Individual plant-level (IPL) soybean biomass estimation and spatial mapping through UAV multisource feature-driven machine learning framework

Abstract

Abstract Rapid and accurate quantification of crop biomass using multisource UAV imagery–derived features, such as plant height, vegetation indices, and texture indices demonstrates strong potential for soybean high-throughput phenotyping. The indeterminate growth habit of soybean, alongside extensive nodulation and intense inter-plant competition, necessitates individual plant level (IPL) monitoring to quantify plant-specific nitrogen fixation and competitive vigor. However, most studies aggregate measurements at multi-plants at the plot level, thereby masking these soybean-specific traits. This study aims to develop and evaluate a UAV imagery-based framework for estimating soybean biomass at the IPL, with the objective of characterizing high-resolution spatial variability and supporting high-throughput phenotyping. Regions of interest (ROIs) for IPL data acquisition were defined as rectangular plots based on planting density and were generated from early-stage imagery before canopy overlap occurred. Using these ROIs, structural information (SIs) including plant height (PH) and vegetation fraction (VF)), vegetation indices (VIs), and texture indices (TIs) were derived for each individual plant from RGB and multispectral imagery and organized into sensor-specific feature groups. Recursive feature elimination was applied to select optimal features, which were then used as inputs for machine learning architectures including support vector regression (SVR) with a linear kernel, Random Forest (RF), and XGBoost (XGB). Among them, the SVR model using fused multisource features (SIs + VIs + Tis) achieved the best performance, with R² = 0.88, RMSE = 55.34 g, and rRMSE = 8.50% on an independent test dataset. The results show that: (1) tree-based models, including XGB and RF may suffer from overfitting due to limited sample size and feature redundancy, whereas linear SVR showed better generalization; (2) fusing RGB and multispectral features consistently improved biomass estimation accuracy. Inparticular, near-infrared and red-edge-based indices such as RECI and NDRE, along with VF and PH, were identified as important predictors, while texture indices were not selected as significant features; and (3) the proposed framework enabled spatially explicit IPL biomass mapping and time-series analysis, revealing variability in growth conditions and distinct growth trajectories. The framework provides a reliable solution for IPL soybean biomass estimation with practical potential for UAV-based agricultural decision-making.

Plant phenotyping relevance

UAV画像由来の特徴量と機械学習により個体レベルのダイズ biomass を推定・評価する枠組みが研究の中心であり、植物表現型の取得・抽出手法に該当する。

Code and data availability

The supplied blocks describe UAV RGB/multispectral imagery, Labelme annotations, and ML models for IPL soybean biomass estimation, but contain no data or code availability statement, no public repository, and no author-provided URL for datasets, images, annotations, or trained models. The only URL present (github.com/…

No evidence-backed public reproduction asset is currently recorded.