Density-based Cluster Analysis Of Fire Hot Spots In Kenya's Wildlife Protected Areas

Abstract

Wild res occurring in Kenya's wildlife protected areas pose a signi cant risk to wildlife conservation since they cause biodiversity loss and habitat degradation. There is a need for the Kenya Wildlife Service (KWS) to identify the regions in the protected areas that are prone to recurring wild re outbreaks during the re season. This study identi ed regions that are re hot spots in Kenya's protected areas by performing a density-based cluster analysis on the Moderate Resolution Imaging Spectroradiometer (MODIS) MCD14ML active re data set for a 12 year period between 2003 and 2014. Feature subset selection was done using an AWK script written to extract the latitude and longitude elds from the data set. QGIS was used to lter re points falling outside protected area boundaries. The Environment for Developing Knowledge Discovery in Databases Applications Supported by Index Structures (ELKI) implementation of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm was used for the clustering. A sorted k-dist graph estimated the initial DBSCAN parameters. 25 trial runs of DBSCAN with di erent parameters were used to select the nal values: MinPts = 7 re points; Eps = 700 meters. A web application with a Google Maps interface was developed to provide an interactive visualization of the re hot spots. 4,968 re incidents were observed in 73% of the protected areas. The initial DBSCAN parameters yielded 29 insigni cant re hot spot clusters from these incidents, while the nal parameters yielded 43 signi cant clusters. The 43 clusters were identi ed in 31% of the protected areas that recorded re activity. 60% of these clusters occurred in four protected areas. The ndings of this study indicate that density-based cluster analysis is a suitable clustering method for identifying hot spots in geospatial data sets. For DBSCAN, the performance of the sorted k-dist graph heuristic is in uenced by the characteristics of a data set. The results also indicate that Chyulu Hills, Dodori, Boni, and Ruma are the protected areas most vulnerable to wild res in Kenya. This study recommends the use of density-based cluster analysis for identifying hot spots in geospatial data sets. Experimentation with a wide range of DBSCAN parameters values is advisable. KWS should focus re management e orts on the identi ed re hot spot regions. In addition, it should investigate the impact of wild re damage in the ecological zones surrounding the hot spots.