Key points are not available for this paper at this time.
Landslides pose significant threats in regions with complex topography and frequent seismic or rainfall activity, requiring accurate and transferable mapping methods across heterogeneous remote sensing domains. Existing methods for landslide mapping primarily focus on optimizing extraction algorithms, but their generalizability is often limited by spectral, geometric, and radiometric inconsistencies across sensors and regions. Instead of developing a new landslide extraction model, this study proposes an approach to enhance the transferability of existing models by unifying data representations through a style harmonization framework. We introduce Landslide Mapping Generative Adversarial Network (LMGAN), which uses a dual-branch architecture that integrates spatial and frequency feature learning. The framework incorporates a Wavelet Pyramid Pooling Module (WPPM) for multiscale spectral decomposition and an Interleaved-Cross-Fan (ICF) block to preserve boundary details with directional convolution operations. Experiments on six landslide regions using multisource remote sensing data demonstrate that LMGAN outperforms three recent methods, including CycleGAN, AttentionGAN, and StegoGAN, achieving approximately a 50% improvement in visual quality and content fidelity of translated images. In downstream landslide extraction using recently proposed models SCDUNet++, DeepLabv3+, and SwinUnet, LMGAN-translated images consistently improve extraction accuracy, with increases of about 20% in IoU and 25% in F1 score. The results demonstrate that LMGAN effectively reduces appearance discrepancies across heterogeneous remote-sensing domains and provides a practical style-harmonization strategy for improving the transferability and fidelity of downstream cross-domain landslide extraction.
Yu et al. (Wed,) studied this question.