig. 4 Patterning of temporal changes in macroinvertebrate communities collected at site A from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters.

116 ECOLOGICAL INFORMATICS 1 (2006) different habitat types. The site representing cluster IIb was not chosen since the cluster was mixed with various sample sites, and a majority of the samples were observed in the temporally unstable period of summer in this subcluster (Fig. 2a). Community abundance data were subsequently trained with SOM (Figs. 4 6) for each site. Temporal patterns of macroinvertebrates were identified in the map. At site A representing cluster IIa, the sample sites were further divided to two main clusters with each cluster producing two more subclusters (Fig. 4a, b). Clusters A1 and A2 were separated according to different sampling periods. Subclusters showed further temporal changes sequentially from A1a (April 1996 January 1997), A1b (February May 1997), A2a (June August 1997) to A2b (September November 1997) in the order of sampling time (Fig. 4a, b). In the last phase, however, the samples collected in this period belonged again to A2a (January March 1998) after A2b (September November 1997). Community abundance patterns were correspondingly different in various subclusters. Different taxa appeared in a diverse manner as time progressed (Fig. 4c). In subcluster A1a in the earliest period, abundant taxa were not observed. The following subcluster A1b, mainly observed in March 1997, was selectively matched to Enchytraeus sp. Both tolerant genus (e.g., Chironomus sp.) and intolerant genus (e.g., Orthocladius sp., Cricotopus sp.) were also commonly grouped in this cluster. Two genera, however, were further associated with subcluster A2a in the next phase. The appearance of recovering species such as Orthocladius sp. and Cricotopus sp., indicated the recovery of water quality in the periods corresponding to A1b A2a (February 1997 March 1998). A large number of taxa were additionally present in the following subcluster A2a. Especially, Baetis, a well known genus appearing in recovering water (Hellawell, 1986), was observed in this subcluster. The appearance of Baetis sp. confirmed the phase of water recovery in the period matching to cluster A2a (Fig. 4c). In the following subcluster A2b, even a larger number of taxa were associated (10 genus) including various aquatic insects and other invertebrates (e.g., Erpobdella sp. and Physa sp.). L. hoffmeisteri, which was the dominant species in the survey area, showed different patterns of occurrence compared with other species. The species was highly associated with subcluster A2b, but was also abundant over a broad area of the map (Fig. 4c). Biotic indices such as SR and BMWP changed accordingly with time (Fig. 4d). The clusters listed below x-axis (month) in Fig. 4d indicate the sample communities patterned by SOM Fig. 5 Patterning of temporal changes in macroinvertebrate communities collected at site H from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters (Fig. 2).

117 302 ECOLOGICAL INFORMATICS 1 (2006) (Fig. 4a). Subclusters appeared along with changes in biotic indices as time progressed. SR and BMWP increased gradually, peaking in October 1997 (Fig. 4d). The changes in indices represented the trend of water quality improvement after the restoration project. Main clusters A1 and A2 were separated according to the sampling time, June The samples grouped in cluster A1 were mainly collected in the early sampling period up to May 1997, whereas the samples in cluster A2 were in the later sampling period starting from June Fig. 5 shows temporal changes in macroinvertebrate communities collected at the sample site H, which represents the downstream zone of the sampling area defined in cluster Ib (Fig. 2a). Similar to site A, the samples were grouped to two main clusters according to the sampling periods (Fig. 5a, b). Cluster H1 mainly represented the earlier periods of the survey, whereas the samples collected at the later periods were grouped in cluster H2. Each cluster was further divided to two subclusters. However, the degree of temporal patterning at the level of subclusters was not as strong as shown in the subclustering at site A (Fig. 4). Different taxa were associated with different clusters at site A (Fig. 5c). At the early phase in cluster H1, however, not many taxa were abundant. Subcluster H1a was loosely associated with Baetis sp., while subcluster H1b was grouped with 2 species of Ceratopsche sp. and Tubifex tubifex (Fig. 5c). Subclusters in H2, however, were associated with diverse taxa. Both tolerant (e.g., Chironomus sp.) and intolerant (e.g., Cricotopus sp.) species were grouped in subcluster H2a. Occurrences of Cricotopus sp. along with the tolerant species, Chironomus sp., indicated recovery of water at this phase. Subcluster H2b was also diversely related to various species but were different from species composition observed in cluster H2a. Species in Chironomidae such as Orthocladius sp. and Tanypus sp. were more associated with cluster H2b. Similar to site A, the tolerant species were abundant and showed different patterns in abundance. Chironomus sp. was grouped in subclusters H2a and H2b, while L. hoffmeisteri was broadly present, covering H1b and H2b on the map (Fig. 5c). Biotic indices also varied according to different clusters at site H (Fig. 5d). However, the temporal patterns were not as distinct as shown at site A (Fig. 4d). While SR and BMWP were in the ranges of 0 15 and 0 40, respectively at site A, the indices were in narrower ranges, 0 10 and 0 30 at site H, respectively. The samples were observed in different temporal periods at the level of main clusters H1 and H2 (Fig. 5d). This type of grouping was revealed at the level of subclusters at site Fig. 6 Patterning of temporal changes in macroinvertebrate communities collected at site E from April 1996 to March 1998: a) clustering of the samples in temporal variation, b) Euclidian distance between clusters, c) distribution patterns of species abundance (log-transform of (individuals/m 2 ); indicated along with the vertical bar) in the selected taxa in different clusters. Darker color represents higher values of each variable, d) temporal variation of SR and BMWP and their corresponding clusters.

118 ECOLOGICAL INFORMATICS 1 (2006) Table 2 Comparison of SR and BMWP in different clusters defined by SOM (Figs. 4 6) Index Site 1st cluster (A1, H1, E1) 2nd cluster (A2, H2, E2) SR A 5.0 (1.6) b 9.9 (2.4) H 3.1 (1.3) 5.6 (2.2) E 3.2 (1.3) 6.4 (1.5) BMWP A 8.8 (2.7) 20.0 (8.4) H 4.2 (2.4) 12.2 (6.6) E 4.6 (2.3) 15.0 (6.9) a b Probability based on the nonparametric Kruskall Wallis test. Mean (SD). A(Fig. 4d). Separation of the sampling time was observed in September 1997 at site H. Fig. 6 shows temporal changes in macroinvertebrate communities collected at site E representing cluster Ia (Fig. 2a). Similar to the other sites, the samples were divided into two main clusters with subdivisions (Fig. 6a, b). Grouping patterns, however, appeared differently. Most samples were associated with cluster E1b, while only a few samples were sparsely located in other clusters. Cluster E1 was mainly sampled in the early periods of the survey, whereas cluster E2 was collected in the later periods (Fig. 6a). Subclustering was mostly based on differences in community abundance. Subclusters in cluster E1 were associated with the selected taxa (Fig. 6c). Cluster E1a was related to Tanypus (sp.) and Carabidae (sp.). Although a large number of sample sites were grouped (Fig. 6a), cluster E1b was associated with a limited number of taxa such as Lumbriculus (sp.) and Enchytraeus (sp.). Communities collected at the later periods in cluster E2a, however, were diverse, being grouped with Erpobdella sp., Orthocladius sp., Glossiphora sp., etc. Other taxa such as Physa (sp.), Cricotopus (sp.), and Chironomus (sp.) also appeared in cluster E2a, but they were also associated with E2b. Baetis (sp.), however, was randomly present in cluster E2b (Fig. 6c). Tolerant species, L. hoffmeisteri and T. tubifex showed the patterns similar to site A. They were broadly abundant over clusters E1b and E2b. Biotic indices were also variable at site E in response to temporal changes (Fig. 6d). SR and BMWP showed similar patterns as those of site A (Fig. 4d). However, ranges of biotic indices were not as wide as at site A. The sample sites were mixed at the level of subclusters, but were differentiated by the sampling periods at the level of the main clusters. Samples in cluster E1 were collected mostly at the early sampling period up to July 1997, whereas samples in cluster E2 were present at the later sampling periods beginning in August At sites A, E, and H, both biotic indices SR and BMWP were significantly different between the early sampling periods and the later sampling periods (Table 2), representing recovery of water quality at the sampling sites after the restoration project. P a through SOM. Based on four input variables (depth, velocity, substrate roughness and silt (%)), the clusters were defined in a hierarchical manner depending on the impact of hydromorphological factors. The results confirmed the variation of the community abundance depending upon spatial heterogeneity in the small scale (Table 1, Fig. 3)(Brown, 2003; Roy et al., 2003; Brooks et al., 2005), and in biotic indices along with different time periods (Figs. 4d, 5d, 6d, Table 2). This study demonstrated that improvement of water quality could be differently monitored at the sample sites in a small scale. In general, water quality changes were more clearly addressed at site A than at sites E and H (Figs. 4 6). Communities varied depending upon the hydro-morphological condition of the sample sites even though the locations are about 5 10 m apart. As stated above, large substrates (>30 mm) were artificially planted at site A (Fig. 7). At site H, which was adjacent to A, but not planted with large substrates, substrate in smaller size ranging 0.5 mm 2 mm were accumulated more abundantly (Fig. 7). Meanwhile substrate at site E was mainly composed of smaller particles less than 0.5 mm in diameter. These differences in substrates were important in characterizing the spatial heterogeneity and were therefore, crucial in determining community abundance patterns at micro-habitats. Considering that large substrates were artificially planted for the restoration project at site A, it could be stated that higher amounts of large-size substrates would accommodate diverse community composition and would reveal sensitive changes in communities, especially in the recovery of water. Consequently, large substrates used in the restoration project 4. Discussion The different patterns of the sample sites intensively collected in a polluted stream in the small scale were elucidated Fig. 7 Comparison of substrate volumes in different particle size: a) proportion (%) of substrates among different sample sites and b) nonparametric comparison of volumes between different sample sites in different substrate sizes.

119 304 ECOLOGICAL INFORMATICS 1 (2006) would be useful for monitoring community changes, confirming the previous reports that maximizing heterogeneity in ecological restoration projects may promote diverse communities and may be useful for the management of aquatic communities (Brown, 2003). There were a few samples sites unexpectedly mixed with other clusters. The sites C and D belonged to the upstream zone (Fig. 1), but a majority of the samples at these sites were grouped in cluster Ia for the pool zone along with the sites E, F and G with SOM (Fig. 2). Hydro-morphological characters of the sites C and D, obliquely located across the sites A and B (Fig. 1), however, were closer to the sites in the pool zone, and were consequently characterized by slower water velocity and the increased amount of silt. Considering that biotic indices may be different according to spatial heterogeneity, consistency in selecting habitats for evaluation of biotic indices may be required. The community patterns appeared differently in each habitat (Table 1, Fig. 3), and in different temporal periods within the sample sites (Figs. 4 6). This type of sampling consistency could be further discussed in the future along with the problem of habitat suitability and patchness in spatial distribution. Community development could be observed in regressive succession for recovery (Hawkes, 1979; Sládeèek, 1979; Hellawell, 1986) in the observed data. In recovering water, community compositions become more diverse, and these consequences were observed in community development at the sample sites in the period of water recovery (April 1996 October 1997) (Figs. 4d, 5d, 6d). The genus indicating recovery of water quality such as Orthocladius (sp.), Cricotopus (sp.) and Tanypus (sp.) were collected from February 1997 to March Densities of Baetis (sp.) were high from July to September 1997, while they were not collected in the corresponding periods in At the phase of higher biotic indices (July November 1997), species in Coleoptera, Ephemeroptera and Odonata were diversely collected at low densities (Figs. 4c d, 5c d, 6c d). The occurrence patterns of recovering species, however, were variable according to the location of the sample sites. The indicator genus such as Baetis (sp.), Orthocladius (sp.), Cricotopus (sp.) and Tanypus (sp.) were collected over a longer period at site A starting from February At site E, the indicator species were also selectively present. Certain species appeared in a longer period. Orthocladius (sp.), for instance, was present from February to October 1997 at site E. These differences indicate that studies on community development in recovery should be also related to spatial heterogeneity. Especially in the situation of disturbance with organic pollution, spatial condition would be greatly affected by the changes in composition of substrates. This type of study should be closely checked with the restoration process and water recovery. However, the topic on the restoration project is beyond the scope of the current study and could be further investigated in the future. 5. Conclusion SOM was utilized to extract information from complex data of environmental variables and benthic macroinvertebrate communities residing in different micro-habitats. Although the sampling was carried out in a limited area, the patterns of environmental variables revealed spatial heterogeneity. The clustering of benthic macroinvertebrate communities in the trained SOM was efficient in showing temporal variation and evaluating water quality according to the conditions of different micro-habitats. Consequently, local spatial heterogeneity is important in revealing dynamics of community abundance and biotic indices, especially regarding restoration processes in polluted streams. Additionally, this study showed that SOM could contribute to broadening the scope of artificial neural networks as a means of unsupervised learning in assessing community characters and water quality for sustainable management of aquatic ecosystems. Acknowledgments This work was supported by the Korea Research Foundation Grant (KRF CP0413). REFERENCES Alhoniemi, E., Himberg, J., Parhankangas, J., Vesanto, J., SOM Toolbox. [online] APHA, AWWA, WPCF, Standard Methods for the Examination of Water and Waste, 16th ed. Washington D.C., 1134 pp. Armitage, P.D., Cannan, C.E., Annual changes in summer patterns of mesohabitat distribution and associated macorinvertebrate assemblages. Hydrol. Process. 14, Barbour, M.T., Gerritsen, J., Griffith, G.E., Frydenborg, R., McCarron, E., White, J.S., Bastian, M.L., A framework for biological criteria for Florida streams using benthic macroinvertebrates. J. N. Am. Benthol. Soc. 15, Brighnam, A.R., Brighnam, W.U., Gnika, A., Aquatic Insects and Oligochaetea of North and South Carolina. Midwest Aquatic Enterprise. Brinkhust, R.O., Guide to the freshwater aquatic microdrile Oligachaetes of North America. Canadian Special Publication of Fisheries and Aquatic Sciences pp. Brinkhurst, R.O., Jamieson, B.G.M., Aquatic Oligochaeta of the World. Oliver and Body, Edinburgh. Brooks, A.J., Haeusler, T., Reinfelds, I., Williams, S., Hydraulic microhabitats and the distribution of macroinvertebrate assemblages in riffles. Freshw. Biol. 50, Brown, B.L., Spatial heterogeneity reduces temporal variability in stream insect communities. Ecol. Lett. 6, Butcher, J.T., Stewart, P.M., Simon, T.P., A benthic community index for streams in the northern lakes and forests ecoregion. Ecol. Indic. 3, Chon, T.-S., Park, Y.S., Moon, K.H., Cha, E.Y., Patternizing communities by using an artificial neural network. Ecol. Model. 90, Chon, T.-S., Kwak, I.-S., Song, M.-Y., Park, Y.-S., Cho, H.D, Kim, M.J., Cha, E.Y., Lek, S., Benthic macro invertebrates in streams of South Korea in different levels of pollution and patterning of communities by implementing the self-organizing mapping. In: Lee, D. (Ed.), Ecology of Korea. Bumwoo Publishing Company, Seoul, pp Copper, S.D., Barmuta, L., Sarnelle, O., Kratz, K., Diehl, S., Quantifying spatial heterogeneity in streams. J. N. Am. Benthol. Soc. 16, Cummins, K.W., Lauff, G.H., The influence of substrate particle size on the microdistribution of stream macrobenthos. Hydrobiologia 34,

120 ECOLOGICAL INFORMATICS 1 (2006) Davies, N.M., Norris, R.H., Thomas, M., Prediction and assessment of local stream habitat features using large-scale catchment characteristics. Freshw. Biol. 45, Death, R.G., Spatial patterns in benthic invertebrate community structure: products of habitat stability or are they habitat specific? Freshw. Biol. 33, Downes, B.J., Lake, P.S., Schreiber, E.S.G., Spatial variation in the distribution of stream invertebrates: implications of patchiness for models of community organization. Freshw. Biol. 30, Ferringto, L.C., Crisp, N.H., Water chemistry characteristics of receiving streams and the occurrence of Chironomus riparius and other chironomidae in Kansas. Acta Biol. Debr., Suppl. Oecol. Hung. 3, Gebler, J.B., Mesoscale spatial variability of selected aquatic invertebrate community metrics from a minimally impaired stream segment. J. N. Am. Benthol. Soc. 23, Hawkes, H.A., Chapter 2. Invertebrates as indicators of river water quality. In: James, A., Evision, L. (Eds.), Biological Indicators of Water Quality. John Wiley and Sons, Chishester, Great Britain. Hellawell, J.M., Biological Indicators of Freshwater Pollution and Environmental Management. Elsevier, London. 546 pp. Hildrew, A.G., Townsend, C.R., Henderson, J., Interactions between larval size, microdistribution and substrate in the stoneflies of an iron-rich stream. Oikos 35, Jain, A.K., Dubes, R.C., Algorithms for Clustering Data. Prentice-Hall, Englewood Hills, NJ. KICT (Korea Institute of Construction Technology), Development of Close-Tonature River Improvement Techniques (CTNRIT) Adapted to the Korean Streams. Report, vol. 2. Kohonen, T., Self-Organizing Maps, 3rd ed. Springer, Berlin. Lamouroux, N., Doledec, S., Gayraud, S., Biological traits of stream macroinvertebrate communities: effects of microhabitat, reach, and basin filters. J. N. Am. Benthol. Soc. 23, Lancaster, J., Hildrew, A.G., Flow refugia and the microdistribution of lotic macroinvertebrates. J. N. Am. Benthol. Soc. 12, Merritt, R.W., Cummins, K.W., An Introduction to the Aquatic Insects of North America. Hunt Publishing Company, Dubugue. 722 pp. Minshall, G.W., Stream ecosystem theory: a global perspective. J. N. Am. Benthol. Soc. 7, Palmer, M.A., Poff, N.L., The influence of environmental heterogeneity on patterns and processes in streams. J. N. Am. Benthol. Soc. 16, Palmer, M.A., Hakenkamp, C.C., Nelson-Baker, K., Ecological heterogeneity in streams: why variance matters. J. N. Am. Benthol. Soc. 16, Park, Y.-S., Céréghino, R., Compin, A., Lek, S., 2003a. Applications of artificial neural networks for patterning and predicting aquatic insect species richness in running waters. Ecol. Model. 160, Park, Y.-S., Chang, J., Lek, S., Cao, W., Brosse, S., 2003b. Conservation strategies for endemic fish species threatened by the Three Gorges Dam. Conserv. Biol. 17, Pennak, R.W., Freshwater Invertebrates of the United States. John Wieley and Sons, Inc., New York. 803 pp. Pik, A., Dangerfield, J.M., Bramble, R.A., Angus, C., Nipperess, D.A., The use of invertebrates to detect small-scale habitat heterogeneity and its application to restoration practices. Environ. Monit. Assess. 75, Poff, N.L., Landscape filters and species traits: towards mechanistic understanding and prediction in stream ecology. J. N. Am. Benthol. Soc. 16, Poff, N.L., Ward, J.V., Physical habitat template of lotic systems: recovery in the context of historical pattern of spatiotemporal heterogeneity. Environ. Manage. 14, Poole, G.C., Fluvial landscape ecology: addressing uniqueness within the river discontinuum. Freshw. Biol. 47, Purcell, A.H., Friedrich, C., Resh, V.H., Restoration project in northern California. Restor. Ecol. 10, Quigley, M., Invertebrates of Streams and Rivers. Edward A. (publishers) Ltd, Colchester. 84 pp. Rabeni, C., Doisy, K.E., Galat, D.L., Testing the biological basis of a stream habitat classification using benthic invertebrates. Ecol. Appl. 12, Resh, V.H., Rosenberg, DM., The Ecology of Aquatic Insects. Praeger Publishers, New York. 625 pp. Resh, V.H., Hildrew, A.G., Statzner, B., Townsend, C.R., Theoretical habitat templets, species traits and species richness: a synthesis of long-term ecological research on the upper Rhone River in the context of concurrently developed ecological theory. Freshw. Biol. 31, Reynoldson, T.B., Norris, R.H., Resh, V.H., Day, K.E., Rosenberg, D.M., The reference condition: a comparison of multimetric and multivariate approaches to assess water-quality impairment using benthic macroinvertebrates. J. N. Am. Benthol. Soc. 16, Richards, C., Haro, R.J., Johnson, L.B., Host, G.E., Catchment and reach-scale properties as indicators of macroinvertebrate species traits. Freshw. Biol. 37, Rosenberg, D.M., Resh, V.H., Freshwater Biomonitoring in Benthic Macroinvertebrates. Chapman and Hall, New York. 488 pp. Roy, A.H., Rosemond, A.D., Leigh, D.S., Paul, M.J., Wallace, B., Habitat specific responses of stream insects to land cover disturbance: biological consequences and monitoring implications. J. N. Am. Benthol. Soc. 22, Sládeček, V., Chapter 3. Continental systems for the assessment of river water quality. In: James, A., Evision, L. (Eds.), Biological Indicators of Water Quality. John Wiley and Sons, Chishester, Great Britain. Statzner, B., Gore, J.A., Resh, V.H., Hydraulic stream ecology: observed patterns and potential applications. J. N. Am. Benthol. Soc. 7, The Mathworks, lnc., MATLAB Version 6.1, Massachusetts. Townsend, C.R., Hildrew, A.G., Species traits in relation to a habitat templet for river systems. Freshw. Biol. 31, Walley, WJ., Hawkes, HA., A computer-based development of the Biological Monitoring Working Party score system incorporating abundance rating, biotope type and indicator value. Water Res. 31, Ward, J.H., Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc. 58, Wiederholm, T., Chironomidae of the Holactic Region. Keys and diagnoses. (Part 1. Larvae). Entomol. Scand., Suppl pp. Wright, J.F., Sutcliffe, D.W., Furse, M.T., Assessing the Biological Quality of Fresh Waters: RIVPACS and Other Techniques. Freshwater Biological Association, Ambleside, UK. 400 pp. Yun, I.-B., Illustrated Encyclopedia of Fauna and Flora of Korea. Aquatic Insects, vol. 30. Ministry of Education, Seoul. 840 pp. Zar, J.H., Biostatistical Analysis, 4th ed. Prentice-Hall, Inc., New Jersey. 98 pp.

121 ecological modelling 203 (2007) available at journal homepage: Self-organizing mapping of benthic macroinvertebrate communities implemented to community assessment and water quality evaluation Mi-Young Song a,b, Hyun-Ju Hwang a, Inn-Sil Kwak c, Chang Woo Ji a, Yong-Nam Oh a, Byung Jin Youn a, Tae-Soo Chon a, a Division of Biological Sciences, Pusan National University, Busan , Republic of Korea b West Sea Fisheries Research Institute, NFRDI, Incheon , Republic of Korea c Bioresources Utilization Program, College of Fisheries and Ocean Sciences, Chonnam National University, Yeosu , Republic of Korea article info abstract Article history: Received 11 February 2006 Received in revised form 5 April 2006 Accepted 12 April 2006 Published on line 6 February 2007 Keywords: Patterning community Classification Self-organizing map (SOM) Artificial neural networks Community dynamics Ecosystem assessment Benthic macroinvertebrate communities serve as an efficient indicator group for assessing biological water quality. Communities, however, are difficult to analyze since the data consist of diverse taxa in a non-linear fashion. We implemented the self-organizing map (SOM) to classification of benthic macroinvertebrate communities collected across different levels of disturbances in streams in a large-scale. The trained SOM was feasible in providing a comprehensive view on community patterns, and the clustering by the SOM showed the gradient of pollution accordingly. New data sets sampled regularly for monitoring were further tested for tracing temporal changes in community states based on the trained SOM. Physicochemical and biological indices were correspondingly evaluated according to the trained SOM, and biological water quality indices were differentiated in the clustered communities Elsevier B.V. All rights reserved. 1. Introduction Sustainable management of aquatic ecosystems has been one of the most urgent concerns in environmental issues due to water resource shortage and its contamination. In order to achieve successful management of aquatic ecosystems, the objective assessment of water quality is a prerequisite. Biological organisms residing in the habitats in aquatic ecosystems convey the integrative and continuous characters of water quality and have been considered as suitable indicators (Allan, 1995; Hawkes, 1979; Hellawell, 1986; Rosenberg and Resh, 1993; Sladecek, 1979; Tittizer and Koth, 1979). Among biological communities, benthic macroinvertebrates have been widely used for ecological assessment of water quality. Macroinvertebrates are sedentary and have intermediate life span (from months to a few years). Additionally, benthic macroinvertebrates play a key role in food web dynamics, linking producers and top carnivores. Consequently, macroinvertebrates have been one of the most suitable indicator groups for assessing water quality (Barbour Corresponding author. Tel.: ; fax: address: [email protected] (T.-S. Chon) /$ see front matter 2006 Elsevier B.V. All rights reserved. doi: /j.ecolmodel

122 ecological modelling 203 (2007) et al., 1996; Butcher et al., 2003; Davies et al., 2000; Hawkes, 1979; Hellawell, 1986; Resh et al., 1995; Reynoldson et al., 1997; Richards et al., 1997; Rosenberg and Resh, 1993; Wright et al., 1993, 2000). However, it is a difficult task to develop an objective and quantifiable indicator system from community data. Especially, data for benthic macroinvertebrates are complex since the communities consist of multi-variables (i.e., diverse taxa) varying in a non-linear fashion. There have been numerous accounts of multi-variate statistical analyses regarding characterization of community data in ecology (e.g., Bunn et al., 1986; Legendre and Legendre, 1998; Ludwig and Reynolds, 1988; Quinn et al., 1991). Data for community classification and ordination have been available by measuring degree of association among the sampled communities and taxa (Legendre and Legendre, 1998; Ludwig and Reynolds, 1988). However, conventional multi-variate methods are generally limited in the sense that they are mainly applicable to linear data and have less flexibility in representing ecological data, for instance, handling noise and data management (Chon et al., 1996; Lek and Guégan, 1999, 2000; Recknagel, 2003). The self-organizing map (SOM) is an efficient tool for mining non-linear data and has been extensively used for patterning community data since 1990s (e.g., Chon et al., 1996, 2000, 2002; Kwak et al., 2000; Levine et al., 1996; Park et al., 2001, 2003a,b, 2004). Chon et al. (1996) classified benthic macroinvertebrate communities in polluted streams with the SOM and elucidated community patterning according to anthropogenic disturbances and locality of the sample sites. Chon et al. (2002) further implemented the SOM to a large-scale data collected in different river systems in the Korean Peninsula for 16 years. The large-scale data were accordingly arranged to reveal the impact of environmental disturbances. In this study, we further elaborated to show the trained SOM as a means of providing a comprehensive view on ecological states of the communities and to use the SOM as a map for assessing biological water quality. 2. Materials and methods 2.1. Self-organizing map (SOM) The SOM based on the Kohonen network (Kohonen, 1989) efficiently mines complex data without templates (or teachers) in an unsupervised manner and has been reported as a reliable classifier of ecological data (Chon et al., 1996, 2002; Kwak et al., 2000; Park et al., 2004). In the SOM, a linear array of M 2 artificial neurons (i.e., computation nodes), with each neuron being represented as j, is arranged in two dimensions for convenience of visual understanding. Suppose a community data containing N species, and the density of species, i, is expressed as a vector x i. The vector x i is considered to be an input layer to the SOM. In the network each neuron, j, is supposed to be connected to each node, i, of the input layer. The connectivities are represented as weights, w ij (t), adaptively changing at each iteration of calculation, t. Initially, the weights is randomly assigned in small values. When the input vector is sent through the network, each neuron of the network computes the summed distance between the weight and input as shown below: N 1 d j (t) = (x i w ij (t)) 2 i=0 The input values with greatly different numerical values in densities are avoided for training. The data were transformed by natural logarithm in order to emphasize the differences in low densities. Subsequently the transformed data were proportionally normalized between 0.01 and 0.99 in the range of the maximum and minimum density for each species collected during the survey period. The neuron responding maximally to a given input vector is chosen to be the winning neuron, the weight vector of which has the shortest distance to the input vector. The winning neuron and possibly its neighboring neurons are allowed to learn by changing the connecting weights in the manner to further reduce the distance between the weight and the input vector as shown below: w ij (t + 1) = w ij (t) + (t)(x i w ij (t))z j where Z j is assigned 1 for winning (and its neighboring) neuron(s) while it is assigned 0 for the rest neurons, and (t) (e.g., ) denotes the fractional increment of the correction. The radius-defining neighborhood is usually set to a larger value early in the training process, and is gradually reduced as convergence is reached. Detailed algorithm could be referred to Kohonen (1989), Zurada (1992), and Chon et al. (1996). After training, the Ward s linkage method (Ward, 1963) was applied to the weights of the SOM for further clustering of the patterned nodes. After preliminary training, we used 80 (10 8) SOM output units for patterning community data Community data The community data obtained from the 25 published papers in Korea from 1984 to 2000 were used for patterning the overall community compositions of benthic macroinvertebrates in Korea (Chon et al., 2002). The communities were mainly collected with the Surber net, and the data to the Family level were used as input to the SOM. We extended the input data (Chon et al., 2002) by adding additional sample sites in different degrees of pollution in in this study: the Piagol valley in the Somjin River, the Tokchon stream in the Nam River, the Miryang River in the Nakdong River system. We used 179 cases of the data consisting of 109 families for training. New community data were additionally collected in the natural and urban areas and were used for testing the trained SOM (Fig. 1). The sample sites in the Suyong River were recognized by the trained SOM (Fig. 1a and b). The sample sites in the Suyong River are affected with various levels of pollution and have been used for the long-term monitoring of the urban streams (Kwon and Chon, 1993; Yoon and Chon, 1996). The sample site YCK (28 cases) in the Suyong stream in the Suyong River is located in the agricultural area, while the site THP (35 cases) in the Soktae stream is highly polluted with domestic sewage. We additionally added the sample sites in the Baenae stream (2 cases) and the main stream of the Nakdong River (10 cases), which have been recently

123 20 ecological modelling 203 (2007) Fig. 1 Location of the sample sites: (a) the Korean Peninsula; (b) the Suyong River; (c) the Nakdong River. selected for LTER (National Long-Term Ecological Research Project in Korea) since 2005 (Fig. 1a and c). Water quality indices were obtained according to community data clustered by the SOM. EPT richness (the number of species in Ephemeroptera, Plecoptera and Trichoptera; see Resh et al., 1995) and Biological Monitoring Working Party Index (BMWP; Hawkes, 1997; National Water Council, 1981; Walley and Hawkes, 1996, 1997) were measured from the sample communities of benthic macroinvertebrates. The indices have been reported to correspondingly decrease with the increase in organic pollution (Hellawell, 1986; Rosenberg and Resh, 1993). 3. Results 3.1. Clustering of benthic macroinvertebrate communities Community abundance data collected from the Southern Peninsula of Korea (Fig. 1) were provided for training (Fig. 2a). The sample sites were firstly grouped according to the impact of pollution, and secondly to the location of streams and rivers. This was similar to the previous studies carried out on patterning of benthic macroinvertebrate communities in Korea (Chon et al., 1996). The clusters were vertically arranged according to degree of pollution: the sample sites with severe pollution appeared in the lower area of the map (e.g., cluster IV), while the clean sites were grouped in the upper right area (e.g., clusters I II). For instance, the groups of the sample sites from the natural areas (e.g., the Somjin River and the Tamjin River), and the sample sites from the clean area of the Han River were located in clusters I II, while communities from the urban area located in the Han, Nakdong, Masan and Suyoung Rivers were found in cluster IV. Cluster III was placed in the middle area, being occupied by less-polluted sites. The gradient shown in the mapping indicated that the SOM could serve as a map covering a broad scope of the impact of environmental disturbances. Within the clusters communities were further grouped according to location of the streams and rivers. This indicated that the samples could be further grouped according to localities within the similar level of environmental disturbances. Biological water quality was measured based on clustering by the SOM. Fig. 2b and c show the levels of EPT richness and BMWP corresponding to the clusters shown in the SOM (Fig. 2a). The levels of EPT richness and BMWP accordingly decreased in the clusters shown in the lower areas of the map (Fig. 2b and c). The values of the indices were statistically significant among different clusters based on the non-parametric multiple comparison test (Mann Whitney test, Zar, 1999) (Fig. 2b and c). BMWP was differentiated in all clusters, while the levels of EPT richness were partially in the same range between clusters II and III. This indicated that community clustering by the SOM properly reveals differences in biological water quality. Profiles of Families in benthic macroinvertebrates could be obtained from the trained SOM. Abundance of the Families is visualized in the corresponding clusters (Fig. 3). In the clean zone, cluster I was mainly characterized by Perlidae and

124 ecological modelling 203 (2007) Fig. 2 The map trained by SOM for pattering benthic macroinvertebrates reported from different streams in South Korea from 1984 to 2000: (a) sample sites; (b) mean and S.E. of EPT richness in different clusters defined in the SOM (n = 31 (I), n = 21 (II), n = 63 (III), n = 64 (IV)); (c) mean and S.E. of BMWP scores in different clusters defined in the SOM (n = 31 (I), n =21 (II), n = 63 (III), n = 64 (IV)). The different alphabets indicate significant difference in the Mann Whitney test (p < 0.001). Fig. 3 Profile of abundance of the prevalent taxa matched to clusters based on the trained SOM. The values in the vertical bar indicate densities (individuals/square meter).

125 22 ecological modelling 203 (2007) Lepidostomatidae, while cluster II was dominated by Gomphidae and Corydalidae. In the less-polluted zone, cluster III was associated with Hydropsychidae and Baetidae. Cluster IV in the polluted zone was strongly occupied by Tubificidae and Chironomidae. The association of indicator taxa and clusters based on the SOM were generally in accord with filed observation. For instance, Tubificidae has been reported to be tolerant in polluted water, while Perlidae is frequently collected in clean water (Hellawell, 1986; Rosenberg and Resh, 1993). In the intermediate zone in cluster III, Hydropsychidae and Baetidae were abundant, and these taxa are known to be the indicators for slightly polluted condition (Hellawell, 1986; Rosenberg and Resh, 1993) Monitoring community patterns with the SOM We tested the trained network (Fig. 2a) with new data sets and monitored changes in communities as time progressed. The communities collected monthly at YCK in the Suyoung River from November 1992 to April 1995 were recognized in a sequence on the trained SOM (Fig. 4a). In the early period (November 1992 November 1993), communities were mostly located in clusters III and IV, frequently crossing over the two clusters. Regarding that values of biological indices in clusters III and IV were low (Fig. 2b and c), water quality appeared to be poor at this stage. The fact that the sample communities frequently crossed the border of the clusters III and IV indicated that community patterns were somewhat variable in between strong and intermediate pollution. It was also notable that some communities tend to be located close together. For instance, a group of nodes was formed around the node (3(x), 3(y)) (e.g., (2, 3) and (3, 2)), and these nodes were characterized with Chironomidae and Tubificidae (Fig. 3). As the time progressed, communities moved from the polluted state in cluster IV to the clean state in cluster I (Fig. 4a), indicating temporal recovery of water quality in winter of 1994 and However, the communities returned to the polluted state in cluster IV in the later period of survey (Fig. 4a). The overall tracks recorded on the map demonstrated that the states of communities collected on the regular basis could be continuously monitored by the trained the SOM. Fig. 4b shows differences in biological and physicochemical indices obtained from the newly recognized samples according to the clusters shown in Fig. 2a. Biological indices such as EPT richness and BMWP were clearly differentiated with statistical significance among different clusters. However, physico-chemical indices such as BOD and turbidity were not statistically different, although BOD and turbidity tended to be higher in averages in clusters III and IV. Additionally, we tested community data from the severely polluted site, THP in the Soktae stream (Figs. 1b and 5). The community data were collected monthly from March 1993 to April 1995 and were invariably recognized at one node (5, 1) in cluster IV (Fig. 5a). At this site, the taxa tolerant to organic pollution such as Oligochaetes and Chironomids were dominantly collected. In addition, Psychodidae and Physidae were characteristically observed in THP (Fig. 5). Water quality data such as EPT richness and BMWP were accordingly in the lowest range (Fig. 5b). Especially, EPT richness appeared to be zero all through the sampling period. Physico-chemical indices such as BOD (16.0 ± 17.4 (S.E.) mg/l) and turbidity (10.0 ± 15.3 (S.E.) NTU) were also correspondingly low but showed a higher degree of variability compared with the biological indices during the survey period. We also monitored the sample sites used for LTER (Fig. 1c) on the trained SOM. The sample sites were accordingly recognized in two clusters on the map (Fig. 6a). The sample sites in the Baenae stream, BN, were located in cluster I, while the samples collected in the Nakdong River, NSJ, NKJ, NJP and NMK, were placed in cluster IV (Fig. 6a). The sample sites in the Nakdong River, however, were divided to two groups: one group of the nodes around the node (3(x), 4(y; from bottom Fig. 4 Monitoring of benthic macroinvertebrate communities collected at YCK in the Suyong stream from November 1992 to April 1995 according to the trained SOM (the sample was not collected in December 1994). (a) Recognition of the samples: November 1992 November 1993 (dots); January 1994 January 1995 (solid), and (b) mean and S.E. of biological and physico-chemical indices in different clusters defined in the SOM (n = 4 (I), n = 12 (III), n = 12 (IV)). The different alphabets indicate significant difference in the Mann Whitney test (p < 0.001).

126 ecological modelling 203 (2007) Fig. 5 (a) Monitoring of benthic macroinvertebrate communities collected at THP in the Suyong River from March 1992 to April 1995 according to the trained SOM (n = 35). (b) Mean and S.E. of biological and physico-chemical indices. of the map)) for the samples collected at NSJ and NKJ, etc., and the other single node (6, 1) for the samples collect at NMK-6 and NJP-6. While the dominant taxa, Tubificidae and Chironomidae were abundant in both groups, Psychodidae and Physidae were additionally observed in the node (6, 1) as shown in THP (Fig. 5a). Water quality data was different according to the clusters. Biological indices such as EPT richness and BMWP were invariably low in cluster IV (Fig. 6b). Physico-chemical indices such as BOD (16.0 ± 6.7 (S.E.) mg/l) and turbidity (10.0 ± 4.9 (S.E.) NTU) were also low but showed relatively higher degree of variability. In cluster I where the site BN was recognized, the biological indices were in the highest range, while BOD and turbidity were in the lowest range (Fig. 6b). 4. Discussion This study illustrated that clustering by the SOM would be useful for showing the gradient of water quality and providing a comprehensive view on overall states of community changes in response to environmental disturbances. The trained SOM readily accommodated diverse scope of community states across different degrees of pollution (Fig. 2a). Consequently, the SOM could serve as an alternative tool for monitoring community data to elucidate temporal changes in community states in response to disturbances. In the training data, physico-chemical data were not available for each sample site. As an alternative, we checked the differences in BOD and turbidity from the tested samples (Figs. 4 6). Although the biological indices (BMWP and EPT richness) were clearly differentiated, physico-chemical factors were not statistically different among the clusters. This indicated that biological indices appeared to be more sensitive in representing water quality based on the clustered community data. However, this would not necessarily mean that biological water quality indices are invariably superior to physico-chemical indices, regarding that biological indices were obtained from the same biological community data in this study. The sensitivity study between biological and physico-chemical factors should be more investigated with the sufficient data. In this study, the samples have not been Fig. 6 Monitoring of benthic macroinvertebrate communities collected at LTER sites according to the trained SOM: (a) recognition of the samples (the number in the sample name indicates the month of collection) and (b) mean and S.E. of biological and physico-chemical indices (n = 2 (I), n = 8 (IV)).

127 24 ecological modelling 203 (2007) sufficiently accumulated to evaluate differences in physicochemical factors in the testing data. The sample collection is still on going from the long-term monitoring sites (Fig. 1). Cross-evaluation of biological indices and physico-chemical indices should be carried out with accumulation of the sufficient data in the future. Due to limit of taxonomic information at the species level in the available data, training was carried out at the Family level in this study. The problem of information limit in taxonomy is critical in the Family of Chironomidae: Chironomidae includes diverse taxa but classification at the species level (especially, larvae) is extremely difficult. In the samples collected in the Baenae stream for testing, diverse species were newly identified in Chironomidae. The majority of species in this case were the specimens appearing in clean water. Diverse species in Chironomidae have been reported to be collected in clean water (Hellawell, 1986; Rosenberg and Resh, 1993). In this study, however, the tolerant species in Chironomidae such as Chironomus were mainly collected and were used for training. Consequently Chironomidae was presented as an indicator of polluted water along with Tubificidae in Oligochaetes in the training data. Due to this problem of training in Chironomidae, the samples newly collected with clean water species in the Baenae stream could not be included for testing at the present time. Further information on taxonomic identification at the species level in benthic macroinvertebrates is required to be updated for training data in the future in order to obtain a higher predictability in water quality assessment through community patterning as the long-term field survey continues. 5. Conclusion The SOM was efficient in extracting information from the complex data of benthic macroinvertebrate communities collected across different levels of pollution in the large-scale, and the gradient of water quality appeared according to the impact of pollution. The trained SOM showed the diverse scope of communities and were efficient in tracking temporal changes in community states. Biological water quality indices were correspondingly different in different clusters by the trained SOM. The SOM could be an alternative map for covering diverse scope of community states in water quality monitoring for the large-scale, long-term data, thus would broaden the scope of ecological informatics implemented to ecosystem assessment. Acknowledgment This research was supported by Korea Ministry of Environment as National Long-Term Ecological Research Project. references Allan, J.D., Stream Ecology: Structure and Function of Running Waters. Chapman & Hall. Barbour, M.T., Gerritsen, J., Griffith, G.E., Frydenborg, R., McCarron, E., White, J.S., Bastian, M.L., A framework for biological criteria for Florida streams using benthic macroinvertebrates. J. N. Benthol. Soc. 15, Bunn, S.E., Edward, D.H., Loneragan, N.R., Spatial and temporal variation in the macroinvertebrate fauna of streams of the northern Jarrah forest, Western Australia: community structure. Freshwater Biol. 16, Butcher, J.T., Stewart, P.M., Simon, T.P., A benthic community index for streams in the northern lakes and forests ecoregion. Ecol. Indicat. 3, Chon, T.-S., Park, Y.S., Moon, K.H., Cha, E.Y., Patternizing communities by using an artificial neural network. Ecol. Model. 90, Chon, T.-S., Park, Y.S., Park, J.H., Determining temporal pattern of community dynamics by using unsupervised learning algorithms. Ecol. Model. 132, Chon, T.-S., Kwak, I.S., Song, M.-Y., Park, Y.S., Cho, H.D., Kim, M.J., Cha, E.Y., Lek, S., Characterizing the effects of water quality on benthic stream macroinvertebrates in South Korea using a self-organizing mapping model. In: Lee, D. (Ed.), Ecology of Korea. Bumwoo Publishing Company, Seoul, Korea. Davies, N.M., Norris, R.H., Thomas, M., Prediction and assessment of local stream habitat features using large-scale catchment characteristics. Freshwater Biol. 45, Hawkes, H.A., Invertebrates as indicators of river water quality. In: James, A., Evision, L. (Eds.), Biological Indicators of Water Quality. John Wiley and Sons, Chichester, UK. Hawkes, H.A., Origin and development of the biological monitoring working party score system. Water Res. 32, Hellawell, J.M., Biological Indicators of Freshwater Pollution and Environmental Management. Elsevier, London, p Kohonen, T., Self-Organization and Associative Memory. Springer-Verlag, Berlin, p Kwak, I.S., Liu, G.C., Chon, T.-S., Park, Y.S., Community patterning of benthic macroinvertebrates in streams of South Korea by utilizing an artificial neural network. Kor. J. Limnol. 33, Kwon, T.-S., Chon, T.-S., Ecological studies on benthic macroinvertebrates in the Suyong River. III. Water quality estimations using chemical and biological indices. Kor. J. Limnol. 26, Legendre, P., Legendre, L., Numerical Ecology. Elsevier Science, The Netherlands. Lek, S., Guégan, J.-F., Artificial neuronal network as a tool in ecological modelling: an introduction. Ecol. Model. 120, Lek, S., Guégan, J.-F., Artificial Neuronal Networks: Application to Ecology and Evolution. Springer, Berlin. Levine, E.R., Kimes, D.S., Sigillito, V.G., Classifying soil structure using neural networks. Ecol. Model. 92, Ludwig, J.A., Reynolds, J.F., Statistical Ecology: A Primer of Methods and Computing. John Wiley and Sons, New York. National Water Council, River Quality: The 1980 Survey and Future Out-look. National Water Council, London. Park, Y.-S., Chon, T.-S., Kwak, I.S., Kim, J.-K., Jørgensen, S.E., Implementation of artificial neural networks in patterning and prediction of exergy in response to temporal dynamics of benthic macroinvertebrate communities in streams. Ecol. Model. 146, Park, Y.-S., Céréghino, R., Compin, A., Lek, S., 2003a. Applications of artificial neural networks for patterning and predicting aquatic insect species richness in running waters. Ecol. Model. 160, Park, Y.-S., Chang, J., Lek, S., Cao, W., Brosse, S., 2003b. Conservation strategies for endemic fish species threatened by the Three Gorges Dam. Conserv. Biol. 17, Park, Y.-S., Chon, T.-S., Kwak, I.S., Lek, S., Hierarchical community classification and assessment of aquatic ecosystems using artificial neural networks. Sci. Total Environ. 327,

128 ecological modelling 203 (2007) Quinn, M.A., Halbert, S.E., Williams III, L., Spatial and temporal changes in aphid (Homoptera: Aphididae) species assemblages collected with suction traps in Idaho. J. Econ. Entomol. 84, Recknagel, F., Ecological Informatics: Understanding Ecology by biologically-inspired Computation. Springer, Berlin, p Resh, V.H., Norris, R.H., Barbour, M.T., Design and implementation of rapid assessment approaches for water resource monitoring using benthic macroinvertebrates. Aust. J. Ecol. 20, Reynoldson, T.B., Norris, R.H., Resh, V.H., Day, K.E., Rosenberg, D.M., The reference condition: a comparison of multimetric and multivariate approaches to assess water-quality impairment using benthic macroinvertebrates. J. N. Benthol. Soc. 16, Richards, C., Haro, R.J., Johnson, L.B., Host, G.E., Catchment and reach-scale properties as indicators of macroinvertebrate species traits. Freshwater Biol. 37, Rosenberg, D.M., Resh, V.H., Freshwater Biomonitoring and Benthic Macroinvertebrates. Chapman & Hall, London, p Sladecek, V., Continental systems for the assessment of river water quality. In: James, A., Evison, L. (Eds.), Biological Indicators of Water Quality. John Wiley & Sons, Chichester. Tittizer, T.T., Koth, P., Possibilities and limitations of biological methods of water analysis. In: James, A., Evison, L. (Eds.), Biological Indicators of Water Quality. John Wiley and Sons, Chichester. Walley, W.J., Hawkes, H.A., A computer-based reappraisal of Biological Monitoring Working Party scores using data from the 1990 River Quality Survey of England and Wales. Water Res. 30, Walley, W.J., Hawkes, H.A., A computer-based development of the Biological Monitoring Working Party score system incorporating abundance rating, biotope type and indicator value. Water Res. 31, Ward, J.H., Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc. 58, Wright, J.F., Furse, M.T., Armitage, P.D., RIVPACS: A technique for evaluating the biological quality of rivers in the UK. Eur. Water Pollut. Contr. 3, Wright, J.F., Sutcliffe, D.W., Furse, M.T., Assessing the Biological Quality of Fresh Waters: RIVPACS and Other Techniques. Freshwater Biological Association, Ambleside, UK, p Yoon, B.J., Chon, T.-S., Community analysis in chironomids and biological assessment of water qualities in the Suyong and Soktae streams of the Suyong River. Kor. J. Lim. 29, Zar, J.H., Biostatistical Analysis, 4th ed. Prentice-Hall Inc., New Jersey, p Zurada, J.M., Introduction to Artificial Neural Systems. West Publishing Company, New York, p. 683.