Enhancing corn yield prediction: Optimizing data quality or model complexity?

Authors: ZHOU, YUTING - Oklahoma State University; MA, SHENGFANG - Oklahoma State University; Zhang, Huihui; AAKUR, SATHYANARAYANAN - Auburn University

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/26/2024
Publication Date: 11/28/2024
Citation: Zhou, Y., Ma, S., Zhang, H., Aakur, S. 2024. Enhancing corn yield prediction: Optimizing data quality or model complexity? Computers and Electronics in Agriculture. Volume 9, December 2024, 100671. https://doi.org/10.1016/j.atech.2024.100671.
DOI: https://doi.org/10.1016/j.atech.2024.100671

Interpretive Summary: Farmers need to estimate how much corn they can produce before harvest. Traditionally, this involves time-consuming field surveys, which require a lot of effort. This study explored using "machine learning" – a type of computer program that learns from data – to analyze aerial images taken from drones to predict corn yield. We compared different methods, including simpler ones and more complex techniques. The study found that the proposed simple machine learning method and other methods can predict corn yield quickly and accurately, especially when using multispectral images. This is especially helpful in the early and late growing season. This technology can be a valuable tool for farmers to make informed decisions about their crops and resources, ultimately leading to better harvests and smarter farm management.

Technical Abstract: Field-scale corn yield prediction before harvest can assist farmers in better organizing their resources. Traditional approaches involve time-consuming ground field surveys and immense human effort. Machine learning-based pipelines for analyzing remote sensing imagery offer an efficient solution to this problem. However, the cost of data acquisition and training requirements for machine learning models depend on various factors, such as equipment (multispectral vs. RGB imagery) and the ability to predict yield from observations across growth stages. In this study, we aim to provide a comprehensive analysis of the effectiveness of traditional ensemble learning methods (Random Forest and Gradient Boosting) and deep learning models (ResNet 18, ResNet34, and ViT) in predicting corn yield across deficit and fully irrigated fields using field imagery acquired by both RGB and multispectral sensors onboard a UAV. It examines the performance of these models across early, middle, and late growth stages, considering both computational complexity and accuracy. We also introduce a shallow CNN framework called SimRes, inspired by the ResNet framework but tailored for streamlined efficiency and simplicity for yield prediction. Extensive quantitative analysis demonstrate that the customized SimRes performs as well as deep learning baselines but with faster computing times, while traditional approaches, such as Random Forests and Gradient Boosting exhibited marginally smaller R-squared values. Models utilizing multispectral images outperformed models using RGB images, albeit with variations across growth stages. Deep learning methods performed better than ensemble learning methods in early and late growth stages using RGB images, while performance became comparable in the middle stage. These results underscore the importance of additional information or more complex models to enhance prediction accuracy alongside a trade-off between computational complexity and accuracy. This research provides valuable insights for optimizing corn yield prediction across different growth stages, informing agricultural management and harvest planning decisions.