Ozean Journal of Applied Sciences 5(2), 2012 ISSN 1943-2429 2012 Ozean Publication INVESTIGATION OF EFFECTS OF SPATIAL RESOLUTION ON IMAGE CLASSIFICATION FATIH KARA Fatih University, Department of Geography, Istanbul, Turkey E-mail address for correspondence: fatihkara@fatih.edu.tr Abstract: Spatial resolution is the most important feature of satellite images which affects the accuracy of image classification. To investigate this issue, three different kinds of satellite images were selected and classified. The MODIS, LANDSAT 4/5 TM, and IKONOS-2 images used in this study were of the commercial and industrial center of Turkey, Istanbul. Before the classification stage, pre-processing methods were applied, such as subset images; re-project images, and atmospheric corrections. An unsupervised classification technique is used for classification of MODIS and LANDSAT 4/5 TM images and both unsupervised and supervised classification methods are used for classification of IKONOS-2 images. As expected, classification results were different, as the accuracy reports revealed. Classification of IKONOS-2 images presented the highest accuracy with more than 96%. MODIS image classification presented the lowest accuracy with 88.5%. Furthermore many land use/land cover classes could not be found on the MODIS image which could be discriminated on the IKONOS-2 and LANDSAT 4/5 TM images. Keywords: Remote sensing, land cover, satellite images, high resolution. INTRODUCTION Remote sensing technologies have been used for the determination of urban-rural land use-land cover (LU/LC) change (Treitz et al., 1992) and changes in the natural environment (Collins and Woodcock, 1996; Coppin and Bauer, 1996) for a long time. This application made remote sensing one of the most important elements of urban and regional planning. Since the first civilian satellite was launched, several methods and techniques have been developed for mapping of the earth s surface by using satellite images and monitoring human-induced or natural environmental changes. Data provided by remote sensing techniques has become indispensable to meet the needs of urban planners and managers and is used for formulation of land use and changing patterns (Treitz and Rogan, 2004). Mapping methods of local, regional, and continental surfaces in remote sensing and related technologies have developed dramatically since first used. Besides the technological facilities, spaceborne and airborne remote sensing data has a very important role in monitoring and mapping of natural and anthropogenic changes in the earth s surface (Treitz and Rogan, 2004). Until 1999, high-resolution images could be taken only from planes flying near to the earth s surface. In 1999 the first high-resolution satellite was launched into space and provided 2 m spatial resolution images. Since 1999, several high-resolution satellites have been sent into space and high-resolution images provided by these 169
satellites give opportunities to make detailed and accurate studies. Many previous studies have mentioned that spatial resolution affects the accuracy of the classification of satellite images (Mayaux and Lambin, 1995; Moody and Woodcock, 1994). In this study MODIS, LANDSAT 4/5 TM, and IKONOS-2 images are classified using the same methods, and the effects of spatial resolution on the accuracy of classification is investigated. MODIS, LANDSAT 4/5 TM, and IKONOS-2 images have different spatial resolution, namely 500 m, 30 m, and 1 m. These images also have different spectral bands and radiometric characteristics. For example, LANDSAT 4/5 TM images have 7 spectral bands and 8-bit radiometric resolution, while IKONOS-2 images have 4 multispectral and 1 panchromatic band and 11-bit radiometric resolution. As a result, it is expected that the spatial, spectral, and radiometric characteristics of the satellite images will affect the accuracy of image classification. Accordingly, the aim of this study is to investigate the effects of resolution differences on semi-automated image classification. STUDY AREA The study area, Istanbul, is the biggest metropolitan area of Turkey and has a population of 13,255,685 according to the address registration system (ADNKS)(TurkStat, 2011). The city is located on both the European and Asian sides of the Bosporus in the northwest part of Turkey (Figure 1). Istanbul province is located between 40 48-41 36 latitudes and 27 58-29 56 longitudes. It has 6220 km² area and 39 sub-provinces. Istanbul is surrounded by the Marmara Sea in the south and the Black Sea in the north. Beside these seas, the effects of the Mediterranean Sea influence the climate of Istanbul, which is transitional between the Mediterranean climate and the humid-warm climate seen in the Black Sea Region. Generally summers are warm and humid and winters are cool and rainy (Tayanç, 2000). Average winter temperature is 6.3 C and average summer temperature is 22.4 C with an average annual temperature of 13.7 C (Anteplioğlu, 2002). The hottest months are July (23.3 C) and August (23.7 C) and the coldest months are January (5.3 C) and February (5.5 C) (Demirci, 2001). Average annual precipitation is 734 mm (Tayanç, 2000). Temperature and precipitation amounts vary from south to north in the province. While average annual precipitation is about 1000 mm on the Black Sea shores to the north, it is almost 600 mm on the Marmara shores to the south (Demirci, 2001). Istanbul is located on low-altitude plateau surfaces on both the Asian and European continents. While the average altitude is 150-200 m, there are a few hills with more than 500 m elevation. Rugged places are generally located on the northern side of the province. Rapid population increase and urbanization has changed the land cover of Istanbul greatly especially in the last 50 years. Forests cover 37.3% of the total area of Istanbul and they are mostly located in the northern parts of Istanbul. Settlement areas are usually located in southern parts and cover 14.3% of the total area (Karaburun et al., 2009). Figure1: Location of Istanbul 170
SATELLITE-BASED LAND SURFACE DATA The main objective of this study is to determine how the resolution of satellite-based data affects the accuracy of classification. Therefore images from MODIS, Landsat 4/5 TM, and IKONOS-2 were chosen to determine the land cover classes of the study area. Twenty-one USGS land cover classes (Table 1) are used to create land cover maps of the study area with different resolution. Table 1: USGS land cover classes No Class No Class 11 Open water 51 Shrubland 12 Perennial ice/snow 61 Orchards/Vineyards/Other 21 Low intensity residential 71 Grasslands/Herbaceous 22 High intensity residential 81 Pasture/Hay 23 Commercial/Industrial/Transportation 82 Row Crops 31 Bare Rock/Sand/Clay 83 Small Grains 32 Quarries/Strip Mines/Gravel Pits 84 Fallow 33 Transitional 85 Urban/Recreational Grasses 41 Deciduous Forest 91 Woody Wetlands 42 Evergreen Forest 92 Emergent Herbaceous Wetlands 43 Mixed Forest DATA AND METHOD Land cover maps are created by using MODIS (2009), Landsat (2009), and IKONOS-2 (2008) satellite images with different resolution. MODIS and Landsat images were downloaded from the USGS website and IKONOS-2 images were taken from Istanbul Metropolitan Municipality (IMM). For creating land cover maps from the satellite images, all the images underwent the same procedure before and after classification. Image pre-processing Raw satellite images are subset by using the provincial borders of Istanbul (in.shp format) taken from IMM. This data has a different datum from Landsat images and it was not possible to put it on those images. Therefore georeferencing was performed and spatial differences were corrected on the fly. The transformation had an error of less than 0.5 pixels of a root mean square (RMS), meaning that the image was referenced accurately (Figure 2). 171
Figure 2: Georeferencing of Landsat 4/5 TM images of the study area As the images used in the research were obtained in different time scales, they have haze and dust in different proportions and this camouflages the real changes or may show the same kinds of land cover classes as different. To overcome these kinds of problems, atmospheric correction methods are used (Berberoglu and Akin, 2009). The radiometric enhancement of images is an important stage of pre-processing. The aim of image enhancement is to make the objects more prominent by raising the quality of images so as to differentiate between different objects or land cover classes. So the techniques of haze reduction and noise reduction were applied on the images for a better understanding of the LU/LC classes (Lillesand & Kiefer, 1994). ATCOR 9.1 extension of Erdas Imagine 9.1 software was used for atmospheric correction of the Landsat images (Figure 3). (a) (b) Figure 3: Landsat image (June 2009) of Istanbul (a) before atmospheric correction and (b) after atmospheric correction 172
The pixel values of the image are changed greatly after this operation (Table 2). Table 2: Pixel values (band 1-blue band) of the image (a) before atmospheric correction and (b) after atmospheric correction. (a) (b) The IKONOS-2 images were radiometrically corrected when bought, therefore it was not necessary to use ATCOR 9.1 for them. Moreover, IKONOS-2 images do not cover the whole of Istanbul. The entire Istanbul area can be covered by mosaicking around 70 single parcels of IKONOS-2 images. Because of time limits, only Buyukcekmece sub-province was selected for a sample study area for the IKONOS-2 images. Coverage of Buyukcekmece sub-province makes use of 4 parcels (Figure 4). Figure 4: IKONOS-2 satellite images covering Buyukcekmece area (4 parcels) 173
Atcor 9.1 could not be used for MODIS images because MODIS images have only one band. Furthermore, other radiometric correction techniques could not be performed for the same reason (Figure 5). Figure 5: MODIS image of the study area (Istanbul) in June 2009 False color composite images were produced to better discriminate land cover classes on the MODIS images. For example, false color composites were very useful in the discrimination of farm land. It was very hard to distinguish some classes from each other, like forests and orchards. Visual interpretation, ground observation, and verification methods were applied in the discrimination of these kinds of spectrally similar classes. Image Classification Land use/land cover classes were mapped by using a supervised classification method and digital remote sensing (satellite images) data (Campbell, 1997; Thomas et al., 1987). The main objective of classifying satellite images is to categorize pixel values automatically and transform them directly into the classes or forms of land use (Lillesand & Kiefer, 1994). The supervised and unsupervised classification tools of ERDAS Imagine Software were used for the classification of the satellite images. Parallel piped was selected as the nonparametric rule, while maximum likelihood was selected as the parametric rule while using supervised classification (Bauer et al., 1994). The best grouping of unknown pixels was provided by the use of parameters of the maximum likelihood statistical method. A signature level was chosen over an image between 7 and 35 for each class and then a map, showing only 14 classes of land use, was created by using the recode tool of ERDAS Imagine Software and unifying these classes. Post-classification Post-classification enhancements are used to reduce the classification faults stemming from base fields, cities and classes which have similar responses, like some crop areas and wetlands (Song et al., 2001; Yuan et al., 2005). As the classes which were among the fallow fields and perceived as extraction were accepted as fallow ploughed at the time the image was taken, they were included in the category of farm lands. There are numerous grasslands in the study area and these have the same spectral responses as small grains in agricultural areas. Therefore, they were accepted as small grains most of the time, but were sometimes accepted as grasslands/herbaceous, depending on the growing areas. The spectral responses of evergreen forests are very similar to deciduous forests in the Landsat 4/5 TM images. Orthophotos and maps of land use taken from Istanbul Metropolitan Municipality were used in the discrimination of these trees from each other. Even then it is not possible to differentiate forest types on the MODIS images. A few MODIS images were downloaded from the USGS website (Figure 6a) and the most useful one for us, showing land cover classes better, was chosen and classified. Unsupervised classification was performed with 50 classes to classify the image. After classification, the image was recoded and cut by using the borders of Istanbul province (Figure 6b). 174
RESULTS Having made the enhancements of post-classification mistakes, it is necessary to test the accuracy of the research. Because changes in land use are discovered through the estimation of total squares of land use classes in the LU/LC maps created for each different period (for Landsat and IKONOS images), errors in the classification can lead to the misinterpretation of change analysis. One of the main techniques for accuracy assessment is using change maps for evaluating each class and calculating the expected accuracy (Yuan et al., 1998). Another method leading to more exact results is to check whether these points are accurately classified or not, after the selection of the points where changes are available or not upon the map. An accuracy assessment test was applied by using an accuracy matrix with 100 randomly selected points. The accuracy assessment was performed by using land-use maps, Google Earth, and aerial photos. Great importance was given to the representation of different LU/LC classes by these randomly chosen points. Figure 6a: Raw MODIS image covering Istanbul in August 2009 Figure 6b: Classified MODIS image 175
Table 3 shows the accuracy assessment results for the MODIS image. LU/LC Classes Table 3: Accuracy assessment results of MODIS image of Istanbul. KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 0.9799 100.00% 99.12% High Intensity Residential 0.8464 42.86% 85.71% Commercial/Industrial/Transport 0,0000 0.00% 0.00% Bare Rock/Sand/Clay 0.1066 100.00% 11.11% Quarries/Strip Mines/Gravel Pits 0,0000 0.00% 0.00% Deciduous Forest 0.8257 76.67% 85.19% Evergreen Forest 0,0000 0.00% 0.00% Shrubland 0,0000 0.00% 0.00% Grasslands/Herbaceous/Cultivated 0,0000 0.00% 0.00% Row Crops 0,0000 0.00% 0.00% Fallow 0,0000 0.00% 0.00% Emergent Herbaceous Wetlands 0,0000 0.00% 0.00% Urban/Recreational Grasses 0,0000 0.00% 0.00% Small Grains 0.8989 75.00% 91.30% Landsat 4/5 TM images of Istanbul were classified using unsupervised classification with 120 classes. The Istanbul land cover map was produced by using two different Landsat 4/5 TM images (31_20090617 and 32_20090617) taken on 16 July 2009. The images were radiometrically corrected and classified using unsupervised classification with 120 classes. After classification, the new images were recoded (Figure 7). Figure 7: Recoding Landsat 4/5 TM (31_20090617 and 32_20090617) images of Istanbul 176
A mosaic operation was performed for two new Landsat images after recoding and a new image covering the entire Istanbul area was produced (Figure 8). Figure 8: Land cover map of Istanbul produced using Landsat 4/5 TM images. Table 4 shows the accuracy assessment results for the Landsat 4/5 TM mosaic image. LU/LC Classes Table 4: Accuracy assessment results of Landsat 4/5 TM mosaic image of Istanbul KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 0,0000 0.00% 0.00% High Intensity Residential 1,0000 66.67% 100.00% Commercial/Industrial/Transport 0,0000 0.00% 0.00% Bare Rock/Sand/Clay 1,0000 60.00% 100.00% Quarries/Strip Mines/Gravel Pits 0,0000 0.00% 0.00% Deciduous Forest 0.8813 100.00% 89.47% Evergreen Forest 1,0000 60.00% 100.00% Shrubland 1,0000 100.00% 100.00% Grasslands/Herbaceous/Cultivated 1,0000 100.00% 100.00% Row Crops 0,0000 0.00% 0.00% Fallow 0.7424 100.00% 76.47% Emergent Herbaceous Wetlands 1,0000 50.00% 100.00% Urban/Recreational Grasses 0,0000 0.00% 0.00% Small Grains 0,0000 0.00% 0.00% As mentioned before, 4 IKONOS-2 images covering Buyukcekmece sub-province of Istanbul were selected for classification in this study. After an atmospheric correction operation, unsupervised classification was performed 177
on 2 of the 4 images (parcels of 2251980 and 2251981) because of getting much better results rather than with supervised classification (Figure 9). Figure 9: The results of unsupervised classification Supervised classification was performed on the other 2 of the 4 images (parcels of 2251990 and 2251991) because of getting much better results than with the unsupervised classification (Figure 10). Figure 10: The results of supervised classification Tables 5,6,7,8 show the accuracy assessment results of the IKONOS-2 images. 178
LU/LC Classes Table 5: Accuracy assessment results of IKONOS-2 (2251980) image of Istanbul KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 0.0000 0.00% 0.00% High Intensity Residential 0.0000 0.00% 0.00% Commercial/Industrial/Transport 0.7462 100.00% 75.00% Bare Rock/Sand/Clay 0.4924 33.33% 50.00% Quarries/Strip Mines/Gravel Pits 0.0000 0.00% 0.00% Deciduous Forest 0.0000 0.00% 0.00% Evergreen Forest 0.0000 0.00% 0.00% Shrubland 0.3939 100.00% 40.00% Grasslands/Herbaceous/Cultivated 1,0000 100.00% 100.00% Row Crops 0.0000 0.00% 0.00% Fallow 0.9588 96.55% 97.67% Emergent Herbaceous Wetlands 0.0000 0.00% 0.00% Urban/Recreational Grasses 0.0000 0.00% 0.00% Small Grains 1,0000 100.00% 100.00% LU/LC Classes Table 6: Accuracy assessment results of IKONOS-2 (2251981) image of Istanbul. KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 1,0000 100.00% 100.00% High Intensity Residential 0.5833 75.00% 60.00% Commercial/Industrial/Transport 0,0000 0.00% 0.00% Bare Rock/Sand/Clay 0,0000 0.00% 0.00% Quarries/Strip Mines/Gravel Pits 0,0000 0.00% 0.00% Deciduous Forest 0,0000 0.00% 0.00% Evergreen Forest 0,0000 0.00% 0.00% Shrubland 0.4949 100.00% 50.00% Grasslands/Herbaceous/Cultivated 0,0000 0.00% 0.00% Row Crops 0,0000 0.00% 0.00% Fallow 0.9587 95.56% 97.73% Emergent Herbaceous Wetlands 0,0000 0.00% 0.00% Urban/Recreational Grasses 0,0000 0.00% 0.00% Small Grains 0.9596 100.00% 97.62% 179
LU/LC Classes Table 7: Accuracy assessment results of IKONOS-2 (2251990) image of Istanbul KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 0.8974 100.00% 95.35% High Intensity Residential 1,0000 100.00% 100.00% Commercial/Industrial/Transport 1,0000 33.33% 100.00% Bare Rock/Sand/Clay 1,0000 100.00% 100.00% Quarries/Strip Mines/Gravel Pits 0,0000 0.00% 0.00% Deciduous Forest 0,0000 0.00% 0.00% Evergreen Forest 0,0000 0.00% 0.00% Shrubland 0,0000 0.00% 0.00% Grasslands/Herbaceous/Cultivated 0,0000 0.00% 0.00% Row Crops 0,0000 0.00% 0.00% Fallow 1,0000 100.00% 100.00% Emergent Herbaceous Wetlands 0,0000 0.00% 0.00% Urban/Recreational Grasses 0.4966 100.00% 50.00% Small Grains 1,0000 100.00% 100.00% LU/LC Classes Table 8: Accuracy assessment results of IKONOS-2 (2251991) image of Istanbul KAPPA (K^) Users Accuracy (%) Total Accuracy (%) Open Water 0.8780 100.00% 90.00% High Intensity Residential 0.9490 100.00% 95.65% Commercial/Industrial/Transport 1,0000 81.48% 100.00% Bare Rock/Sand/Clay 0.6622 100.00% 66.67% Quarries/Strip Mines/Gravel Pits 0,0000 0.00% 0.00% Deciduous Forest 0,0000 0.00% 0.00% Evergreen Forest 0,0000 0.00% 0.00% Shrubland 0,0000 0.00% 0.00% Grasslands/Herbaceous/Cultivated 1,0000 100.00% 100.00% Row Crops 0,0000 0.00% 0.00% Fallow 1,0000 100.00% 100.00% Emergent Herbaceous Wetlands 0,0000 0.00% 0.00% Urban/Recreational Grasses 0,0000 0.00% 0.00% Small Grains 1,0000 100.00% 100.00% After the classification operation, the classified images were recoded and combined with the mosaic operation. Then the new mosaic image was cut by using the border shapefile of Buyukcekmece sub-province (Figure 11). 180
Figure 11: Land cover map of Buyukcekmece province (generated using pan-sharpened IKONOS-2 images with 1 m resolution). Table 9 shows the results of the accuracy assessment operation for each satellite image used in this study. Image Table 9: Accuracy Results of Each Classification Resolution (m) Kappa Statistics Overall Classification Accuracy (%) MODIS August-2009 500 0.8194 88.50% Landsat 4/5 TM 31_32_20090617 IKONOS pan-sharpened 2251980 IKONOS pan-sharpened 2251981 IKONOS pan-sharpened 2251990 IKONOS pan-sharpened 2251991 30 0.9171 95.33% 1 0.9407 96.50% 1 0.9200 95.00% 1 0.9354 96.00% 1 0.9576 96.67% According to the accuracy assessment results, IKONOS-2 images are classified more accurately than the others. The overall classification accuracy is 96.67% for IKONOS pan-sharpened parcel 2251991, 96.50% for IKONOS 181
pan-sharpened parcel 2251980, 96.00% for IKONOS pan-sharpened parcel 2251990, and 95.00% for IKONOS pan-sharpened parcel 2251981. The Landsat 4/5 TM image was classified as a little less accurate than the IKONOS-2 images, but it is still classified very well, with 95.33% overall accuracy. The MODIS image has the least accuracy with 88.50%, but this is in fact a very high number for a MODIS image. The main reason for this high accuracy is that only a few land cover classes were determined on that image. CONCLUSIONS AND DISCUSSIONS In remote sensing and GIS applications it is very important to get updated and accurate land cover and change maps. But the expensiveness of data, difficulties acquiring data, and inappropriate methodology and techniques all affect the quality of studies. More accurate land cover maps are obtained from high resolution satellite images. If resolution increases, accuracy increases. In this study, the accuracy was higher than 95% for the IKONOS images but 88.50% for the MODIS image. While the MODIS land cover map of Istanbul has 5 LU/LC classes, the Landsat 4/5 TM land cover map has 12 LU/LC classes. Furthermore, every LU/LC class can be detected on the pan-sharpened IKONOS images (1 m). For this reason, MODIS images should be used for only global studies. They are not appropriate for small-scale study areas such as Istanbul. There is no certain rule about which classification method gives more accurate results. Sometimes unsupervised classification, sometimes supervised classification gives more accurate results. In this study, the best results were achieved with supervised classification for 2 of the IKONOS-2 images, and with unsupervised classification for the other 2 images. It can be said that unsupervised classification is more useful for classification of Landsat 4/5 TM images. Separation of mining areas from residential areas was not possible with supervised classification, but it was possible with unsupervised classification. In addition, the number of classes should be increased when using high resolution images in unsupervised classification. Training data should be well distributed and cover all characteristic landscape features across the scene in a supervised classification. If the study area is large and covers more than 1 satellite image, satellite images should be classified first before being made into a mosaic. This is because the mosaic operation changes the class values and blurs the images, and LU/LC classes can t be defined very well on blurred mosaic images. It can be hard to separate some signatures from each other, especially on MODIS and Landsat 4/5 TM images. For example, shrub lands and forest areas have almost the same spectral signature. Moreover, the spectral signatures of agricultural areas are very similar to grasslands. Therefore the acquisition date of satellite images is very important in classification. 182
REFERENCES Anteplioglu, U., Topçu, S. & Incecik, S. (2002). An application of a photochemical model for urban airshed in Istanbul. In C. Borrego and G. Schayes (Eds.), Air Pollution Modeling and Its Application XV. (p. 168). New York: Kluwer Academic/Plenum Publishers. Bauer, M. E., Burk, T. E., Ek, A. R., Coppin, P. R., Lime, S. D., Walsh, T. A., Walters, D. K., et al. (1994). Satellite inventory of Minnesota forests. Photogrammetric Engineering and Remote Sensing, 60(3), 287 298. Berberoglu, S., & Akin, A., (2009). Assessing different remote sensing techniques to detect land use/cover changes in the eastern Mediterranean. International Journal of Applied Earth Observation and Geoinformation, 11 (2009) 46 53. Campbell, J. B. (1997). Introduction to remote sensing. New York: Guilford Press. Collins, J. B., & Woodcock, C. E. (1996). An assessment of several linear change detection techniques for mapping forest mortality using multitemporal Landsat TM data. Remote Sensing of Environment, 56, 66 77. Coppin, P. R., & Bauer, M. E. (1996). Digital change detection in forest ecosystems with remote sensing imagery. Remote Sensing Reviews, 13, 207 234. Demirci, A. (2001). Types and distribution of landslides in the eastern parts of Buyukcekmece Lake, using GIS [Unpublished Master Thesis]. Fatih University, Social Science Institute, Istanbul. Karaburun, A., Demirci, A., & Suen, I-S. (2009). Impacts of urban growth on forest cover in Istanbul (1987-2007). Environmental Monitoring and Assessment. doi 10.1007/s10661-009-1000-z. Lillesand, T. M., & Kiefer, R. W. (1994). Remote sensing and image interpretation (4th ed.). New York: Wiley. Mayaux, P., & Lambin, E. (1995). Estimation of tropical forest area from coarse spatial resolution data: a two step correction function for proportional errors. Remote Sensing of Environment, 53, 1 16. Moody, A., & Woodcock, C. E. (1994). Scale-dependent errors in the estimation of land-cover proportions implications for global land-cover datasets. Photogrammatic Engineering and Remote Sensing, 60, 585 594. Song, C., Woodcock, C. E., Seto, K. C., Lenney, M. P., & Macomber, S. A. (2001). Classification and change detection using Landsat TM data: When and how to correct atmospheric effects? Remote Sensing of Environment, 75, 230 244. Tayanç, M., Im, U., Doğruel, M., & Karaca, M. (2000). Climate change in Turkey for the last half century. Climatic Change, 94, 483-502. Thomas, I. L., Benning, V. M., & Ching, N. P. (1987). Classification of remotely sensed images. Bristol: Adam Hilger. Treitz, P. M., Howarth, P. J., Gong, P. (1992). Application of satellite and GIS technologies for land-cover and landuse mapping at the rural urban fringe: a case study. Photogrammetric Engineering and Remote Sensing, 58, 439 448. Treitz, P., & Rogan, J. (2004). Remote sensing for mapping and monitoring land-cover and land-use change an introduction. Progress in Planning, 61, 269 279. TurkStat (2011). Turkey in Statistics, 2011. Turkish Statistical Institute, Printing Division, Ankara. 183
Yuan, D., Elvidge, C. D., & Lunetta, R. S. (1998). Survey of multispectral methods for land cover change analysis. Remote sensing change detection: Environmental monitoring methods and applications (pp. 21 39). Michigan: Ann Arbor Press. Yuan, F., Bauer, M. E., Heinert, N. J., & Holden, G. (2005). Multi-level land cover mapping of the Twin Cities (Minnesota) metropolitan area with multi-seasonal Landsat TM/ETM+data. Geocarto International, 20(2), 5 14. 184