A lightweight Context-aware adaptive fusion network for automatic identification of active landslides

Timely identification of active landslides is critical for disaster early warning and risk management. Interferometric Synthetic Aperture Radar (InSAR) technology, which can capture subtle displacements of active landslides over large areas, has become a key tool for landslide identification. The development of deep learning provides new opportunities to improve InSAR-based landslide identification. However, existing approaches often struggle to balance identification accuracy and computational efficiency. In this study, we propose the Context-aware Adaptive Fusion Network (CAFNet), a lightweight encoder-decoder framework that optimizes multi-scale feature learning from color-mapped deformation. In the encoder, the Wavelet-based Down-sampling Block (WDB) is introduced to perform down-sampling while preserving fine-grained details. Additionally, we develop a Multi-branch Scale-aware Aggregation (MSA) module to adaptively select and integrate multi-scale features based on target characteristics, ensuring flexible feature alignment. The decoder employs an efficient Conv-based Up-sampling Block (CUB) to progressively restore spatial resolution while refining boundaries. Experimental results demonstrate that CAFNet outperforms existing deep learning models such as DeepLabV3+ and ResUNet, achieving 90.8 % precision and 83.3 % IoU at the pixel level, and 89.5 % correct detection and 7.3 % false alarm rate at the object level. Notably, CAFNet achieves these results a 20 × reduction in parameters and a 50 % decrease in computational costs, while maintaining robust generalization abilities. These findings highlight the potential of CAFNet for the establishment and periodic updating of active landslide inventories, which is essential for minimizing losses caused by landslide disasters.