Smoke plume detection using U-Net model: an application in the Mexican territory

Forest fires are a global problem that risks human health, the environment, and the economy. The smoke produced by forest fires generates emissions of gases such as CO, CO2, NOx, and SOx, which can deteriorate air quality, cause respiratory diseases, and contribute significantly to climate change. In this concern, smoke plume detection becomes an essential task. In this sense, we propose a semantic segmentation model of smoke plumes applied to forest fires in Mexico using GOES-16 satellite images. To do that, first, a review of the images from the Advanced Baseline Imager (ABI) sensor was made, and also, a dataset of hand-segmented images was created to generate a semantic segmentation model based on deep learning. As a result of the first qualitative analysis of ABI images, we found that bands 1 (0.47 μm), 2 (0.64 μm), and the true color composite are the most relevant ones for the detection of smoke plumes because they presented the greatest ability to distinguish smoke plumes from the rest of the image effectively, at least visually. A dataset of 1061 human-segmented smoke plume images was generated to build and test the model. The segmentation model is done with a slight modification of a U-Net architecture; this model achieved intersection over union metrics of 0.8252, Dice coefficient of 0.9042, and accuracy of 0.9898 on the test set. The results demonstrate that developing and generating a smoke plume semantic segmentation model using satellite images and deep learning is feasible. This model can be used to monitor and detect forest fires from space, improving the management of these disasters.