| Abstract | Digital elevation models (DEMs) extracted from high-resolution stereo images (SPOT-5, EROS-A, IKONOS-II and QuickBird) using a three-dimensional multi-sensor physical model developed at the Canada
Centre for Remote Sensing, Natural Resources Canada were evaluated. In a first step, the photogrammetric bundle adjustment was set-up for the stereo-images with few accurate ground control points. In a second step, DEMs were generated using an
area-based multi-scale image matching method and then compared to 0.2-m accurate LIDAR elevation data. Elevation linear errors with 68% confidence level (LE68) of 6.5 m, 20 m, 6.4 m and 6.7 m were achieved for SPOT, EROS, IKONOS and QuickBird,
respectively. The poor results for EROS are mainly due to its asynchronous low orbit, which generated large geometric and radiometric differences. However, when such differences were not large, LE68 of 10 m (four pixels) was achieved. Since the
SPOT, IKONOS and QuickBird DEMs were in fact digital terrain surface models, where the height of land covers was included, elevation accuracy was performed only on bare surfaces (soils and lakes), where there was also no difference between the
stereo-extracted elevations and the LIDAR data. LE68 of 2.2 m, 1.5 m and 1.2 m were then obtained for SPOT, IKONOS and QuickBird, respectively. When compared to sensor resolution, multi-date across-track SPOT achieved three to four times better
results than same-date in-track IKONOS and QuickBird: half-pixel versus 1.5 or two pixels. |