A Bayesian approach for mapping event landslides using optical remote sensing imagery and digital terrain data

Event landside inventory maps can be prepared using conventional or new mapping methods. Conventional methods, including field mapping and the visual interpretation of stereoscopic aerial photographs, are time consuming and resource intensive, restricting the ability to prepare event inventory maps rapidly, repeatedly, and for large and very large areas. This is a significant drawback for regional landslide studies and post event remedial efforts. Investigators are currently experimenting new methods for preparing landslide event inventories exploiting remotely sensed data, including qualitative (visual) and quantitative (numerical) analysis of very-high resolution (VHR) digital elevation models obtained chiefly through LiDAR surveys, and the interpretation and analysis of satellite images, including panchromatic, multispectral, and synthetic aperture radar images. We devised a stepwise, semi-automatic approach to detect, map, and classify internally rainfall-induced shallow landslides exploiting multispectral satellite images taken shortly after a landslide-triggering event, and information on the topographic signature of landslides obtained from a pre-event digital elevation model. In a Bayesian framework, the approach combines a standard image classification obtained by a supervised classifier (e.g., the Mahalanobis Distance classifier) applied to a post-event image, with information on the morphometric landslide signature measured by statistics of terrain slope and cross section convexity in landslide and stable areas. The semi-automatic approach is applied in two steps. First, the rainfall-induced landslides are detected and mapped, separating them from the stable areas. Next, the mapped landslides are classified internally, separating the source from the run out areas. We have applied the approach in a 117 km2 study area in Taiwan, where shallow landslides triggered by high intensity rainfall brought by typhoon Morakot in august 2009 were abundant. Comparison in a GIS of the event landslide inventory produced by the proposed semi-automatic method with a similar event inventory obtained through the visual interpretation of post-event ortho-photographs reveals a degree of matching > 90%. The new approach is flexible, it can exploit different satellite imagery and terrain elevation data, and can be used to map and classify landslides caused by various triggers in different physiographical environments. We expect our new method to contribute to the production of event landslide inventory maps.