Hyperspectral Inversion of Apple Leaf Nitrogen Across Phenological Stages Based on an Optimized XGBoost Model

Precision monitoring of leaf nitrogen content (LNC) in fruit trees is critical for optimizing fertilization and fruit quality. In this study, 1120 apple-leaf samples spanning two phenological stages were collected. Characteristic wavelengths were selected using competitive adaptive reweighted sampling and the successive projection algorithm (CARS–SPA). To mitigate inefficient exploration during population initialization and iterations, we propose a collaborative enhancement strategy integrating Sobol-sequence sampling and elite opposition-based learning (EOBL), termed SEO, which simultaneously refines initialization and iterative updating in swarm-based optimization algorithms. Four machine learning algorithms were trained to construct cross-phenological-stage LNC inversion models. Results indicated characteristic wavelengths lay within the visible region. The combined SEO strategy improved search capability and efficiency, with SEO-BKA achieving the best performance. Consequently, the SEO-BKA-XGBoost model yielded the highest accuracy in the bloom and fruit-set stage (R2 = 0.883; RMSE = 0.124) and fruit-enlargement stage (R2 = 0.897; RMSE = 0.069). These findings provide robust technical support for LNC hyperspectral inversion in apple trees.