Climate-vegetation intersection in determining the burn rate in an area of the Brazilian Cerrado

This study aimed to develop a predictive model, using climatic data and vegetation indices collected via satellite from 2001 to 2021, to estimate the daily burn rate in a protected area of the Brazilian Cerrado. For this purpose, data from the MODIS sensor was used, covering variables such as temperature, evapotranspiration, active fire, and burned area, complemented by climatic information from CHIRPS and Terraclimate products. The determination of the burn rate involved the polygonization of burned areas and the accounting of the duration of fires. In this context, machine learning algorithms such as Random Forest, Multilayer Perceptron, and SVM were explored, implemented through the Scikit-learn library. The focus of the study was to evaluate which of these algorithms would show the most suitable performance in predictions. The results point to an annual average of 50,000 hectares affected by wildfires in the protected area and its vicinity, with notable variations over the years analyzed. A correlation was identified between the daily burn rate and variables such as wind speed and temperature, while NDVI and evapotranspiration showed an inversely proportional relationship. The results suggest that, among the tested models, Random Forest may have a relatively more efficient performance, standing out in the accuracy of predictions. However, it is emphasized that caution and additional studies are necessary to confirm these results. In summary, our foundings indicates the potential of machine learning techniques in environmental management of wildfires, potentially contributing to the protection of the Cerrado ecosystems, although further investigations are necessary for a more comprehensive understanding.