Processing Framework for Landslide Detection Based on Synthetic Aperture Radar (SAR) Intensity-Image Analysis

We present an object-based image analysis (OBIA) approach to identify temporal changes in radar-intensity images and to locate land-cover changes caused by mass-wasting processes at small to large scales, such as landslides. Our approach is based upon change detection in SAR intensity images that remain in their original imaging coordinate system rather than being georeferenced and map-projected, in order to reduce accumulation of filtering artifacts and other unwanted effects that would deteriorate the detection efficiency. Intensity images in their native slant-range coordinate frame allow for a consistent level of detection of land-cover changes. By analyzing intensity images, a much faster response can be achieved and images can be processed as soon as they are made publicly available. In this study, OBIA was introduced to systematically and semiautomatically detect landslides in image pairs with an overall accuracy of at least 60% when compared to in-situ landslide inventory data. In this process, the OBIA feature extraction component was supported by derived data from a polarimetric decomposition as well as by texture indices derived from the original image data. The results shown here indicate that most of the landslide events could be detected when compared to a closer visual inspection and to established inventories, and that the method could therefore be considered as a robust detection tool. Significant deviations are caused by the limited geometric resolution when compared to field data and by an additional detection of stream-related sediment redeposition in our approach. This overdetection, however, turns out to be potentially beneficial for assessing the risk situation after landslide events.

Automated summary

An object-based image analysis approach was presented to identify temporal changes in radar-intensity images and locate land-cover changes caused by mass-wasting processes, achieving at least 60% overall accuracy. The method offers faster response times and can detect most landslide events, making it a robust detection tool despite significant deviations caused by limited geometric resolution.