Automation of landslide detection using optical images and lidar data

Both of optical images and airborne LiDAR data are used to recognize landslide features in this study. NDVI and green index derived from spectral information of optical images are used at the first to obtain a preliminary distribution of landslides. Subsequently, parameters derived from DSM and DEM of airborne LiDAR data are applied, including slope gradients, OHM and surface roughness. The study area is in Shimen Watershed of northern Taiwan. After a comparison of the automated-recognized landslides and those interpreted manually by aerial photo-interpretation, an accuracy of 63.32 % is achieved when only NDVI is applied whereas it is 84.79% when slope gradient is included. An accuracy of 94.5 % is achieved after manual editing with additional thematic information is included such as known agriculture land-use, roads and rivers. Four parameters are included in an automated recognition of landslides, namely greenness, slope, OHM and roughness. The foremost one is derived from orthophoto and the other three are derived from LiDAR data. They can be applied by selecting training areas for automatically setting of default values for region-growing and adoptive thresholds. This can effectively recognize local distribution of landslides.