|
ABSTRACT
KURZFASSUNG
(Form-)Charakteristiken definiert.
i 1 BASIC IDEA
The following experiences and ideas are the starting point of
the project:
e The human interpreter, analysing remotely sensed im-
ages visually or interactively on an image processing
system, uses knowledge about the physical mechanism
of the image generation process. The main advantage
of this is the fact that scene-independent knowledge
(e.g. on spectral reflectance characteristics of surfaces)
can be used in the interpretation process, and that dis-
turbing factors (e.g. atmospheric influences) can be
accounted for.
e |n automatic analysis, this knowledge up to now is
used, if at all, in a very coarse and implicit way only.
For example, land use classification of an optical satel-
lite image may be based on the vegetation index being
defined as the ratio of an infrared and a red channel.
In this case, one makes use of 2 pieces of knowledge:
(a) that differences in the terrain vegetation cover can
be recognized in terms of differences of the ratio of
infrared and red reflectance, and (b) that disturbing
multiplicative influences from the atmosphere and from
the illumination on uneven terrain cancel out by taking
the ratio.
Fully automatic analysis of remotely sensed image data
is highly desirable because of the large earth obser-
vation data volumes, the high expenditure in visual
*' This work is financed by the Austrian "Fonds zur Fórderung der wis-
senschaftlichen Forschung" (project S7003).
REMOTE SENSING IMAGE UNDERSTANDING BASED ON PHYSICAL MODEL INVERSION *
Werner Schneider
Institute for Surveying and Remote Sensing
Universitat für Bodenkultur, Wien
(University of Agriculture, Forestry and Renewable Natural Resources, Vienna)
Austria
e-mail: schneiwe@mail.boku.ac.at
764
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Commission lll, Working Group 3
KEY WORDS: Classification, Automation, Vision, Satellite, Image Understanding, Radiometric Model, Land Use Mapping
The application of computer vision methods for the analysis of remotely sensed images is studied within the framework of
the Research Programme "Theory and Applications of Image Processing and Pattern Recognition" of the Austrian Science
Foundation (“Fonds zur Förderung der wissenschaftlichen Forschung"). This contribution gives the system overview of a pro-
posed image understanding system based on the inversion of a physical (radiometric) model of image acquisition. The physical
model, sufficiently simplified for practical implementation, is formulated and discussed in detail. The image understanding
system is devised to perform automatic land use mapping from optical satellite images. The land use categories are defined in
terms of their spectral reflectance on the ground and geometric (shape) characteristics.
Die Anwendung von Computer-Vision-Methoden für die Auswertung von Fernerkundungsbildern wird im Rahmen des
Forschungsschwerpunkts "Theory and Applications of Image Processing and Pattern Recognition” des österreichischen Fonds
zur Förderung der wissenschaftlichen Forschung untersucht. Dieser Beitrag gibt den Systemüberblick über ein vorgeschla-
genes System zum automatischen Bildverstehen, das auf der Inversion eines physikalischen (radiometrischen) Modells der
Bildaufnahme beruht. Das physikalische Modell, das für die praktische Ausführung hinreichend vereinfacht ist, wird formuliert
und im Detail besprochen. Das Bildanalysesystem ist für die automatische Landnutzungskartierung aus optischen Satel-
litenbildern konzipiert. Die Landnutzungskategorien sind durch ihre spektrale Reflexion im Gelände und durch geometrische
or semiautomatic, interactive interpretation and the
shortage of expert interpreters.
The idea suggests itself to formalize the expert knowl-
edge about the physical mechanism of remote sensing
image acquisition and to use this knowledge in an au-
tomatic analysis procedure. A model of image acquisi-
tion transforms a scene in the real world (more exactly:
a description of a scene) into an image. Image analy-
sis is nothing else than the reversal of this process: In
image analysis, a scene description is derived from an
image. The basic idea of this project therefore is to
analyse images by inverting a physical model of image
acquisition.
As a byproduct of image analysis, a physical model con-
taining a quantitative description of the radiometry of
image acquisition can provide a radiometric calibration
of the images, i.e. the exact transformation parame-
ters between pixel values in the images and reflectance
values on the ground.
2 SYSTEM OVERVIEW
A description of a remotely sensed scene is a thematic map
consisting of cartographic objects such as regions, line ob-
jects, point objects etc. In the project reported here, the
attention is restricted to regions as the most frequent ob-
jects. The task of image analysis is simplified considerably
if regions can be identified in the image in a segmentation
process before the physical model of image acquisition is ap-
plied. The overall information flow in an analysis system
with a separated segmentation process is illustrated in Fig. 1
^
/
1 ie
Figu
ing
phys
[Sch
Star
NS
The
flect
rain
imag
ters
cess
para
The
obje
face
valu
cert:
ters
to tl
of tl
gic
geor
imag
mul: