Collapse susceptibility evaluation based on an improved two-step sampling strategy and a convolutional neural network

Abstract

Objective Machine learning has been widely applied in the fields of collapse, landslide and debris flow susceptibility analysis. The selection of nonhazard samples is a key issue in landslide susceptibility analysis. Traditional random sampling and manual labelling methods may involve randomness and subjectivity. Methods In view of the potential randomness and representativeness of noncollapse samples, this paper considered soil collapse susceptibility evaluation a positive-unlabelled (PU) learning problem and proposes a two-step convolutional neural network framework (ISpy-CNN) that combines an information value model and the Spy technique. First, 15 collapse-related factors were selected for modelling based on the geomorphological, geological, hydrological, and artificial environmental conditions of the study area. Low-information-value samples that were able to map the distribution structure of noncollapsing samples were screened by the information value model. Then, through the Spy technique and training the CNN model, negative samples with high confidence were identified from low-information-value samples that were classified as noncollapsed samples. Finally, based on the framework and traditional random sampling, we used support vector machine (SVM) and random forest (RF) models to compare and verify the reliability, prediction accuracy and data sensitivity of the proposed learning framework and other models. Results The results illustrate that the proposed ISpy-CNN method can improve the accuracy, F1 value, sensitivity and specificity on the validation set by 6.82%, 6.82%, 6.82%, 8.23%, respectively compared to random sampling and 2.86%, 2.89%, 2.86%, 2.31%, respectively compared to the traditional Spy technique. The prediction accuracy of step 2 in PU learning using the CNN model is higher than that of the RF and SVM models. The sample set screened by the ISpy-CNN framework exhibited greater stability, prediction accuracy and growth rate than those screened by the traditional Spy technique by adding the same number of training samples. Conclusion The ISpy-CNN framework proposed in this paper can better assist in the selection of nonhazard samples and real collapse spatial distribution maps, and the results of the framework are more consistent with the actual collapse distributions.

Keywords

• collapse
• susceptibility evaluation
• positive and unlabeled (pu) learning
• spy technique
• information value
• convolutional neural network
• random forest
• support vector machine