- the spatial distribution of the fixed points, that must be
homogeneous over the entire network;
- the spatial density of the fixed points, that must satisfy the
technical specifications of the Cadastre for a second order
network;
- the number of polygons having in common the same fixed
points.
According to these conditions, three configurations of fixed
points have been identified:
Configuration. l: constituted by 9 fixed points, that is 1.59
points per km? , partially distributed along the perimeter of the
network and partially | in the internal part;
Configuration 2: constituted by 7 fixed points, that is 1.19
points per km”, distributed along the perimeter of the network;
Configuration 3 : constituted by 4 points, that is 0.70 points per
km”, distributed along the perimeter of the network.
4.2 Error attribution for the simulated network
Before the adjustment, particular error models have been
applied to the original file containing the true coordinates of the
various polygons: in this way a new set of data files simulating
the effective measurements has been obtained. According to the
kind of errors applied, five different classes and subclasses of
analysis have been generated:
1) measurements affected only by systematic errors (S):
2) measurements contemporarily affected by systematic and
random errors (S and R), which can be partitioned into three
different subclasses:
2a) prevalence of the systematic error component with
respect to the random one (S>R)
2b) equivalence between the systematic, and the random
error components (S=R)
2c) prevalence of the random error component with respect
to the systematic one (R>S)
3) measurements affected only by random error components
(R).
The systematic errors have been generated by scaling the single
fiducial polygons of a random amount. The scale variations are
randomly distributed in the maximum range of +100 ppm.
These systematic components have only a local effect:
considering the general network their overall averaged
contribution becomes practically negligible.
For each class and subclass of analysis, three different error
simulations have been executed. The total number of datasets is
therefore equal to 15; in this way a significant number of
possible situations has been originated. Every dataset has been
then adjusted applying both the traditional and the Procrustes
methods, considering, for every computation, the scheme of
fixed points characterising one of the three mentioned
configurations. In this way, 90 different computations have been
performed, comparing the obtained results with the original true
values, for each of the two adjustment procedures (Clerici,
2002).
4.3 Numerical results
The numerical experiments have been executed with the aim to
verify the capability of the Procrustes algorithm to compensate
for the various kinds of error. To verify the capability of this
original method, it is necessary to compare the obtained results
with those furnished by the classical least squares adjustment of
the individual polygon sides. For this purpose, the mean values,
the medians and the root mean square errors for a specific error
simulation referring to each of the fixed point configurations
have been collected and compared. Analogous investigations
have been carried out for the other two cases of error
generation.
34
An analysis of the results, in particular for what is referred to
the Configuration 1, allows to conclude that for the Procrustes
analysis the mean and the median values for the x and y
coordinate components never differ for values greater than 6
mm in absolute value. For the traditional adjustment technique
this difference reaches 5 cm; this fact indicates a better
symmetry of the residual distribution when the Procrustes
algorithm is applied. Furthermore, it is possible to observe that
in the case of the only presence of systematic errors, Procrustes
adjustment leads to mean and median values equal to zero for
all of the three configurations of fixed points. In this case, the
residuals follow a perfectly symmetrical distribution, without
any systematic error component; vice versa, by applying the
traditional adjustment procedure, these values vary between 2
and 11 cm. When applying the Procrustes algorithm, the root
mean square error is always near to zero, while for the classical
adjustment procedure this has a magnitude of 10 centimetres.
However, due to the particular systematic local random error
generation, the residual values after the adjustment of the whole
network are always contained within the rms error. Finally, if
only random errors are present, the values of the differences
between means and medians are always comparable for the two
methods; this remains true also for the values of the root mean
square error, but is limited to the configurations of fixed points
I and 2 that consider a sufficient number of fixed points for the
whole network.
The graphs reported show, for the Configuration 1, first, the
differences between the true values of the point coordinates and
the values obtained by applying the Procrustes algorithm (see
proper symbol), and second, the differences between the true
coordinate values and those ones obtained by a traditional
adjustment procedure of the polygon sides (see proper symbol).
Figure 6 refers to the case in which the systematic error is only
present. The results obtained by applying the Procrustes
technique are almost corresponding to the true values, while the
classical adjustment leads to an error dispersion of 20-30 cm.
X COORDINATE DIFFERENCES
E €^099999097v^o99909^99 EHE
Point ID
Y COORDINATE DIFFERENCES
E | 00900.0000600000000000000000,
Point ID
Figure 6. Residuals for pure systematic error conditions
(Procrustes €, traditional . )
Figu
The gra
random
signific
Procrus
of men
with th
the cap
system:
superim
Fi
The abc
random