Optimizing corn yield prediction: Integrating multi-temporal UAS data and machine learning

Authors: Zhang, Huihui; ZHOU, YUTING - Oklahoma State University; MA, SHENGFANG - Oklahoma State University; Yemoto, Kevin

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/19/2025
Publication Date: 8/21/2025
Citation: Zhang, H., Zhou, Y., Ma, S., Yemoto, K.K. 2025. Optimizing corn yield prediction: Integrating multi-temporal UAS data and machine learning. Smart Agricultural Technology. 12. Article e101344. https://doi.org/10.1016/j.atech.2025.101344.
DOI: https://doi.org/10.1016/j.atech.2025.101344

Interpretive Summary: Accurate and timely prediction of corn yield is essential for optimizing agricultural practices. This study investigates the potential of multispectral and thermal UAV-based remote sensing to enhance yield estimation in both well-watered and water-stressed corn fields. By comparing machine learning models (LASSO, Random Forest, Gradient Boosting) and sensor configurations (RGB, reflectance, thermal), we explored the optimal approach for yield prediction across different growth stages. Results indicate that multispectral data, particularly when combined with thermal information, significantly improved yield prediction accuracy compared to RGB-only models, especially in water-stressed conditions and during reproductive growth stages. These findings highlight the value of advanced remote sensing technologies and complex modeling techniques for improving agricultural decision-making.

Technical Abstract: The United States plays a pivotal role in global corn production, necessitating accurate corn yield estimation at the field scale for effective and timely field management. This study investigates a novel approach for estimating corn yield in well-watered and water-stressed fields using multisource data acquired by Unmanned Aerial Systems (UAS). The approach integrates high-resolution Red-Green-Blue (RGB), multispectral reflectance (Ref), and thermal (LWIR) imagery with machine learning techniques. Machine learning models, including LASSO regression, Random Forest (RF), and Gradient Boosting (GB), were applied to predict yield based on data across multiple phenological stages. Results from one of the reproductive stages (R5 stage, at Day of Year (DOY) 259) showed that Ref data can predict corn yield better than the high-resolution RGB, regardless of the model used. The more complex RF and GB models outperformed the simpler LASSO regression method. Adding thermal data (LWIR) could further improve the performance with RF and GB models, while LASSO did not benefit from this additional information. Moreover, LWIR data helped overcome the overestimation of corn yield in both fully- and deficit-irrigated fields, particularly when using GB. The time series of model performance indicates that the Ref data could provide stable and accurate corn yield estimation as early as the V9 stage (DOY 189) when using RF. In contrast, RGB data exhibited limited effectiveness in predicting corn yield during the reproductive growth stages, emphasizing the necessity of a multispectral sensor during these periods. Across all growth stages, combining LWIR data with Ref data enhanced the accuracy of corn yield predictions in the deficit irrigated field with RF, especially during the reproductive stages, highlighting the importance of thermal sensors in water-stressed fields. This study introduces a novel approach for estimating corn yield in well-watered and water-stressed fields using multisource data from UAS combined with machine learning models. The integration of multispectral reflectance and thermal imagery significantly improves yield prediction accuracy, particularly in water-stressed conditions, emphasizing the importance of advanced sensing technologies and complex models such as RF and GB. The potential of deep learning models needs to be explored to see if they could further increase the accuracy of corn yield prediction.