Comparison of the airborne datasets

The main dissimilarity of the two airborne surveying methods lies in the density of the resulting XYZ point. Where laserscanning data can have a constant sampling density of

almost 2.5 point per m² the automatically matched mass point from stereo aerial images has an inhomogeneous distribution depending on the texture and the spectral information of the stereo images. The resulting mass points showed in relatively well structured areas 0.8 point per m² density however in less correlating areas it dropped down to 0.2 point per m². This sampling difference accounts for the more detailed description of the surface using LIDAR data and a lower resolution one from stereo image matching

The classical limitation of automated 3D information extraction from stereo images lies in the dependency on texturing and spectral information of the data. In case the image has lower textured spectral information neither can automatic techniques nor can manual interpreter match data points over the image pairs. This is well reflected in the analysed dataset where on large open green grass fields no points were found to be matched thus a lack of elevation information appeared (Figure 2.4-4).

Figure 2.4-4.: Density of matched elevation mass points marked with black dots in open grass fields around King’s college on upper left image and its interpolated TIN model on the upper right image. No matched points were found on lower textured open grass fields marked with blue ellipse. Higher density was found around buildings marked with blue square from aerial images. The same area from raw last pulse LIDAR data on the two lower images

After determining the spatial distribution of elevation mass points to measure difference the algebraic subtraction of the datasets was calculated (equation (2-6.)). The filtered LIDAR digital surface model without vegetation was set as a master source for elevation information.

The stereo matched data set was subtracted from it as following:

ΔDSM = DSM lidar – DSM photo (2-6.)

The difference was quantified for the whole image by calculating the histogram of the difference image (Figure 2.4-5). The histogram has a Gaussian distribution with a mean at 6.19 m and a standard deviation of 8.55 m. Its extreme values are high; its minimum is at - 41 m and maximum at 66 m. This reflects a great uncertainty in elevation information and a lower elevation accuracy of image matching when compared with laser survey. Furthermore its generally higher elevation information reflects the presence of vegetation cover that was not eliminated after the matching process in lack of sufficient point density to serve as a basis for filtering.

Figure 2.4-5.: Histogram of elevation differences in meters between stereo matched elevation points and laser scanned height information

The spatial distribution of the height differences was visualised around the town (Figure 2.4-6). The results of the elevation differences revealed great height measurement dissimilarity in shadowed areas next to high buildings in the centre of the town. This is to be explained by the generally lower matching quality in shadows due to changed spectral information and the geometrical shadowing on aerial images described in chapter 2.4.1.

resulting in a low accuracy of elevation information. On the other hand as the edge of shadows has a high contrast jump usually matching is performed along the border area as visible in Figure 2.4-4. The temporal delay of acquisition between the stereo image pairs can cause a displacement in the shadow. Consequently the spatial intersection along the shadow might result in incorrect elevation measurement as visible on the same figure.

Figure 2.4-6.: Spatial distribution of height differences between stereo matched elevation points and laserscanned data.

Another huge difference appeared between the two datasets in the description of urban canyons. The sudden elevation drop over streets between high buildings in the city centre was modelled differently (Figure 2.4-7). In some cases the radial-symmetric displacement of objects in the aerial images was geometrically hindering the matching of the stereo points on the street elevation. In other cases the lack of texture on street pavement was hindering the search for conjugating points on the streets between high buildings. LIDAR data showed a lower information lack that was caused only by the geometrical shading depending on the scan angle and the flying height. Thus surface structures with great elevation differences of urban canyons were better described using laser scanning.

Figure 2.4-7.: Urban canyons in elevation profiles across Cambridge town from different surface models.

Left image shows the location of the profile on the LIDAR elevation data. Right image shows the extracted profile where red arrows refer to urban canyons of streets or courts.

Location of the profile

Summarising the higher sampling density of LIDAR data was effecting the quality of the resulting DSM when compared with data from stereoscopic elevation extraction. Data

processing after acquisition is faster to achieve first results than in photogrammetry. However photogrammetry can get higher elevation accuracy with manual interpretation then laser scanned data. Especially it can be important in urban areas where built-up structures include breaklines in the elevation. The effect of elevation uncertainty on steady state flow calculation will be analysed and elaborated in the following main chapter.