Tropical Wildfires Analyzed Through Remote Sensing and Machine Learning

Abstract

This study provides a new contribution to understanding wildfire dynamics in Alagoas, Northeastern Brazil, by applying machine learning techniques to identify stable spatial and temporal patterns of fire activity between 2012 and 2024. Active fire clusters were detected using the Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm applied to VIIRS satellite data (S-NPP and NOAA-20). Associated atmospheric emissions of CO2, PM2.(5), and black carbon were assessed using data from the Copernicus Atmosphere Monitoring Service (CAMS), while vegetation variability was characterized through the Normalized Difference Vegetation Index (NDVI). The results reveal distinct regional fire regimes: the Litoral and Zona da Mata exhibited the highest peaks of Fire Radiative Power (> 100 MW/h), largely linked to sugarcane burning, whereas the Sert & atilde;o and S & atilde;o Francisco Sert & atilde;o regions experienced more persistent events lasting up to 90 days, reflecting semiarid conditions and low soil moisture. Strong and highly significant correlations were found between FRP and atmospheric pollutants, particularly in the Litoral (r = 0.96; p = 0.0000), confirming the direct role of fire intensity as a determinant of emissions. The integration of clustering techniques with atmospheric and vegetation indicators highlights critical areas of vulnerability, providing new insights to guide land management strategies, fire prevention, and mitigation efforts aimed at containing air quality deterioration and environmental degradation in Alagoas.