Abstract: To address the low efficiency of conventional geological mapping and the challenges associated with field surveys in high-altitude areas with rugged terrain, this study selected the Aketao region of Xinjiang, China, as the study area. ASTER visible–near-infrared and thermal-infrared data were integrated with Sentinel-2 multispectral imagery, and principal component analysis was applied for dimensionality reduction to construct a fused multi-source remote-sensing feature set. A one-dimensional–two-dimensional dual-branch convolutional neural network (1D–2DCNN) and a three-dimensional–two-dimensional hybrid convolutional neural network (HybridSN) were then developed for automated lithological classification. These models represent two spectral-spatial learning strategies: branch-based feature fusion and joint convolutional feature extraction. Single-branch 1DCNN and 2DCNN models were also implemented for comparison. To mitigate accuracy overestimation caused by spatial autocorrelation, the samples were partitioned into spatial blocks separated by buffer zones. The partitioning procedure was repeated using three random seeds to assess the robustness of the results. HybridSN achieved the best overall performance, with an overall accuracy of 95.58%, a Kappa coefficient of 0.945±0.061, and macro- and weighted-average F1-scores of 0.955±0.048 and 0.951±0.056, respectively. The 1D–2DCNN achieved an overall accuracy of 92.51%, comparable to that of the 2DCNN and higher than that of the 1DCNN. Single-source comparison experiments showed that the fused dataset improved overall accuracy by approximately 7.3% relative to the ASTER-only dataset. Ablation experiments further showed that removing the thermal-infrared bands reduced overall accuracy by approximately 4.9%, highlighting the importance of thermal-infrared information for identifying silicate lithological units in the study area. In addition, the random pixel-based split overestimated accuracy by approximately 7.5% compared with the spatial block-based split, demonstrating the necessity of spatially independent sample partitioning for objective model evaluation. These results indicate that integrating ASTER and Sentinel-2 data with deep spectral-spatial feature learning has considerable potential for automated lithological mapping in the high-altitude, geologically complex terrain of the West Kunlun region and can provide a valuable technical reference for regional geological surveys and strategic mineral exploration in similar areas.