Landslide susceptibility evaluation based on active deformation and graph convolutional network algorithm

Abstract

Disastrous landslides have become a focus of the world’s attention. Landslide susceptibility evaluation (LSE) can predict where landslides may occur and has caught the attention of scientists all over the world. This work establishes integrated criteria of potential landslide recognition and combines the historical landslides and newly-identified potential landslides to improve the accuracy, rationality, and practicability of a LSE map. Moreover, slope units can well reflect the topographic constraint to landslide occurrence and development, and Graph Convolutional Network (GCN) can well portray the topological and feature relation among various slope units. The combination of slope units and GCN is for the first time employed in LSE. This work focuses on Wanzhou District, a famous landslide-serious region in the Three Gorges reservoir area, and employs multisource data to conduct potential landslide recognition and LSE and to reveal the distribution characteristics of high landslide susceptibility. Some new viewpoints are suggested as follows. 1) The established criteria of potential landslide recognition consist of the characteristics of active deformation, stratum and lithology, tectonics, topography, micro-geomorphology, environment, meteorology, earthquakes, and human engineering activity. These criteria can well eliminate 4 types of false alarm regions and is successfully validated by field survey. 2) 34 potential landslides are newly discovered, and the movement of these potential landslides were controlled or induced by the combined action of soft-hard interbedding rock mass, steep topography, frequent tectonic movement, strong fluvial erosion, abundant precipitation, and intensive road and building construction. 3) The GCN algorithm reaches a relatively high accuracy (AUC: 0.941) and outperforms the other representative machine learning algorithms of Convolutional Neural Network (AUC: 0.926), Support Vector Machine (AUC: 0.835), and CART Tree (AUC: 0.762). 4) High landslide susceptibility is caused by the coupled action of weathered rock cavities, soft rock and swelling soil, strong river erosion, abundant rainfall, and intensive human engineering activity.

Landslide Deformation Prediction and Automatic Warning by Coupling Machine Learning and Physical Models

“…Moreover, currently, GCN, as an excellent deep learning algorithm, has rarely been adopted in the landslide area. Recently, owing to precision and universality advantages, a few studies applied it to predict landslide displacements (Jiang et al, 2022;Khalili et al, 2023;Ma et al, 2021), to evaluate landslide susceptibility (Du et al, 2021;Wang, Du, et al, 2023;Xia et al, 2024), or to detect landslides (Li et al, 2023). In this work, GCN is for the first time employed to forecast deformation stages of a landslide.…”

Machine learning for high-resolution landslide susceptibility mapping: case study in Inje County, South Korea

“…Recently, ML techniques have emerged as powerful tools in LSM due to their ability to analyze large volumes of data, identify hidden patterns and relationships that may not be evident to human analysts (Lv et al, 2022;Wang et al, 2023). Common models employed for LSM include decision trees (DTs), RF, SVM, and neural networks (Azarafza et al, 2021;Wang et al, 2021).…”