Work

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

Year: 2026

Type: article

Abstract: Crop yield prediction is crucial for ensuring food security and informed agricultural decision-making. Conventional approaches usually follow model-driven or data-driven approaches. Model-driven approaches offer biophysical interpretability but rely on several site-specific parameters that are challenging to obtain. Data-driven approaches afford flexible modeling of multi-source inputs but require a large number of samples and typically operates as uninterpretable black boxes. To address this, a novel process-aware yield-prediction framework is proposed to integrate the interpretability of crop growth modeling with the flexibility of deep learning. A Knowledge-Informed Fusion Network (KIFN) is introduced to transcend shallow fusion by deeply transferring mechanistic knowledge from process-driven models into dynamic data-driven architectures, which incorporates a three-stage design: (1) a pretraining stage in which an expert network is trained on synthetic daily growth trajectories generated by simulation of the crop growth model; (2) a fine-tune stage that aligns the expert network with historical yield data; and (3) a dual-stream fusion network that uses remote sensing observations to calibrate the expert network output via a cross-attention mechanism, thereby dynamically adjusting and refining in-season predictions. Experimental results from the USDA's multi-year corn yield dataset show that our model achieves high accuracy (R2 = 0.793, RMSE = 1.039 t/ha). Furthermore, the model's prediction improves progressively throughout the growing season as dynamic satellite observation is used for calibration, validating its effectiveness for dynamic in-season yield prediction. 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.

Source: International Journal of Applied Earth Observation and Geoinformation

DOI: 10.1016/j.jag.2026.105154