DEM generation with worldview-2 images

Abstract

For planning purposes 42km coast line of the Black Sea, starting at the Bosporus going in West direction, with a width of approximately 5km, was imaged by WorldView-2. Three stereo scenes have been oriented at first by 3D-affine transformation and later by bias corrected RPC solution. The result is nearly the same, but it is limited by identification of the control points in the images. Nevertheless after blunder elimination by data snooping root mean square discrepancies below 1 pixel have been reached. The root mean square discrepancy at control point height reached 0.5m up to 1.3m with a base to height relation between 1:1.26 and 1:1.80. Digital Surface models (DSM) with 4m spacing have been generated by least squares matching with region growing, supported by image pyramids. A higher percentage of the mountainous area is covered by forest, requiring the approximation based on image pyramids. In the forest area the approximation just by region growing leads to larger gaps in the DSM. Caused by the good image quality of WorldView-2 the correlation coefficients reached by least squares matching are high and even in most forest areas a satisfying density of accepted points was reached. Two stereo models have an overlapping area of 1.6 km times 6.7km allowing an accuracy evaluation. Small, but nevertheless significant differences in scene orientation have been eliminated by least squares shift of both overlapping height models to each other. The root mean square differences of both independent DSM are 1.06m or as a function of terrain inclination 0.74m + 0.55m * tangent (slope). The terrain inclination in the average is 7 degrees with 12% exceeding 17 degrees. The frequency distribution of height discrepancies is not far away from normal distribution, but as usual, larger discrepancies are more often available as corresponding to normal distribution. This also can be seen by the normalized medium absolute deviation (NMAS) related to 68% probability level of 0.83m being significant smaller as the root mean square differences. Nevertheless the results indicate a standard deviation of the single height models of 0.75m or 0.52m + 0.39* tangent (slope), corresponding to approximately 0.6 pixels for the x-parallax in flat terrain, being very satisfying for the available land cover. An interpolation over 10m enlarged the root mean square differences of both height models nearly by 50%.