Modelo preditivo de incêndios florestais para o semiárido nordestino baseado em redes neurais artificiais

Abstract

Forest fires pose a serious threat to the ecological and socioeconomic balance, especially in the northeastern semiarid region of Brazil, an area characterized by long dry seasons and fire-susceptible vegetation. The ability to predict the occurrence of fire hotspots is, therefore, an essential tool for planning mitigation actions and rapid response. This study aimed to develop a predictive model for forest fire hotspots in the Sertão de Crateús region, Ceará, using Artificial Neural Networks (ANN). For this purpose, data on fire hotspots from satellites (MODIS and VIIRS) and meteorological variables (temperature, solar radiation, humidity, precipitation, wind, and atmospheric pressure) from INMET were collected, covering the period from 2019 to 2024. The methodology involved processing and integrating these data to train, validate, and test a Multi-Layer Perceptron (MLP) type ANN. The analysis of variable importance, carried out using feature permutation, revealed that maximum atmospheric pressure (PATMAX) and global solar radiation (RADGLB) were the factors with the greatest impact on the prediction. The final model achieved an overall accuracy of 72% on the test data, with a sensitivity of 77%, indicating a high capacity to correctly identify actual fire occurrences. The analysis of the learning curves demonstrated stable convergence with controlled overfitting. The results indicate that ANN modeling is a promising and effective approach for predicting fire hotspots, providing a solid basis for the development of early warning systems and assisting in resource management for firefighting in the semiarid region.