Previsão de incêndios no centro-oeste brasileiro com modelos de aprendizado de máquina

Abstract

This study investigates the application of Support Vector Machine (SVM) algorithms for forest fire prediction, aiming to improve early detection and reduce environmental damage in the Central-West region of Brazil. The research integrates three complementary datasets: Firewatch (2024), which contains information on affected cities and Fire Radiative Power (FRP); Climate Weather (1986–2017), which provides historical climatic variables; and a set of socioeconomic indicators that describe the human context of vulnerable areas. The integration of these datasets enables the development of machine learning models focused on predicting the occurrence and intensity of forest fires. The combined use of climatic, radiative, and socioeconomic factors aims to increase prediction accuracy and provide technical support for prevention and control policies. The SVM model was trained using one-hot encoding, also incorporating meteorological variables, and achieved an F1-score between 98% and 99%. The results indicate that SVM models integrating multiple data sources can serve as a useful tool for environmental monitoring and decision-making in forest fire management, contributing to the protection of ecosystems and at-risk communities.