A knowledge-informed fusion network for process-aware crop yield prediction

Crop yield prediction is crucial for food security and agricultural decision-making. Conventional approaches struggle to reconcile the heavy parameter reliance of process-based models with the extensive label dependencies of data-driven models. To address this, we propose a process-aware crop yield prediction framework that integrates the mechanistic interpretability of crop growth models with the flexibility of deep learning. A Knowledge-Informed Fusion Network (KIFN) is introduced to transcend conventional shallow fusion by transferring mechanistic knowledge of biomass accumulation dynamics from process-based models into data-driven architectures. KIFN incorporates a three-stage design. (1) In the knowledge-informed pre-training stage, an expert network is trained on synthetic daily growth trajectories generated through crop growth model simulation. (2) In the process-aware fine-tuning stage, the expert network is aligned with historical yield data. (3) In the dual-stream fusion stage, the expert network output is calibrated using remote sensing observations via a cross-attention mechanism, thereby dynamically refining in-season predictions. Experimental results from the USDA’s multi-year corn yield dataset show that our model achieves high accuracy ( R 2 = 0.793, RMSE = 1.039 t/ha). The model’s predictions improve progressively throughout the growing season as satellite observations are used for calibration, demonstrating its effectiveness for dynamic in-season yield prediction. Additionally, interpretability analysis indicates that the cross-attention mechanism in KIFN assists the network in identifying more informative temporal windows throughout the crop growth process. This work provides a principled and interpretable pathway for crop yield prediction, offering new insights into integrating domain knowledge with deep learning in agricultural applications. • Knowledge-informed network is proposed for integrating CGMs and deep learning. • Process-aware cross-attention is introduced to fuse meteorological and RS features. • Our dual-stream fusion model improved the accuracy for in-season yield prediction.