Satellite remote sensing enables systematic observation of desert boundaries across vast geographic areas inaccessible to ground surveys. Accurate automated extraction from Landsat imagery is essential for determining desert spatial extent, which constitutes the foundation for desertification assessment-a global environmental challenge affecting 40% of Earth's land surface. This study presents a deep learning framework integrating multidimensional dynamic convolution, parallel Mamba2 encoding with structural state-space duality, deformable large kernel attention, and cascaded spatial-channel attention mechanisms. Validated on Landsat 8 imagery covering major desert systems, the model achieves 99.27% overall accuracy and 98.54% mean IoU, outperforming 13 representative methods across diverse boundary types: desert-water interfaces (99.28% accuracy), desert-gobi transitions (99.25%), and desert-mountain-urban boundaries (99.44%). Multiregional testing confirms robust generalization (98.87% average accuracy) across different geographical conditions. The linear computational complexity O(T) enables efficient large-area processing. The framework demonstrates operational feasibility for automated desert boundary extraction, advancing satellite remote sensing capabilities for desertification assessment and ecological restoration monitoring.