Mapping vegetated landslides using LiDAR derivatives and object- oriented analysis

Rapid mapping of fresh landslides is a necessity for efficient post-disaster response and updating of landslide inventories and susceptibility and hazard maps, and the increasing availability of very high resolution (VHR) remote sensing data has been facilitating such efforts. Fresh landslide scars are predominantly detected based on their strong spectral contrast to their surroundings, in particular the absence of vegetation (Martha et al. in press; Martha et al. 2010). Many currently inactive or reactivated landslides also pose a significant hazard (Van Den Eeckhaut et al. 2007), and thus need to be inventoried/identified/mapped and monitored (Kääb 2002; Metternicht et al. 2005). Such slide features, however, are usually either revegetated or at least still contain partial vegetation cover. Passive optical image data quickly lose their utility in such cases. Light Detection and Ranging (LiDAR) and its wide range of derivatives has become a powerful tool in landslide research, particularly for landslide morphology analysis (Glenn et al. 2006) and landslide identification and inventory mapping (Razak et al. 2011). Such data have been mainly used in expert-based visual analysis, while recently also a number of pixel-based computer-aided detection methods have been proposed (e.g. McKean and Roering 2004). Given the recent success of object-oriented analysis (OOA) approaches in landslide mapping and monitoring with optical data (e.g. Lu et al. 2011; Martha et al. 2010; Stumpf and Kerle 2011), we here test to what extent (semi-)automatic detection and monitoring of (partly) vegetated landslides with LiDAR data are also possible. Many recent OOA studies on landslide recognition have included elevation information, yet only for the classification stage (e.g. Martha et al. 2010). In our work we exclusively rely on the LiDAR data, which form the basis for both segmentation and classification. Features that have been identified as critical in previous studies, such as NDVI to detect absence of vegetation, or grey level cooccurrence matrix (GLCM)-based texture measures, were thus not considered.