11.07.2015 Views

Exploring the Spatial Distribution of Bird Habitat with ... - IEEE Xplore

Exploring the Spatial Distribution of Bird Habitat with ... - IEEE Xplore

Exploring the Spatial Distribution of Bird Habitat with ... - IEEE Xplore

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.2 Hierarchical Cluster and <strong>Spatial</strong>-Tree buildingbased on DBSCANAlbeit we have found suitable parameter <strong>of</strong> Eps and MinPts, this bringsabout o<strong>the</strong>r problems. If a user defines <strong>the</strong> neighborhood <strong>of</strong> a point byspecifying a particular radius and looks for core points that have apre-defined number <strong>of</strong> points <strong>with</strong>in that radius, <strong>the</strong>n ei<strong>the</strong>r <strong>the</strong> tight clusteron <strong>the</strong> map will be picked up as one cluster, and <strong>the</strong> rest will be marked asnoise, or else every point will belong to a single cluster. We <strong>the</strong>n canassume that <strong>the</strong> cluster area <strong>with</strong> small Eps would tend to be <strong>the</strong> childrennode for <strong>the</strong> bigger cluster area.Also, for MCP range area commonly includes areas never visited byanimal. In order to get <strong>the</strong> actual area used <strong>with</strong>in bar-headed goose homeranges, core areas were calculated. The core area usually defined as areasreceiving concentrated by individual at each wetland. For example, <strong>the</strong>fostering place would be core region and <strong>the</strong> foraging area would be <strong>the</strong>out core region. To find <strong>the</strong> core area in <strong>the</strong> breeding MCP range area oro<strong>the</strong>r bigger ones, we have to run <strong>the</strong> cluster approach again on this area<strong>with</strong> much smaller Eps. Meanwhile, <strong>the</strong>re is hundreds <strong>of</strong> core areascattering on <strong>the</strong> map. It is necessary to find an efficient and easy way forfur<strong>the</strong>r management <strong>of</strong> <strong>the</strong>se results. Motivated by this requirement, wedevelop a Hierarchical DBSCAN cluster method , which can reflect <strong>the</strong>relationship between different clusters, where <strong>the</strong> algorithm can build<strong>Spatial</strong>-Tree from top (large cluster) to bottom (several clusters), andcoding <strong>the</strong> index for every cluster as <strong>the</strong> <strong>Spatial</strong>-Tree growing.The problem we hope to solve out is how to cluster hierarchically,<strong>the</strong>se methods can be ei<strong>the</strong>r agglomerative or divisive. Agglomerativemethods (eg. AGNES) starts <strong>with</strong> as many clusters as variables and buildsa tree by successive join <strong>the</strong> two nearest clusters. On <strong>the</strong> o<strong>the</strong>r hand,divisive methods(eg. DIANA ) starts <strong>with</strong> one big cluster and successivesplits <strong>the</strong> largest cluster by <strong>the</strong> most dissimilar variable. AGNES andDIANA are fully described in chapters 5 and 6 <strong>of</strong> [12].Hierarchical cluster Approach:Different from <strong>the</strong> DIANA which select <strong>the</strong> cluster <strong>with</strong> largestdiameter and divided into two new clusters until all clusters contain onlyone variable. We utilize <strong>the</strong> breadth-first search (BFS), in graph <strong>the</strong>ory,BFS is a graph search algorithm that begins at <strong>the</strong> root node and exploresall <strong>the</strong> neighboring nodes to build <strong>the</strong> <strong>Spatial</strong>-tree from top to bottom. Ourhierarchical cluster approach employs <strong>the</strong> similar strategy like <strong>the</strong> BFS. Atfirst <strong>the</strong> only one cluster which contains all datasets was selected andDBSCAN <strong>with</strong> <strong>the</strong> predefined parameter would split this cluster into subcluster. Those cluster results were pulled into one queue, and <strong>the</strong>n <strong>the</strong> datasets from <strong>the</strong> queue was selected to run DBSCAN again. Our splittingphase will stop until <strong>the</strong> cluster level reach <strong>the</strong> defined level from top tobottom. When <strong>the</strong> cluster was spitted into smaller ones, that cluster wasindexed as well. Take <strong>the</strong> space into consideration we omit a detaileddescription <strong>of</strong> <strong>the</strong> hierarchy algorithm. But we ought to note <strong>the</strong> followingpoints:A. We need to initial two pointers: first and last one to identify <strong>the</strong>tree level in <strong>the</strong> QueueB. If <strong>the</strong> number <strong>of</strong> cluster result from his parents data node is lowerthan 2, we will not pull this cluster results into <strong>the</strong> queue.C. Members are in <strong>the</strong> same level <strong>of</strong> tree, <strong>the</strong>y will share <strong>the</strong> samecluster parameter Eps and MinPts we predefined before.D. We use one garbage collector to save <strong>the</strong> cluster which will be <strong>the</strong>noise member in <strong>the</strong> next tree level. Because <strong>the</strong> parameter willchange as <strong>the</strong> tree level change.E. As <strong>the</strong> <strong>Spatial</strong>-Tree growing, every cluster node will be encodedby <strong>the</strong>ir fa<strong>the</strong>r cluster id joining <strong>with</strong> its own cluster number.Consider <strong>the</strong> Fig 3, <strong>the</strong> depth <strong>of</strong> <strong>Spatial</strong>-Tree is four and <strong>the</strong> clusterarea also was encoded. For example: <strong>the</strong> left most leave node in tree wasencoded <strong>with</strong> 0/0/1/0.thus, we can know its parent node is 0/0/1 and itsgrandfa<strong>the</strong>r is 0/0.Fig 3: <strong>Spatial</strong>-TreeThe hierarchy cluster procedure has deployed <strong>the</strong> mainframe todiscover <strong>the</strong> habitat, and <strong>the</strong>n <strong>the</strong> next step is to define object similarityand cluster parameter for every tree level.Similarity based on <strong>the</strong> Geographical distanceAs previously stated, only a dissimilarity matrix is needed toperform a divisive cluster analysis, so we should concentrate ourefforts on this matrix efficiency. The location points in our paper aregeographic coordinates (latitude and longitude), geographicaldistance formula [7] was deployed to calculate <strong>the</strong> great-circledistance. In addition, bird migration route overlaps broadly <strong>with</strong> TheEast Asian Flyway <strong>of</strong> Anatidate and <strong>the</strong> East Asian-AustralasianFlyway <strong>of</strong> shorebirds[9]. We set <strong>the</strong> (sphere <strong>of</strong> radius) to 6378140meters in <strong>the</strong> distance formula 1:rd r . Where is <strong>the</strong>central angle, d is <strong>the</strong> geographical distance.Determining <strong>the</strong> Parameter <strong>of</strong> Hierarchy ClusterWe employed <strong>the</strong> following procedures which could allow us tocompute <strong>the</strong> approximately cluster parameters Eps and MinPts forhierarchy cluster method.1. In general, we only interest in location points that bird speciestend to stay longer ra<strong>the</strong>r than <strong>the</strong> location <strong>the</strong>y fly over occasionally,thus we used <strong>the</strong>TMinpts T / Tsum ,which was associated <strong>with</strong> <strong>the</strong> time birdmoving in one certain scope <strong>of</strong> area to define this phenomena.Therefore, <strong>the</strong> number <strong>of</strong> MinPts could be computed by formula 2:sum p . T p is <strong>the</strong> time interval which location132


signal we received. In our paper,Tsumis set as two days, because astopover event as <strong>the</strong> use <strong>of</strong> a migration area for more than two daysafter departure from ei<strong>the</strong>r <strong>the</strong> breeding area or <strong>the</strong> wintering area [9].And <strong>the</strong> average time T p for each bird from our data sets was twohours. Then <strong>the</strong> resulting MinPts can be 25.2. To get <strong>the</strong> parameter Eps, we develop a simple but effectiveheuristic to determine <strong>the</strong> parameters Eps. The heuristic is based on<strong>the</strong> following observation. The wintering and breeding area usuallycover at least ten thousand square kilometers wide region. We canassume <strong>the</strong> Eps would be at most ten thousand meters. O<strong>the</strong>rwise, if<strong>the</strong> Eps was much higher, <strong>the</strong> cluster area would be over estimatedand would not promise us as hoped. In addition, we have to reduce<strong>the</strong> amount <strong>of</strong> Eps value <strong>with</strong> multiples <strong>of</strong> 10 from <strong>the</strong> 10,000 m,1000 m, 100 m to 10 m to get a smaller region <strong>of</strong> space.3. The stay duration <strong>of</strong> one wild bird in its own habitat isnaturally determined. Look at from ano<strong>the</strong>r way, <strong>the</strong> temporaldistribution <strong>of</strong> bird location also can reflect <strong>the</strong> spatial distributionfrom our cluster approach. So we utilize <strong>the</strong> migration time recorded[9] and our observation to rectify <strong>the</strong> parameter Eps several times togain our ends.3.3 Minimum Convex Polygon Home RangeTo know <strong>the</strong> precise range <strong>of</strong> <strong>the</strong> habitat region, Kernel densityestimation has become prevalent in <strong>the</strong> home range analysis [3]. In our4 Discovery and analysis <strong>of</strong> cluster areaIn this section, we describe <strong>the</strong> results <strong>of</strong> discovering habitat on 2007survey and 2008 survey data sets using our hierarchy cluster approach.Because <strong>of</strong> space constraints, we mostly focus on <strong>the</strong> 2007 survey, andsome in 2008 survey also were presented. As one part <strong>of</strong> our discussions,we will analyze site-fidelity, implication for o<strong>the</strong>r research works as well.4.1 System framework <strong>of</strong> HAMapWe are developing HAMap (<strong>Habitat</strong> Area Map), a web GIS systembased on <strong>the</strong> Google map which greatly improves a user's Webvisualization experience. Fig 4 shows <strong>the</strong> system architecture <strong>of</strong> <strong>the</strong>Cluster Visual Map. Here, Google Map is a popular web based GISdeveloped by Google. The main tasks <strong>of</strong> <strong>the</strong> HAMap are listed as follows:.Our web server will download <strong>the</strong> latest data from <strong>the</strong> mail box and <strong>the</strong>nsave <strong>the</strong>m to <strong>the</strong> data base automatically..When a web Brower input <strong>the</strong> parameters in our .jsp page (Java serve netpage). The host analyzes <strong>the</strong> input parameter and translates <strong>the</strong>m to <strong>the</strong>cluster model.cluster model at first run <strong>the</strong> cluster algorithm on <strong>the</strong> datasets fromdatabase, and <strong>the</strong>n respond <strong>the</strong> user <strong>with</strong> <strong>the</strong> document stream as <strong>the</strong> .xmltype, which are detailed by <strong>the</strong> different cluster result such as <strong>the</strong> locationinformation, <strong>the</strong> area <strong>of</strong> convex hull and so on..The xml stream will be sent to Google Server net and <strong>the</strong> Brower willdisplay <strong>the</strong> cluster information based on <strong>the</strong> Google Map.work, MCP (Minimum Convex Polygon Home Range) would be moreproper in our home range estimation [10]. We first go to compute <strong>the</strong>polygon in this area which would be a convex . There are two algorithmsthat compute <strong>the</strong> convex hull <strong>of</strong> a set <strong>of</strong> n points. Graham’s scan runs inO(nlgn)time, and <strong>the</strong> Jarvis’s march runs in O(nh)times, where h is <strong>the</strong>number <strong>of</strong> vertices <strong>of</strong> <strong>the</strong> convex hull. Both Graham’s scan and Jarvis’smarch use rotational sweep, which processing vertices in <strong>the</strong> order <strong>of</strong> <strong>the</strong>polar angles <strong>the</strong>y form <strong>with</strong> a reference vertex. At first, Point <strong>with</strong> Maximor Minimum latitude was computed, because only <strong>the</strong>y will be <strong>the</strong> pointon <strong>the</strong> convex. And <strong>the</strong>n we utilize Graham-Scan to get <strong>the</strong> MCP. The runtime will be limited to O(n) time. A much more technical descriptioncould be seen in [11].Home range was computed based on spherical geometry from <strong>the</strong>convex results. A closed geometric figure on <strong>the</strong> surface <strong>of</strong> a sphere isformed by <strong>the</strong> arcs <strong>of</strong> great circles. The spherical polygon is ageneralization <strong>of</strong> <strong>the</strong> spherical triangle [12]. Ifs [ ( n2) ]* Ris <strong>the</strong> sum <strong>of</strong> <strong>the</strong>radian angles <strong>of</strong> a spherical polygon on a sphere <strong>of</strong> radius R, <strong>the</strong>n <strong>the</strong> areais:2Fig 4: The system architecture <strong>of</strong> <strong>Habitat</strong> Area Map4.2 <strong>Spatial</strong>-Tree for bar-head gooseThe different annual migration route would not overlap <strong>with</strong> eacho<strong>the</strong>r. This involves dividing those data sets into two parts: 2007 surveyand 2008 survey.2007 <strong>Spatial</strong>-Tree was built from top to bottom on <strong>the</strong> annualmigration data from 2007-03 to 2008-03. In <strong>the</strong> Fig 5, <strong>the</strong> left slide is <strong>the</strong>tree and <strong>the</strong> right is <strong>the</strong> spatial distribution associated <strong>with</strong> that node. Theconvex home range also was depicted <strong>with</strong> different color <strong>of</strong> polygon and<strong>the</strong> description would be shown when <strong>the</strong> user clicks <strong>the</strong> marker <strong>with</strong>certain index.When <strong>the</strong> <strong>Spatial</strong>-Tree depth is 3, <strong>the</strong>re are 6 clusters in <strong>the</strong> level 1<strong>with</strong> Eps is 35000 meters and MinPts is 25, and <strong>the</strong> average range <strong>of</strong>habitats area is 29045.38 Km 2 . In <strong>the</strong> map, we can clearly find <strong>the</strong>133


eeding area Qinghai lake <strong>with</strong> index 3, post breeding area Zhalin-Elinglake <strong>with</strong> index 4. The maximum one is <strong>the</strong> wintering area in Tibet rivervalley in China <strong>with</strong> 9254 Km 2 , and it is interesting to find that onespecies(No. BH07_67693) moved to cluster <strong>with</strong> index 1 <strong>with</strong>in Mongoliara<strong>the</strong>r than stay in Qinghai Lake for breeding. The movement <strong>of</strong>bar-headed goose depicted in this study conformed to <strong>the</strong> Central AsianFlyway [8]. There are eight species migrated to Tibet river valley, and stay<strong>the</strong>re from 2007-10-21 to 2008-02-25, <strong>with</strong> a total <strong>of</strong> 127 days for overwinter ra<strong>the</strong>r than fly to north-eastern India and Bangladesh.232.9636 Km 2 , and it mainly covers <strong>the</strong> salt lake illustrating <strong>the</strong>importance <strong>of</strong> salt lake. These salt lakes are integral to <strong>the</strong> hydrologydefining <strong>the</strong> landscape and vegetation characteristics <strong>of</strong> <strong>the</strong> region.In <strong>the</strong> third level <strong>of</strong> <strong>Spatial</strong>-Tree, 146 nodes are computed <strong>with</strong><strong>the</strong> Eps 300 meters and MinPts 25. We also focus <strong>the</strong> cluster result inQinghai Lake. The core area colored red which indexed as 2 in Fig 7precisely covers one famous small island named San Kuai Si <strong>with</strong>in<strong>the</strong> Qinghai Lake, a famous breeding place. Seven bar-heed geesehave stayed <strong>the</strong>re from 2007-04-10 to 2007-06-24 for 74 days. Thisillustrated that our cluster method could detect <strong>the</strong> core range areaprecisely.We also can discover <strong>the</strong> main movements <strong>of</strong> <strong>the</strong> fourteenbar-headed geese in 2007 in Fig.5. They depart <strong>the</strong> breeding place inQinghai Lake and <strong>the</strong>n arrive at <strong>the</strong> post breeding area inZhalin-Eling lake area or Huangheyuan wetland and stay <strong>the</strong>re abouttwo months before heading to south. Then <strong>the</strong>y follow <strong>the</strong> cluster 2and cluster 5 which served as <strong>the</strong> stop over area, and finally arrivingat <strong>the</strong> winter area. Mean fall migration time was about 7 days.Fig 5: Overview <strong>of</strong> 2007 bar-headed goose <strong>Spatial</strong> TreeFig 8: Overview <strong>of</strong> 2008 bar-headed goose <strong>Spatial</strong> Tree2008 <strong>Spatial</strong>-Tree was built on <strong>the</strong> location data from 2008 surveyFig 6: Core area <strong>of</strong> Qinghai lake breeding areaduring <strong>the</strong> period from 2008-03-01 to 2008-11-01 in <strong>the</strong> Fig 8. The sizeand distribution <strong>of</strong> post breeding and breeding area approximatelyconformed to <strong>the</strong> cluster result from 2007 survey. This suggests thatbar-headed goose did not move randomly during <strong>the</strong> whole process <strong>of</strong>migration. At last, seven birds have arrived at <strong>the</strong> wintering place inTibet river valley. The fall migration distance is about 1600 km.4.3 Discussion and Implications for habitatconservation and avian influenzaThere are two migration seasons for bar-headed goose: springFig 7: San kuai Si island core area <strong>with</strong>in Qinghai LakeThe Eps and MinPts for <strong>the</strong> second level <strong>of</strong> <strong>Spatial</strong>-Tree are 3000meters and 25. As a result, <strong>the</strong>re are 34 cluster nodes in <strong>the</strong> second level<strong>of</strong> <strong>Spatial</strong>-Tree. The average size is 282.96 Km 2 . Qinghai lake breedingarea is one <strong>of</strong> cluster area in <strong>the</strong> first level, and <strong>the</strong>re are six sub clusterarea <strong>with</strong>in its MCP area in <strong>the</strong> Fig 6. There are 14 birds staying at thisrange <strong>of</strong> area for 231 days. The average size <strong>of</strong> core home range ismigration and fall migration. Biological scientist always aspired to knowwhe<strong>the</strong>r those species would fly back <strong>the</strong>ir original habitat in <strong>the</strong> next year.This is <strong>the</strong> so called site fidelity. On o<strong>the</strong>r hand, <strong>the</strong> site fidelity willreflect <strong>the</strong> quality <strong>of</strong> our cluster area result more or less. Three birds in2007 survey still active even on Nov 1. We choose <strong>the</strong> species(NO.BH07_6795) from <strong>the</strong>m and draw its location data from <strong>the</strong>2008-04-01 to 2008-11-1 on our 2007 discovered habitat in Fig 9. Those134


location points <strong>with</strong> green color cover most <strong>of</strong> our discovered MCP in2007. This not only proves this species would like to return <strong>the</strong>ir habitatnext year, but also shows <strong>the</strong> outcomes <strong>of</strong> our research to some degree.possible solution is to extract <strong>the</strong> possible candidate sequence based onsome famous sequence mining approach. Also, we have begun toinvestigate comparing <strong>the</strong> cluster area spatial distribution betweendifferent years, and hope to discover <strong>the</strong> habitat changing trend <strong>of</strong> birdspecies. Finally, we intend to extend our analyses to o<strong>the</strong>r species in <strong>the</strong>Qinghai Lake to identify <strong>the</strong> cross habitat for different species.Fig 9: Site-Fidelity <strong>of</strong> Spring Migration in 2007Our cluster based approach for discovering <strong>the</strong> bar headed gooseapproximately depicts <strong>the</strong> geographical distribution <strong>of</strong> this wild water bird.Both <strong>of</strong> <strong>the</strong> cluster results in 2007 and in 2008 matched greatly, whichshowed us that some certain habitat such as <strong>the</strong> Qinghai Lake, DaLingLake and <strong>the</strong> Tibet river valley are <strong>of</strong> vital importance for some species tolive. The wide area <strong>of</strong> MCP proves that it is necessary to build one broadnetwork to cover <strong>the</strong> different core region area. The cluster resultdisplayed in <strong>the</strong> GIS paves <strong>the</strong> way for human beings to construct onesystematic nature reserve in future. In addition, scientists would do muchmore research works such as virus, plant and environment quality surveyto discover <strong>the</strong> way <strong>of</strong> highly pathogenic avian influenza disperse in <strong>the</strong>wild bird species' MCP.5 Conclusion and Future WorkSatellite tracking has been used successfully to determine <strong>the</strong>migration routes and stopover sites <strong>of</strong> a number <strong>of</strong> birds. Suchinformation allowed <strong>the</strong> development <strong>of</strong> a future plan for protecting <strong>the</strong>breeding and stopover sites. However, <strong>the</strong> lack <strong>of</strong> an efficient dataanalyzing approach presents an obstacle to help researchers to do thiswork automatically and precisely. In this paper we suggest to exploit <strong>the</strong>field <strong>of</strong> using <strong>the</strong> location data information as a supplement clusteringprocess which can provide an alternative approach for visual observationin counting <strong>the</strong> location plots for finding <strong>the</strong> wild bird species’ habitat.Meanwhile, in order to discover <strong>the</strong> core rang, a new cluster strategy wasproposed: A hierarchy cluster based on <strong>the</strong> BFS. Our initial analysis <strong>of</strong> <strong>the</strong>clustering results suggests that this clustering strategy can effectivelymanage <strong>the</strong> different cluster area and discover <strong>the</strong> spatial relationship <strong>of</strong>cluster area. It does provide an effective assistance for biologicalresearchers in discovering new habitat.In <strong>the</strong> future, we plan to extend our current work to address severalunresolved issues. Specifically, we want to figure out if <strong>the</strong>re is anyfrequent and interesting visiting sequence for different wild birds. OneAcknowledgementsThis work was supported in part by Chinese Academy <strong>of</strong> Sciences(No.INF105-SDB) and Special Project <strong>of</strong> Informatization <strong>of</strong> ChineseAcademy <strong>of</strong> Sciences in" <strong>the</strong> Eleventh Five-Year Plan",No.INFO-115-D02. The authors would like to thank <strong>the</strong> researchers inWestern Ecological Research Center, Patuxent Wildlife Research Centerand United States Geological Survey for sharing <strong>the</strong> data sets <strong>with</strong> us.Reference[1] Liu, J., Xiao, H., Lei, F., Zhu, Q., Qin, K., Zhang, X. W., Zhang, X. L., Zhao, D., Wang,G. & o<strong>the</strong>r authors (2005). Highly pathogenic H5N1 influenza virus infection inmigratory birds. Science 309, 1206.[2] Li, Z. W. D. and Mundkur, T. (2007) Numbers and distribution <strong>of</strong> waterbirds andwetlands in <strong>the</strong> Asia-Pacific region: results <strong>of</strong> <strong>the</strong> Asian Waterbird Census 2002–2004.Kuala Lumpur, Malaysia: Wetlands International.[3] Worton.B.J.. 1989 kernel methods for estimating <strong>the</strong> utilization distribution inhome-range studies.Ecology 70. 164-168[4] Ester, M., Kriegel, H., Sander, J. and Xu, X. A density-based algorithm fordiscovering clusters in large spatial databases <strong>with</strong> noise. Proc. 2nd Int. Conf. onKnowledge Discovery and Data Mining. Portland, OR. 1996, pp. 226-231.[5] Ng R.T., and Han J. 1994. Efficient and Effective Clustering Methods for <strong>Spatial</strong>Data Mining, Proc. 20th Int. Conf. on Very Large Data Bases, 144-155. Santiago, Chile.[6] L. Ert¨oz, M. Steinbach, and V. Kumar. Finding topics in collections <strong>of</strong> documents:A shared nearest neighbor approach. In Proceedings <strong>of</strong> Text Mine’01, First SIAMInternational Conference on Data Mining, Chicago, IL,USA, 2001.[7] Great-circle distance http://mathworld.wolfram.com/GreatCircle.html[8] Miyabayashi, Y. and Mundkur, T. (1999) Atlas <strong>of</strong> key sites for Anatidae in <strong>the</strong> EastAsian Flyway. Tokyo: Japan, and Kuala Lumpur, Malaysia: WetlandsInternational—AsiaPacific.Available at http://www.jawgp.org/anet/aaa1999/aaaendx.htm.Accessed 11 March2008.[9] Muzaffar S.B, Johny.Takekawa,. Seasonal movements and migration <strong>of</strong> Pallas’sGulls Larus ichthyaetus from Qinghai Lake, China Forktail 24(2008):100-107[10] Hooge, P. N. And B. Eichenlaub. 2000. Animal movement extension to ArcViewver. 2.0 and later. USGS Alaska Biological Science Center, Anchorage,Alaska.[11] Guo Hao Shan, Li JunYu, Lin QiYing. Introduction to ACM internationalCollegiate Programming Contest Secnod Edition. 100-102 (in Chinese)[12] Weisstein, Eric W. "Spherical Polygon." From MathWorld--A Wolfram WebResource. http://mathworld.wolfram.com/SphericalPolygon.html[13] Kaufman, L., Rousseeuw, P.J.: Finding Groups in Data: An Introduction to ClusterAnalysis. John Wiley and Sons, Inc., New York (1990)135

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!