Medicinal Plant Mapping Based on UAV Hyperspectral Imagery via an Interpretable Framework

Accurate mapping of medicinal plants is essential for resource inventory and sustainable management. While Unmanned Aerial Vehicle (UAV) hyperspectral sensing provides a powerful monitoring tool, the high-dimensional spectral ambiguities and the black-box nature of conventional global classifiers severely restrict mapping accuracy and interpretability. To overcome this, we developed an interpretable, class-specific framework for fine-scale crop identification. A multi-feature pool was distilled through a hierarchical Successive Projections Algorithm (SPA)–Competitive Adaptive Reweighted Sampling (CARS)–Genetic Algorithm (GA)–Recursive Feature Elimination (RFE) strategy, isolating class-specific diagnostic fingerprints that transition mapping from black-box modeling to an explainable mechanism. The tailored feature subsets were supplied to parallel Random Forest binary classifiers to generate class-specific probability outputs, which were subsequently fed into a Logistic Regression decision-level fusion layer that calibrated prediction probabilities and resolved inter-class conflicts to yield the definitive classification results. Applied in Yunfu, Guangdong, the framework mapped target species (Murraya exotica, Zanthoxylum nitidum, and Melicope pteleifolia) with an overall accuracy of 97.35% (Kappa = 0.97), outperforming conventional benchmark models such as standard multi-class Random Forest, Extreme Gradient Boosting, and spectral angle mapper. These findings demonstrate that coupling class-specific feature distillation with probability-calibrated decision fusion provides a robust, interpretable paradigm for precision agricultural mapping in heterogeneous environments.