Application of convolutional neural networks featuring Bayesian optimization for landslide susceptibility assessment

This study developed a deep learning based technique for the assessment of landslide susceptibility through a one-dimensional convolutional network (1D-CNN) and Bayesian optimisation in Southern Yangyang Province, South Korea. A total of 219 slide inventories and 17 slide conditioning variables were obtained for modelling. The data showed a complex scenario. Some past slides have spread over steep lands, while others have spread through flat terrain. Random forest (RF) served to keep only important factors for further analysis as a preprocessing measure. To select CNN hyperparameters, Bayesian optimization was used. Three methods contributed to overcoming the overfitting issue owing to small training data in our research. The selection of key factors by RF helped first of all to reduce information dimensionality. Second, the CNN model with 1D convolutions was intended to considerably decrease the number of its parameters. Third, a high rate of drop-out (0.66) helped reduce the CNN parameters. Overall accuracy, area under the receiver operating characteristics curve (AUROC) and 5-fold cross-validation were used to evaluate the models. CNN performance was compared to ANN and SVM. CNN achieved the highest accuracy on testing dataset (83.11%) and AUROC (0.880, 0.893, using testing and 5-fold CV, respectively). Bayesian optimization enhanced CNN accuracy by similar to 3% (compared with default configuration). CNN could outperform ANN and SVM owing to its complicated architecture and handling of spatial correlations through convolution and pooling operations. In complex situations where some variables make a non-linear contribution to the occurrence of landslides, the method suggested could thus help develop landslide susceptibility maps.