Authors

Abstract

Soybean (Glycine max) grain moisture content (MC) at harvest affects yield, storage, and processing quality, but traditional measurements are laborious and unsuitable for large-scale monitoring. This study aimed to develop an efficient method for estimating soybean (Glycine max) grain moisture content (MC) at maturity, addressing the limitations of traditional labor-intensive measurements. UAV-based multispectral imagery from a DJI Mavic 3M was used to extract spectral reflectance and vegetation indices (VIs). Three feature selection techniques (SHAP, RFA, ReliefF) and six regression models (PLSR, SVR, MLR, RFR, XGBoost, RR) were applied to identify key predictors and optimize performance. Results showed that SVR using spectral reflectance achieved the highest accuracy (R² = 0.763, RMSE = 1.473), while RFR performed best for combined spectral and VI features. The RE and NIR bands were the most sensitive to MC. The findings demonstrate that integrating UAV multispectral data with machine learning and feature selection enables accurate, rapid, and non-destructive prediction of soybean MC, supporting precision harvest and crop management.

Abstract in Chinese

大豆（Glycine max）收获时的籽粒含水量（MC）影响产量、储存和加工质量，但传统测量费时费力，不适用于大规模监测。本研究旨在开发一种高效的方法，用于估算大豆（Glycine max）在成熟期的籽粒含水量（MC），以解决传统人工测量方法的局限性。使用大疆Mavic 3M无人机获取的多光谱影像提取了光谱反射率和植被指数（VIs）。应用了三种特征选择技术（SHAP、RFA、ReliefF）和六种回归模型（PLSR、SVR、MLR、RFR、XGBoost、RR）来识别关键预测因子并优化性能。结果表明，使用光谱反射率的SVR模型获得了最高精度（R² =0.763，RMSE=1.473），而RFR模型在结合光谱和VI特征时表现最佳。红边（RE）和近红外（NIR）波段对 MC 最敏感。研究结果表明，将无人机多光谱数据与机器学习和特征选择相结合，能够实现对大豆 MC 的准确、快速、无损预测，为精准收获和作物管理提供支持。