Detecting translational landslide scars using segmentation of Landsat ETM+ and DEM data in the northern Cascade Mountains, British Columbia

Extensive landslide inventories are often utilized for hazard assessment studies and when investigating medium- to long-term evolution of alpine terrain. The predominant methodology for collecting these databases is aerial photographic interpretation, which can be time-consuming and expensive. Earlier work has demonstrated that spectral response patterns for satellite images, when used alone, are unreliable at detecting most types of landslides. Principal difficulties are related to inadequate image resolution and spectral methods of classifying image data that are not sensitive to the characteristics that identify landslide features such as their shape and topographic expression. This study in the Cascade Mountains of coastal British Columbia attempts to overcome the latter problem through image segmentation and the use of geomorphometric data derived from a digital elevation model (DEM). Image segmentation involved grouping pixels into discrete objects based on similarities and differences in their reflectance and the use of shape criteria. A hierarchical classification system was then developed such that the normalized difference vegetation index (NDVI) and slope data eliminated all areas in the image that were both vegetated and on a gradient of less than 15°. The remaining "unvegetated steeplands" were classified using a supervised classification based on spectral, geomorphic, and shape criteria. The technique produced an overall accuracy of 75% in the detection of landslides that were over 1 ha in area.