TCN–LSTMs: A robust and interpretable framework for winter wheat yield prediction using multi-source time-series data

Accurate and timely crop yield estimation is essential for ensuring food security. Although deep learning methods integrating multi-source remote sensing and meteorological data have been widely applied in yield prediction, existing studies still face limitations in capturing crop growth temporal dynamics, quantifying the full-season temporal contributions of multi-source features, and assessing prediction uncertainty induced by climate variability. These limitations constrain model accuracy, interpretability, and cross-regional generalization capability. To address these challenges, this study proposes an interpretable yield estimation framework that combines a Temporal Convolutional Network (TCN) with Long Short-Term Memory Networks (LSTM) in a serial architecture (TCN–LSTMs). Using multi-source time-series data, we conduct county-level winter wheat yield prediction and attribution analysis across Henan Province, revealing the temporal cumulative effects of key growth-related variables and the uncertainty introduced by late-season climate fluctuations. The results show that the TCN–LSTMs model ( R 2 = 0 . 81 , RMSE = 574.26 kg/ha) achieves notably better predictive performance than baseline models. Specifically, it outperforms CNN–LSTM ( R 2 = 0 . 78 ) and TCN–LSTMp ( R 2 = 0 . 75 ), as well as attention-based architectures like Temporal Fusion Transformer(TFT) ( R 2 = 0 . 74 ) and Transformer ( R 2 = 0 . 69 ), reducing the RMSE by 6.1% to 21.1%. Leave-one-out cross-validation and additional experiments in the Huang–Huai–Hai region further verify the robustness and generalization ability of the proposed model. The relative prediction error (RE) is mainly concentrated between -10% and 10%, and accurate yield estimation can be achieved one month in advance. The attribution analysis based on Integrated Gradients (IG) indicates that early-season soil temperature and moisture affect emergence rates; mid-season canopy photosynthesis and soil moisture influence spike number and grain-setting rate; and late-season photosynthetic efficiency and water availability are critical for grain filling and thousand-kernel weight. The period from mid-March to late April provides the most informative signals for yield prediction. Uncertainty analysis further shows that early-stage stable signals, such as vegetation indices, help reduce predictive uncertainty, whereas mid- and late-season uncertainty fluctuates due to climate variability and management differences. Overall, the proposed TCN–LSTMs framework not only improves the accuracy and robustness of winter wheat yield prediction but also enhances interpretability and early warning capability. This study provides an effective pathway for regional crop yield estimation and food security assessment.