Uso de índices espectrais na estimativa da produtividade do cafeeiro por meio de técnicas de aprendizado de máquina

Abstract

Coffee cultivation (Coffea arabica L.) exhibits high spatial and temporal variability, making yield estimation a complex challenge due to the biennial bearing phenomenon and edaphoclimatic variations. The use of remote sensing associated with machine learning emerges as a promising alternative to overcome the inefficiency of manual sampling. The objective of this study was to estimate the spatial variability of coffee yield using spectral indices derived from high-resolution orbital imagery integrated with the Random Forest (RF) algorithm. The study was conducted in São João Evangelista - MG, in a three-year-old plantation of the Arara cultivar. Yield data were collected from 100 georeferenced plants through manual stripping. Multispectral images from the CBERS-4A satellite (WPM sensor) were acquired in September 2024 (pre-flowering period) and subjected to pansharpening to obtain a spatial resolution of 2 meters, mitigating noise and spectral mixing from the inter-rows. NDVI, SAVI, NDWI, and VARI indices were calculated. For predictive modeling, the RF algorithm was employed and validated using the Leave-One-Out Cross-Validation (LOOCV) method. The results demonstrated high yield variability in the area (CV of 51.77%), associated with altitude and initial stand establishment dynamics. Indices based on the near-infrared (NDVI and SAVI) showed strong spatial coherence with yield, while VARI (restricted to the visible spectrum) lacked the ability to discriminate between management zones. The RF model achieved a coefficient of determination (R²) of 0.61 and a Mean Absolute Error (MAE) of 0.51 L plant⁻¹ (~6.3 bags ha⁻¹). NDWI stood out as the most important predictor variable. It is concluded that the use of high-resolution imagery during the pre-flowering period, combined with the Random Forest algorithm, is a robust and effective tool for early coffee yield forecasting, providing detailed maps that enable the transition to precision agriculture through variable rate fertilization and site-specific pruning.