Pixel-based classification method for earthquake-induced landslide mapping using remotely sensed imagery, geospatial data and temporal change information

A series of earthquakes occurred in Kumamoto, Japan, in April 2016, which caused numerous landslides. In this study, high-resolution pre-event and post-event optical imagery, plus bi-temporal Synthetic Aperture Radar (SAR) data are paired with geospatial data to train a pixel-based machine learning classification algorithm using logistic regression to identify landslides occurred because of the Kumamoto earthquakes. The geospatial data used include a categorical variable (surficial geology), and six continuous variables including elevation, slope, aspect, curvature, annual precipitation, and landslide probability derived by the USGS preferred geospatial model which incorporates ground shaking in the input parameters. Grayscale index change and vegetation index change are also calculated from the optical imagery and used as input variables, in addition to temporal differences in HH (horizontally transmitted and horizontally received polarization) and HV (horizontally transmitted and vertically received polarization) amplitudes of SAR data. A detailed human-drawn landslide occurrence inventory was used as ground-truth for model development and testing. The selection of optimal features was done using a supervised feature ranking method based on the Receiver Operating Characteristic (ROC) curve. To weigh the benefit of combining different types of imagery, temporal change information and geospatial environmental indicators for landslide mapping after earthquakes, five different combinations of features were tested, and the results showed that adding data of selected geospatial parameters (landslide probability, slope, curvature, precipitation, and geology) plus selected change indices (grayscale index change, vegetation index change, and HV amplitude difference of SAR data) to the imagery (post event optical) lead to the highest classification accuracy of 86.5% on class-balanced independent testing data. A comparative analysis was conducted to evaluate the performance of the proposed method with five other commonly used machine learning classification methods, and the results have shown the superiority of the logistic regression method, followed by support vector machines.