Detection of forest fires using artificial intelligence

Abstract

With the advent of climate change comes the fear wildfires will become a rising concern in the near future as is hinted by several environmental studies. This fear has already become a reality for some parts of the globe. This work implements and compares different deep learning architectures for flame semantic segmentation on RGB images of fires occurring in a natural environment taken from the ground or from an unmanned aerial vehicle (UAV). The Corsican Fire Database is exploited after comparing it to other candidate public datasets. Results are compared in terms of the intersection over union (IoU), the mean squared error (MSE), the binary accuracy and the recall metrics as well as their number of network parameters. The implemented architectures are the FLAME U-Net, the DeepLabv3+ architecture considering the EfficientNet-B4 and the ResNet-50 backbones, the Squeeze U-Net as well as the ATT Squeeze U-Net. Notable among the evaluated architectures, the DeepLabV3+ with an EfficientNet backbone was the one that achieved the best results with an IoU of 0.93 and a recall of 0.967 while exploiting 22M parameters; and the ATT Squeeze U-Net that scored very decently with an IoU of 0.893, a recall of 0.928 and the least amount of network parameters (885K). All implementations were made public.