Yang et al. (2026) Knowledge-guided graph machine learning improves corn yield mapping in the U.S. Midwest

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Authors: Jie Yang, Licheng Liu, Qi Yang, Xiaowei Jia, Bin Peng, Kaiyu Guan, Zhenong Jin
• DOI: 10.1016/j.rse.2026.115287

Research Groups

• Department of Bioproducts and Biosystems Engineering, University of Minnesota Twin Cities, USA.
• Agroecosystem Sustainability Center, Institute for Sustainability, Energy, and Environment, University of Illinois Urbana-Champaign, USA.
• National Center for Supercomputing Applications, University of Illinois Urbana-Champaign, USA.
• Department of Biogeochemical Integration, Max Planck Institute for Biogeochemistry, Germany.
• Department of Computer Science, University of Pittsburgh, USA.
• Institute of Ecology, College of Urban and Environmental Science, Peking University, China.
• State Key Laboratory for Vegetation Structure, Function and Construction (VegLab), Peking University, China.

Short Summary

The study develops KGML-Graph, a machine learning framework that integrates spatial graph neural networks with temporal deep learning to improve corn yield mapping. By incorporating historical yield correlations as knowledge-guided edge weights, the model significantly outperforms standard temporal models, especially under extreme climatic conditions and in spatial transferability.

Objective

• To develop and evaluate a knowledge-guided graph machine learning framework (KGML-Graph) that explicitly captures spatial dependencies (e.g., yield autocorrelations and time-invariant environmental factors) alongside temporal features for large-scale crop yield mapping.

Study Configuration

• Spatial Scale: Regional; covering 627 counties across the U.S. Corn Belt (Midwest).
• Temporal Scale: 2000–2020 (21 years).

Methodology and Data

• Models used: KGML-Graph (integrating Graph Neural Networks with temporal structures), benchmarked against Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Neural Network (TempCNN).
• Data sources: Remote sensing time-series data, historical yield records (USDA NASS), soil property databases (e.g., soil organic carbon), topography data, and meteorological datasets.

Main Results

• Accuracy Improvement: KGML-Graph reduced RMSE by at least 10.8% and improved $R^2$ by at least 9.3% in temporal extrapolation (2017–2020) compared to benchmark models.
• Extreme Climate Performance: In out-of-distribution testing under extreme weather, the model achieved a $\ge$ 14.4% improvement in $R^2$ and reduced the mean estimation residual from -0.413 to -0.074 t/ha.
• Spatial Transferability: The framework demonstrated superior performance in unseen regions and significantly reduced spatial autocorrelation in estimation residuals.
• Feature Attribution: Analysis confirmed that the graph structure and knowledge-guided edge weights effectively captured key static variables, such as soil organic carbon content, which are often underutilized in purely temporal models.

Contributions

• Methodological Innovation: Unifies spatial and temporal learning by incorporating domain knowledge (historical yield correlations) into graph edge weights.
• Improved Robustness: Demonstrates that capturing spatial dependencies mitigates systematic yield overestimation during extreme climatic events.
• Scalability: Provides a robust framework for large-scale agricultural monitoring that accounts for both environmental heterogeneity and climatic variability.

Funding

• Not explicitly detailed in the provided text snippet.

Citation

@article{Yang2026Knowledgeguided,
  author = {Yang, Jie and Liu, Licheng and Yang, Qi and Jia, Xiaowei and Peng, Bin and Guan, Kaiyu and Jin, Zhenong},
  title = {Knowledge-guided graph machine learning improves corn yield mapping in the U.S. Midwest},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115287},
  url = {https://doi.org/10.1016/j.rse.2026.115287}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)