Modelos preditivos aplicados ao risco de incêndios no Cerrado: desempenho e importância de variáveis ambientais

Abstract

The recurrence of wildfires in the Brazilian Cerrado severely threatens the region’s fauna, flora, and overall biodiversity. Between 2001 and 2019, studies reported a decline in the photosynthetic activity of tree foliage in the area, attributed to the frequent occurrence of these fires. In this context, the present study aimed to assess the impact of environmental and meteorological variables on fire risk estimation. Three machine learning models with distinct approaches were applied (Random Forest, XGBoost, and MLPRegressor) to replicate the fire risk index provided by INPE, using a reduced set of five variables: air temperature, accumulated precipitation, number of consecutive dry days, relative humidity, and atmospheric smoke concentration. Over 600 daily files covering the period from August 2023 to May 2025 were analyzed. After training and evaluating the models using cross-validation and regression metrics, the Random Forest model achieved the best predictive performance, with a final coefficient of determination (R2) of 0.8166. Based on this model, further analyses were conducted, including permutation and correlationbased feature importance, residual diagnostics, sensitivity tests, and spatial comparisons of predictions. Three variables—relative humidity, number of dry days, and precipitation— accounted for approximately 90% of the model’s predictive power. The sensitivity analysis confirmed this influence, as increased dry days and reduced humidity and precipitation led to higher predicted fire risk. Residual analysis showed errors centered around zero, with low variability and no systematic bias, while monthly spatial maps demonstrated strong alignment between predicted and observed values. Despite promising results, the study has notable limitations: the reduced number of variables, the absence of data on vegetation cover, wind, atmospheric pressure, and gas concentrations. Moreover, the mismatch between the temporal resolution of fire risk data (daily) and burned area data (monthly) hinders precise correlation between predicted risk and actual impact. Future work should incorporate additional environmental variables and complementary datasets to enhance model robustness and generalizability.