Authors

Abstract

To achieve high-precision non-destructive monitoring of Soil and Plant Analyzer Development (SPAD) values for winter wheat, this study proposes a fusion method integrating multi-feature selection and machine learning for data inversion. High-resolution remote sensing imagery was acquired using a drone-mounted multispectral camera during the jointing, heading, and grain filling stages, with ground SPAD values measured simultaneously. The Pearson Correlation Coefficient-Random Forest-Cross Validation (PCC-RF-CV) feature fusion optimization method was employed to determine optimal feature combinations for each growth stage. Six machine learning inversion models were constructed for systematic comparison. Results demonstrated that the multi-source feature fusion strategy exhibited superior predictive capability across all growth stages. The PCC-RF-CV method effectively optimized feature input dimensions, establishing optimal feature sets for each growth stage. The XGBoost model for the grain filling stage, constructed using this method, achieved the best inversion results with validation set R² = 0.92, RMSE = 0.36, and MAE = 0.30. This method enables precise inversion of winter wheat SPAD values, reveals the dynamic patterns of canopy spectral response across growth stages, and provides reliable technical support for crop growth monitoring and precision agriculture.

Abstract in Chinese

为实现冬小麦Soil and Plant Analyzer Development (SPAD)值的高精度无损监测，本研究提出一种融合多特征优选与机器学习的反演方法。基于无人机多光谱相机获取拔节期、抽穗期和灌浆期高分辨率遥感影像，同步测量地面SPAD值。研究采用Pearson Correlation coefficient- Random Forest - Cross Validation(PCC-RF-CV)特征融合优选方法确定各生育期最优特征组合，构建六种机器学习反演模型进行系统对比。研究显示，多源特征融合策略在全生育期均展现出优越的预测能力；PCC-RF-CV方法有效优化了特征输入维度，为各生育期构建了最优特征集；基于该方法构建的灌浆期XGBoost模型取得最佳反演效果，验证集R²=0.92，RMSE=0.36，MAE=0.30。该方法能精准反演冬小麦SPAD值，揭示冠层光谱响应的生育期动态规律，为作物生长监测与精准农业提供可靠技术支撑。