Multi-scale landslide detection in complex scenarios based on adaptive sparsity regulation strategy

Landslides, as a type of geological disaster with significant destructive potential, endanger human lives and cause substantial property damage. To reduce the losses caused by landslide disasters and improve emergency response efficiency, effective landslide detection is critical. However, landslide size variability and complex environmental interference challenge detection. To overcome these limitations, this paper proposes ACSLDNet, which significantly improves boundary texture blurring caused by UAV imaging by incorporating the ManbaIRv2 super-resolution technique. To overcome the challenges of multi-scale detection, a dual residual structure combined with dynamic Inception convolution is used to enhance the shallow sensory field successfully. In addition, the multi-scale fusion framework GIEP, which integrates a novel shallow feature introduction mechanism, strengthens the ability to fuse multi-scale features. An Adaptive Hybrid Sparsity Attention mechanism is introduced to combat background interference in complex scenarios to improve the balance between high-response and global information. These methods significantly enhance multi-scale landslide detection performance in complex scenarios, with ACSLDNet achieving an AP50(%) of 82.0 on our constructed UAV low-altitude view landslide dataset for complex scenarios. Furthermore, experimental results on a public dataset show that ACSLDNet outperforms other methods in small-target detection, highlighting its tremendous potential for landslide detection in real-world complex scenarios.