OBJECT-ORIENTED CLASSIFICATION FOR EXTRACTING LANDSLIDES FROM DMC AERIAL IMAGES

摘要

Collection of landslide data is important for land conservation and disaster management. Aerial ortho-images and geo-referenced satellite images have been used in detection of landslides, however, generation of those products is time-consuming and thus can be inefficient for landslide analysis. In this paper, an "object-oriented classification method" for landslide extraction from raw DMC (Digital Mapping Camera) images is proposed. Processing of each raw DMC image consists of four steps: (1) Segment the image into individual regions-"image objects"-using multi-resolution segmentation algorithm. (2) Categorize image objects into three subsets-darker-area, normal-area,and lighter-area-based their brightness values (BVs), then apply different rules to extract landslide areas from each subset. The image classification results are then exported in shapefile format, one vector layer for each raw image. (3) Convert spatial reference of exported landslide data from "image" coordinate system into "map" (TWD97 TM2) coordinate system using ray-tracing algorithm. (4) Overlay landslide data with map coordinates on ancillary topographic data, such as slope and aspect data, to further filter and refine the initial classification results. Test results show that both user's accuracy and producer's accuracy of the landslide classification can be higher than 82%.