Queimadas e fatores associados a produtividade primária bruta no cerrado: uma abordagem de machine learning explicável

Abstract

Primary Growth Production (GPP) is one of the fundamental components for the planet. It is related to climate, vegetation, agribusiness, water cycle, health, as well as fires and burning, and consequent climate changes, etc., and is also the base of the food chain. Objective: To analyze the most influential variables of Growth Primary Production (GPP), through Explainable Artificial Intelligence (XAI) and techniques and SHapley Additive ExPlanations (SHAP) in the Brazilian Cerrado. Methodology: Eleven meteorological stations from three states (Goiás, Mato Grosso and Mato Grosso do Sul) and the Federal District, composed the sample. The data related to the variables Weather, Climate and Vegetation and Fire outbreaks, of the 2003-2020 time series were obtained from the Moderate Resolution Imaging Spectroradiometer (MODIS)/National Aeronautics and Space Administration (NASA) remote sensor, from the National Institute of Meteorology (INMET) and from the Queimadas Program of the National Institute for Space Research (INPE). An exploratory analysis, pre-processing, modeling and inference of the data was carried out. The analyses were made using Ensemble Models, Random Forest (RF), Extra Trees (ET) and Adaptive Boosting. Results: The models that presented the best performance in relation to GPP were ET, RF and AdaBoost. The most important variables (feature importance) were Surface Temperature and Fires. Conclusion: SHAP made it possible to overcome the limitations of traditional predictive models, favoring a deeper analysis of the factors influencing GPP. Variables such as fire outbreaks and temperature had a negative impact, unlike vegetation, which had a positive impact on the sustainability of the Brazilian Cerrado ecosystem. Future expansion is proposed to include more variables and models, improving environmental sustainability analyses.