Michele Crosetto 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Slope Number | Standard Slope Number | Standard 
Range of pixels | Deviation Range of pixels | Deviation 
[%] [ 7c] [m] [ 76 ] [70] [m] 
073 10.6 8.30 3+0 10.7 8.15 
3+6 9.2 9.02 -6 + -3 9.3 8.18 
6=9 7.7 9.97 -9 + -6 7.0 9.20 
9=12 5.7 11.01 -12+-9 5.3 9.91 
12 + 15 4.5 12.17 -15+-12 4.0 10.77 
15+ 18 3.5 13.30 -18 7-15 3.1 11.46 
18 +21 2.6 14.83 -21 +-18 24 11.85 
21 — 24 2.0 16.23 -24 + -21 1.9 12.05 
24 +27 1.4 17.58 -27 + -24 15 12.97 
27 +30 1.0 19.01 -30 = -27 1.1 13.71 
30 = 33 0.7 20.62 -33 = -30 0.8 14.32 
33 +36 0.5 22.53 -36 = -33 0.6 15.53 
> 36 1.2 24.41 <-36 1.6 19.53 
  
  
  
  
  
  
  
  
Table 4: Ascending InSAR DEM. Statistics of the height differences (InSAR versus reference DEM) computed for 
different terrain slope classes. The slopes are calculated along the ground range direction, i.e. perpendicular to the 
satellite track. Slopes facing the SAR antenna have positive values. 
2.1.3 Atmospheric effects. Atmospheric inhomogeneities during the SAR image acquisition cause distortion effects 
in the generated DEMs. Such effects, that are independent of the terrain topography and the coherence, can be noticed 
in the ascending DEM. In fact, it is affected by important systematic errors with low spatial frequency characteristics 
and magnitude up to 30735 m. The effect of such errors appears evident considering the autocovariance function of the 
height differences between the IDSAR DEM and the reference one (see Figure 3). The correlation length is about 505 m 
and the correlation decreases to zero very slowly, i.e. the height differences are spatially highly correlated. 
In order to assess the importance of atmospheric distortions, we adopted the data fusion procedure for atmospheric 
effect compensation described in (Crosetto and Pérez, 1999). We used as low resolution auxiliary data two coarse 
resolution grids (250 m spacing) of the ORFEAS data set generated using stereoscopic techniques (i.e. they are not 
affected by atmospheric effects). The first grid was interpolated from a 90 m DEM derived through a radargrammetric 
procedure implemented at ICC (Crosetto and Pérez, 1999) processing a pair of Radarsat images. The 90 m DEM has 
RMS error of 26.5 m (i.e. it is less precise and less dense than the InSAR DEM), but the errors are evenly distributed in 
the entire scene, i.e. they do not show systematic trends. This characteristic, confirmed by the correlation length of the 
height differences of about 40 m, is very important for the purpose of the data fusion procedure. The second grid used 
as auxiliary data is a DEM derived from optical images with 250 m spacing and RMS error of 23.1 m. 
The two coarse resolution grids were fused separately with the ascending InSAR grid, obtaining two new DEMs (see 
the corresponding statistics in Table 2). Most of the systematic effects on the original InSAR DEM were properly 
removed through the data fusion. In both cases there is an important improvement of the DEM precision (the global 
standard deviation drops from 18.1 m to 13.8 and 11.4 m for radargrammetry and optical data respectively). The 
correlation length of the height differences (see Figure 3) is 55 m and 120 m for optical and radargrammetry data 
respectively. These values confirm the effectiveness of the artefact correction. In fact, the errors of the new DEM are 
almost spatially decorrelated because the systematic errors caused by atmospheric heterogeneity were properly 
removed. The atmospheric distortion compensation with optical data gives the best results due to the more 
homogeneous quality of the optical grid in the flat and mountainous areas. 
2.2 Descending Image Pair 
A pair of descending ERS-1 images covering the same area analysed in previous sections was processed (see Table 5). 
The low coherence of the filtered images (the mean over the entire scene is 0.41), due to the long time interval of the 
interferogram (35 days), made the phase unwrapping very difficult. Even with an image compression of two times in 
range and eight times in azimuth (pixel footprint size of about 32 by 40 m, while for the ascending images it was 16 by 
20 m), only one third of the processed scene was correctly unwrapped. For the INSAR geometry refinement we adopted 
the joint calibration proposed in (Crosetto, 2000). 
  
50 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 
ft^ 
€) €. 5 09 "e Qu, "P