FULLY AUTOMATED GENERATION OF ACCURATE DIGITAL SURFACE MODELS
WITH SUB-METER RESOLUTION FROM SATELLITE IMAGERY
J. Wohlfeil ', H. Hirschmüller ? B. Piltz ! A. Börner | M. Suppa 2
German Aerospace Center (DLR), Institute of Robotics and Mechatronics,
| Dept. of Data Processing for Optical Systems, Rutherfordstr. 2, 12489 Berlin, Germany
? Dept. of Perception and Cognition, Münchner Str. 20, 82234 Wessling, Germany
(Juergen.wohlfeil, anko.boerner, heiko.hirschmueller, michael. suppa)@dlr.de
Commission III, WG III/1
KEY WORDS: Camera Orientation, Digital Surface Model, Satellite Images, Bundle Adjustment
ABSTRACT:
Modem pixel-wise image matching algorithms like Semi-Global Matching (SGM) are able to compute high resolution digital surface
models from airborne and spaceborne stereo imagery. Although image matching itself can be performed automatically, there are
prerequisites, like high geometric accuracy, which are essential for ensuring the high quality of resulting surface models. Especially
for line cameras, these prerequisites currently require laborious manual interaction using standard tools, which is a growing problem
due to continually increasing demand for such surface models. The tedious work includes partly or fully manual selection of tie-
and/or ground control points for ensuring the required accuracy of the relative orientation of images for stereo matching. It also
includes masking of large water areas that seriously reduce the quality of the results. Furthermore, a good estimate of the depth range
is required, since accurate estimates can seriously reduce the processing time for stereo matching. In this paper an approach is
presented that allows performing all these steps fully automated. It includes very robust and precise tie point selection, enabling the
accurate calculation of the images' relative orientation via bundle adjustment. It is also shown how water masking and elevation
range estimation can be performed automatically on the base of freely available SRTM data. Extensive tests with a large number of
different satellite images from QuickBird and WorldView are presented as proof of the robustness and reliability of the proposed
method.
1. INTRODUCTION While the computation time was drastically reduced, the time
consumption for these laborious steps became the critical part
Data from imaging remote sensing systems is a primary of the whole procedure. Therefore, during the last years,
source for a huge variety of geo-spatial products and services. many different components have been developed to solve this
New technologies permit the generation of high resolution problem (Hirschmüller, 2005 and 2008; Wohlfeil, 2010 and
digital surface models (DSM). These products allow the 2012). Their combination leads to a very operational solution
generation of highly accurate orthophotos. In combination, for highly automated generation of high resolution digital
these two types of products provide a standardized three- surface models with high throughput. As most of the current
dimensional spatial reference for each pixel including all imaging sensors with spatial high resolution are line scanners,
spectral information (e.g. red, green, blue, height). This using the pushbroom principle (e.g. WorldView 1/2, GeoEye,
information is a prerequisite for most geo-spatial data QuickBird, Pleiades), it is focused on this type of sensors.
products and services. Especially if their spatial resolution is
in the sub-meter range, they offer a huge number of
applications in the fields of change detection, urban noise
Manual preparation
modeling, radio propagation, flooding simulation, opencast | Selection of image data |
mining, etc. Moreover, the new and quickly growing market v
of public on-line geo-information services (e.g. Google Earth, Automatic processing
Bing Maps 3D, etc.) currently creates a very strong demand
on these products in order to provide detailed landscape and
city models.
Height Range Determination
Homologous Points Selection
| |
| |
For the creation of such high resolution DSMs, Semi-Global | Bundle Adjustment |
Matching (SGM) (Hirschmüller, 2005 and 2008) turned out
to achieve better results in many cases than other stereo | Water Masking |
matching methods (Hirschmüller and Bucher, 2010) and | SGM Image Matching |
other technologies, e.g. LIDAR (Gehrke, 2010) or Radar. [ DSM and orthophoto generation |
Moreover, stereo matching can be performed much more
economically than its technological alternatives, as it does not Figure 1: Overview of the processing steps
require additional sensors. The images used for matching are
required anyway in most cases. Despite of its relatively high Since dense matching is typically slow, the search range
computational complexity, the computation time can be should be limited. The search range in the images can be
handled very well by parallelization and/or optimization for computed from the height range of the terrain. An automatic
special hardware like graphic cards (Ernst and Hirschmüller, solution of this problem is presented in Section 3.6.
2008) and FPGAs (Gehrig et al., 2009; Hirschmüller, 2011). For computing geometrical reconstructions from images, the
But in order to achieve optimal results there are some prepro- intrinsic and extrinsic geometry of cameras must be known.
cessing steps, needed to meet the algorithm's prerequisites. The intrinsic camera geometry (interior orientation) is
74
typi
the
con:
line.
geoi
if pe
poir
orie
relat
poir
soft
ani
this
If th
met,
also
mea
Ano
ima;
ine
shoi
inte
An:
only
suit:
exte
folk
For
are
are
groi
If it
proc
of t
use]
The
imp
ous
foui
the
bet
to u
ang
occ]
the
this
toa
3.1
One
occ
mat
the
mat
is St
