Rice Yield Prediction Model at Pixel Level Using Machine Learning and Multi-Temporal Sentinel-2 Data in Valencia, Spain

Rice yield prediction at high spatial resolution is essential to support precision management and sustainable intensification in irrigated systems. While many remote sensing studies provide yield estimates at the field scale, pixel-level predictions are required to characterize within-field variability. This study assesses the potential of multitemporal Sentinel-2 imagery and machine learning to estimate rice yield at pixel level in the Albufera rice area (Valencia, Spain). Yield data from combine harvester maps were collected for ‘JSendra’ and ‘Bomba’ Japonica varieties over five growing seasons (2020–2024) and linked to 10 m Sentinel-2 bands in the visible, near-infrared (NIR) and short-wave infrared (SWIR) regions. Random Forest (RF) and XGBoost (XGB) models were trained with 2020–2023 data and independently validated in 2024. XGB systematically outperformed RF, achieving at 110 and 130 DAS (days after showing), R2 values of 0.74 and 0.85 and RMSE values of 0.63 and 0.28 t·ha−1 for ‘JSendra’ and ‘Bomba’. Prediction accuracy increased as the season progressed, and models using all spectral bands clearly outperformed configurations based only on spectral indices, confirming the dominant contribution of NIR reflectance. Spatial error analysis revealed errors at field edges and headlands, while central pixels were more accurately predicted. Overall, the proposed approach provides accurate, spatially explicit rice yield maps that capture within-field variability and support both end-of-season yield estimation and early season forecasting, enabling the identification of potentially low-yield zones to support targeted management decisions.