High‐throughput phenotyping and machine learning techniques in soybean breeding: Exploring the potential of aerial imaging and vegetation indices

Abstract

Abstract Soybean ( Glycine max (L.) Merr.) breeding programs face challenges in evaluating large progeny populations, which is labor‐ and resource‐intensive. This study addresses these challenges using high‐throughput phenotyping and machine learning (ML) models to predict phenotypic traits in soybeans. We developed and validated ML models using vegetation indices and canopy images from aerial imagery. A total of 275 soybean genotypes were characterized across two environments and management practices. A total of 11 classical traits were measured, and five vegetation indices were calculated from aerial images at different growth stages. ML algorithms, including support vector machine for regression, random forest (RF), multilayer perceptron (MLP), and adaptive boosting, were employed. Additionally, convolutional neural networks with transfer learning were used to extract features from the images. Significant correlations were found between agronomic traits, vegetation indices, and canopy characteristics. The high heritability of the red–green–blue vegetation index and green leaf index (mean broad‐sense heritability of 0.56) compared to other RGB‐based indices indicates their potential usefulness in genetic evaluations. Advanced ML techniques, particularly transfer learning with ResNet 50, enhanced the prediction of phenotypic traits such as days to the R7 growth stage (DR7) and plant height at maturation (PHM). The integration of ResNet 50 with RF achieved a prediction accuracy of 0.64 for DR7, while ResNet 50 with MLP reached an accuracy of 0.68 for PHM. These findings highlight the potential of these techniques to improve decision‐making in soybean breeding. Lastly, principal component analysis identified genotypes with desirable trait combinations, advancing soybean development.

Plant phenotyping relevance

航空画像・植生指数・機械学習を用いた植物形質予測モデルを開発・検証しており、フェノタイピング手法が研究の中心である。

Code and data availability

The supplied blocks contain only the article's references, acknowledgments, and boilerplate. No public phenotype dataset, aerial imagery, author analysis code, or trained model is deposited with an authors' URL; the Supporting Information note gives no link or content description, and all cited DOIs are prior work or,

No evidence-backed public reproduction asset is currently recorded.