Modelos de regressão aplicados a dados climatológicos de reanálise e índices vegetativos para predição de produtividade da soja

Abstract

Soybean is one of the most important agricultural commodities globally, serving as one of the pillars of the Brazilian economy. The unpredictability of its productivity generates significant risks for the logistical and financial planning of the sector. In this context, early yield prediction becomes a strategic tool. This work aimed to develop and evaluate regression models for predicting soybean productivity in the Western Region of Paraná, correlating climatological reanalysis data and satellite-derived vegetation indices. The methodology was executed on the Google Colab platform, where a datasset was built integrating vegetation indices (NDVI and EVI) and clima tological variables (temperature, precipitation, radiation). Supervised regression algorithms, namely Random Forest, XGBoost, LightGBM, and LSTM, were tested and compared using three approaches: Temporal, Spatial, and Spatio-Temporal. The analysis of the results demonstrated that the purely temporal approach was ineffective, presenting a negative R2. In contrast, the spatio-temporal approach proved to be the most accurate, with the XGBoost model obtaining the highest coefficient of determination (R2) of 0.71 and a Mean Absolute Error of 249.40 kg/ha. It is concluded that integrating satellite and climate data into spatio-temporal models is a viable strategy and significantly more accurate than isolated temporal analysis. The validated model presents itself as a promising tool for producer planning, contributing to the modernization of digital and sustainable agriculture.