LanPPT: Enhancing landslide crack detection through pyramid pooling transformers

Landslide cracks, also known as tension cracks, are a major part of landslides. Based on the distribution of landslide fractures, the location of a landslide can be broadly described, and the stress distribution of the sliding mass can be inferred. Cracks at a landslide’s head, which are a key indicator of the displacement of the landslide body, can provide early warning signs for landslide hazards. The complete knowledge of the crack development features is still lacking because of many influencing elements and intricate reasons for fracture creation. Even though some early interventions reported models for automated crack identification utilizing advanced machine learning techniques, the problem still has not been solved to its full potential. Thus, an effective deep architecture for landslide crack segmentation is suggested to address these issues, utilizing a synergistic blend of vision transformers and the pyramid pooling concept. In this work, we use the universal vision transformer backbone called the Pyramid Pooling Transformer, and plug it into our pooling-based multi-head spatial attention to build a deep architecture that identifies and segments the landslide cracks, namely LanPPT. Experiments revealed that when a pyramid pooling transformer is used as the backbone network, it performs significantly better than many earlier convolutional neural networks and normal transformer-based networks in various vision tasks. Systematic experiments show that the proposed model achieved superior performance in terms of mIoU and FPS when compared with the chosen state-of-the-art baselines in the landslide crack detection task.