Forest fire risk prediction from satellite data with convolutional neural networks

Forest fires cost the world an estimated value of 200 Billion dollars annually in damages. Furthermore, the main concerns are not only monetary as the vanishment of the carbon-dioxide soaking forests further exacerbates climate change. This paper presents a predictor system based on deep convolutional neural network to predict the risk level of wildfire from satellite data. The proposed Neural Network has an encoder-decoder architecture that allows to provide emergency operators with a pixel-wise fire risk prediction of a given area, allowing precise preventive interventions. The dataset utilised for the training has been generated from publicly available sources as a set of raster images, including several of the most significant satellite products. The paper also proposes a customised loss function for the training of the network and several statistical metrics to establish its performances and validate the reliability of the system. A proof of concept demonstration is discussed for two different case studies: the island of Sicily and an area in California.