Application of Classification Algorithms in Data Mining for Hotspots Occurrence Prediction in Riau Province Indonesia

High fire occurrence in Riau Province, Indonesia has been going on in the recent years with large areas occurring in the peat soil. In this paper a data mining technique namely classification was applied on forest fire data to develop classification models for hotspots occurrence in Riau Province. The models provide characteristics of areas where active fires occurred. We studied physical data including land cover, road, river, city centers, industrial timber plantation, logging concession, peatland depth and peatland types to classify 2693 target objects. Target objects are true alarm data namely hotspots distribution in 2008 and false alarm data which are randomly generated within the areas at least 1 km away from any true alarm data. We applied three classification algorithms that are available in the data mining toolkit Weka 3.6.2: J48 module as Java implementation of C4.5 algorithm, SimpleCart and NaïveBayes. The result shows that the classifier generated from the J48 has highest accuracy i.e. 69.59 % compared to two other algorithms. Our results based on the J48 classifier show that hotspots are predicted to take place in areas that (1) are non logging concession areas, (2) are plantation and dryland forest, and (3) have peatland type: Very deep Hemists/Saprists (> 400 cm). Additionally, hotspot occurrence probability is higher in areas located 10 km from roads, 3 km from rivers and within 5 km to 20 km of city centers where the areas are accessible to humans.