e CO-
f the
ation
ue to
on of
t any
995
153
A
Qu 0
Co(apriori)
4 ni (t
Jn ook
The partial redundancies are called r;. The values for d and K are determined with the help of Q.
i =
3) Proposal of Huber:
Vi or vj| «c
d R vi) T El 2e e
The value of c can be chosen by the user.
4) Proposal of Hampel:
vj for [vi] <a
a-sign(v;) for a <|vi| <b
nens a (v -c-sign(v;))/(b - c) for b «l|vilsc e
0 for vil >C
The values of a,b,c can be chosen by the user.
Iterative reweighting procedures with objective function orientation
The methods with objective function orientation have to be distinguished from the above mentioned trial and error methods
because they describe an alternative algorithmic procedure to the method of linear programming. In particular there is
1) Approximation of L; -norm-method:
pi(t+1)=1/|v;(t)+c| with c » 0 m
2) Lp-norm-estimates with 1<p<2:
pitt +D =1/(jy ff +c) withec>0;1< p<2 (8)
3.2. Least Median of Squares Procedure
One further robust estimation method is the Least Median of Squares (LMS) procedure (ref. Rousseeuw, Leroy 1987), which
minimizes the median of the square sum of the residuals. Die objective function is given by:
median; (v2) — min (9)
With this strategy Rousseeuw was able to equalize the influence of the geometry of the observations. Unfavourable geometries can
led to the case that single observations become restrictions and influence the parameter estimation result like leverage points (ref.
chapter 4).
Besides the high computation time the LMS-method can generate wrong results. An example out of the literature (table 1) (ref.
Fellbaum 1994) demonstrates this fact. Through the introduction of a single error in the y-co-ordinate of point 2 the LMS estimator
breaks down. The error is not detectable while looking at the residuals. The parameter estimation was done with the Program
PROGRESS (Program for Robust reGRESSion) (Rousseeuw, Leroy 1987).
The example shows clearly that the LMS-method is barely suitable for error detection. In this case even the non robust L;-norm-
method was able to detect the error. Therefore a different strategy for equalization of the geometry of the observations is used in
the next chapter.
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995