Incremental Clustering on Hotspot Data as Forest and Land Fires Indicator in Sumatra

Abstract

Hotspot as one of forest and land fires indicators is regularly monitored for early detection the fires. Large fires are difficult to detect based on a single occurrence of hotspot. However, hotspots that are occurred in high density clusters can be used as indicator of large forest and land fires. Our previous studies have applied several clustering algorithms on hotspot datasets in Sumatra and Kalimantan. However, those studies have not included new hotspot data in the clusters of hotspots. This study aims to implement incremental clustering on hotspot datasets in Sumatra. Clustering was performed using the modified spatio-temporal density based clustering algorithm namely ST-DBSCAN. The main process of the incremental clustering in hotspot datasets is to update the initial clusters based on the existence of new hotspot data without doing clustering from the beginning. The incremental clustering algorithm has three parameters namely Eps1, Eps2 and MinPts. Eps1 is the distance parameter for spatial attributes whereas Eps2 is the distance parameter for non-spatial attributes. Eps represents maximum radius of the neighbourhood. MinPts is the minimum number of objects within Eps1 and Eps2 distance of an object. The algorithm was applied on the hotspot dataset in Sumatra in the period of September to October 2017. As many 2659 hotspots were used in initial clustering and 1023 new hotspots were used for updating the initial clusters. The initial clustering at the parameter eps1 of 0.1, eps of 7, and MinPts of 5 results 16 clusters and 23 outliers. The incremental clustering at those parameters results 7 new clusters and 6 new outliers.