关键词: 高光谱遥感, 冬小麦, 磷含量估算, 光谱降维, 集成机器学习

Abstract:

【Objective】Phosphorus is a critical nutrient element for crop growth and development, directly influencing photosynthesis and physiological functions. Accurate monitoring of leaf phosphorus content is essential for efficient crop management and yield prediction. In this study, an ensemble hyperspectral dimensionality reduction model was proposed for estimating leaf phosphorus content based on different spectral preprocessing methods combined with various spectral dimensionality reduction techniques, in order to provide a theoretical basis for hyperspectral diagnosis of crop phosphorus nutrition.【Method】Through two years of field experiments, canopy spectral reflectance and leaf phosphorus content were collected for winter wheat during four key growth stages (jointing, heading, flowering, and grain filling) under three nitrogen application levels. Derived spectral features were generated using De-trending transformation, standard normal variate transformation, and first-order derivative reflectance transformation. Spectral dimensionality reduction was performed using successive projections algorithm (SPA), least absolute shrinkage and selection operator (LASSO), and competitive adaptive reweighted sampling (CARS). Single estimation models and integrated estimation models were constructed by combining random forest regression (RF), support vector regression (SVR), and partial least squares regression (PLSR).【Result】The leaf phosphorus content-sensitive spectral features were primarily concentrated in the visible, near-infrared, and short-wave infrared bands. Compared to original spectral reflectance and other derived spectral features, first-order derivative-derived spectral features demonstrated significant advantages in detecting leaf phosphorus content. Among the single models, the highest estimation accuracy was achieved when CARS was used for dimensionality reduction to select sensitive first-order derivative-derived features, driving the RF algorithm, with R²=0.843 and RMSE=0.038% in the training dataset, and R²=0.756 and RMSE=0.057% in the testing dataset. The integrated hyperspectral dimensionality reduction estimation further improved model accuracy. The integrated model driven by first-order derivative-derived spectral features and constructed using RF regression achieved the best estimation performance, with R²=0.932 and RMSE=0.025% in the training dataset, and R²=0.817 and RMSE=0.049% in the testing dataset.【Conclusion】Ensemble hyperspectral dimensionality reduction method effectively enhanced the estimation accuracy of leaf phosphorus content in winter wheat, providing theoretical support for large-scale monitoring of crop nutrition and growth under different hyperspectral remote sensing platforms, crop types, and growth conditions in the future.

Key words: hyperspectral remote sensing, winter wheat, estimation of phosphorus content, hyperspectral dimensionality reduction, ensemble machine learning