Wildfire Susceptibility Mapping through Machine Learning and Remote Sensing Data with Distance Based Sampling for Fire-Free Points

Abstract

Wildfire is a common form of natural disaster present in Southeast Asia due to the high temperature and availability of “fuel” during the dry season especially in the form of peat. The negative impact of wildfire can be long lasting to the economy, and environment. Its occurrences are hard to predict given the number of variables that governs it. Thus, due to complex nature of wildfire, a machine learning based approach had seemed like the viable solution to the problem. An ANN model was developed for this study to predict and map out the wildfire susceptibility of the study area, which was Sibu, Sarawak, with data from remote sensing providers sampled through a distance-based approach. Variables chosen for this study to develop the ANN model was aspect, elevation, lithology type, land use and land cover, normalised difference vegetation index, proximity to rivers, and topographic wetness index. The machine learning model was evaluated to have a prediction rate area under the curve score of 0.89, and a precision score of 0.75, making it a viable solution to predict wildfire susceptibility.