Efficient Recognition of Potential Landslides using Open Access Multi-Source Remotely Sensed Images with Deep Learning Approaches

A landslide is a severe and frequent natural hazard that seriously threatens the safety of human lives and properties. Under a changing climate context, landslides are likely to occur more frequently in the future. It is critical to acquire spatial knowledge about where landslides might occur as a pathway to prepare for future occurrences and mitigate devastating outcomes.

The employment of high-resolution Remote Sensing (RS) imageries made the remote investigation of potential landslides possible. Traditional methods, LSM, cannot pinpoint specific locations with the potential for landslide occurrence, and the data requirement is also demanding. In addition, most of the factors used in the LSM won't updated frequently, and thus the generated susceptibility map can easily become outdated. With the development of Deep learning (DL) and the increased availability of RS imageries, DL-based methods emerged. However, almost all the proposed DL-based methods focused on improving the efficiency and accuracy of LSM, while the aforementioned issues embedded in LSM per se have not been improved.

Under this context, this presentation provides an overlook of three studies that contribute to developing improved DL methods to accurately recognize potential landslides with the support of public multi-source RS data. The first work proposed OpenLandslide, a complex multi-source RS dataset with high-quality potential landslide labels that can not only be used for traditional LSM-related studies but can also facilitate DL-based potential landslide recognition studies. The second work proposed a multi-scale residual classification model based on OpenLandslide, named the MSRNet, fine-tuned to extract potential landslide features efficiently and accurately from multi-source input to better recognize potential landslides. The third work proposes a DL-based potential landslide segmentation model, also based on OpenLandslide, to accurately detect potential landslides' exact locations and boundaries. The proposed model, coined as the APNU-Deeplab, is inspired by DeeplabV3 and the self-attention mechanism. A hybrid random position sampling strategy is proposed in this study with verified effectiveness. The proposed model reaches around 85% in F1-score and 75% in mIoU, which indicates high performance and usability.