Accurate discrimination of exposed bedrock (EB) and gravel surfaces is essential for quantifying soil resources, understanding erosional controls on pedogenesis, and guiding conservation strategies in bedrock-dominated mountain ecosystems. Conventional methods based on manual visual interpretation are labor-intensive, costly, and typically classify both as a single "mixed bedrock-gravel surface," leading to misestimation of soil resources. Here, we present a framework that integrates topographic features, remote sensing spectral indices, and interpretable machine learning to classify EB and gravel in the high-elevation, geomorphically complex mountains of southern Tibet, China (average elevation >4,500 m). A total of 7,798 samples were generated from Google Earth Pro high-resolution imagery. By combining Sentinel-2 spectral bands, soil- and vegetation-related indices, and DEM-derived topographic variables, a recursive feature elimination-random forest (RFE-RF) model achieved an overall accuracy of 95.64 %, significantly exceeding that of the legacy approach (overall accuracy = 88 %). Independent field validation confirmed the robustness of the predictions. Shapley analysis revealed slope height and topographic position index as the primary drivers of EB-gravel differentiation, reflecting denudation processes on ridges and deposition in valleys. Shortwave infrared bands (B11, B12) and derived indices (clay index, geological index) further enhanced separation. The resulting maps aligned closely with Gaofen imagery and manual interpretations. This study establishes a transferable paradigm for high-precision surface classification in alpine environments, enabling fine-scale identification of potential soil resources.