GeoAI for Disaster Mitigation: Fire Severity Prediction Models using Sentinel-2 and ANN Regression

Abstract

Wildfire is a common disaster that hits Indonesia every dry season, especially on the islands of Kalimantan and Sumatra. In order to reduce the impact of fire hazards, preventive measures are needed before the occurrence of fires. One of them is by setting up an information system such as EWS. The aim of this study is to create an effective image- and machine learning-based predictive model of the severity of forest and land fires based on vegetation conditions prior to burning. Three parameters of prefire vegetation conditions, namely vegetation greenness indices, vegetation moisture, and vegetation senescence, were selected as independent variables to predict the postfire dependent variable, i.e., fire severity. There are 25 vegetation greenness index options tested, using either ANN regression or multiple linear regression. The vegetation moisture information is represented by the Normalized Difference Moisture Index (NDMI). The vegetation senescence information is extracted using the Plant Senescence Reflectance Index (PSRI). Meanwhile, the wildfire severity is measured using the Burned Area Index for Sentinel-2 (BAIS2). All vegetation conditions and wildfire severity information were extracted from Sentinel-2 imageries. The topology of ANN regression models is configured from one to six hidden layers. More than 100,000 pixels are used as samples, which are then separated into training samples and validation samples. The results of model development and testing show that ANN regression with Inverted Red-Edge Chlorophyll Index (IRECI) as a vegetation greenness parameter is the model that has the highest accuracy in predicting wildfire severity.