DEM generation with Cartosat-1 Stereo Imagery

By the Shuttle Radar Topography Mission (SRTM) a nearly worldwide digital height model has been generated. The SRTM height models can be used for several applications, but it has gaps in mountainous regions, caused by the radar layover, and is not so accurate in steep areas. In addition the grid width of 3 arcsec, corresponding to approximately 92m at the equator, is limiting the morphologic details. From the very high resolution optical space sensors only a limited number of stereo pairs, taken from the same orbit, are available. The automatic matching of images taken with significant time interval is very difficult and not leading to satisfying results. In addition stereo pairs from IKONOS, QuickBird and the former OrbView-3 are expensive. The stereo pairs from the SPOT HRS only have been available for test purposes, usually only the derived height models can be bought. As alternative solution the stereo sensors Cartosat-1 and ALOS/PRISM are now available; both systems with 2.5m ground sampling distance (GSD).
In the frame of the ISPRS-ISRO Cartosat-1 Scientific Assessment Programme (C-SAP) digital height models have been generated by automatic image matching with least squares method. The generated digital surface models (DSMs) are describing the height of the visible surface, but the reference digital elevation models (DEMs) the height of the bare ground. By this reason the generated DSMs have been filtered with the Hannover program DEMCOR to the points describing a DEM. For the comparison with the reference DEMs the horizontal fit of the height models to each other has been checked by adjustment. In addition partially a correction by geoid undulation was required. For a comparison of generated height models with a reference height model, the orientation of the height models has to be checked, because often datum problems exist. So as standard procedure before analysis, the orientation of the height models to each other is checked by adjustment to avoid unrealistic accuracy information just caused by a shift of the height models. After fitting and filtering for open and flat terrain a standard deviation of the generated DEM in relation to a reference DEM of up to 2.4m has been reached. This corresponds to a standard deviation of 0.7 GSD for the x-parallax. The 2.5m GSD justifies a point spacing of 7.5m for the DEM. Of course the automatic image matching fails if the object has no contrast like in the case of snow coverage, partially existing in the Warsaw test area, but even in a rolling up to mountainous area, nearly completely covered by forest, less than 3% of the area could not be matched in a Cartosat-1 stereo scene.