propagation. Field check of the end product often 
reveals large local systematic errors. These errors 
have their roots in the long chain of processes. The 
question which arises here is whether a better control 
of the various quality parameters could provide an 
explanation of these systematic errors and whether 
eventually it could lead to corrective actions or at 
least predict in which model area field control is 
necessary. For this purpose, a total quality factor can 
be computed for each model, summing up the quality 
factors of the individual products (e.g. photos, control 
points, etc.) : 
Q (model) = (Q1 + Q2 + ...Qn)/n 
Qi : expressed in percentages 
It is expected that quality information at feature level 
and model level will support efficiently the field 
completion process. 
Some of the quality parameters and standards are 
summarised in figure 2. 
3. METHODOLOGIE FOR QUALITY CONTROL 
OF A PHOTOGRAMMETRIC DATA SET 
A photogrammetric control process has to be applied 
to the new data set before proceeding to the editing 
phase. The following quality components will be 
assessed: positional (relative) accuracy, classi- 
fication accuracy and completeness. Some quality 
attributes are normally present in the data set at 
feature level in the form of reliability codes, indicating 
how good features could be identified and measured. 
A sample of check points must be created by an 
independent process of higher accuracy: for this 
purpose an analytical plotter can be used. A reliable 
checking of completeness requires a superimpoition 
system. If larger scale photographs are not available, 
the same photographs will be used for feature 
extraction and control measurements. 
3.1 Positional accuracy 
The various classes of features are not 
homogeneous in terms of accuracy; therefore it can 
be recommended to group features in accuracy 
classes : 
e.g. acl: road, railway, canal, 
ac2 : building 
ac3 : vegetation boundary 
For point features and buildings, it is easy to identify 
homologous check points ; it is more difficult if not 
impossible to identify such points on linear features 
like roads, vegetation boundaries etc. 
An interesting method has been developed at IGN- 
Paris, based on the Hausdorff distance which allows 
the evaluation of planimetric accuracy of a line 
feature with respect to a reference line (Hottier et al., 
1994). 
A less rigorous approach consists of taking distances 
between a sample of check points and a measured 
line; a RMSE (root mean square error) can be 
14 
computed and used as a rough accuracy estimator. 
One can also compute "pseudo" homologous points 
on aline for a given sample of check points. In this 
way, the same procedure can be applied, whether 
one deals with point features or line features. In both 
cases two sets of homologous points are used: check 
points (X, Y, Z) and points from the observed data set 
(X, Y', Z). RMSEs can be computed for each 
coordinate separatly, as well as for planimetry : 
DS, = / DX? « DY? 
c 
RMSEP= 22S; 
P (planimetry) 
a 
RMSEH = A (height) 
DX, , DY, , DZ are the differences between the two 
data sets 
n : number of check points. 
DS, represents planimetric discrepancies (not errors) 
in a bivariate normal distribution, while the height 
discrepancies DZ, follow a linear normal distribution. 
The steps of the sampling and testing procedure can 
be summarized as follows : 
measure a sample of check points per accuracy 
class (n = 50 points) 
compute the mean values of discrepancies: 
sx-62M gy. DH g€2€r 
n FK isl 7 
compute the standard errors 
Oc 0, 0, 
apply a test for significant bias based on Student's 
t distribution at a 196 significance level 
compute the disrepancies DS, in planimetry 
detect and eliminate gross errors in planimetry; a 
robust estimator of the RMSEP can be computed 
with the help of the median : 
RMSEP =1.201'm (fii: median) 
compute a tolerance value at a 196 level of 
significance : 
A 
T = 2,14 RMSEP 
recompute RMSEP = y £25 
7 
n1 : new sample size 
apply a test of precision based on a chi-squared 
distribution (test of goodness of fit). 
A similar distribution can be applied for the height 
discrepancies, keeping in mind that we deal with a 
normal gaussian distribution. 
Experiments carried out on two tests areas show the 
following results in planimetry : 
Project (A) : 46 um at 1:30,000 photo scale (check 
points measured from 1:10,000 photo scale). 
Project (B) : 53 um at 1:40,000 
photo scale. 
Larger values are expected with field checks as can 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
és m à (rnm OO OO O0) "T 
CRM CN 
cU m "nv 
-—— m CN. 1 
CD CY Te] 
~~ 
MP