, pp.229-240 http://dx.do.org/10.14257/gdc.2014.7.6.19 A Survey on Clusterng based Meteorologcal Data Mnng We Tan 1, Yuhu Zheng 1, Runzh Yang 2, Sa J 1 and Jn Wang 1 1 College of Computer and Software, Nanng Unversty of Informaton Scence & Technology, Nanng 210044, Chna 2 Natonal Meteorologcal Informaton Center, Chna Meteorologcal Admnstraton, Beng 100080, Chna Abstract Data mnng s an mportant tool n meteorologcal problems solved. Cluster analyss technques n data mnng play an mportant role n the study of meteorologcal applcatons. The research progress of the clusterng algorthms n meteorology n recent years s summarzed n ths paper. Frst, we gve a bref ntroducton of the prncples and characterstcs of the clusterng algorthms that are commonly used n meteorology. On the other hand, the applcatons of clusterng algorthms n meteorology are analyzed, and the relatonshp between the varous clusterng algorthms and meteorologcal applcatons are summarzed. Then we nterpret the relatonshp from the perspectves of algorthms characterstcs and practcal applcatons. Fnally, some man research ssues and drectons of the clusterng algorthms n meteorologcal applcatons are ponted out. Keywords: data mnng; clusterng; meteorologcal applcatons 1. Introducton Clmate change and ts consequences are ncreasngly beng recognzed as one of the most sgnfcant challenges whch faced by humanty. Managng and takng advantage of large mass of meteorologcal data s the bass for studyng clmate change as well as the key to mprove the accuracy of predcton of dsastrous weather. A growng number of technologes have been or wll be ntroduced nto the feld of meteorology, n whch the cluster analyss n data mnng has occuped an mportant poston n the study of meteorology. K-means and herarchcal methods are classcal algorthms n cluster analyss and they are wdely used n meteorology. Some other clusterng algorthms, such as DBSCAN (Densty-based Specal Clusterng of Applcatons wth Nose), EM (Expectaton Maxmzaton), SOM (Selforganzng Feature Maps), FCM (Fuzzy C-means Clusterng), SA (Smulated Annealng), WaveCluster and graph clusterng, also perform well n some specal meteorologcal applcatons. In order to make better use of clusterng algorthms n meteorology, we analyze the applcatons of commonly used clusterng algorthms n meteorology, and summarze the dstrbuton and characterstcs of clusterng algorthms n meteorologcal applcatons. Then we nterpret the results from the perspectves of characterstcs of the algorthms and practcal applcatons. 2. Problem Defnton The cluster analyss of meteorologcal data, both observed and model-generated, poses a number of unque challenges: () the data s spatally and temporally correlated, ISSN: 2005-4262 IJGDC Copyrght c 2014 SERSC
() the data s potentally nosy, () massve quanttes of data are avalable for mnng, etc. Common data types n cluster analyss manly contan nterval-scale varables, bnary varables, categorcal varables, ordnal varables, rato-scaled varables, varables of mxed types and vector obects. There are some commonly used dstance measures, for example, Eucldean dstance, Manhattan dstance, Mnkowsk dstance, Chebyshev dstance, Mahalanobs dstance, Hammng dstance and correlaton coeffcent [1]. Dfferent measure methods have dfferent features and advantages. Thus, n a specfc meteorologcal applcaton, combned wth the data characterstcs, choosng the approprate clusterng algorthm s the premse of successful experments. 3. The Commonly used Clusterng Algorthms n Meteorology At present, many clusterng algorthms are dscussed [2]. But few clusterng algorthms can be used n meteorology. In the rest of ths secton, clusterng algorthms that are commonly used n meteorologcal applcatons are summarzed brefly from the perspectves of prncples and characterstcs. 3.1 Parttonng clusterng Gven a dataset wth n obects, a parttonng method classfes the data nto k groups. Then an teratve relocaton technque s used to mprove the parttonng. The process stops untl the crteron functon converges. The k-means algorthm [3] s a promnent parttonng clusterng. It chooses k ntal centrods where k s a user-specfed parameter. The k-means proceeds as follows: (1) Select k ponts as ntal centrods. (2) Repeat. (3) Form k clusters by assgnng each pont to ts closest centrod. (4) Recomputed the centrod of each cluster. (5) Untl Centrods do not change. The k-means algorthm s very smple and can be easly mplemented n many practcal problems. It can work very well for compact and hypersphercal clusters. But there s no effcent and unversal method for dentfyng the ntal parttons and the number of clusters k. The teratvely optmal procedure cannot guarantee convergence to a global optmum. It s also senstve to outlers and nose. 3.2 Herarchcal clusterng Herarchcal methods organze data nto a herarchcal structure accordng to the proxmty matrx. They can be further classfed as agglomeratve methods or dvsve methods, based on how the herarchcal decomposton s formed. Agglomeratve clusterng [4] starts wth N clusters and each of them ncludes exactly one obect. Based on dfferent defntons of dstances between two clusters, there are manly fve agglomeratve methods ncludng sngle lnkage, complete lnkage, average lnkage, medan lnkage and ward s method. The dfferent dstances for the fve methods are defned as follows. Sngle lnkage method Dmn ( C, C ) MIN x C, y C x y (1) 230 Copyrght c 2014 SERSC
Complete lnkage method Dm ax ( C, C ) MAX x C, y C x y (2) Average lnkage method D avg ( C, C ) x C y C x y n n (3) Medan lnkage method D ( C, C ) (1/ C( n n,2)) x y (4) mean x, y ( C, C ) Ward s method ( C, C ) (1/( n n )) x n 2 D ward (5) x ( C, C ) where n s the merged cluster center. x-y s the dstance between two obects x and y. n s the number of obects n cluster C. And C( n + n,2) s the total number of the methods for dfferent combnatons of two elements extracted from the n + n elements. 3.3 Model-based clusterng Model-based clusterng methods attempt to optmze the ft between the gven data and some mathematcal model. Such methods are often based on the assumpton that the data are generated by a mxture of underlyng probablty dstrbuton. EM and SOM are two popular methods of model-based clusterng. (1) EM method: EM [5] s a popular teratve refnement algorthm for fndng maxmum lkelhood estmate of parameters n statstcal models, where the model depends on unobserved latent varables. EM can be vewed as an extenson of the k-means paradgm. However, t assgns each obect to a cluster accordng to a weght representng the probablty of membershp rather than cluster mean. EM teraton alternates between performng an expectaton (E) step, whch creates a functon for the expectaton of the log-lkelhood evaluated usng the current estmate for the parameters, and a maxmzaton (M) step, whch computes parameters maxmzng the expected log-lkelhood found on the E step. These parameter-estmates are then used to determne the dstrbuton of the latent varables n the next E step. EM s smple and easy to mplement. In practce, t converges fast but may not reach the global optma. It s senstve to the ntal parameters too. (2)SOM method: SOM [6] s a type of the artfcal neural network that s traned usng unsupervsed learnng to produce a low-dmensonal, dscretzed representaton of the nput space of the tranng samples, called a map. It s one of the most popular neural network methods for clusterng analyss. The method s partcularly useful when a nonlnear mappng s nherent n the problem tself. Just lke the k-means, SOM also needs to predefne the number of the clusters, whch s unknown for most crcumstances. 3.4 Fuzzy clusterng In fuzzy clusterng, data elements can belong to more than one cluster, and assocated wth each element s a set of membershp levels. Fuzzy clusterng s a process of assgnng these membershp levels, and then usng them to assgn data elements to one or more clusters. FCM [7] s one of the most wdely used fuzzy clusterng algorthms. Lke the k-means algorthm, the FCM ams to mnmze the followng obectve functon: Copyrght c 2014 SERSC 231
U [ u, ] where c N J( U, M ) c N 1 1 ( m u, ) D u, s a fuzzy partton matrx. s the membershp value of M [ m1,..., mc ] D s a cluster centers matrx. s the dstance element x n the cluster. between element x and cluster center m. The functon dffers from the k-means obectve u, functon by the addton of the membershp value and the fuzzy coeffcent m. The fuzzy coeffcent m determnes the level of cluster fuzzness. A large m results n smaller membershp values, hence, fuzzer clusters. In the absence of expermentaton or doman knowledge, m s commonly set to 2. FCM s partcularly useful when the boundares among the clusters are not well separated or ambguous. Moreover, the membershp values may help us dscover more sophstcated relatons between a gven obect and the dsclosed clusters. However, FCM suffers from the presence of nose and outlers. It s also dffcult to dentfy the ntal parttons. 3.5 Combnatoral search technques-based clusterng The basc obectve of search technques s to fnd the global or approxmate global optmum for combnatoral optmzaton problems, whch usually has NP-hard complexty and need to search an exponentally large soluton space. Clusterng can be regarded as a category of optmzaton problem. Gve a set of data ponts, clusterng algorthms am to organze them nto K subsets that optmze some crteron functon. 4. The Applcatons of Clusterng Algorthms n Meteorology 4.1 Clmate change Clmate change and ts mpact s one of the greatest challenges facng humanty and the earth. Investgatng the clmate change and explorng the law for the beneft of manknd s everyone s responsblty. Cluster analyss technques study the clmate change from three aspects of meteorologcal elements classfcaton, weather type classfcaton and atmospherc crculaton. 4.1.1 Meteorologcal elements classfcaton: The classfcaton s a task of classfyng regons based on smlar basc characterstcs and changes of meteorologcal elements. Cluster analyss technques classfy regons to study regonal weather, clmate change, predcton and preventon of extreme weather from aspects of wnd, precptaton, temperature, clouds, pressure etc. At present the study of classfcaton based on meteorologcal elements focus on k-means method, herarchcal method, DBSCAN method, SOM method and FCM method. (1) K-means method: In 2009, Jménez, P. A. et al. [8] classfed the daly surface wnd felds nto wnd pattern types wth the combnaton of the complete lnkage algorthm and an algorthm smlar to the k-means accordng to ther spatal smlarty over the Comundad Foral de Navarra regon. In the same year, Chrstos J. Lols [9] appled the k-means to the factor scores tme seres classfyng the 56 years nto sx dstnct clusters that descrbe the man modes of spatal dstrbuton of cloudness. Then n 2011, İ brahm Sönmez et al. [10] (6) 232 Copyrght c 2014 SERSC
appled the k-means to reclassfy ranfall regons of Turkey [11] and nvestgated ther temporal varablty n relaton to North Atlantc Oscllaton. (2) Herarchcal method: Robeson, S.M. et al. and J. Kysely et al. [12] classfed homogeneous regons accordng to the temperature and precptaton by dfferent herarchcal clusterng methods respectvely. In 2008, M. Burlando et al. [13] appled 15 dfferent clusterng technques resultng from the combnaton of three dstance measures and fve agglomeratve methods to study the wnd clmate of Corsca. (3) DBSCAN method: In 2005, a data mnng applcaton based on DBSCAN was carred out on ar temperature database and obtaned clusters that have smlar temperature trends [14]. (4) SOM method: In 2007, Reusch, D.B. et al. [15] analyzed the mean sea level pressure data from 1957 to 2002 wth SOM method n order to study North Atlantc clmate varablty n general, especally the North Atlantc oscllaton. 4.1.2 Atmospherc crculaton: Atmospherc crculaton has been topcs of nterest to clmatologsts for years [16]. It s the domnant factor n global clmate and a wde range of weather stuaton as well as the background of the weather system actvty on a varety of scales. Cluster analyss technques analyze the characterstcs and long-term changes of the atmospherc crculaton as well as ts relatonshp wth the clmate change. The clusterng algorthms used n ths feld contan k-means method, herarchcal method, SA method and graph theory-based method. (1) K-means method: In 2006, Esteban, P. et al. [17] appled PCA and the k-means to characterze the daly surface synoptc crculaton patterns. The results are consstent wth the subectve knowledge of the daly atmospherc crculaton over the area. (2) Herarchcal method: In 2007, EM and a tradtonal herarchcal agglomeratve clusterng method (HAC) were appled and generated dstnctly dfferent atmospherc patterns characterzng seasonal crculaton over eastern North Amerca. (3) SA method: In 2007, Phlpp, A et al. [18] proposed a new clusterng scheme combnng the concepts of smulated annealng and dversfed randomzaton (ASNDRA) when explorng long-term varablty of daly North Atlantc-European pressure patterns snce 1850. The method s able to reduce substantally the nfluence of chance n the cluster assgnment, leadng to parttons that are notceably nearer to the global optmum and more stable. Based on the former method, a representatve set of patterns that fully characterze the dfferent crculaton types appearng n each season was generated [19], and n 2009 [20], t was used to classfy daly pressure-pattern sequences and evaluate whether sequence classfcaton s more sutable to descrbe surface ar-temperature condtons n Europe. 4.2. Urban meteorology Urban meteorology, ncludng the cty meteorologcal observaton, urban atmospherc polluton, urban fne weather forecastng, s an mportant part of human lfe. Hgh-mpact weather research, ar qualty predcton and weather forecastng are the man applcatons of clusterng algorthms n the feld of urban meteorology. Copyrght c 2014 SERSC 233
4.2.1 Ar qualty: Ar pollutants, such as SO2, CO, O3, PM2.5, PM10, acd ran, acd fog, have a defnte mpact on human lvng condtons. At present, cluster analyss technques are manly used to study the mpact of ar polluton as well as the relatonshp between weather condtons and ar polluton. The man clusterng technques n ths feld are the k-means, herarchcal method, the SOM and the FCM. (1) K-means method: Cervone, G. et al. [21] and L, L. et al. [22] nvestgated the mpact of ar polluton by k-means. In 2011, Alex Mace et al. [23] proposed adaptve k- means clusterng algorthm that used both the traectory varables and the assocated chemcal value. The ar mass traectores were clustered to dentfy source regons of certan chemcal speces. (2) Herarchcal method: Joseph H. Casola et al. [24] dentfed weather regmes through dfferent herarchcal clusterng methods. S. Yonemura et al. [25] and Charbel Aff et al. [26] analyzed the characterzaton of pollutant gas concentratons by dfferent herarchcal clusterng methods. In 2011, Seungmn Lee et al. [27] appled an agglomeratve herarchcal clusterng algorthm based on the Ward's mnmum varance clusterng crteron to the back traectores to examne the orgn of and favorable meteorologcal condtons for hgh concentratons of PM10 n Seoul, Korea. (3) SOM method: Consderng the nfluence of synoptc-scale meteorology on ar qualty, Ignaco J. Turas et al. [28] and John L. Pearce et al. [29] have done somethng by SOM method respectvely. Although the SOM has proven ts effcency n meteorologcal parameters and ar polluton concentraton clusterng, t s dffcult to clearly dentfy cluster's borders f the SOM s very populated. Moreover, usng only k- means generates hgh computatonal tme when clusterng a large multdmensonal tme seres data. In 2012, Soufane Khedara et al. [30] proposed a two stages clusterng approach based on the SOM and the k-means algorthm to the characterzaton of meteorologcal condtons. 4.2.2 Weather forecastng: Weather forecastng s made based on analyss of satellte cloud mages and weather maps, combned wth relevant meteorologcal data, topography, seasonal characterstcs and the mass experences. At present, forecastng s stll not accurate because of the uncertan data of nature. On the other hand, t s tme-consumng to process large amount of data. All these defcences make cluster algorthms more applcable n weather forecastng feld. The man clusterng algorthms n ths feld are herarchcal method and FCM method. 4.3 Hydrometeorology Hydrometeorology s a subect that apples the prncples and methods of meteorology to study the ssues related to precptaton and evaporaton n hydrologcal cycles and water balance. Clusterng algorthms are appled n hydrometeorology to tackle hydrologcal forecastng and water resources management. 4.3.1 Hydrologcal forecastng: Hydrologcal forecastng s an mportant aspect of hydrologcal servces on economy and socety, especally makng predcton of dsastrous hydrologcal phenomena. Currently, the man applcatons of cluster algorthms n hydrologcal forecastng nclude ranfall montorng, flood forecastng, 234 Copyrght c 2014 SERSC
water qualty montorng, nflow and outflow calculaton. The man clusterng methods contan k-means method, herarchcal method, SOM method and FCM method. (1) K-means method: In 2010, Wang Mn et al. [31] appled k-means clusterng algorthm to segment the SAR water mage. In the same year, the k-means was used to categorze the data ponts of the dfferent floodng characterstcs n the study area and dentfy the control pont(s) from ndvdual floodng cluster(s) when forecastng flood nundaton depths [32]. (2) Herarchcal method: In 2006, Sabu Paul et al. [33] grouped the water bodes nto clusters havng smlar watershed characterstcs wth Ward s mnmum varance method. Ths approach s helpful to dentfy possble sources and determne approprate models. Later PCA and ward s method were used to determne surface water orgn and ts nteractons wth groundwater n Medterranean streams n 2008 [34]. (3) SOM method: In 2007, Gwo-Fong Ln et al. [35] estmated desgn hyetographs of ungauged stes through a SOM-based approach and the number of clusters can be obectvely decded by vsual nspecton. Later, they proposed a hybrd neural network model combned the SOM and the multlayer perceptron network to forecast the typhoon ranfall [36]. In 2011, they used the SOM to analyze the nput data and reveal the topologcal relatonshps among nput data when developng a reservor nflow forecastng model [37]. 4.3.2 Water resources management: Clmate change and human actvtes have great mpacts on water resources. It manly ncludes precptaton change, runoff decrease, flood dsaster and so on. Therefore, reasonable confguraton and ntegrated plannng of water resources are mportant parts of the water resources management. Clusterng algorthms classfy dfferent watersheds wth smlar hydrologcal characterstcs. And under certan clmatc condtons, changes of precptaton, basn evaporaton and snow densty have been studed. At present, the clusterng methods n ths feld contan the k- means, herarchcal method, SOM method and FCM method. (1) K-means method: In 2005, k-means was performed amng to solate the weather regmes responsble for the nter-annual varablty of the wnter precptaton over Portugal [38]. In 2008, Mzukam, N. et al. [39] appled k-means clusterng to obtan qualtatve nformaton on spatal patterns of snowpack densty and densfcaton rates. Densty s one of the fundamental propertes of snowpack because t drectly affects many physcal propertes of snow. (2) Herarchcal method: In 2008, V. Guldal et al. [40] appled a herarchcal clusterng algorthm to cluster the monthly evaporaton losses wth the monthly wnds speed and wnd blow number of Eğrdr Lake n Turkey. Later, Danel G. Kngston et al. [41] dentfed hydrologc regons for the northern North Atlantc perphery, based on the ward s method of nter-annual varablty n 112 basns across ths regon. The prmary clmatc drvers of hydrologcal varaton between regonal classes have also been explored. (3) SOM method: In 2008, F-John Chang et al. [42] appled k-means and SOM algorthm to group watersheds wth smlar values for hydrologcal statstcs. In 2010, S. R. Fassnacht et al. [43] used SOM to defne regons of homogenety n the Colorado Copyrght c 2014 SERSC 235
Rver Basn usng snow telemetry snow water equvalent data. In the same year, Chang, F.-J. et al. [44] accessed the effort of meteorologcal varables for evaporaton estmaton by the SOM. The results demonstrated that the topologcal structures of SOM could gve a meanngful map to present the clusters of meteorologcal varables and the networks could well estmate the daly evaporaton. 4.4 Energy meteorology Wnd and solar energy are renewable resources, clean and envronmentally frendly, wth a broad development space and huge value. Ratonal use of meteorologcal resources, strengthenng the development and utlzaton of wnd and solar energy are some of the mportant ways to solve the energy problem as well as for energy savng. Cluster analyss technques may be used for short-term wnd power predcton. They can also assess the qualty of energy so as to dentfy the locatons whch are optmal for energy power plant. The man clusterng technques n ths feld are k-means method and herarchcal method. 5. Dscusson From the prevous secton, we can see the man dstrbuton of clusterng algorthms n meteorologcal applcatons. Then we gve a smple summary of the dstrbuton. The relatonshp between the clusterng algorthms and ts applcatons n meteorology are shown n Fgure 1. K-means Herarchcal clusterng DBSCAN EM SOM WaveCluster FCM SA Graph clusterng meteorologcal elements classfcaton Weather types classfcaton Atmospherc crculaton Hgh-mpact weather Ar qualty Weather forecastng Hydrologcal forecastng Water sources management Energy meteorology Clmate change Urban meteorology Hydrometeorology Energy meteorology Fgure 1. Algorthms/applcatons relatonshp map Combned the research n secton 3 and Fgure 1, we see summary from two aspects of clusterng algorthms and meteorologcal applcatons. In the term of the clusterng algorthms, we can see that k-means and herarchcal methods are wdely used n almost all the meteorologcal applcatons mentoned above for ther typcalty and dstnct advantages. The applcatons are manly n clmate change, urban meteorology, hydrometeorology and energy meteorology. FCM s a typcal fuzzy clusterng algorthm. Compared to the hard dvson, t fts many specal meteorologcal applcatons based on fuzzy partton. These specal applcatons are 236 Copyrght c 2014 SERSC
manly dstrbuted n urban meteorology and hydrometeorology. It s also appled to classfy Indan meteorologcal statons nto homogeneous groups due to ts advantage of assgnng every staton wth partal membershp n each group. As a method of neural network, SOM s wdely appled n hgh-dmensonal data vsualzaton applcaton for t can represent all ponts n a hgh-dmensonal source space by ponts n a lowdmensonal target space and try to mantan the topology of the nput space. Its applcatons are manly n clmate change, ar qualty and hydrometeorology. For some other algorthms, such as DBSCAN, EM, WaveCluster, SA and graph clusterng method, ther applcatons are lmted n solvng specfc meteorology ssues. DBSCAN s now manly used for dentfyng and trackng storm cell and studyng the lght flash characterstcs. EM s partcularly used to determne the typcal weather types, whle WaveCluster dentfes the outlers n meteorologcal data. SA and Graph clusterng are manly appled to nvestgate atmospherc crculaton and ts mpact on clmate change. In terms of meteorologcal applcatons and the real world, the clusterng algorthms are mostly used n applcatons whch are closely related to human lfe, such as urban meteorology and clmate change. Wth the rapd development of socety and technology, people are ncreasngly concerned about the qualty of lfe and scentfc producton, forecastng, montorng, etc. Ths s also verfed that people propose or create a varety of technques to meet the requrements of the humans, whch s n lne wth the law of human and socal development. Acknowledgements The authors wsh to thank the Natonal Natural Scence Foundaton of Chna (41475091, 91337102, 41275093, 61402234, 61402235), the Natural Scence Foundaton of Jangsu provnce (BK2012461, BK2011825), the Proect Funded by the Prorty Academc Program Development of Jangsu Hgher Educaton Insttutons and the College graduate research and nnovaton proects of Jangsu Provnce (2012). Dr. We Tan s the correspondng author. References [1] http://users.csc.calpoly.edu/~dekhtyar/466-sprng2009/lectures/lec09.466.pdf, Accessed on (2013) Aprl 15. [2] R. Xu and D. Wunsch, IEEE Transactons on Neural Networks, vol. 16, no. 3, (2005). [3] J. MacQueen, L. M. L. Cam and J. Neyman, Some methods for classfcaton and analyss of multvarate observatons, Proceedngs of the ffth Berkeley symposum on mathematcal statstcs and probablty, (1967) June 21-July 18, (1965) December 27, (1966) January 7, Berkeley, Calforna [4] F. Murtagh, The Computer Journal, vol. 26, no. 4, (1983). [5] G. McLachlan and T. Krshnan, Wley, vol. 1, (1997), pp. 997. [6] T. Kohonen, H. J. Trussell and V. Damle, Proceedngs of the IEEE, vol. 78, no. 9, (1990). [7] Y. H. Lu, T. H. Ma, C. H. Yn, X.Y. Xe, W. Tan and S. M. Zhong, Internatonal Journal of Database Theory and Applcaton, vol. 6, no. 6, (2013). [8] P. A. Jménez, J. F. González-Rouco, J. P. Montávez, E. García-Bustamante and J. Navarro, Internatonal Journal of Clmatology, vol. 29, no. 4, (2009). [9] C. Lols, Theoretcal and Appled Clmatology, vol. 96, no. 3, (2009). [10] İ. Sönmez and A. Ü. Kömüşcü, Theoretcal and Appled Clmatology, vol. 106, no. 3, (2011). [11] S. M. Robeson and J. A. Doty, Journal of Clmate, vol. 18, no. 8, (2005). [12] J. Kysely, J. Pcek and R. Huth, Studa Geophysca et Geodaetca, vol. 51, no. 2, (2007). [13] M. Burlando, M. Antonell and C. F. Ratto, Internatonal Journal of Clmatology, vol. 28, no. 5, (2008). [14] T. Blgn and A. Çamurcu, (Eds.), Advances n Informaton Systems, Sprnger Berln Hedelberg, New York, (2005). [15] D. B. Reusch, R. B. Alley and B. C. Hewtson, Journal of Geophyscal Research: Atmospheres, (1984-2012), vol. 112, no. D2, (2007). [16] R. Huth, C. Beck, A. Phlpp, M. Demuzere, Z. Ustrnul, M. Cahynov a, J. Kysel y and O. E. Tveto, Annals Copyrght c 2014 SERSC 237
of the New York Academy of Scences, vol. 1146, no. 1, (2008). [17] P. Esteban, J. Martn-Vde and M. Mases, Internatonal Journal of Clmatology, vol. 26, no. 11, (2006). [18] A. Phlpp, P. M. Della-Marta, J. Jacobet, D. R. Fereday, P. D. Jones, A. Moberg and H. Wanner, Journal of Clmate, vol. 20, no. 16, (2007). [19] D. R. Fereday, J. R. Knght, A. A. Scafe, C. K. Folland and A. Phlpp, Journal of Clmate, vol. 21, no. 15, (2008). [20] A. Phlpp, Theoretcal and Appled Clmatology, vol. 96, no. 1, (2009). [21] G. Cervone, P. Franzese, Y. Ezber, Z. Boybey, F. Bonch, B. Berendt, F. Gannott, D. Gunopulos, F. Turn, C. Zanolo, N. Ramakrshnan and X. Wu, Rsk Assessment of Atmospherc Hazard Releases usng K-means Clusterng, Proceedngs of IEEE Internatonal Conference on In Data Mnng Workshops, (2008) December 15-19; Psa, Italy. [22] L. L and S. Cheng, A Calculated Methodology of Regonal Contrbutons Based on MM5-CAMx n Typcal Cty: A 2006 Case Study of SO2 and Sulfate, Proceedngs of the 4th Internatonal Conference on Bonformatcs and Bomedcal Engneerng, (2010) June 18-20; Chengdu, Chna. [23] A. Mace, R. Sommarva, Z. Flemng and W. Wang, n Adaptve K-means for clusterng ar mass traectores, Edted H. Yn, W. Wang and V. Rayward-Smth, Sprnger Berln, Hedelberg, vol. 6936, (2011), pp. 1-8. [24] J. H. Casola and J. M. Wallace, Journal of Appled Meteorology and Clmatology, vol. 46, (2007). [25] S. Yonemura, S. Kawashma, H. Matsueda, Y. Sawa, S. Inoue and H. Tanmoto, Theoretcal and Appled Clmatology, vol. 92, no. 1, (2008). [26] C. Aff, A. L. Dutot, C. Jambert, M. Abboud, J. Adzan-Gerard, W. Farah, P. E. Perros and T. Rzk, Ar Qualty, Atmosphere & Health, vol. 2, no. 2, (2009). [27] S. Lee, C. H. Ho and Y. S. Cho, Atmospherc Envronment, vol. 45, no. 39, (2011). [28] I. J. Turas, F. J. González, M. L. Martín and P. L. Galndo, Atmospherc Envronment, vol. 40, no. 3, (2006). [29] J. L. Pearce, J. Bernger, N. Ncholls, R. J. Hyndman, P. Uotla and N. J. Tapper, Atmospherc Envronment, vol. 45, no. 1, (2011). [30] S. Khedara and M. T. Khadr, Atmospherc Research, vol. 113, (2012). [31] M. Wang, S. Zhou, B. Heng, M. Nng and Y. Song, SAR Water Image Segmentaton Based on GLCM and Wavelet Textures, Proceedngs of the 6th Internatonal Conference on Wreless Communcatons, Networkng and Moble Computng, (2010) September 23-25; Chengdu, Chna. [32] L. C. Chang, H. Y. Shen, Y. F. Wang, J. Y. Huang and Y. T. Ln, Journal of Hydrology, vol. 385, no. 1, (2010). [33] S. Paul, R. Srnvasan, J. Sanabra, P. K. Haan, S. Mukhtar and K. Nemann, Journal of The Amercan Water Resources Assocaton, (2006). [34] A. Mencó and J. Mas-Pla, Journal of Hydrology, vol. 352, no. 3, (2008). [35] G. F. Ln and M. C. Wu, Journal of Hydrology, vol. 339, no. 3, (2007). [36] G. F. Ln and M. C. Wu, Journal of Hydrology, vol. 375, no. 3, (2009). [37] G. F. Ln and M. C. Wu, Journal of Hydrology, vol. 405, no. 3, (2011). [38] J. A. Santos, J. Corte-Real and S. M. Lete, Internatonal Journal of Clmatology, vol. 25, no. 1, (2005). [39] N. Mzukam and S. Perca, Journal of Hydrometeorology, vol. 9, no. 6, (2008). [40] V. Guldal and H. Tongal, Appled Ecology and Envronmental Research, vol. 6, no. 4, (2008). [41] D. G. Kngston, D. M. Hannah, D. M. Lawler and G. R. McGregor, Hydrologcal Processes, vol. 25, no. 7, (2011). [42] F. J. Chang, M. J. Tsa, W. P. Tsa and E. E. Herrcks, Journal of Hydrology, vol. 354, no. 1, (2008). [43] S. R. Fassnacht and J. E. Derry, Water resources research, vol. 46, no. 4, (2010). [44] F. J. Chang, L. C. Chang, H. S. Kao and G. R. Wu, Journal of Hydrology, vol. 384, no. 1, (2010). Authors We Tan, Lecturer at the School of Computer and Software, also a doctoral canddate of atmosphere scence at School of Atmospherc Scence, Nanng Unversty of Informaton Scence and Technology. Hs research nterest covers data mnng, meteorologcal Data Processng, remote sensng mage processng, typhoon track mutaton mechansm. 238 Copyrght c 2014 SERSC
Yuhu Zheng, Assosate Processor at the School of Computer and Software, Nanng Unversty of Informaton Scence and Technology. Hs research nterest covers mage processng, pattern recognton, and remote sensng mage restoraton. Runzh Yang, Engneer at Natonal Meteorologcal Informaton Center of Chna. She receved her master degree from Behang Unversty,Chna, 2006. Her research nterest covers meteorologcal data processng, cloud computaton, and parallel computng. Sa J, Professor at the School of Computer and Software, Nanng Unversty of Informaton Scence and Technology. He receved hs Bachelor (NUIST, Chna, 1999), Master (NUAA, Chna, 2006). Hs research nterests are n the areas of Data Mnng, Computer Measurement and Control and Wreless sensor networks. Jn Wang, Professor at the School of Computer and Software, Nanng Unversty of Informaton Scence and Technology. Hs research nterests manly nclude routng protocol and algorthm desgn, performance evaluaton and optmzaton for wreless ad hoc and sensor networks. He s a member of the IEEE and ACM. Copyrght c 2014 SERSC 239
240 Copyrght c 2014 SERSC