Automatic image classification of landslides improved with terrain roughness indices in various kernel sizes

Using spectral-only information for landslides classification is usually confusing with houses, roads, and other bare lands because these ground features have similar spectral patterns on images. The terrain roughness can be measured by significant wavelengths; some studies have linked the relationships between terrain roughness and the landslide by using numerical analyses of topography data. In this study, airborne LiDAR data of 1m grid are used to explore the possibility of improvement of landslide classification, the LiDAR-derived data include DEM slope and terrain roughness indices including diversity, dominance and relative richness with different grid size data are used to improvement classification accuracy. The improvement of accuracy when including DEM slope is 22% in producer's accuracy and 27% in user's accuracy. The accuracy of diversity, dominance and relative richness indices all are improved when kernel sizes enlarge in Maximum Likelihood and Mahalanobis Distance algorithms.