Joshi, A., Pradhan, B., Chakraborty, S., Varatharajoo, R., Alamri, A., Gite, S., & Lee, C. W.(2025). An Explainable Bi-LSTM Model for Winter Wheat Yield Prediction. Frontiers in Plant Science, 15, 1491493. Click (a) Absolute error map of the Bi-LSTM model: Spatial error distribution of the yield predictions, demonstrating the model's superior spatial generalizability with the lowest Mean Absolute Error (MAE) of 0.48. (b) SHAP summary and dependence plots: Detailed visualizations of the non-linear impacts and contributions of key remote sensing and meteorological variables (EVI, precipitation, maximum temperature, etc.) on yield predictions within the Bi-LSTM model. This study proposes a deep learning framework combined with Explainable Artificial Intelligence (XAI) to accurately predict winter wheat yield at a regional scale across 10 U.S. states while ensuring model interpretability. Utilizing multi-temporal remote sensing data (MODIS-derived EVI) and meteorological data (maximum temperature, precipitation, evapotranspiration, wind speed) from 2008 to 2021, the performance of 1D-CNN, LSTM, Bi-LSTM, and Random Forest (RF) models was comparatively evaluated.The results demonstrated that the Bi-LSTM model, which processes sequential data in both directions, outperformed the others with the highest predictive performance (R2 up to 0.88) and the lowest Mean Absolute Error (MAE), proving its excellent spatio-temporal generalizability. Furthermore, to overcome the 'black box' limitation of deep learning models, LIME, Integrated Gradients (IG), and SHAP techniques were applied. The XAI analysis quantitatively revealed that EVI, precipitation, and maximum temperature during the later stages of crop growth (especially in June, pre-harvest) are the most dominant factors determining yield. By providing both high predictive accuracy and intuitive model interpretability, this study presents a reliable decision-support tool for food security and agricultural policy-making. Continue reading...
