Hyperspectral data-driven corn nitrogen monitoring: application and interpretability analysis of multi-source feature optimization and stacked ensemble learning methods

Introduction Accurate monitoring of canopy nitrogen content is essential for sustainable nitrogen management, yield improvement, and environmental protection in industrial maize production. However, the high dimensionality of hyperspectral data and the limited accuracy and interpretability of existing models hinder practical applications. Methods This study was conducted in Heilongjiang Province, China, using the maize cultivar Jinboshi. Genetic Algorithm (GA), Successive Projections Algorithm (SPA), and their hybrid strategy were compared for spectral band optimization. Sensitive vegetation indices were selected using multiple evaluation criteria, and a 0–2 order fractional-order derivative (FOD) method was applied to construct optimal two-dimensional (2D) and three-dimensional (3D) spectral indices. A stacked ensemble learning model was developed using XGBoost, GBDT, and Ridge as base learners and Bayesian Ridge as the meta-learner. Interpretability techniques were applied to analyze feature contributions. Results The GA–SPA hybrid strategy effectively improved key spectral band selection. The 3D spectral index based on FOD achieved superior performance compared to vegetation indices and 2D indices (R 2 p = 0.801, RMSEP = 0.481). The optimized multi-source feature set combined with the stacked ensemble model yielded the best performance (R 2 p = 0.826, RMSEP = 0.450). Features from the red-edge and near-infrared regions, along with the 3D index, were the primary contributors to model predictions, consistent with plant nitrogen physiology. Discussion The proposed framework, integrating feature optimization, advanced modeling, and interpretability analysis, provides an effective tool for precise nitrogen management in industrial maize and supports improved production efficiency with reduced environmental impact.