Segmentation-assisted image matching in high resolution satellite images
Digital elevation models are one of the most important sources of information for many spatial analyses. Traditionally, DEMs have been produced by manual extraction of heights by a human operator which is quite a time-consuming and tedious procedure. Thus developing algorithms for automatic extraction of heights from air-borne and space-borne images has been an active research area for many years in Remote Sensing and Photogrammetry. Today, many sensors aboard newly launched satellites like IKONOS, QuickBird, SPOT and IRS are able to produce high resolution along-track and across-track stereo images to be used in large scale topographic mapping. With the vast amount of data captured by these sensors, the need for algorithms capable of accurate generation of DEMs is increasing. Many algorithms have been proposed for automatic extraction of heights from satellite imagery, but most of them fail to produce acceptable results and need a lot of post-processing work to eliminate the false heights mainly generated by false image matching. This is as a result of the fact that traditional methods of satellite image matching do not pay any attention to land cover of the working region and perform a fixed matching procedure for all kinds of the earth surface. Here we propose an adaptive algorithm for image based matching which benefits from a preliminary segmentation of images to find the conjugate pixels more accurately. While images are segmented, textural region information are used to detect problematic regions like low-textured ones. This is in turn used to guide the matching process to be tuned with respect to different land cover types. Our experimental results show that the algorithm is robust when dealing with various land types and can drastically decrease the number of gross errors in the final DEM product.