Overview

Brazil faces an increasing wildfire crisis, with 278,229 heat spots recorded in 2024 alone, the highest number in 14 years according to INPE. The Amazon region was the most affected biome, with over 140,000 spots representing a 77% increase compared to 2023. This project addresses the challenge of reactive fire detection by using machine learning to predict where fires are most likely to occur based on environmental conditions, enabling preventive action before fires spread. The system uses data from SISAM (Sistema de Informações de Saúde Ambiental), containing heat spot records across Brazil since 2003, with 14 environmental and atmospheric variables including gas concentrations (CO, NO₂, O₃, PM2.5, SO₂), meteorological data (precipitation, temperature, humidity, wind) and geographic coordinates. Three ML models were trained with K-Fold cross validation: Random Forest, MLP (Neural Network with 100x100 layers) and XGBoost, all with threshold optimization to maximize F1 score. The architecture includes a FastAPI backend for model inference and a React frontend for interactive predictions, model comparison and metrics visualization.