Peatland Data Fusion for Forest Fire Susceptibility Prediction Using Machine Learning

Abstract

Forest fires have been a severe hydrometeorological hazard during the dry season in Indonesia. Pulang Pisau Regency in Central Kalimantan has become one of the most forest fires affected areas during the 2015 El Nino event. Based on MODIS data, more than 120.000 hotspots have been recorded between 2014 and 2019. Previous studies concluded that peatlands act as contributing factor to forest fires in this country. This study proposed the peat-effect on the development of machine learning models for forest fire susceptibility (FFS), which can be alternative tool to support forest fire disaster management. In addition to the peat effect, such as elevation, slope, Normalized Difference Vegetation Index (NDVI), rainfall, distance from the road network, and distance from the residents also analyzed. Those variables were divided into training (2014 – 2018) and testing (2019). Random Forest (RF), Support Vector Classifications (SVC), and Gradient Boosting Classification (GBC) models were used to build the FFS map. The experiment results showed an increase in Area Under Curve (AUC) from 0.84 – 0.87 to 0.87 – 0.88 with the addition of the peat depth variable. The complete test resulted in the highest accuracy of 0.80 in the RF and SVC.