Análise comparativa do desempenho de modelos de machine learning na previsão de focos de incêndio no cerrado utilizando variáveis climáticas

Abstract

Predicting fire outbreaks in the Cerrado biome using meteorological variables is a relevant problem for environmental monitoring and supporting alert systems. In this work, a comparative analysis of classical machine learning models for hourly prediction of fire outbreak occurrence was carried out, based on the integration between INMET data and records from INPE's BDQueimadas. The Logistic Regression, Naive Bayes, linear SVM, Random Forest, and XGBoost models were evaluated in different data preparation scenarios, including original datasets, datasets with derived variables, KNN imputation, and imbalance treatment strategies such as SMOTE and weight balancing. The results showed that tree ensemble-based models were the most suitable for the problem, especially XGBoost, and that feature engineering and explicit imbalance treatment contributed decisively to increased performance, especially in metrics more sensitive to the detection of the positive class, such as PR - AUC and F1-score.