DRCD-TL: a novel co-seismic landslide mapping approach by deep representation-driven unsupervised change detection based on transfer learning

Most existing unsupervised change detection methods for co-seismic landslide mapping rely on handcrafted shallow landslide features and fail to exploit deep information, which results in limited performance in complex scenarios. To address this limitation, we propose a novel deep representation-driven unsupervised change detection method based on transfer learning (DRCD-TL), which employs adaptive deep feature mining to comprehensively highlight landslide information. Experiments in Mainling, China, and Hokkaido, Japan, show that DRCD-TL outperformed six other methods, with F1-score improvements of 5.54~20.88% and 1.25~27.46%, respectively. These results confirm the potential of DRCD-TL for accurate co-seismic landslide mapping.