Integrating vegetation phenology and SWOT-derived surface water dynamics for improved county-scale winter wheat yield prediction

Winter wheat yield is shaped by crop growth status and surface hydrological conditions. However, remote-sensing approaches based primarily on vegetation index (VI) time series may not fully represent the surrounding surface-water environment in county-scale yield prediction. We integrated growth-stage VI features with surface-water dynamics derived from the Surface Water and Ocean Topography (SWOT) mission to predict county-level winter wheat yield in Henan Province, China. We evaluated direct feature fusion, same-year residual correction, and target-year residual correction following cross-year VI extrapolation using outer ten-fold cross-validation. Inner five-fold cross-fitting generated same-year VI residual targets exclusively within each outer training set. The VI-only model achieved R2 = 0.772, RMSE = 0.688 t ha−1, and MAE = 0.510 t ha−1. Direct incorporation of SWOT features reduced predictive accuracy, with R2 = 0.749, RMSE = 0.721 t ha−1, and MAE = 0.547 t ha−1. By contrast, SWOT-based residual correction yielded a clearer improvement in same-year prediction, increasing R2 to 0.795 and reducing RMSE to 0.652 t ha−1. The historical VI model was trained on 2010–2021, independently tested on 2022–2023, and extrapolated to 2024, where it achieved R2 = 0.676, RMSE = 0.819 t ha−1, and MAE = 0.611 t ha−1. Calibrated using observed 2024 yields from training counties and evaluated on held-out 2024 counties, subsequent target-year residual correction with concurrent SWOT features increased R2 to 0.749 and reduced RMSE and MAE to 0.722 and 0.524 t ha−1, respectively. SHAP analysis showed that SWOT contributions were concentrated during seeding, emergence, overwintering, green, heading–flowering, and filling and varied by spatial scale and prediction scenario. Overall, SWOT-derived surface-water observations provided complementary predictive information relative to the evaluated VI baselines, particularly for target-year spatial calibration. Their broader applicability across regions, crops, and years remains to be established through independent validation.