Manfred Sties
to transform the radar data to the local system, and this was performed by colleagues at the Fachhochschule Karlsruhe.
The Intermap preference was to transform the laser data into the radar reference system. In-house software was developed
by Intermap that utilized the transformation parameters provided by LVA.
4.2 Comparison of ’Original’ Measurement Data in Karlsruhe
The comparison of ’original’ measurement data was done by the authors in Karlsruhe. The method was based upon
selection of laser measurement points within a limited x-y-distance from radar cell centers. Elevation differences were
calculated between both, the laser and the radar data sets for those point pairs. A distance threshold of approx. 0.6 m was
chosen; a smaller distance limit threshold led to selection of very small numbers of point pairs, which would not produce
reliable statistical figures. For the 46 test areas of different relief and vegetation coverage, all sets of less than 20 point
pairs were deleted as statistically unreliable. In the case of steep slopes or vertical break lines, elevation values taken at
points which were selected through this method will differ substantially; these effects were taken into account by deleting
evident statistical outliers.
4.3 Comparison of Co-located Elevation Rasters
The second method was to compare rasterized grid data. The Karlsruhe University and Intermap approaches were similar,
4.3.1 Areas of Pre-selected Land Cover Type (Karlsruhe) The authors from Karlsruhe chose a grid spacing of
4.5 m, intermediate between the 5 m spacing of the regular radar grid and the average 4m spacing of the irregularly
spaced laser data set. Both data sets were interpolated using the ARC/INFO IDW (inverse distance weighting) function
with parameters set at n = 8 neighboring points, and exponent exp = 3 for the distance weighting. The approach here
was to classify the test area into six sub-classes (urban, forest, water, vineyards, farmland with no or very low ground
cover, farmland with vegetation), based upon interpretation of the digital ortho-photos (0.5 m resolution) and a 1 : 50.000
topographic map of the area. Polygons were digitized with respect to homogenous classes as visually interpreted. Sta-
tistical tests were performed on the elevation difference grids (radar minus laser) with respect to each of the classes. All
interpolated elevation rasters for test areas which were based on too small a number of original’ measurement points
were deleted. This - evidently - refers to all test areas of the class "water' and class 'agriculuture' and to some test areas
of class 'vineyard' where the data subsets "vegetation! of the laser scanner generated measurements were empty. This is
presumably due to the mid-winter acquisition date for the laser data.
4.3.2 Areas of Post-selected Land Cover Type (Intermap) The Intermap method was to grid the laser points into 5 m
cells matching the radar grid. In this case, the MapInfo/Vertical Mapper software package was used. An inverse distance
weighting interpolator was used, with a search radius of 15 m and a weighting exponent of 2. The radar data comprised
both the radar DEM and the ortho-rectified radar image (ORI) that was produced simultaneously with the DEM. The 2.5 m
spacial resolution of the ORI is useful for interpretation support. An overview of the test area is presented in Fig. 4a. The
test region contains a mixture of rolling hills, broad valleys, agricultural land (crops and grassland), forests, villages and
infrastructure.
Because the principle thrust in this work was to observe the error magnitude of the radar DSM (relative to the laser),
it was necessary to eliminate the effects of terrain coverage differences from the statistical analysis. This was done
through the process of classifying and sampling sub-areas that were believed to be representative of the bald-earth. The
classification was performed visually and was based on a combination of DSM, ORI and elevation difference surface. It
was easy to eliminate forests, villages and other significant objects from the sample. The challenge lay in differentiating
elevated crops from bare earth and low-lying crops. This cannot be done merely from inspection of the ORI, because
some low-lying crops (e.g., cabbage) have bright radar returns and therefore show up well in the ORI, but are only a few
centimeters above the ground. Fortunately, the crop patterns for elevated crops were quite visible in the difference surface
visualization. Therefore, the bald-earth classification consisted of visually searching for the lowest (height-wise) local
field patterns in the difference surface, subject to minimum areal size constraints. A total of 67 'bald-earth' sections were
selected in this manner, with an average area of about 5 hectars. This sampling method will tend to over-estimate errors
because of failure to remove all low-lying crop effects. The spatial distribution, although not uniform, was adequate to
ensure at least one sample per km?. In order to obtain a global result, the difference values from the 67 sections were
analysed.
5 RESULTS
5.1 Results Comparing "Original? Elevation Measurements (Karlsruhe)
Fig. 1 shows a graphical representation of elevation differences derived from the selected coincident point pairs of the
SAR and the laser ' ground' data sets. In total, 19.390 point pairs from all 46 test areas were included in this calculation of
868 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
Fi
el