Hyperspectral inversion of winter wheat nitrogen content based on feature optimization and stacking ensemble learning.

Rapid, non-destructive crop nitrogen monitoring is a core pillar of precision fertilization and mass-balance nutrient management, which requires real-time uptake data to balance fertilizer inputs, crop demand and environmental nitrogen losses. But hyperspectral data redundancy, background noise and black-box machine learning models limit inversion generalization performance and interpretability, hampering field-scale mass-balance management. Using jointing-to-filling winter wheat canopy hyperspectral data, we built a three-tier progressive framework of preprocessing, feature selection and model architecture, systematically evaluating 2 preprocessing, 5 feature selection and 5 modeling schemes, with SHAP-based analysis to reveal the physiological mechanisms underlying band contributions. The main findings are summarized as follows: (1) First-derivative (FD) transformation achieved selective enhancement of nitrogen signals. By removing indirect coupling confounding signals, the test-set R² of all models increased from 0.57-0.73 under Savitzky-Golay (SG) full-spectrum input to 0.71-0.76. (2) Feature selection delivered a significantly greater accuracy gain than model optimization, and a clear feature-model compatibility pattern was identified. Specifically, the sparse high-purity features selected by LASSO were highly aligned with the similarity measurement mechanism of the GPR kernel function. (3) The FD-LASSO-stacking ensemble scheme achieved the optimal inversion performance, with a test-set R² of 0.875, RMSE of 0.281%, and RPD of 2.834. This model also achieved favorable performance with an R2 of 0.781 in cross-growth-stage inversion, which verifies its generalization ability and stability to a certain extent.(4) SHAP analysis quantified the contribution levels of features screened by each algorithm, and identified cross-algorithm common core bands including the 757 nm red edge and 939 nm near-infrared bands, verifying the biological rationality of model decisions. This study not only realizes high-accuracy nitrogen inversion in winter wheat, but also deepens the mechanistic understanding of hyperspectral inversion from the perspectives of feature-model compatibility and model interpretability. It provides theoretical and technical support for lightweight sensor band design, field precision nitrogen diagnosis, and mass-balance-based nitrogen management.