Modelagem preditiva de produtividade agrícola a partir de fontes de dados espaciais (FLDAS) para a cultura da soja na Mesorregião Norte do estado do Mato Grosso

Abstract

Forecasting agricultural productivity is a strategic challenge for production planning, rural credit, and public policy formulation. In Brazil, such forecasts have traditionally relied on meteorological data from conventional weather stations, which are limited by inconsistencies, low spatial coverage, and temporal gaps. This study proposes the development of a crop yield prediction model based on climate data obtained from multisatellite remote sensing, integrating information from FLDAS/NOAH (NASA/Giovanni) for the historical period from 1989 to 2022. Climate variables, processed in NetCDF format, were spatially extracted from municipal centroids and combined with official soybean yield records provided by IBGE/SIDRA. The method applied was based on machine learning through ensemble decision trees (gradient boosting), exploring 216 climate variables derived from energy balance, hydrological fluxes, radiation, and atmospheric parameters. The results achieved a coefficient of determination (R²) of 0.823, with consistent RMSE and MAE values, demonstrating the model’s strong ability to reproduce observed yield variability, even without including agronomic variables such as soil fertility and management practices. Variable importance analysis using SHAP values highlighted that hydrological balance and radiation components were decisive in predicting yield during critical phenological stages of soybean. The findings indicate that the model is robust and scalable, allowing validation for other crops and agricultural regions. In addition to providing reliable forecasts, the proposed approach expands the potential for agricultural monitoring in areas lacking conventional meteorological data, representing a valuable tool for farm management, public policy, rural credit, insurance, and logistics planning.