Landslide spatial susceptibility mapping by using deep belief network

More than 3800 landslide locations have been reported in the area near the Three Gorges Reservoir (TGR) along the Yangtze River in China, which poses a serious threat to the socioeconomic stability of the region. An efficient and accurate method of generating landslide susceptibility maps is very important to mitigate the loss of lives and properties caused by these landslides. In this paper, a deep belief network (DBN) with datasets developed via a geographic information system (GIS) and remotely sensed data was used to create a landslide spatial susceptibility map for the TGR region on the Yangtze River in Zigui County. The landslide inventory map was initially constructed using field surveys, aerial photographs, and a literature search of historical landslide records. The twelve causative factors were evaluated in different ways: elevation, topographic slope, topographic aspect, curvature, distance from drainage, distance from road, schedule performance index and topographic wetness index were derived from a digital topographical map at 1:10,000 scale; engineering petrofabric, slope structure and distance from faults were obtained from a geological map at 1: 50,000 scale; normalized difference vegetation index were generated from CBERS (China-Brazil Earth Resources Satellite) data. All the causative factors were resampled to a grid cell size of 10 * 10 m by using the grid analysis function of ArcGIS software. Initially, 30% of the landslides pixels and equal number of non-landslides pixels were randomly selected as training data. Then these twelve factors were used as the input to DBN. By integrating the twelve factor maps in the GIS via pixel-based computing, the landslide spatial susceptibility map was obtained. Then the study area was reclassified into four categories of landslide susceptibility: high, moderate, low, and very low by using natural breaks classification methods. Approximately 13.6% of the study area was identified as severe susceptibility, and moderate, low, very low susceptibility zones covered 1.3%, 1.4%, and 83.7% of the study area, respectively. To show the performance of the DBN, the results were then compared with a logistic regression model (LRM) by using the receiver operating characteristics (ROC). The area under the ROC curve was 0.949 for the DBN and 0.859 for the LRM. The results showed that, the DBN proposed in this study outperforms the LSM. Based on the efficiency and accuracy of DBN, the proposed approach can be employed for rapid response to natural hazards in the Three Gorges area.