Optimizing on-farm corn yield prediction by a multi-source data fusion approach using remote sensing and machine learning

Authors: RAZA, AAMIR - University Of Minnesota; MIAO, YUXIN - University Of Minnesota; Huang, Yanbo; STUEVE, KIRK - Ceres Ai; LU, JUNJUN - University Of Minnesota; YANG, ZHENGWEI - National Agricultural Statistical Service (NASS, USDA); BINDLISH, RAJAT - National Aeronautics And Space Administration (NASA)

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/12/2025
Publication Date: 11/13/2025
Citation: Raza, A., Miao, Y., Huang, Y., Stueve, K., Lu, J., Yang, Z., Bindlish, R. 2025. Optimizing on-farm corn yield prediction by a multi-source data fusion approach using remote sensing and machine learning. Smart Agricultural Technology. 12(2025)101630:1-16. https://doi.org/10.1016/j.atech.2025.101630.
DOI: https://doi.org/10.1016/j.atech.2025.101630

Interpretive Summary: Accurate and reliable on-farm yield prediction is therefore crucial for optimizing corn production management. This study uniquely developed a scheme to identify the optimal time window of the contribution of remote sensing information for best corn yield prediction during corn growth. This study assessed machine learning algorithms for satellite image processing and fusion of soil property data and filed topographic variables to improve the yield prediction. The results of the study highlights the integration of multiple sources of data with machine learning to provide a robust, scalable, and operational framework for corn yield prediction. Such approaches offer a promising pathway to enhance precision agriculture and the development of more sustainable crop production systems.

Technical Abstract: Corn (Zea mays L.) is a cornerstone crop for global food security, providing nearly 30% of daily calories to over 4.5 billion people. Accurate and reliable on-farm yield prediction is therefore crucial for optimizing corn production management. Remote sensing data integrated with machine learning (ML) models has been widely used for corn yield prediction. However, gaps remain regarding the identification of the optimal time window for the contribution of vegetation indices (VIs) to yield estimation, the comparative predictive power of different ML algorithms, the improvement offered by multi-source data fusion approaches, and the relative effectiveness of satellite imaging platforms like Sentinel-2 (S2) and Landsat 8 (L8) imagery. To address these gaps, a study was conducted on three rainfed corn fields in Minnesota, USA, using observed yield monitor data. Multiple VIs derived from imagery of S2 and L8 satellites indicating plant health, biomass, and growth status were used alongside POLARIS soil property data and topographic variables extracted from LiDAR-derived digital elevation models. Eight ML algorithms—random forest, decision trees, extra trees, gradient boosting, extreme gradient boosting, K-nearest neighbors, support vector machine, and histogram-based gradient boosting were evaluated. The results showed that the random forest model consistently outperformed other algorithms, achieving R² values greater than 75% when VIs alone from S2 and L8 imagery were used during the optimal yield prediction window (August 10–31). Multi-source data fusion significantly enhanced prediction accuracy, reducing the mean absolute error by over 40% compared to using VIs alone. Although S2 data offered slight advantages over L8 due to its finer spatial and spectral resolution, L8 also delivered viable results when used with a multi-source data fusion approach, particularly benefiting from its higher radiometric resolution. This study highlights that integrating multiple VIs, soil, and topographic data using ensemble ML models like random forest provides a robust, scalable, and operational framework for corn yield prediction. Such approaches offer a promising pathway to enhance precision agriculture and the development of more sustainable food systems.