KNN-GCN: A Deep Learning Approach for Slope-Unit-Based Landslide Susceptibility Mapping Incorporating Spatial Correlations

Ding Xia, Huiming Tang (Korresp. Autor*in), Thomas Glade, Chunyan Tang, Qianyun Wang

Veröffentlichungen: Beitrag in FachzeitschriftArtikel


Landslides pose a significant risk to human life and property, making landslide susceptibility mapping (LSM) a crucial component of landslide risk assessment. However, spatial correlations among mapping units are often neglected in statistical or machine learning models proposed for LSM. This study proposes KNN-GCN, a deep learning model for slope-unit-based LSM based on a graph convolutional network (GCN) and the K-nearest neighbor (KNN) algorithm. The model was experimentally applied to the Lueyang region and validated through the following steps. Firstly, we collected data for 15 landslide causal factors and from landslide inventories and established a slope unit map (SUM) through slope unit division. Next, we performed a multicollinearity analysis of landslide causal factors and divided the training and test sets at a 7:3 ratio. We then constructed a GCN model based on a slope unit graph (SUG) generated from the SUM using the KNN algorithm. The proposed KNN-GCN model was tuned using a grid search with fivefold cross-validation on the training set, and then trained and validated on training and test sets separately. Finally, the performance of the KNN-GCN model was compared with that of six other models which were categorized into two groups: CG#1 was the traditional KNN, support vector regression (SVC), and automated machine learning (AutoML), and CG#2 included KNN-G, SVC-G and AutoML-G with additional spatial information. Our results demonstrate that the proposed model achieves superior performance (area under the curve [AUC] = 0.8351) and generates the most comprehensible susceptibility map with distinct boundaries between different susceptibility levels. Notably, while the proposed KNN-GCN model displays exceptional performance in slope-unit-based LSM, its implementation requires high-level computing resources, and it is not recommended for small datasets.
FachzeitschriftMathematical Geosciences
PublikationsstatusVeröffentlicht - 6 Feb. 2024

ÖFOS 2012

  • 105408 Physische Geographie