IMPROVEMENT OF THE STABILITY SOLVING RATIONAL POLYNOMIAL COEFFICIENTS

Abstract

The rational function model (RFM) utilized for high resolution satellit e imagery (HRSI) provides a transformation from image to object space coordinates in a geographic reference system. Compared with the rigorous model based on the collinearity condition equation or the affine model, the RFM with 80 coefficients would be over parameterized. That would result in an ill-conditioned normal equation. Tikhonov regularization is often used to resolve this problem, and many applications have verified its serviceability. This paper will detail the method for regularization parameter selection. However, Tikhonov regularization makes the two sides of equation unequal, resulting in a biased solution. An unbiased method - The Iteration by Correcting Characteristic Value (ICCV) was introduced, and a strategy to resolve the ill-conditioned problem for solving rational polynomial coefficients (RPCs) was discussed in this paper. The tests with SPOT-5 and QuickBird imagery were accomplished. The empirical results have shown that our methodology can effectively improve the condition of the normal equations.

Full text

IMPROVEMENT OF THE STABILITY SOLVING RATIONAL POLYNOMIAL
COEFFICIENTS
Xianyong Lin a,*, Xiuxiao Yuan a
a School of Remote Sensing and Information Engineering, Wuhan University, 129 Luoyu Road, Wuhan 430079, China
winterlinny@163.com
Commission I, WG I/5
KEY WORDS: High Resolution Satellite Imagery; Sensor Model; Orientation; Regularization; Accuracy
ABSTRACT:
The rational function model (RFM) utilized for high resolution satellite imagery (HRSI) provides a transformation from image to
object space coordinates in a geographic reference system. Compared with the rigorous model based on the collinearity condition
equation or the affine model, the RFM with 80 coefficients would be over parameterized. That would result in an ill-conditioned
normal equation. Tikhonov regularization is often used to resolve this problem, and many applications have verified its serviceability.
This paper will detail the method for regularization parameter selection. However, Tikhonov regularization makes the two sides of
equation unequal, resulting in a biased solution. An unbiased method - The Iteration by Correcting Characteristic Value (ICCV) was
introduced, and a strategy to resolve the ill-conditioned problem for solving rational polynomial coefficients (RPCs) was discussed in
this paper. The tests with SPOT-5 and QuickBird imagery were accomplished. The empirical results have shown that our
methodology can effectively improve the condition of the normal equations.
1. INTRODUCTION
Since the launch of the IKONOS II satellite, the rational
function model (RFM) has gained considerable interests in
photogrammetric community. SpaceImaging Company provides
the RFM to users instead of the physical sensor model,
subsequently, DigitalGlobe Corporation provides the RFM
together with the strict geometric model in order to satisfy
different users. The RFM has been universally accepted, and
validated, as an alternative sensor orientation model for high
resolution satellite imagery (HRSI). The RFM is an
approximation of the rigorous sensor model, via a number of
control points. Then it could be utilized in the photogrammetric
process instead of the complex rigorous sensor model. It would
be a part of the standard image transfer format, and it is
becoming a standard way for economical and fast mapping from
remotely sensed imagery.
The key of the RFM is to gain accurate rational polynomial
coefficients (RPCs). Compared with the rigorous model based
on the collinearity model or the affine model, the RFM with 80
coefficients would be over parameterized (Fraser et al., 2005).
That may cause the design matrix to become almost rank
deficient because of the complex correlation among RPCs. It
may result in numerical instability in the least squares
adjustment, or even producing wrong solutions. The
regularization technique was often suggested to tackle the
possible ill-conditioned problem during the adjustment (Tao and
Hu, 2001a). It has been proved to effectively improve the
condition of the normal equations. But the determination of the
regulation parameter has still been considered to be a challenge.
Regularization parameter selection is crucial to the
regularization technique. There are several methods for the
optimal parameter determination, including ridge trace method,
L-curve criterion, generalized cross validation (GCV) method,
ordinary cross validation (OCV) method, and so on. The effects
may be totally different when we use different methods. The
L-curve, GCV and OCV were compared by Choi et al. (2007).
In practice, ridge trace method is widely used for its simpleness,
where solutions are computed for a large number of different
regulation parameters, selecting the best one by suitable
heuristics (Tao and Hu, 2001a). However, ridge trace method
can not obtain the optimal parameter, and it is inconvenient for
automatic computation. The authors try different regulations for
the RPCs computation. The L-curve criterion has been proved
to be efficacious.
Despite regularization technique gives a good result, it makes
the two sides of equation unequal by imposing constraints to the
diagonal elements of the normal equation matrix, resulting in a
biased solution. So, we will introduce an unbiased method, the
Iteration by Correcting Characteristic Value (ICCV). This
method is simple, and it was put forward more than ten years
ago, but still not widely used. The initial values will be the main
factor that affects the result. And this paper will suggest two
ways to set initial values, just for RPCs computation.
Accurate RPCs are crucial to the RFM model, which directly
determines whether it could replace the physical sensor model
to accomplish the photogrammetric process. And the
ill-conditioned normal equation would be the main problem.
This paper aimed at finding a proper method to resolve the
possible ill-conditioned problem, getting an accurate solution.
Initially, we will review the basic model and the methods
including the RFM solution, the terrain-independent and
terrain-dependent computational scenarios. The regulation
technique and the method of regularization parameter selection
are then addressed, focusing on the L-curve method. The
unbiased ICCV is followed. The results of experimental tests
with SPOT-5 and QuickBird imagery are then discussed.
Finally, we will suggest a strategy for RFM computation
according to the experiments and comprehensive analysis of the
characteristics of the various methods.
711

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
2.2 2. THE RATIONAL FUNCTION MODEL
The RFM Model 2.1
The RFM relates object point coordinates to image pixel
coordinates in the form of rational functions that are ratios of
polynomials. For the ground-to-image transformation, the
defined ratios have the forward form (OGC, 1999):
),,(
),,(
),,(
),,(
WVUDen
WVUNum
s
WVUDen
WVUNum
l
S
S
n
L
L
n
=
= (1)
Where
3
20
2
1921
3
20
2
1921
3
20
2
1921
3
20
2
19
2
18
2
17
3
16
2
15
2
14
2
13
3
1211
2
10
2
9
2
876
54321
),,(
),,(
),,(
),,(
WdWUdVddWVUDen
WcWUcVccWVUNum
WbWUbVbbWVUDen
WaWUa
WVaUWaUa
UVaVWaVUa
VaUVWaWa
UaVaUWaVWa
VUaWaUaVaaWVUNum
S
S
L
L
++++=
++++=
++++=
++
+++
+++
+++
++++
++
+
+=
K
K
K
(2)
Here, and are the 80 RPCs; and are commonly
set to 1. are the normalized line and sample index of
the pixels in image space, while are normalized object
point coordinates. That is:
iii cba ,, i
d1
b1
d
),( nn sl
),,( WVU
eHeightScal
etHeightOffsh
W
caleLongitudeS
ffsetLongitudeO
V
aleLatitudeSc
fsetLatitudeOf
U
eSampleScal
etSampleOffss
s
LineScale
LineOffsetl
l
n
n
−
=
−
=
−
=
−
=
−
=
λ
φ
(3)
Here, are the image line and sample coordinates;
represent latitude, longitude, height; the offsets and scales
normalize the coordinates to [-1,1], minimizes the introduction
of errors during computation.
),( sl ),,( h
λφ
The RFM has nine configurations with some variations, such as
subset of polynomial coefficients, equal or unequal
denominators. Also, it has forward and backward forms (Tao
and Hu, 2001a). Generally speaking, the RFM refers to a
specific case that is in forward form, has third-order
polynomials with unequal denominators, and is usually solved
by the terrain-independent scenario.
RFM Solution
Two methods have been developed to solve for the RFM, direct
and iterative least-squares solutions (Tao and Hu, 2001a). Here,
the direct least-squares solution of RFM is given as follows:
0),,(),,(
0),,(),,(
=⋅−=
=⋅
−
=
WVUDenlWVUNumF
WVUDensWVUNumF
LLl
SSs (4)
PL,BXV
−
=
(5)
(6)
PLBPBBX T1T )( −
=
where
)20,,2;20,,1( LL ==
⎥
⎥
⎥
⎥
⎦
⎤
⎢
⎢
⎢
⎢
⎣
⎡
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
=
ji
d
F
c
F
b
F
a
F
d
F
c
F
b
F
a
F
j
l
i
l
j
l
i
l
j
s
i
s
j
s
i
s
B，
[
]
T
jiji dcba=X,
[
]
[
]
TT
11 lslbsd ==L.
P
is the weight matrix, and it is usually set as identity
matrix.
2.3
3.1
Approaches of Determining RPCs
There are terrain-independent scenario using known physical
sensor model and terrain-dependent scenarios using ground
control points. The terrain-independent scenario is to use the
onboard ephemeris and attitude data. With the physical model
available, a virtual control grid covering the full extent of the
image and the entire elevation is generated. The RPCs are
estimated using a least-square solution with the image grid
points and the corresponding object grid points (Tao and Hu,
2001a, 2001b).
For the terrain-dependent scenario, a number of ground control
points are collected for the RPCs computation. At least 39
ground control points are needed per image to solve 78 RPC
coefficients, excluding the constant parameters and . And
the solution is highly dependent on the actual terrain relief, the
distribution and the number of GCPs (Tao and Hu, 2001a,
2001b).
1
b1
d
3. REGULARIZATION TECHNIQUE
The RPCs may display very high correlation between
coefficients. That would be a potential problem for obtaining a
stable solution. The design matrix is usually ill conditioned in
the experiments (Tao and Hu, 2000). Even for the well
conditioned observation equations, regularization can improve
the accuracy of the RPCs, and help produce well-structured
RPCs, especially for the third-order RFM (Hu and Tao, 2004).
Ridge Regression
Ridge regression (Ridge estimate), a part of regularization
technique, is a biased estimation for nonorthogonal problems
(Hoerl and Kennard, 1970). It carries out by adding a small
positive quantity to the diagonal of
B
B
T. Ridge regression
712

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
obtains biased estimates with smaller mean square error. Ridge
regression is defined as follow:
PLBIPBBX T1T )()(
ˆ−
+= kk (7)
Where
k is ridge parameter or regularization parameter, usually a
small positive quantity;
I is identity matrix;
)(
ˆkX is ridge regression estimation.
3.2
3.3
Ridge trace method for parameter determination
Solutions are computed for a set of different k values. And the
best k is selected by suitable heuristics, for the least error at
check points (Tao and Hu, 1970). This method is very simple,
and it is widely used.
L-curve criterion for parameter determination
The L-curve is a log-log plot of the norm of a regularized
solution versus the norm of the corresponding residual (fitting
error) as the regularization parameter is varied (Hansen, 1992;
Rodriguez and Theis, 2005). L-curve is presented as:
)
ˆ
lg,
ˆ
(lg))(),(( kk XLXBkk −=
ξη
(8)
The curve is L-shaped: approximately vertical for small k, and
approximately horizontal for large k, with the corner providing
the optimal regularization parameter. So the object is to find out
the point with biggest curvature:
322'2'
''''''
)(
maxarg
ηξ
ηξηξ
+
−
=k (9)
Where
'
ξ
、= the first and second derivative of
''
ξ
ξ
on k;
'
η
、= the first and second derivative of
''
η
η
on k.
In practical computation, curve fitting is often used to obtain the
L-curve.
The L-curve criterion is able to recognize correlated errors,
while the GCV method may fail to do so. That is essentially
because the L-curve criterion combines information about the
residual norm with information about the solution norm,
whereas the GCV method only uses the information about the
residual norm. The research done by Choi et al (2007) shows us
that the L-curve method performed better than OCV or GCV,
particularly for high noise levels. The L-curve method is found
to be less susceptible to producing large reconstruction errors
but it tends to over-regularize the solution in the presence of
low noise, leading to under-estimates of the forces.
4. THE ITERATION BY CORRECTING
CHARACTERISTIC VALUE
Regularization technique imposes constraints to the diagonal
elements of the normal equation, resulting in a biased solution.
So, we will introduce an unbiased method - the Iteration by
Correcting Characteristic Value (Wang et al. 2001) for RPCs
computation.
Here is the norm function,
PLBXPBB TT ˆ
=
Add to both sides，
X
ˆ
XPLBXI)PBB ˆˆ
(TT +=+
There are on both sides, so it should be resolved in iterative
mode: X
ˆ
)
ˆ
()(
ˆ)1(T1T)( −− ++= kk XPLBIPBBX (13)
If we set ,
1T )−
+= IPB(Bq
Then the (13) could be written as:
)0(kT2)( ˆ
)(
ˆXqPLBqqqX ++++= kk L (14)
Where
)0(
ˆ
X= initial values of the solutions.
Eqs. (13) and (14) are the expressions of the iteration by
correcting characteristic value. The convergent and unbiased
properties are discussed by Wang Xinzhou et al. (2001).
The ICCV carried by iteration, initial values should be offered
for the iteration, and they have an important impact on the result
or even determine the success of the method. The direct
least-squares solutions are usually used as the initial values.
Unfortunately, when the ill-condition happens, it is possible that
the LS solution is so bad that the iteration is unconvergence. So,
here we suggest another way special for the RPCs solution.
Considering that the third-order RPCs are closed to zero, the
initial values may set to zero. We will test it in the experiments.
5. TEST RESULTS AND EVALUATION
5.1 Design and Tests
The tests have been designed for these purposes:
To evaluate the numerical stability of the direct least squares
solution. The condition number is cursorily employed to
measure the condition of the design matrix. The number is
much bigger when the function is ill-conditioned.
To compare the performances of the regularizations for the
RPCs computation. We choose the widely-used ridge trace
method and the L-curve criterion.
Mainly to evaluate the potential of the unbiased ICCV
method for the RPCs computation, and to test the impact of the
initial values. Initial values are set by zero and least square
solution respectively.
To find out an effective strategy to tackle the possible
ill-conditioned problem.
Here we confine the experiments to the third-order RFM with
80 coefficients, based on the terrain-independent scenario. With
the rigorous sensor model established and the elevation range
obtained from a cursory DEM, the 3-D grid of object points was
generated, with 5 constant elevation planes each with 10 by 10
grid points. While the check grid consists of 10 constant
elevation planes each with 20 by 20 grid points. So there are
500 control points and 4000 check points.
The fitting accuracy is measured in image both at control points
and check points. Firstly, the image position of the grid points is
713

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
calculated by the obtained RFM. Then the differences between
the pixel coordinates of the original grid points and those from
the RFM are calculated for evaluation. The accuracy
determination is quite the same as mentioned by Grodecki and
Dial (2001).
Fig. 1 3D object grid generated for solving RPCs
5.2 Test data sets
Tao and Hu (2001a) tested with the aerial photograph data and
SPOT data with sizes of 6000 by 6000. In order to evaluate the
fitting accuracy of the different methods for HRSI, we choose
SPOT-5 and QuickBird imagery. Respective ground pixel sizes
for testfield imagery were 5 m for SPOT-5, and 70 cm for the
QuickBird. Further details regarding the test-range are given in
Table 1.
Data set Ground pixel
Size (m) Image size
(pixel) Elevation range
(m)
SPOT-5 5 12 000×12 000 -2~327
QuickBird 0.7 27 552×22 700 340~1194
Table 1. Information of the data sets
5.3 Results and evaluation
All the methods are tested on both the SPOT-5 and QuickBird
imagery. The RMSE and the maximal errors in the imagery at
the control points and the check points are listed in Table 2 for
SPOT-5 data, Table 3 for QuickBird data. The condition
numbers of the norm function, before and after regulation, are
also listed in the tables.
There is not an absolute criterion for exactly judging that the
norm function is ill-conditioned or not, and how ill-conditioned
it is. Generally speaking, the condition number is helpful, the
bigger it is, the worse the condition is. Based on Table 2 and
Table 3, the condition numbers are big for both images,
for SPOT-5 data, and for QuickBird
data. The direct least square solutions are not very good, out of
sub-pixel, especially for the QuickBird data, the RMSE at check
points arrives at 115.937 pixels, and the maximal error is as
bigger as 7280.348 pixels. Therefore, the direct least square
solutions here could not be the final RPCs which would
substitute the physical sensor model.
14
1091.7 ×11
1013.1 ×
Determining regularization parameter using ridge trace method
is shown in Fig. 2, and L-curve method is shown in Fig. 3,
where SPOT-5 imagery is employed as an example. In the
experiment, RPCs are computed for a number of k with
different orders of magnitude varying from to , to
determine the order of magnitude. Then employ more k around
the order of magnitude, and choose the one that has the smallest
error at check points. This method can not select the best
parameter, and it is not convenient for automatic computation.
10
10−1
10−
For the L-curve method, the curve is shaped like “L”. And the
corner point on the L-curve that has maximum curvature
corresponds to the optimal parameter. This method can offer an
exact parameter automatically, without the need to plot the
L-curves. In the experiments, the parameter determined by
L-curve criterion is for SPOT-5 data, and
for QuickBird data.
7
1004.2 −
×
6
1004.9 −
×
Fig. 2 Determining ridge parameter using ridge trace method
Fig. 3 Determining ridge parameter using L-curve method
By comparison, the regulation by L-curve criterion, made very
significant improvements in terms of accuracy. After the use of
regulation, the condition numbers are smaller than the original
one, for the SPOT-5 data, reducing from to , and for
the QuickBird data, reducing from to . Except the
high accuracy, L-curve method shows very strong stability
based on more data sets.
14
10 9
10
13
10 8
10
The ICCV carries out by iteration. In the experiments, the initial
values are set by zero and least square solution respectively.
The iterative threshold value set as . For the zero initial
values condition, the results are pretty good based on the tables
that the accuracy is so close to and even better than the results
6
10−
714

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
of the L-curve method. It converged after a few times iteration,
even the iterative threshold value set to , it converges
quickly. For the least squares solution initial value condition,
we can note that it is good for SPOT-5 data but invalid for the
QuickBird data. It converged only after one time iteration even
that the iterative threshold value is strict. Coercive iteration is
also invalid for the improvement of the accuracy. More tests
should be done with the ICCV. It is worth of pointing out that
the computation of the ICCV is very simple and fast, that is
because there is no need to determine the regulation parameter,
and no need to inverse the matrix every time during the iteration.
And the structure of the solution by ICCV is as good as that by
regulation. Anyway ICCV is a potential way to overcome the
ill-conditioned problem for the RFM solution.
9
10−
Errors at CNPs (pixels) Errors at CKPs (pixels)
Approaches Condition
number Iteration
times RMSE Max RMSE Max
Least squares (LS) 14
1091.7 ×- 1.774 4.190 1.609 4.188
Ridge
Estimate L-curve criterion 9
1026.4 ×- 0.000 0.001 0.000 0.001
LS solutions as initial values - 7 0.000 0.001 0.000 0.001
ICCV
Zero as initial values - 12 0.000 0.001 0.001 0.001
Table 2. RMSE and Max errors in image with the SPOT-5 data
Errors at CNPs (pixels) Errors at CKPs (pixels)
Approaches Condition
number Iteration
times RMSE Max RMSE Max
Least squares (LS) 13
1011.1 ×- 4.893 77.168 115.937 7280.348
Ridge
Estimate L-curve criterion 8
1007.1 ×- 0.358 0.745 0.335 0.734
LS solutions as initial values - 1 4.895 77.334 115.684 7264.072
ICCV
Zero as initial values - 8 0.357 0.728 0.335 0.710
Table 3. RMS and Max errors in image with the QuickBird data
The DigitalGlobal Corporation provide the RPC file to the users.
As a comparison, we choose 9 ground control points to
checkout the RPCs, calculating the differencec between the
image coordinates of the GCPs and that from the RPCs. And the
errors at the GCPs are listed in Table 4. From the table we can
see the RPCs by the L-curve and the ICCV are so close
according to the accuracy, and they are slightly better than the
RPC provided by the corporation.
Errors at GCPs (pixels)
RPC mlmsmls
RMSE 9.310 9.082 13.006
By L-curve Max 11.556 13.334 16.856
RMSE 9.309 9.081 13.005
By ICCV Max 11.550 13.335 16.855
RMSE 9.391 9.592 13.423
Provided by
DigitalGlobal Max 11.563 13.867 17.646
Table 4. RMS and Max errors in image at GCPs
6. CONCLUSIONS
Since the SpaceImaging Company provided the RPCs to the end
users and the service providers, the RFM has been with us for
eight years, and a lot of researches show us that it is a useful
tool for exploiting high resolution satellite images.
Subsequently, the DigitalGlobal Corporation provides the RPCs
together with the physical sensor model, and more imagery
vendors may adopt the RFM, providing a way for economical
and fast mapping from HRSI.
The aim of this paper is to suggest a proper way to resolve the
ill-conditioned problem for the RFM solution. Regulations
improve the stability of the inverse matrix evidently and
produce a well structure RPCs. And the L-curve method
performs well in the experiments, being accurate and stable.
The ICCV is unbiased, simple, fast, and accurate, and the idea
that set the initial values as zero acts well for the RFM solution.
Both the methods show good effects, improving the accuracy of
the solutions, and ameliorating the RPCs’ structure. Considering
that the L-curve method has the risk of over-regularizing the
solution when the ill-condition is slight, though not happened in
the experiments, the ICCV should be the first choice. Finally,
we suggest a strategy, that L-curve method work for high level
ill-condition and ICCV for low. Even for the well conditioned
715

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
design matrix, the ICCV is really helpful to improve the
accuracy.
ACKNOWLEDGEMENTS
Thanks for the supporting from the 973 Program of the People’s
Republic of China under Grant 2006CB701302 and the
National Natural Science of China under Grant 407721001.
REFERENCES
Choi, H.G., Thite, A.N., Thompson, D.J. 2007. Comparison of
methods for parameter selection in Tikhonov regularization
with application to inverse force determination. Journal of
Sound and Vibration, 304, (3-5), pp. 894-917.
Hoerl, A.E., Kennard, R.W., 1970. Biased Estimation for
Non-orthogonal Problems. Technometrics, 12(1), pp. 55-67.
Hansen, P.C., 1992. Analysis of Discrete Ill-posed Problems by
Means of the L-curve. SIAM Review, 34(4), pp. 561-580.
Hu, Y., Tao, C.V., Croitoru, A., 2004. Understanding the
rational function model: methods and applications, IAPRS,
12-23 July, Istanbul, vol. XX, 6 p.
OpenGIS Consortium (OGC), 1999. The openGIS abstract
specification-Topic 7: Earth imagery.
Rodriguez, G., Theis, D., 2005. An Algorithm for Estimating
the Optimal Regularization Parameter by the L-curve.
Rendiconti di Matematica, 25(1), pp. 69-84.
Fraser, C.S., Dial, G., Grodecki, 2005. Sensor orientation via
RPCs. ISPRS Journal of PRS, 60, pp. 182-194.
Tao, C.V., Hu, Y., 2000. Investigation of the rational function
model. Proceedings of ASPRS Annual Conference, Washington,
DC. May 22-26. 11 pages. http://www.gis.usu.edu/docs/
protected/procs/asprs/asprs2000/pdffiles/papers/039.pdf
(accessed 31 Mar. 2008).
Tao, C.V., Hu, Y., 2001a. A comprehensive study on the
rational function model for photogrammetric processing,
Photogrammetry Engineering and Remote Sensing, 67(12), pp.
1347-1357
Tao, C.V., Hu, Y., 2001b. The rational function model: a tool
for processing high resolution imagery. Earth Observation
Magazine (EOM). 10(1). pp. 13-16
Grodecki, J., Dial, G., 2001. Ikonos geometric accuracy, Joint
ISPRS Workshop on HRM from Space, 19-21 Sept., pp. 77-86.
Wang Xinzhou, Liu Dingyou, Zhang Qianyong, Huang Hailan.
2001. The Iteration by Correcting Characteristic Value and its
application in surveying data processing. Journal of
Heilongjiang institute technology. 15(2), pp. 3-6.
716