Precise Mapping of Medicinal Crops Using UAV Hyperspectral Image: A Strategy Driven by Crop-Specific Feature Selection and Decision-Level Fusion Classifier

Precision management of medicinal plant resources is critical for the sustainability of traditional medicine industries. However, accurate identification of genetically similar herbs in heterogeneous environments remains challenging due to high spectral similarity and the "curse of dimensionality" in UAV hyperspectral data. To address these issues, this study challenges the conventional "global optimization" paradigm by proposing a hierarchical Class-Specific Feature Selection (CSFS) strategy. Integrating SPA, CARS, GA, and RFE, this strategy extracts parsimonious diagnostic features tailored to the unique separability of each species, rather than a uniform subset. Furthermore, a probability-calibrated Stacking Ensemble model (RF-LR) was constructed to resolve decision ambiguity in transition zones. The results demonstrate that the CSFS strategy successfully mitigated data redundancy, achieving a dimensionality reduction rate of 96%–98% (reducing 321 features to 5–14 key variables). Mechanistic analysis revealed distinct bio-optical drivers for separability: Melicope pteleifolia is distinguished by pigment-induced spectral shifts in visible bands, Murraya exotica by chlorophyll-sensitive red-edge traits, and Zanthoxylum nitidum by morphology-driven canopy textures. Consequently, the RF-LR model achieved an Overall Accuracy of 97% and a Kappa of 0.96, significantly outperforming traditional classifiers (RF, XGBoost, SAM) in terms of stability and generalization. This study validates the effectiveness of coupling class-specific optimization with decision-level fusion, providing a robust, interpretable, and lightweight technical solution for the operational monitoring of medicinal plant resources.