TOWARDS AUTOMATED DEM GENERATION FROM HIGH RESOLUTION STEREO 
SATELLITE IMAGES 
Pablo d'Angelo, Manfred Lehner, Thomas Krauss, Danielle Hoja and Peter Reinartz 
German Aerospace Center (DLR), Remote Sensing Technology Institute, D-82234 Wessling, Germany- 
(Pablo.Angelo, Manfred.Lehner, Thomas.Krauss, Danielle.Hoja, Peter.Reinartz)@dlr.de 
Commission IV, WG IV/9 
KEY WORDS: Spacebome Scanner Systems, Digital Elevation Models (DEM), Image Matching, CARTOSAT-1, Orthoimage, 
Accuracy Analysis 
ABSTRACT: 
High resolution stereo satellite imagery is well suited for the creation of digital surface models (DSM). In this paper, a system for 
highly automated DSM and orthoimage generation based on CARTOSAT-1 imagery is presented. The proposed system processes 
photometrically corrected level-1 stereo scenes using the rational polynomial coefficients (RPC) universal sensor model. The RPC 
are derived from orbit and attitude information and have a much lower accuracy than the ground resolution of approximately 2.5 m. 
Ground control points are used to estimate affine RPC correction. Accurate GCP are not always available, especially for remote 
areas and large scale reconstruction. In this paper, GCP are automatically derived from lower resolution reference images (Landsat 
ETM+ Geocover and SRTM DSM). It is worthwhile to note that SRTM has a much higher lateral accuracy than the Landsat ETM+ 
mosaic, which limits the accuracy of both DSM and orthorectified images. Thus, affine RPC correction parameters are estimated by 
aligning a preliminary DSM to the SRTM DSM, resulting in significantly improved geolocation of both DSM and orthoimages. 
Robust stereo matching and outlier removal techniques and prior information such as cloud masks are used during this process. DSM 
with a grid spacing of 10 m are generated for 9 CARTOSAT-1 scenes in Catalonia. Checks against independent ground truth 
indicate a lateral error of 3-4 meters and a height accuracy of 4-5 meters. Independently processed scenes align at subpixel level and 
are well suited for mosaicing. 
1. INTRODUCTION 
In May 2005 India launched its IRS-P5 satellite with 
CARTOSAT-1 instrument which is a dual-optics 2-line along- 
track stereoscopic pushbroom scanner with a stereo angle of 
31° and the very interesting resolution of 2.5 m. The 
operational use of the data is described in (Srivastava et al, 
2007). The CARTOSAT-1 high resolution stereo satellite 
imagery is well suited for the creation of digital surface models 
(DSM). In this paper, a system for highly automated DSM 
generation based on CARTOSAT-1 stereo scenes is presented. 
CARTOSAT-1 stereo scenes are provided with rational 
polynomial functions (RPC) sensor model, derived from orbit 
and attitude information. The RPC have a much lower accuracy 
than the ground resolution of approximately 2.5 m. 
Traditionally, subpixel accurate ground control points (GCP) 
are used in previous studies to estimate bias or affine RPC 
correction parameters required for high quality geolocation of 
HRSI images. Such highly accurate GCP are usually derived 
from a DGPS ground survey or high resolution orthoimages and 
digital elevation models. For many applications, especially ones 
demanding near real-time results, such as desaster assessment 
tasks in remote regions, highly accurate GCP data is often not 
available. Without accurate GCPs, CARTOSAT-1 scenes are 
of limited use, since bias or affine RPC correction is required 
before CARTOSAT-1 scenes can be used for DSM extraction 
and orthorectification (Lehner et al, 2007). 
models (DSM) are derived from dense stereo matching and 
forward intersection and subsequent interpolation into a regular 
grid. Since stereo matching is unreliable in large, homogeneous 
image areas, such as fields, meadows and water bodies, as well 
as in complicated terrain with occlusions and shadows, strict 
consistency checks are used during matching. The first section 
of the paper describes the process used for DSM generation. 
The second part evaluates the processor using 9 CARTOSAT-1 
stereo pairs. 
2. DSM GENERATION 
The DSM generation process consists of the following main 
steps, implemented as part of the DLR XDibias image 
processing system. 
1. Stereo matching in epipolar geometry 
2. Affine RPC correction and alignment to reference 
DEM 
3. Forward intersection and outlier removal 
4. Interpolation 
5. Orthorectification 
6. Quality inspection and manual editing 
A CARTOSAT-1 stereo scene consists of a nadir looking image 
with a along track tilt of -1°, a forward looking image with a 
along track tilt of 26°. They are named Aft and Fore throughout 
this paper. 
We propose the use of widely available lower resolution 
satellite data, such as the Landsat ETM+ and SRTM DSM 
datasets as a reference for RPC correction. Digital surface 
1137