2:1 Data Reduction
The comparison between ADS-based SGM and LiDAR point clouds by Gehrke et al. (2010) has shown that the
resolution (ADS < 50 points/m2
, LiDAR < 10 points/m2
) and the horizontal accuracy (ADS: ~ 0.5 GSD, LiDAR: 10-
30 cm) are better than LiDAR for medium and high-resolution ADS data (GSD < 50 cm), while the vertical
accuracy is comparable to LiDAR (~ 5 cm) only for high resolution ADS imagery (GSD ~ 5 cm) – along with the
much higher point density. These differences in accuracy as well as the experience that such a huge amount of data
cannot be handled properly in the interactive post-processing as described below, we decided to reduce the amount
of data along with increasing the vertical accuracy.
This is achieved by joining the disparities of 2x2 neighboring base image pixels, similar to a 2:1 minification of
the disparity map. The consistency of respective values is verified: disparities that differ by more than 1 from the
(initial) average are disregarded; at least two disparities are required to agree for the final result. Such averaging
removes noise and doubles the vertical accuracy (for results based on 4 input pixels) and reduces the data by up to
75%. The result is considered the high-resolution output of the XPro DSM Extractor. Depending on the input image
GSD, it can still exceed the density of LiDAR data and theoretically achieve more comparable accuracy properties.
Note that the horizontal resolution of this final result is (almost) identical to an SGM output from the 2:1
pyramid level but the vertical accuracy is much better for two reasons: the use of full resolution imagery and, thus,all available information; and the averaging of neighboring disparities. The advantage of pyramid level output is
performance, a factor of (up to) 8 per level. In that regard, the XPro DSM Extractor provides the option of a ‘quick’
overview based on very fast computation using 8:1 minified imagery.