Beyond data-driven: Integrating scientific knowledge and machine learning for crop prediction

Context Accurate and robust prediction of crop traits remains a critical challenge in agriculture. While machine learning approaches have gained significant popularity, they often lack generalizability and interpretability. Recently, a promising trend has emerged that integrates scientific prior knowledge, primarily in the form of physics-based models, with machine learning to enhance predictions in agriculture. Objective This review provides a timely synthesis of the state-of-the-art in applying this integrated approach to predict crop traits, which broadly include structural, biochemical, physiological, phenological, and productivity variables. Methods We summarized different strategies for integrating machine learning with scientific knowledge and identified two overarching approaches: hybrid modeling and knowledge-guided machine learning. Hybrid modeling was further divided into cascade coupling, surrogate modeling, transfer learning, and residual modeling. Knowledge-guided machine learning refers to leveraging prior knowledge to modify machine learning components such as model architectures and loss functions for improved predictive performance. Results and conclusions Our synthesis showed that most existing studies have adopted the hybrid modeling strategies, whereas the knowledge-guided machine learning remains underexplored. Most of the studies reported reduced prediction errors by using the integrated approaches. In addition, this emerging approach holds great potential for improved robustness, generalizability, transferability, and interpretability. Significance Finally, we discussed the current challenges and applicability of the two integration categories. To advance this emerging field, we proposed a conceptual framework for model integration and highlighted its potential applications for agricultural crop systems.