Optimizing Multidimensional Spectral Indices and Ensemble Learning Methods for Estimating Nitrogen Content in Torreya grandis Leaves Based on UAV Hyperspectral

Ensuring sufficient nitrogen intake during the early growth stages of Torreya grandis is crucial for improving future fruit yield and quality. Hyperspectral remote sensing, enabled by unmanned aerial vehicle (UAV) platforms, provides extensive spectral information on forest canopies across large areas. However, the potential of combining multidimensional optimized spectral features with advanced machine learning models to estimate leaf nutrient stress has not yet been fully exploited. This study aims to combine optimized spectral indices and ensemble learning methods to enhance the accuracy and robustness of estimating leaf nitrogen content (LNC) in Torreya grandis. Initially, based on full-band spectral information, five spectral transformations were applied to the original spectra. Then, nine two-band spectral indices and twelve three-band spectral indices were optimized based on published formulas. This process created a total of 27 spectral features across three dimensions. Subsequently, spectral features of varying dimensions were combined with multiple linear regression (MLR), decision tree regression (DTR), random forest (RF), and eXtreme Gradient Boosting (XGBoost) to train base estimators for ensemble models. Using a stacking strategy, various modeling combinations were experimented with, resulting in the construction of 22 LNC estimation models. The results indicate that combining two-band and three-band spectral features can more comprehensively capture the subtle changes in the nitrogen status of Torreya grandis, with the optimized spectral index mNDVIblue (555, 569, 572) showing the highest correlation with LNC at −0.820. In the modeling phase, the base estimators used MLR, RF, and XGBoost, while the meta estimator employed MLR’s stacking model to achieve the highest accuracy and relatively high stability on the validation set (R2 = 0.846, RMSE = 1.231%, MRE = 3.186%). This study provides a reference for the efficient and non-destructive detection of LNC or other phenotypic traits in large-scale economic forest crops using UAV hyperspectral technology.