Abstract: Gravel-sliding slopes are among the most frequent and hazardous geohazards along the China-Pakistan Karakoram Highway. To address feature confusion, complex textures, and background interference in remote-sensing identification of gravel-sliding slopes, this study proposes an optimized SRNet model for plateau mountainous areas. With Mask R-CNN as the baseline, a ResNet50-FPN-SR backbone network is constructed: the SCConv module suppresses redundant background information and enhances texture features, while the RGA module strengthens contextual perception of the overall gravel-sliding slope morphology. The Sost-Khunjerab Pass section of the China-Pakistan Highway is selected as the study area. A dataset of 800 Sentinel-2 remote-sensing image patches is constructed, divided into training and test sets at a ratio of 8:2, and augmented for model training. The proposed method achieves a precision of 87.37%, recall of 88.24%, and F1-score of 87.80%, improving on the baseline Mask R-CNN model by 7.18, 4.91, and 6.07 percentage points, respectively. Compared with Faster R-CNN, Mask Scoring R-CNN, Cascade Mask R-CNN, and other models, the proposed method performs best across all evaluation metrics. Grad-CAM heatmap visualization further verifies the synergistic feature-enhancement effect of the SCConv and RGA modules. This study provides technical support for automatic remote-sensing identification of granular-slope hazards in high-altitude cold regions.