Presentation Information
[U15-P09]Landslide Susceptibility Mapping Using Vision Transformer with Pre-Event Topographic DEM: A Case Study of Deep-Seated Landslides in the Kii Peninsula, Japan★Invited Papers
*Teruyuki Kikuchi1, Tetsuyasu Yamada1 (1.Suwa University of Science)
Keywords:
Vision Transformer,Deep-seated landslide,Landslide susceptibility,DSGSD
1 Introduction
Understanding past landslide events is essential for predicting future slope failures under similar environmental conditions. Landslide occurrence is controlled by multiple interacting factors including geology topography and rainfall intensity. However because these factors vary regionally and interact in complex ways a unified framework for landslide susceptibility assessment remains limited. A key precursor of deep seated landslides is deep seated gravitational slope deformation DSGSD which represents slow bedrock deformation accompanied by subtle geomorphic features such as double ridges scarps irregular relief and valley head incision. These features may precede catastrophic failures and can be quantitatively represented using high resolution LiDAR derived digital elevation models DEM.
2 Study Area and Data
The study area is the Kii Mountains Japan where Typhoon Talas in September 2011 brought more than 2000 mm of rainfall and triggered over 50 deep seated landslides. Failures occurred when cumulative rainfall exceeded approximately 600 mm within 48 to 72 hours. The region consists of steep mountainous terrain from 220 m to 1915 m elevation mainly developed within the Shimanto accretionary complex.
Pre and post event airborne LiDAR surveys provided 1 m resolution DEM data. Only the pre event DEM was used as model input. Thirty eight deep seated landslides larger than 1000 m2 were defined as landslide class y0. Additionally 63 DSGSD sites that did not fail during the 2011 event were defined as deformation class y1 while stable terrain was defined as class y2.
Eight DEM derived topographic indices were prepared as multi channel input data including slope angle eigenvalue ratio curvature positive openness negative openness topographic wetness index TWI wavelet coefficient and elevation. These indices represent terrain roughness convex concave morphology ridge valley configuration and localized flow concentration associated with gravitational deformation.
The images were divided into 50 x 50 pixel tiles corresponding to 50x 50 m areas. A total of 36985 tiles were generated and split into training 70 percent and validation 30 percent datasets.
3 Methods
Two deep learning architectures were compared. The CNN model extracts local micro topographic features through convolution and pooling layers and contains about 1.6e5 parameters emphasizing computational efficiency. The Vision Transformer ViT divides each image into 5 x 5 pixel patches and processes them using a 16 layer Transformer encoder with multi head self attention allowing modeling of long range spatial relationships. The ViT model contains about 1.3e8 parameters.
4 Results
Validation accuracy reached 0.867 for ViT and 0.837 for CNN. Recall for the landslide class exceeded 0.95 in ViT and 0.92 in CNN indicating strong detection capability. The DSGSD class showed lower recall in both models reflecting its transitional geomorphic characteristics. Overall ViT demonstrated slightly superior performance in capturing broader spatial context.
5 Discussion
Model behavior was examined in a 15 km2 test domain including unseen terrain. Both CNN and ViT identified coherent clusters in areas characterized by double ridges irregular relief and step like morphology. Some predicted areas did not fail during the 2011 rainfall event but exhibit terrain characteristics similar to trained landslides. These areas may indicate slopes with latent instability that could fail under comparable extreme rainfall conditions.
6 Conclusion
ViT based susceptibility mapping using only pre event DEM data effectively captured geomorphic structures associated with deep seated landslides. The results suggest that attention based architectures can support large scale identification of potentially unstable slopes while further improvement of generalization performance remains necessary.
Understanding past landslide events is essential for predicting future slope failures under similar environmental conditions. Landslide occurrence is controlled by multiple interacting factors including geology topography and rainfall intensity. However because these factors vary regionally and interact in complex ways a unified framework for landslide susceptibility assessment remains limited. A key precursor of deep seated landslides is deep seated gravitational slope deformation DSGSD which represents slow bedrock deformation accompanied by subtle geomorphic features such as double ridges scarps irregular relief and valley head incision. These features may precede catastrophic failures and can be quantitatively represented using high resolution LiDAR derived digital elevation models DEM.
2 Study Area and Data
The study area is the Kii Mountains Japan where Typhoon Talas in September 2011 brought more than 2000 mm of rainfall and triggered over 50 deep seated landslides. Failures occurred when cumulative rainfall exceeded approximately 600 mm within 48 to 72 hours. The region consists of steep mountainous terrain from 220 m to 1915 m elevation mainly developed within the Shimanto accretionary complex.
Pre and post event airborne LiDAR surveys provided 1 m resolution DEM data. Only the pre event DEM was used as model input. Thirty eight deep seated landslides larger than 1000 m2 were defined as landslide class y0. Additionally 63 DSGSD sites that did not fail during the 2011 event were defined as deformation class y1 while stable terrain was defined as class y2.
Eight DEM derived topographic indices were prepared as multi channel input data including slope angle eigenvalue ratio curvature positive openness negative openness topographic wetness index TWI wavelet coefficient and elevation. These indices represent terrain roughness convex concave morphology ridge valley configuration and localized flow concentration associated with gravitational deformation.
The images were divided into 50 x 50 pixel tiles corresponding to 50x 50 m areas. A total of 36985 tiles were generated and split into training 70 percent and validation 30 percent datasets.
3 Methods
Two deep learning architectures were compared. The CNN model extracts local micro topographic features through convolution and pooling layers and contains about 1.6e5 parameters emphasizing computational efficiency. The Vision Transformer ViT divides each image into 5 x 5 pixel patches and processes them using a 16 layer Transformer encoder with multi head self attention allowing modeling of long range spatial relationships. The ViT model contains about 1.3e8 parameters.
4 Results
Validation accuracy reached 0.867 for ViT and 0.837 for CNN. Recall for the landslide class exceeded 0.95 in ViT and 0.92 in CNN indicating strong detection capability. The DSGSD class showed lower recall in both models reflecting its transitional geomorphic characteristics. Overall ViT demonstrated slightly superior performance in capturing broader spatial context.
5 Discussion
Model behavior was examined in a 15 km2 test domain including unseen terrain. Both CNN and ViT identified coherent clusters in areas characterized by double ridges irregular relief and step like morphology. Some predicted areas did not fail during the 2011 rainfall event but exhibit terrain characteristics similar to trained landslides. These areas may indicate slopes with latent instability that could fail under comparable extreme rainfall conditions.
6 Conclusion
ViT based susceptibility mapping using only pre event DEM data effectively captured geomorphic structures associated with deep seated landslides. The results suggest that attention based architectures can support large scale identification of potentially unstable slopes while further improvement of generalization performance remains necessary.