CS-DeepLabV3+: A Fine-Grained Semantic Segmentation Method for Mining Land Use in the Kunlun Mountain Region Using High-Resolution Remote Sensing Imagery

Mining areas in high-altitude cold and arid mountains exhibit heterogeneous land-cover types, large spatial extent, and fragmented boundaries, which makes large-area monitoring difficult with manual interpretation. This study proposes CS-DeepLabV3+, an enhanced semantic segmentation framework built upon DeepLabV3+ for 1-m optical imagery in the Kunlun Mountains. A contextual modeling block is inserted between the encoder output and the atrous spatial pyramid pooling module to strengthen long-range dependency modeling under complex backgrounds. In the decoder, a channel attention block is applied to fused features to suppress redundant responses and improve separability among confusing categories. Experiments on a self-built dataset (Kunlun-Set) demonstrate improved boundary delineation and region consistency for typical mining-related classes (e.g., tailings ponds, stockpiles, and industrial yards). CS-DeepLabV3+ achieved an 81.83% mean intersection-over-union on the test set, outperforming the DeepLabV3+ baseline by 3.52 percentage points. Ablation studies verify that contextual modeling and channel recalibration provide complementary gains.