Machine Learning-based Coffee Yield Prediction using Multi-Temporal Sentinel-2 Data

Abstract

Accurate prediction of coffee yield is essential for effective crop management and sustainable production in Vietnam’s Central Highlands. This study proposes a remote sensing-based machine learning framework for predicting annual coffee yield at the commune level using multi-temporal Sentinel-2 imagery in the newly established Dak Ha communes of Quang Ngai province. Five vegetation indices—NDVI, EVI, NDMI, NDRE, and GNDVI—were com-puted and aggregated across key coffee phenological stages to capture canopy vigor and moisture dynamics. Four regression models, including Linear Regression, Random Forest, XG-Boost, and CatBoost, were evaluated using a Leave-One-Year-Out cross-validation strategy to ensure temporal robustness. Results from 2020 to 2025 show that ensemble models substantially outperform the linear baseline, with CatBoost achieving the best performance (MAE = 0.18 t/ha, RMSE = 0.24 t/ha). The framework demonstrates that Sentinel- 2 time series data alone provide reliable information for commune-level coffee yield forecasting and can be effectively integrated into Web-GIS systems for operational monitoring and decision support.