Estimation of safety factor in landslide through image processing with recurrence plot and oversampling techniques

Abstract

Abstract The convolutional neural network (CNN) algorithm as one of image processing techniques has been applied to reveal whether the safety factor is higher or lower than the standard value. However, the existing methods have limitations in quantitatively revealing the safety factor across various ranges. The objective of this study is to quantitatively assess the safety factor with the CNN algorithm through an oversampling technique for reflecting various ranges of the safety factor. Eight geotechnical parameters are selected as independent variables and are obtained through experimental studies. The numeric data in each grid are converted into images using the Recurrence Plot (RP) algorithm to carry out the CNN algorithm. The converted images are matched with the safety factor as the true value calculated by the infinite slope stability model, and the synthetic minority oversampling technique (SMOTE) is applied to solve imbalances in the data, which is the case for a relatively small amount of data in each safety factor. The constructed image data are trained and tested using the ResNet 50 algorithm, and the data oversampled by SMOTE showed higher accuracy than the imbalanced data. This study demonstrated that the suggested strategy may be used as an alternative method to find various ranges of safety factors using numeric data with an oversampling technique.