Real-Time Active Fire Detection in the Pantanal Biome, Brazil, Using Convolutional Neural Networks

The use of artificial intelligence in identifying active fire zones serves as an essential methodology for monitoring and quantifying the impacts of forest fires, especially in geographically isolated or difficult-to-access regions, such as the Pantanal biome in Brazil. This research analyzed active fire detection using remote sensing datasets combined with machine learning algorithms within a tropical river basin in the Encontro das Águas State Park—the second-largest reserve in the Brazilian Pantanal. In this study, data acquired from Sentinel-2 satellite imagery and a Convolutional Neural Network (CNN) model facilitated the identification of active fire pixels. A total of 15 images, with a spatial resolution of 20 m and dimensions of 2496 × 2489 pixels, captured between August and October 2020, served as the experimental dataset. Following data preprocessing, 1500 images, each measuring 256 × 256 pixels, were analyzed using the U-Net architecture to identify active fire zones. Subsequent processing yielded 15 masks, each measuring 2560 × 2560 pixels and corresponding to specific dates within the observational period. These masks served as binary indicators of active fire presence and were validated against data from the BDQueimadas platform. Observations indicated that most fires occurred in September, with 84.7% of all active pixels detected during this month. Peak activity was observed on the final day, with 20,177 activated pixels, representing roughly one-third of the cumulative total. A marked decline in detected fires followed this peak, continuing until the end of the monitoring period. In summary, the CNN model demonstrated efficacy in monitoring and assessing wildfire progression on a large geographical scale, specifically within a tropical river basin in the Pantanal biome of Brazil.