The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1289
1 = (T t W 2 T)-'T t W 2 G
3. REGULARIZATION METHOD
3.1 Regularization Model
I m =0,W m =W(I (0) ) = E
C> = V.) + (T'K-J + A’sr'T’XflVo
^,=^(/ w ).v w =G-r/ (S)
(10)
The denominators 5 ( . and Z) (i = 1,...«) change quickly in
quantity when the GPS input distributed unevenly in calculation,
which makes matrix T ill-condition in the equation, and thus the
matrix T T W 2 T singularity. This case often happens when the
rank of RPC polynomial is higher (eg, more than 2), which may
result in not convergent during iteration.
A unit matrix E could be added by the regularization method to
improve the condition number of matrix T T W 2 T .
T T W 7 T is a symmetric nonnegative definite matrix, and the
eigenvalue of matrix T T W 2 T + h 2 E is in the range
so the condition number of which
it will reduce with the increase of h 2 . A regularization rule
method is employed in this paper to improve the condition
number of matrix to obtain stable numerical solution. The
normal equation is calculated through iteration and can be
ended when the condition be met (Neumaier,1998) .
(T t W 2 T + h 2 E)I - T t W 2 G = 0 (9)
Where:
E is the unit matrix, h is the regularization parameter, s is the
number of iteration.
When the pixel measurement error is known, the covariance
matrix P of parameter I could be calculated from the following
equation:
P = (T t W 2 T + h 2 E)-' T t W 2 R c W 2 T(T t W 2 T + h 2 E) r
(ii)
There are many ways to obtain the regularization parameter h.
Different h will get different result, of which the optimal value
is attained by trial method, here L curve method (Neumaier,
1998) is employed. To get the optimal value of h, different
value was tested in formula (9). The third power RPC was
employed in the test, 50 GCP and their corresponding image
points were selected for calculation of RPC, while 29 CkP were
used for accuracy assessment. The result was shown as figure 1.
Based on the test, less than 10 iteration times may get good
convergence under most conditions when h was in the range of
0.009 and 0.1. We also found that as long as h was in the range,
the accuracy was not sensitive to specific h, that is to say the
results were within 0.01 pixel (see table 1). So h=0.05 was used
in the following text.
Figure 1 Calculation for h—L Curve Figure