LandslideCL: towards robust landslide analysis guided by contrastive learning

Accurate and automatic landslide detection plays a vital role in keeping abreast of disaster situations and supporting rescue-related decision-making. Currently, deep learning has brought innovation to landslide detection techniques. However, previous studies did not consider the nested connection between low-level and high-level feature maps, resulting in coarse landslide segmentation boundaries. In addition, due to the instability of processing noise, existing models significantly degrade the performance when faced with complex landslide scenes. In this study, we present a novel robust rainfall-induced landslide detection model guided by contrastive learning. Specifically, we embed the residual block and channel attention module into U-Net+++ to adequately exploit semantic details and focus on vital information. Meanwhile, we implement effective data augmentation strategies to obtain two different image views, and feed them into a dual-branch model to predict landslide locations. Afterwards, we develop contrastive dice similarity coefficient loss to maintain the consistency of landslide region pairs, which stimulates the model to further mine invariance characteristics. We successfully fuse the modified U-Net+++ and contrastive learning to solve the coarse boundary and poor robustness problem. Numerous experiments are implemented to demonstrate that our model generates excellent performance with mean intersect over union over 0.80 and outperforms other classic segmentation methods in crucial criteria.

Automated summary

This study presents a novel robust rainfall-induced landslide detection model guided by contrastive learning. The model combines deep learning techniques, such as residual blocks and channel attention modules, to accurately predict landslide locations. It also utilizes contrastive dice similarity coefficient loss to maintain consistency in landslide regions. Experimental results demonstrate that the proposed model performs excellently, outperforming other classic segmentation methods in crucial criteria.