A GLOBAL TOPOGRAPHIC NORMALISATION ALGORITHM FOR SATELLITE IMAGES
Josef Jansa, Vienna University of Technology, Austria
email: Josef.Jansa @tuwien.ac.at
Commission VII, Working Group 1
KEYWORDS: Topographic Normalisation, Radiometric Correction, Minnaert Model
ABSTRACT:
Topographic normalisation, ie the elimination of shading due to sun illumination and terrain elevation is regarded as a useful
pre-processing step to image analysis, in particular to multispectral classification. Though a strict solution has to concentrate
on the individual object classes because of their class-related directional reflectances, we suggest to perform an improved
global topographic correction of the whole image scene in order, for instance, to facilitate the pre-classification that may be
input to a more refined class-specific normalisation process. The model presented here is based on Minnaert's theory, as
many others do, but extended by a simple skylight term. The algorithm has been developed for highly automatic
determination of the model's parameter by a least squares approach assuming a similar distribution of reflectance intensities
within all categories of incident lighting angles. Though rather simple this augmentation is of particular importance for short
wave channels where atmospheric scattering predominantly influences the image contrast und brightness and where models
without skylight term fail. Examples applied to LANDSAT TM pictures of mountainous, heavily shaded areas of the Alps
prove the suitability of the approach.
KURZFASSUNG:
Die topographische Normalisierung, d.h. die Beseitigung von Schatteneffekten, die durch Sonnenbeleuchtung und
Gelàndeform hervorgerufen werden, wird als sehr nüztlicher Vorverarbeitungsschritt für die Bildanalyse, im besonderen für
die multispektrale Klassifizierung, angesehen. Obwohl man eine strenge Lósung auf die einzelnen Objektklassen ins Auge
fassen müßte, da die Reflexionseigenschaften richtungsabhängig und klassenspezifisch sind, wird in dieser Arbeit
vorgeschlagen, einen verbesserten Korrekturansatz über das Gesamtbild anzuwenden, um zum Beispiel eine
Vorklassifzierung zu erleichtern, die wiederum vor einem verfeinerten, klassenspezifischen Normalisierungsprozef?
verwendet werden kónnte. Dem vorgestellten Modell liegt, wie bei vielen anderen bekannten Ansátzen, die Minnaert'sche
Theorie zugrunde, wobei jedoch eine einfache Erweiterung durch einen Himmelslichtterm erfolgt. Der Algorithmus wurde so
entwickelt, da& er unter Verwendung eines Kleinste-Quadrate-Ausgleiches móglichst automatisch ablaufen kann.
Angenommen wird lediglich, daB die Helligkeitswerte im Bild der Einfallswinkelkategorien etwa áhnlich verteilt sind. Obwohl
diese Erweiterung sehr einfach gehalten ist, wird sie besonders in kurzwelligen Spektralbereichen wichtig, wo
atmosphárische Streuung vorwiegend wirksam wird und Bildhelligkeit und Kontrast beeinflußt, und wo daher Modelle ohne
Himmelslichtberücksichtigung versagen. Beispiele von korrigierten LANDSAT TM Bildern aus dem alpinen und stark von
Schatten betroffenen Bereich zeigen die Brauchbarkeit des vorgestellten Ansatzes.
1. INTRODUCTORY REMARKS The predominant obstacle for an accurate solution lies in
the undefined or at least not well defined reflection
The interpretation of satellite imagery is quite often heavily
influenced by shading effects caused by the terrain relief. In
case of supervised classification, for instance, the image
analysts are forced to select several training areas for the
same object class in order to take into consideration the
various spectral signatures caused by the different
illumination. Theoretically one training sample per class for
each slope category are to be found. Besides the
impossibility of choosing that many samples, the interpreter
faces the problem of being unable to find the correct areas
even visually. The ideal image for interpretation would be
one of a surface diffusely and uniformly illuminated, without
any atmospheric influence, observed perpendicularly to a
horizontal reference plane and indenpent of the terrain
slope. The aim of topographic normalisation is to create that
sort of ideal image out of the actual given image.
This problem brought up the idea to develop a method for a
radiometric correction as a function of the sun incidence
angle to the terrain surface as well as of the observation
angle. Many publications are the result of those
investigations and research [e.g: Colby 1991, Meyer et al.
1993, Conese et al. 1993, Ekstrand 1996, Sandmeier 1997]
behaviour of the object classes. The theortically ideal
reflection property is described by the reflectance p. The
actual reflectance is dependent on the wavelength (i.e. o
becomes p(A)) and more severely, also on the lighting and
observation direction. The reflectance properties are
therefore best described by the so-called bidirectional
reflectance distribution function (BRDF) p(A, ©,, ,, 6,, @),
where the indeces ; and e denote the incident and exitant
ray, respectively, and (6, ¢) the incidence angle and the
incidence azimuth. As the BRDF is usually unkown or
hardly determinable in practice the directional reflectance is
more feasible o,(A, 9,, «,) [Kraus et al., 1988]. In any case
all reflectance functions are class dependent. That means,
if applied for performing radiometric corrections, the object
classes have to be known in advance as they are input to
the correction algorithm. On the other hand the radiometric
correction should help to facilitate the determination of the
object classes. This basic contradiction demonstrates that
the task of radiometric correction due to illumination effects
is not trivial at all. Even the most sophisticated model will
not be able to perform the procedure in one single step, it
will always be an iterative approach.
8 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
In practice
able to
important
approxim
actual ref
imagery t:
not able
Minnaert |
the shad
slightly r
Lamberti:
where g
corrected
constant,
incidence
exitance
assuming
direction
area and
the comn
TM, SPC
position i
as well a
model. C
correspor
Though :
common!
on the ot
clearly se
and phy:
satisfying
Too man)
such as r
like skylig!
surface
character
values wi
e
C
=
In order t
though sti
Minnaert
from the «
model, a
the first
fulfilled in
Firstly, a f
Minnaert
a. Satell
chanr
regior
than 9
not th
b. Theal
(in otl
