Interpretable machine learning and uncertainty quantification for high-precision fractional vegetation cover inversion across scales in alpine grasslands

Fractional Vegetation Cover (FVC) is a critical indicator for grassland monitoring and ecological assessment, playing a vital role in advancing ecological sustainability research. However, high-resolution FVC inversion faces two persistent challenges: (1) unquantified uncertainty arising from spatial misalignment between UAV measurement footprints and satellite imagery pixels (SMUS), and (2) heavy reliance on concurrent in-situ measurements and satellite imagery acquisitions, which limits robust mapping capabilities across scales and regions. To address these gaps, this study develops a multi-source remote sensing framework that integrates three key components. First, we introduced PSO-SHAP, an interpretable feature selection algorithm that integrates Particle Swarm Optimization (PSO) and SHapley Additive exPlanations (SHAP). Second, we established an uncertainty quantification framework using a 5%-interval overlap gradient to quantify SMUS-induced inversion uncertainty. Finally, we constructed a cross-scale transfer learning scheme for high-accuracy mapping and elucidated the feature transfer mechanisms. The results demonstrated that: (1) PSO-SHAP effectively balanced feature importance and interpretability; (2) misalignment-induced uncertainty decreased nonlinearly with increasing overlap degree, stabilizing beyond approximately 85% overlap, with maximum uncertainties of 27.88% for Landsat 8/9 and 7.74% for Sentinel-2A/B; and (3) machine learning models exhibited cross-scale transferability, while SHAP analysis interpreted the transfer processes and revealed the scale-invariant advantages of key features. Additionally, Sentinel-2A/B outperformed Landsat 8/9 in 30-m FVC mapping accuracy and spatial continuity. The proposed framework presents a novel methodology for multi-source synergistic FVC inversion and provides a reference for inverting and validating diverse ecological parameters.