Automatic Early Detection of Wildfire Smoke With Visible Light Cameras Using Deep Learning and Visual Explanation

The present work focus was developing a system for early automatic detection of smoke plumes in visible-light images. The system used a realistic dataset gathered in 274 different days from a total of nine real surveillance cameras, with most smoke plumes being viewed from afar and 85% of them occupying less than 5% of the image area. We employed the innovative strategy of using the whole image for classification but "asking" the neural networks to indicate, in a multidimensional output, which image regions contained a smoke plume. The multidimensional output helped to focus the detector on the smoke regions. At the same time, the use of the whole image prevented wrong image classification caused by a constrained view of the landscape under analysis. Another strategy used was to rectify the detection results using a visual explanation algorithm, Gradient-weighted Class Activation Mapping (Grad-CAM), to ensure that detections corresponded to the smoke regions in an image. The detection algorithms tested were residual neural networks (ResNet) and EfficientNet of various sizes because these two types have given good results in the past in multiple domains. The training was done using transfer learning. Our dataset contained a total of 14125 and 21203 images with and without smoke, respectively, making it, to the best of the author's knowledge, one of the largest or even the largest reported dataset in the scientific literature in terms of the number of images with smoke collected from large distances of various kilometers. This dataset was fundamental to achieve realistic results concerning smoke detection efficiency. Our best result in the test set was an Area Under Receiver Operating Characteristic curve (AUROC) of 0.949 obtained with an EfficientNet-B0.