Lightweight semantic segmentation for Co-seismic landslide identification using adaptive transfer learning

Rapid and accurate detection of co-seismic landslides is essential for post-earthquake emergency response, yet remains challenging due to scarce labeled data, class imbalance, and overfitting in existing models. To address these issues, the Stepwise Lightweight Adaptive Transfer Learning (SLATL) framework was proposed, a compact semantic segmentation model that balances accuracy and efficiency. The SLATL integrates a Stepwise Lightweight Decoder (SLD) that progressively prunes shallow feature channels to reduce computation, a Unified Spatial–Channel Attention Module (USCAM) that strengthens spatial and channel feature fusion, and a Compact Pyramid Pooling Module (CPPM) that captures multi-scale landslide patterns with minimal overhead. Furthermore, an adaptive transfer learning strategy leverages an intermediate domain to mitigate domain gaps and improve generalization under limited data conditions. Validated on the Hokkaido and Wenchuan datasets, the SLATL achieves mean Intersection-over-Union scores of 0.833 and 0.800, respectively, while requiring only 1.418 Gigaflops (GFLOPs, billions of floating-point operations per second) and 3.578 parameters (in millions, M). These results outperform other lightweight baselines and demonstrate SLATL's potential for rapid and reliable disaster assessment in resource-constrained environments.