; 
s 
| € 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
At first, as mentioned before in paragraph 3, an image 
preprocessing step to reduce speckle noise was carried out, after 
the orientation performed by the rigorous model and before 
starting the matching procedure; to this aim, imagery has been 
enhanced with a 7 x 7 window Lee filter. 
The accuracy assessment has been performed using 
DEMANAL software, developed by Prof. K. Jacobsen— 
Leibniz University Hannover, allowing for a full 3D 
comparison to remove possible horizontal biases too. The 
height differences are computed interpolating, with a bilinear 
method, the analyzed DSM over the reference LIDAR DSM; 
the value is negative when the extracted DSM is above the 
reference LIDAR DSM. The accuracy in terms of RMSE was 
computed at the 95% probability level (RMSE LE95), after 
having evaluated the LE95; therefore, the largest outliers were 
filtered out. 
Two tiles (herein called Tile 1 and Tile 2) have been selected 
for the analysis, considering different morphological situations. 
In particular, the first one is mainly an urban area, the second 
includes both an urban area and forested and cultivated areas 
displaying a hilly topography. 
The results of the accuracy assessment are presented in Table 2. 
At first, the point clouds derived from the ascending and 
descending stereo pairs, directly produced by matching 
procedure without any further post-processing, have been 
analyzed. The accuracy is around 5.0 — 5.5 m for Tile 1, around 
7.0 — 9.0 for Tile 2. Analyzing the results, some outliers have 
been detected in the point clouds, probably due to mismatching 
causing incorrect morphological reconstruction in small areas. 
To remove these outliers, a free available low resolution DSM 
(Shuttle Radar Topography Mission — SRTM DEM, 3' grid 
posting) has been used as reference. In details, the point clouds 
have been compared with SRTM, and the height differences 
have been computed; when the difference was greater than a 
fixed threshold, the corresponding point has been rejected. Two 
tests have been performed, using two thresholds respectively 
fixed at 20 m and 15 m. As shown in Table 2, an accuracy level 
of 4.5 m and 6 m respectively for Tile 1 and Tile 2 were 
obtained; the results achieved using 20 m and 15 m thresholds 
were not found significantly different, so that the 20 m 
threshold can already be considered effective for filtering out 
the highest errors, with a loss of matched points of about 10%. 
Subsequently, starting from the point clouds, three DSMs have 
been generated and assessed on a 2 m grid posting by a rough 
linear interpolation (Delaunay triangulation). In Table 3 the 
results of the interpolated DSMs are shown; in particular the 
ascending DSM is generated using the points cloud of the 
ascending stereo pair, the descending DSM using the points 
cloud of the descending one, and finally a merged DSM has 
been generated using a combination of the point clouds that 
have been previously filtered in order to remove the matched 
points with lower correlation (Figures 1 and 2). 
In Tile 1 the accuracy is around 11 m and 6 m for the ascending 
and the descending DSMs respectively, and around 8 m for the 
merged product; after the SRTM filtering, the accuracy 
increases to approximately 5 m. 
In Tile 2 the accuracy is around 11 m and 13 m for the 
ascending and the descending DSMs respectively, and around 
9.5 m for the merged product; after the SRTM filtering, the 
accuracy increases to approximately 7.5 m. 
The merging does not improve the results significantly, with the 
exception of the not processed product of Tile 2. Overall Tile 2 
presented worse results than Tile 1, probably due to more 
complex morphologies of the area. 
30 
  
  
F igure 1. Radargrammetric DSM of Trento — Tile 1, descending 
(above), ascending (centre), merged (below) 
6. CONCLUSIONS AND FUTURE PROSPECTS 
A new model for radargrammetric processing (orientation and 
DSM generation) of high resolution satellite SAR stereo pairs 
was defined and implemented in the scientific software SISAR. 
An experiment was carried out over the test site of Trento 
(Northern Italy), where two same-side SpotLight stereo pairs 
have been acquired on ascending and descending orbits 
respectively by TerraSAR-X. 
Two tiles have been selected for the accuracy analysis 
considering different morphological situations. 
At first, the ascending and descending stereo pairs have been 
processed separately and the correspondingly generated point 
clouds have been assessed with respect to a LIDAR DSM used 
as reference; the accuracy was estimated around 5.0 — 5.5 m for 
Tile 1 and around 7.0 — 9.0 m for Tile 2. 
Further, the generated point clouds have been filtered, using 
SRTM DEM (3' grid posting) in order to mostly remove large 
errors; the contribution was found effective and the accuracy 
increased to 4.5 m and 6 m for Tile 1 and Tile 2 respectively. 
Finally, complete DSMs with 2 m grid posting were generated 
both for Tile 1 and Tile 2, applying a rough linear interpolation 
(Delaunay triangulation), obtaining an accuracy of 5 m and 7.5 
m respectively.