The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
h D {x i ,y i ) = z i (1) 
A 3D affine transformation is used to align the initial stereo 
point cloud to the SRTM DSM: 
Pa = A Pi 
(2) 
where 
p j =(x i y ; Z- l) r is the original point, 
A is a 3x4 matrix, 
Pu ~ ( x „ y,i z „ Y ' s transformed point. 
The affine transformation matrix A is estimated using an 
iterative least mean squares algorithm. Using Eq. (1) and (2), 
the following observation equation is obtained. 
v i = h D (x li ,y li )-z i 
(3) 
Since the model is non-linear, the solution is obtained 
iteratively. An identity transform is used as initial 
approximation, since the stereo points are not far from the 
reference. It is likely that the stereo point cloud, and to a 
smaller extend the DSM contains outliers, which cannot be 
handled by a standard least mean squares algorithm. After the 
initial estimation, points with a residual larger than 3 times the 
standard deviation are removed and a new transformation is 
estimated. This procedure is repeated until less than 0.3% 
outliers are detected and the squared sum of the outlier 
residuals accounts for less than 5% of the squared sum of all 
residuals. 
The estimated affine transformation could be used to align the 
final DSM to the SRTM reference and thus improve its 
accuracy. Orthoimages would however still be limited by the 
ETM+ accuracy. It is desirable to include the correction in the 
RPC models, too. This is done by aligning the 3D stereo points 
to SRTM and using them as GCP for a second RPC correction 
which yields the final affine RPC correction used in all 
subsequent steps. 
2.3 Forward intersection and outlier removal 
Forward intersection is done via iterative least squares 
adjustment using 4 observation equations and derives object 
space coordinates in Geographic coordinates in WGS84 datum. 
(Grodecki et al, 2004, Lehner et al, 2007). The residuals in 
image space are used for a further blunder reduction step. 
Points with a residual larger than 0.5 pixels are rejected. Of 
course, only residuals in cross track direction will be effective 
because wrong row coordinates of tie points are translated into 
wrong height values if only two stereo partners are available 
(stereo imaging direction). 
The forward intersected points still contain a small amount of 
blunders due to matching errors in regions with sparse texture. 
To eliminate gross outliers, a reference check against the 
SRTM DSM is performed. All points whose height deviates 
more than 3 times the height error of the SRTM are rejected. 
The SRTM height error map was found to be a good 
approximation of the true height error (Rodriguez et al. 2005), 
and is used to dynamically adjust the height difference 
threshold. Typical thresholds are 24 m in flat areas, and 75 m in 
mountainous areas. 
2.4 DSM interpolation 
Result of matching and forward intersection is a set of 3D 
points representing the Earth surface (including f.e.. tree tops) 
acquired by the stereo images. To ease further applications, the 
irregular point cloud is transferred to a regularly spaced grid 
with a spacing of 10 m. If multiple points fall into the same grid 
cell, their heights are averaged to form a new point. The points 
are connected by Delauney triangulation into a triangulated 
irregular network (TIN). Finally, the triangles are superimposed 
on the regularly spaced grid of the resulting DSM. For each 
triangle the plane defined by the three vertices is calculated. To 
each pixel inside the triangle the height value interpolated on 
this plane is assigned (Hoja et al., 2005). 
2.5 Orthorectification 
Orthoimages with user defined datum and projection are 
created by orthorectification of the Aft image with the 
generated DSM and the affine corrected RPC. 
3. EVALUATION 
The DSM creation process described above is evaluated with 9 
CARTOSAT-1 scenes of Catalonia. Scene Cat was part of the 
Cartosat scientific assessment programme, while the remaining 
8 scenes have been provided by Euromap. The Landsat ETM+ 
Geocover mosaic and the SRTM C band DSM have been used 
as sources for GCP collection. Two reference DTM with a GSD 
of 15 m and 10 orthoimages with a resolution of 0.5 m have 
been provided by the Institut Cartographic de Catalunya (ICC) 
and are only used as ground truth during the evaluation. The 
location of the scenes and ground truth data is shown in Figure 
1. The scenes are mostly cloudless. Scene 117/207 contains two 
large clouds in the upper left comer, covering most of the 
overlap between with scene 116/207. The scenes located near 
the coastline contain both flat areas along the coast as well as 
the Montseny mountain range with peaks of over 1600 meters. 
As shown in Table 1, the scenes were aquired early in the year, 
leading to large shadows in the mountainous areas. 
Abbreviation for 
the paper for aft 
and fore scenes 
Imaging date 
Cat-A/F 
01 Feb. 2006 
115/207-A/F 
16 Feb. 2008 
115/208-A/F 
16 Feb. 2008 
116/207-A/F 
05 March 2008 
116/208-A/F 
05 March 2008 
116/209-A/F 
05 March 2008 
117/207-A/F 
25 Jan. 2008 
117/208-A/F 
25 Jan. 2008 
117/209-A/F 
25 Jan. 2008