Objectives  Landslides are very harmful natural geological phenomenon.They are widely developed in mountainous areas in China, especially the Three Gorges Reservoir area. At present, more than 3 800 landslides have been reported. However, it is difficult to evaluate the landslide susceptibility and improve the accuracy of the model for areas with the same/similar geological background scientifically and accurately. The main objective of this study is to produce landslide susceptibility maps for the Guojiaba Town of Zigui County at the Three Gorges area, China by using deep belief network (DBN) method, and to explore its application effect in landslide susceptibility mapping.

  Methods  The landslide susceptibility mapping method is DBN. It is a superposition of multiple restricted Boltzmann machines (RBMs), which can better reflect the relationship among factors in order to obtain better results. Firstly, multi-source data is used to extract the factors affecting the landslide development in the study area. By analyzing the correlation between each factor and historical landslides, 12 evaluation factors that have a large impact on landslide development are selected, and they are quantified and normalized. Then, the grid unit is selected as the smallest unit for landslide susceptibility mapping, and the obtained sample units are selected and divided. 30% of all landslide data and the same amount of non-landslide data is selected randomly to form a training sample set. 70% of all landslide data and the same amount of non-landslide data as the validation sample set; third, the obtained training sample set is input to the deep belief network for training, and the validation sample set is used for model verification. In order to test the stability of the proposed model, with the same sample and parameter settings, five experiments are carried out. Finally, the same training set and validation set are used to evaluate the landslide susceptibility using shallow neural networks and logistic regression models.The results of the three methods are evaluated by using ROC(receiver opera- ting characteristics) curve, OA (overall accuracy) and Kappa coefficient.

  Results  The proposed DBN model is used to calculate based on the evaluation units in the study area, and the probability values corresponding to the evaluation units are obtained. The susceptibility of landslides in the study area is divided into four classes which are non-prone, low-prone, moderate-prone and high-prone area. The zonation map of landslide susceptibility based on DBN is obtained. From the results of the stability test of DBN, the average accuracy of DBN constructed in the mapping of landslide susceptibility is 91.14%, and the mean square error is 1.94%, which proves that the method has better performance of stability. By comparing with the mapping results obtained by the shallow neural network and logistic regression model, it is found that DBN has higher accuracy on the three evaluation indexes of ROC curve, OA and Kappa coefficient, which are 0.95, 90% and 0.81, followed by the shallow neural network model, of which are 0.92, 88%, and 0.76. The accuracies of logistic regression model are the lowest, of which are 0.86, 78%, and 0.55, respectively.

  Conclusions  A DBN is used to quantitatively mapping the landslide susceptibility in Guojiaba Town, Zigui County, the Three Gorges Reservoir area, and a regional landslide susceptibility map is generated.Compared with the shallow neural network and traditional logistic regression method, the accuracy assessment of the results is performed using three accuracy evaluation indexes, which are the receiver's working characteristic curve, OA, and Kappa coefficient. The results show that DBN obtains good stability, and the area under the curve, OA, and Kappa values are the highest among the three models, indicating that DBN performs a good predictive ability and stability in the mapping of landslide susceptibility.