A unified network of information considering superimposed landslide factors sequence and pixel spatial neighbourhood for landslide susceptibility mapping

• Published in: International journal of applied earth observation and geoinformation Vol. 104; p. 102508
• Main Authors: He, Yi, Zhao, Zhan'ao, Yang, Wang, Yan, Haowen, Wang, Wenhui, Yao, Sheng, Zhang, Lifeng, Liu, Tao
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15-12-2021; Elsevier
• Subjects: GRU; Landslide susceptibility; Lanzhou city; MSCNN; RS and GIS; Lanzhou city; GRU; Landslide susceptibility; MSCNN; RS and GIS
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2021.102508

•The proportion of landslide binary pixels was used as label data input in network models.•MSCNN model was carried out for LSM the first time.•A unified network in parallel with GRU and MSCNN was proposed for LSM.
Landslide susceptibility mapping (LSM) is very important for hazard risk identification and prevention. Most of existing neural network models extract a pixel neighborhood feature or a pixel sequence feature of landslide factors on one side, which leads to the generalization ability of the network models difficultly, and had a low prediction accuracy in complex scenes. In this paper, a new unified network of information considering superimposed landslide factors sequence and pixel spatial neighbourhood is proposed for LSM. Different from the traditional prediction model framework, the landslide conditioning factors are merged into a unified network model in parallel with the pixel sequence features and pixel neighbourhood features. In the experiment, we take the proportion of landslide binary pixels as label data, which represents the landslide possibility in the neighbourhood. We propose a pixel sequence feature extraction algorithm based on a gated recurrent unit (GRU) network and a pixel neighbourhood feature extraction algorithm based on a multi-scale convolution neural network (MSCNN). In this study, the landslide conditioning factors were analysed by multicollinearity analysis and the frequency ratio (FR) method. The performance of the modes was evaluated by statistical indexes and the correlation analysis. The LSM results were verified by google earth images and field investigation. Our research shows that the proposed model can greatly improve the accuracy of LSM compared with the individual GRU and MSCNN, especially, the proposed model had 6.1% more improvement than the GRU model in terms of the area under curve (AUC). Therefore, we suggest that the proposed model is a suitable technology for use in early identification and landslide prediction.