Estimation of rice yield using multi-source remote sensing data combined with crop growth model and deep learning algorithm

• Integrated multi-source remote sensing data (MODIS and Sentinel-2) to capture crop growth, enhancing yield estimation accuracy. • Assimilated Sentinel-2 LAI data into the WOFOST model using EnKF, improving model simulation precision. • Developed the BCLA deep learning model with Bayesian optimization, outperforming traditional models in yield prediction. • Identified key yield-influencing factors, such as LAI and PsnNet, using SHAP analysis. • Generated regional yield maps, demonstrating the model's potential and highlighting regional accuracy discrepancies. Accurate rice yield estimation is vital for agricultural planning and food security, especially in Northeast China, a key rice-producing region. This study presents an integrated framework combining multi-source remote sensing data, crop growth modeling, and deep learning techniques to enhance rice yield prediction accuracy. We utilized Moderate Resolution Imaging Spectroradiometer (MODIS) and Sentinel-2 satellite data to capture both temporal and spatial crop dynamics. High-resolution Leaf Area Index (LAI) data from Sentinel-2 were assimilated into the World Food Studies (WOFOST) crop growth model using the Ensemble Kalman Filter (EnKF), improving the model’s simulation precision. To further refine yield estimates, we developed the Bayesian-optimized Convolutional Long Short-Term Memory with Attention (BCLA) model, which integrates Residual Convolutional Neural Networks (ResNet-CNN), Long Short-Term Memory (LSTM) networks, and Multi-Head Attention mechanisms, optimized through Bayesian optimization. The proposed hybrid framework was applied to rice growing seasons from 2019 to 2021, demonstrating significant improvements in prediction accuracy compared to traditional models such as Random Forest and XGBoost. The BCLA model achieved higher R 2 and lower Root Mean Square Error (RMSE) values, indicating its superior ability to capture complex spatial and temporal patterns. SHapley Additive exPlanations (SHAP)-based feature importance analysis identified key factors influencing yield predictions, including LAI, Net Photosynthesis (PsnNet), and Kernel Noramlized Difference Vegetation Index (kNDVI). Regional yield maps validated against statistical data showcased the model’s robustness, although some regional discrepancies highlighted areas for further refinement. This comprehensive approach offers a scalable and accurate solution for high-resolution rice yield estimation, supporting precision agriculture and sustainable food security initiatives.