After a landslide, swift and precise identification of the affected area is paramount for facilitating urgent rescue operations and damage assessments. This is particularly vital for land use planners and policymakers, enabling them to efficiently address hazard mitigation, the resettlement of those affected by the hazards, and to strategize land planning in the impacted regions. Despite the importance, conventional methods of monitoring landslides often fall short due to their restricted scope and the challenges associated with data acquisition. This study proposes a landslide detection method based on unsupervised multisource and target domain adaptive image segmentation (LUDAS) that is capable of achieving robust and generalized landslide mapping across multiple sources and target domains. Specifically, LUDAS consists of two phases. In the first phase, we introduce an unsupervised interdomain translation network to align the styles of multiple source domains to multiple target domains, generating pseudotarget domain data. Our interdomain translation network is capable of style transfer between any two domains. Through careful design of the network structure and loss functions, we ensure effective style transfer while preserving the content structure of the source domain images. In the second phase, the landslide segmentation model is trained in a supervised manner using annotated data from multiple source domains and multiple pseudotarget domains, resulting in a model with strong generalization capabilities that can adapt to multiple source and target domains. Finally, through extensive qualitative and quantitative analysis experiments, our study confirms that the proposed domain-adaptive segmentation model not only achieves exceptional landslide segmentation performance across multiple target domains but also, due to its good generalizability and transferability, has great potential for application in the emergency response to landslide. This capability can provide strong support for post-disaster emergency rescue, disaster assessment, and land planning in areas with scarce data.
