-
= WV
Tell ^^ Ww we
IESUS uA
SPATIAL DISTRIBUTION OF CONTROL POINTS
Ricardo Vergara
Alejandro C. Frery
Julio d'Alge
INPE - DPI
Av. dos Astronautas, 1758
12201 - Sao Jose dos Campos, SP, BRAZIL
ISPRS - COMISSION III
ABSTRACT:
We present some Monte Carlo results about the
spatial distribution of the
Control Points used in the
geocodification of satellite images in Brazilian Amazonas.
This information aims at stablishing criteria for the choice of number, quality and spatial distribution of
Control Points (CPs) to be used in image matching techiques.
Some theory about the statistical analysis
of spatial point patterns is recalled, as an aid to a future
formulation of this problem in terms of an experiments design problem.
KEY WORDS: Accuracy, Algorithm, Image Matching, Image Processing, Registration, Simulation, Theory.
1. INTRODUCTION
The geometric accuracy of remote sensing satellite
images plays a central role in most applications
(as in Cartography, to name one), so these images
are geometrically corrected before being used.
Moreover, when digital images are used, their
geometric accuracy is improved by the registration
with respect to a map.
In this image processing techique, clearly
identifiable points (the "Control Points", or CPs)
are sought in both the image and the map. Taking
the second as a reference, and using the CPs, it is
possible to estimate the -mathematical- mapping
that corrects the digital image. Usually, a
polinomial transformation is used, and the degree
of the polinomial determines the minimum number of
CPs needed to estimate its coefficients
(Mascarenhas & Velasco, 1989).
The registration accuracy between an image and a
map depends on the location accuracy and on the
number and the spatial distribution of the CPs used
to perform it (Ford & Zanelli, 1985). The same
happens with the attainable precision in the
evaluation of the geometric accuracy of an image.
Suppose that is possible to have CPs with
arbitrarily high location accuracy, either to
evaluate the geometric precision of an image or to
accomplish the registration between an image and a
map. Precision will be highly dependent on the
number and spatial distribution of the CPs.
Concerning to the number, it is possible to say
that the higher the quantity of CPs, the higher the
achieved precision (Ford & Zanelli, 1985... Orti,
1981).
Relating to the spatial distribution, a
specification of the adopted criterion has not been
detailed reported in the literature, but in general
is considered that a uniform distribution of -
evenly spaced- CPs is the most appropriate choice
(Ford & Zanelli, 1985). However, it usually is
difficult to find CPs distributed in this way;
sometimes the CPs are grouped in a small part of
109
the area under study or a very small number of
dependable CPs is available. Obviously, all CPs
must be used when the last situation occurs; in
this case it is not possible to make a choice among
them.
the case of having plenty of CPs of
good location accuracy. In this situation the
following question appears: Which should be
selected in order to obtain a spatial distribution
that makes possible the achievement of an accurate
geometric correction? Or, if the registration has
already been made with no good accuracy: How to
know if that is due to a inadequate spatial
distribution of the CPs used to perform it? And
when this is the reason: Is there some manner to
detect those CPs whose locations damage the quality
of the geometric transformation? Note that the same
questions are valid when the evaluation of the
geometric quality of an image has to be assessed.
But suppose
The aim of this paper is proposing some partial
answers to these questions, providing a tool for
either doing an adequate selection -concerning to
the spatial distribution- of CPs to be used for
performing an image-map registration or for
evaluating the geometric accuracy of an image. This
procedure allows the identification of CPs whose
locations damage the quality of the spatial
distribution of the set of points; so, their
substitution for other CPs located in more
appropriate places is made possible. The proposed
metodology consists in selecting CPs whose plane
coordinates are known with precision, and then in
submitting these points to one or more spatial
distribution tests, For the evaluation of this
procedure, different registered images of brazilian
legal Amazonas were taken and their respective sets
of CPs were submitted to the tests above, The
amazonian region was chosen for this study because
it has special caracteristics that make hard the
finding of CPs spatially well distributed. To
compare results, the same kind of analysis was
carried out using a registered image of the
Buritama region, Sao Paulo State, where the
relative abundance of CPs allowed the selection of
those whose spatial distribution is nearer to the
optimal model.