DTM Generation from Ikonos In-Track Stereo Images Using a 3D Physical Model

A digital elevation model (DEM) extracted from Ikonos in-track stereo images using a 3D physical model developed at the Canada Centre for Remote Sensing, Natural Resources Canada was evaluated. First, the stereo photogrammetric bundle adjustment was set up with about ten accurate ground control points. The DEM was then generated using an area-based multiscale image matching method and 3D semiautomatic editing tools and then compared to lidar elevation data with a 0.2-m accuracy. Because the DEM is, in fact, a digital terrain surface model where the height of land cover (trees, houses) is included, the accuracy varies depending on land-cover types. Using 3D visual classification of the stereo Ikonos images, different classes (forests, residential, bare soil, lakes) were generated to take into account the height of the surface (natural and human-made) in the accuracy evaluation. An elevation linear error with 68 percent confidence level (LE68) of 1.5 m was obtained for bare surfaces while an LE68 of 6.4 m was achieved over the full area. Five-meter contour lines could thus be derived, compliant with the highest topographic standard. Better results could thus be expected when using stereo-images acquired in the summertime. On the other hand, an LE68 of 2.5 m to 6.6 m was obtained depending on the land-cover type and its surface height. For residential areas, the surface height did not affect the errors very much (2.5-m LE68) when compared to bare surface results because one- and two-story houses were sparse in the test area. Because the images were unfortunately acquired in wintertime and the lidar data in summertime, elevation errors (LE68 and bias) also depended on the type of forest (deciduous, coniferous, mixed, sparse). An evaluation based on terrain slope and azimuth showed that the DEM error was linearly correlated with slope and that elevations on sun-facing slopes were 1-m more accurate than elevations on slopes facing away from the sun.