SHiPA-LLM: Interpretable Regional Crop Production Estimation and Driver Analysis by Coupling Deep Learning with a Knowledge-Grounded Large Language Model

<jats:p>Under climate change, regional-scale crop production estimation requires not only stable numerical results, but also a traceable evidence chain that can support mechanism diagnosis and decision-making. Existing regional monitoring systems usually rely on stitching multi-source heterogeneous data with complex alignment procedures, and prediction and explanation are often separated, making it difficult to form an auditable closed loop. Focusing on the main grain-producing region of Northeast China, this study proposes SHiPA-LLM, which couples deep learning estimation with knowledge-constrained large language model interpretation. The framework uses a single Earth observation input, applies a multi-task model to estimate total production and harvested area, derives yield, and reconstructs meteorological variables to form an evidence package. The interpretation module then generates driver explanations based on a knowledge base. Experimental results show that in the regional-scale evaluation over Northeast China, the proposed method achieves R2=0.7287 and RMSE =43,049.70 kg for production, and R2=0.7482 and RMSE =19.34 ha for harvested area. We further test the accuracy of physics-derived yield under different thresholds, and at the 90% threshold R2 reaches 0.63. In addition, the LLM demonstrations show that the framework can provide traceable, evidence-driven explanations for yield variability in anomaly years, offering a practical decision-support path for regional production monitoring and risk assessment.</jats:p>