Finegrained UAV Hyperspectral Grassland Mapping via Mixed Spatial Downsampling and Ecologically Interpretable Band Selection

Grasslands in the Hulun Lake Basin have faced degradation for decades. Therefore, accurately discriminating grass species at a fine scale is vital for regional environmental protection. To overcome mixed-pixel interference and high spectral dimensionality, this study proposed a dual-domain decoupling framework, which spectrally employs a combined principal component analysis and pretrained random forest (RF) feature importance band selection (BS) method, and spatially integrates a mixed spatial downsampling (MDS) mechanism. Multiple machine learning models were applied to classify key dominant forage species (Leymus chinensis; Stipa krylovii) and degradation indicator species (Allium polyrhizum; Allium ramosum). SHapley Additive ExPlanation (SHAP) model interpretability analysis was then employed to identify high-contribution bands, and the proposed method was systematically benchmarked against state-of-the-art (SOTA) deep learning architectures. Results showed that: 1) Key bands with high influence on model training were primarily located in the red-edge region (650–720 nm) and near-infrared (NIR) plateau (900–1000 nm) regions; 2) the proposed BS & MDS method significantly improved classification performance, with both RF and artificial neural network achieving the highest overall accuracy of 0.97, which was higher than those of the evaluated SOTA models; 3) sensitivity analysis indicated that our framework maintained consistent results under various training sample ratios while requiring minimal computational time; 4) ecologically, dominant forages (LC, SK) had better separability, while grazing disturbance drove distinct intraclass multicenter spectral clustering; 5) red-edge bands dominated species discrimination in SHAP analysis with higher SHAP values than those bands at the NIR plateau. This study provided a highly robust, computationally efficient, and ecologically explainable method to accurately discriminate grass species in Hulun Lake Basin at fine scale, supporting the precision management of grassland degradation.