Previsão de séries temporais de queimadas no Pantanal com uso de aprendizado de máquina

Abstract

The Brazilian Pantanal comprises a biome responsible for an enormous diversity of fauna and flora. However, wildfires affecting this biome have become a major problem in recent years, generating impacts that may become irreversible in the future. Therefore, fire prediction methods prove to be essential and extremely important for the prevention and control of wildfires throughout this territory. Nevertheless, the main fire risk indices currently in use present limitations, such as: (1) not being adapted to the characteristics of each biome; (2) being restricted to specific climatic variables; and (3) not being able to predict forest fire risk over a defined multi-day horizon. This last aspect, in particular, is of utmost importance, as its improvement enables planning and coordinated action by environmental authorities with adequate advance notice. In order to investigate this limitation, this study developed a comparative approach between two computational models based on Machine Learning (XGBoost and LSTM), evaluating which one produced results closer to the observed reality. The results obtained indicated that the XGBoost model showed superior performance in predicting fire occurrences, achieving lower values of mean absolute error (MAE) and root mean square error (RMSE), in addition to a greater ability to explain the variability observed in the time series. The LSTM model, in turn, demonstrated the ability to capture seasonality and the average behavior of the series; however, it showed limitations in reproducing the magnitude of extreme events, especially due to univariate modeling and the effective volume of data available for training. Overall, the findings highlight that approaches based on boosting techniques combined with temporal feature engineering were more suitable for the analyzed dataset, although both methodologies confirmed the existence of learnable temporal patterns in the historical series of the Brazilian Pantanal.