County-Level Rice Yield Estimation in Guangdong Province, China, Using a Prior-Guided Environmental Fusion Network

Accurate county-level rice yield estimation is important for regional production monitoring, yet dense field-level observations are often unavailable and prediction across unseen years remains challenging. This study involved the development of a prior-guided environmental fusion network (PGEF-Net) for annual rice yield estimation in Guangdong Province, China. PGEF-Net uses a Random Forest-derived feature-importance prior to guide environmental attention and integrates environmental and general feature representations through gated fusion. Under leave-one-year-out (LOYO) validation of 447 county–year observations from 2020 to 2024, PGEF-Net achieved the highest pooled R2 (0.883) and lowest RMSE (0.263 t ha−1) among five learning-based models, but its RMSE advantage over the training-fold county-mean historical reference (0.268 t ha−1) was not statistically significant. Ridge yielded the lowest MAE (0.154 t ha−1) and MAPE (2.58%) among the learning-based models. Among environmental and remote-sensing predictors, thermal variables showed the strongest cross-method signal, with maximum temperature and GDD10 among the leading variables. Ablation analysis showed the largest numerical decreases in R2 and increases in RMSE after removing Environmental Attention or the Environmental Branch. An additional yield anomaly analysis showed limited and year-dependent prediction performance after removing training-fold county means, indicating that persistent county-level yield differences were an important source of predictability under the present LOYO design.