Optical River Ice Spectral Subclassification on the Tibetan Plateau: A Landsat 5–9 and Sentinel-2 Benchmark with Interpretable Machine Learning

River ice products from optical satellites are still dominated by binary ice–water or ice–snow discrimination, leaving within-ice spectral heterogeneity largely unresolved. This study benchmarks how far river ice can be subclassified from multispectral reflectance alone on the Tibetan Plateau using Landsat 5/7, Landsat 8/9, and Sentinel-2 surface-reflectance imagery. We compiled 356 winter scenes acquired between 2000 and 2024 across eight Tibetan Plateau basins, delineated river ice using NDSI and RDRI, and extracted 24,674 pixel-level spectra. To define reproducible subclasses, we applied K-means clustering guided by the Silhouette Coefficient, Davies–Bouldin index, Calinski–Harabasz index, and Gap Statistic. Combined with stratified visual interpretation, this approach consistently supported four optical spectral subclasses: thin-snow-covered ice, thick ice cover, thin ice, and frazil ice. Within-sensor classification accuracy remained extremely high (overall accuracy ≥ 0.948; kappa ≥ 0.929), with the Backpropagation Neural Network (BPNN) and tree ensembles performing best. Crucially, evaluating the optimal BPNN architecture revealed exceptional multi-dimensional generalizability: a Leave-One-Basin-Out spatial cross-validation yielded a stable average OA > 99% with an average Kappa > 0.98, while a unified multi-sensor model achieved a robust OA of 90.14% and a Kappa of 0.86. The most stable discriminative cues were visible-band brightness, reflectance turnover near ~0.7 μm, and shortwave-infrared sensitivity to effective thickness and surface wetness. These results provide a sensor-aware benchmark for practical optical river ice spectral subclassification and clarify which multispectral bands most strongly constrain subclass separability.