Field-Scale Rice Yield Prediction in Northern Coastal Region of Peru Using Sentinel-2 Vegetation Indices and Machine Learning Models

Accurate rice yield prediction is essential for optimizing water management and supporting decision-making in agricultural systems, particularly in arid environments where irrigation efficiency is critical. This study assessed five machine learning algorithms—Multiple Linear Regression (MLR), Support Vector Regression (SVR, linear and RBF), Partial Least Squares Regression (PLSR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—for plot-scale rice yield estimation using Sentinel-2 vegetation indices (VIs) during the 2022 and 2023 seasons in the Chancay–Lambayeque Valley, Peru. VIs sensitive to canopy vigor, water status, and structure were derived in Google Earth Engine and optimized via Sequential Forward Selection to identify the most relevant predictors per phenological stage. Models were trained and validated against field yields using leave-one-out cross-validation (LOOCV). Intermediate stages (Flowering, Milk, Dough) yielded the strongest relationships, with water-sensitive indices (NDMI, MSI) consistently ranked as key predictors. MLR and PLSR achieved the highest generalization (R 2 _CV up to 0.68; RMSE_CV ≈ 1.3 t ha −1 ), while RF and XGBoost showed high training accuracy but lower validation performance, indicating overfitting. Model accuracy decreased in 2023 due to climatic variability and limited satellite observations. Findings confirm that Sentinel-2–based VI modeling offers a cost-effective, scalable alternative to UAV data for operational rice yield monitoring, supporting water resource management and decision-making in data-scarce agricultural regions.