The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1317
The other parameter which was used to evaluate the test results
is critical difference (CD) to the reference value. This value is a
value less than or equal to which the absolute difference
between particular observation and reference data is expected to
be at the 95% of confidence (ISO 5725). The critical difference
as comparison with reference value for more than one operator
can be calculated as below (3):
(eq. 3)
Where p = no of operators
n = no of observations
a R = standard deviation under reproducibility
conditions
o r = standard deviation under repeatability conditions
3. MEASUREMENT
The test site was situated on the region of Maussane in France
and covered 10km by 10km area. All data, i.e. stereo pair of
Cartosat-1 Aft and Fore and digital orthophoto UltraCamD,
were projected in UTM 31N (WGS 1984). The set of parcels
(as polygons) was acquired using ARCGIS software as the
measuring environment.
The number of parcels to be measured was probabilistically
determined according to previous results (Pluto-Kossakowska et
al., 2007). After field verification (Feb 2008), 203 parcels were
digitised on UltracamD RGB composition, then verified and
corrected by independent operator. This set of parcels was
checked also on Cartosat-1 images, and 18 parcels were
rejected in terms of invisible borders. This means that only 8%
of 203 parcels were not detectable on Cartosat-1. Finally, 185
parcels were used for the study. The land cover classes and
parcel borders were precisely identified on these selected
parcels during the field campaign in February 2008.
Three replicates per parcel were obtained from five operators on
two images of Cartosat-1 (Aft and Fore) and on the UltraCamD
orthophoto. The parcels appeared on the screen randomly and
were enlarged at the maximum zoom within the viewer window.
On screen, the parcel to be measured was centrally identified
(see figure 1). After the parcels measurement, the area and
perimeter were calculated for each observation. The buffer
value was derived using equation (1) and was used for
following statistical analysis.
4. STATISTICAL ANALYSIS AND RESULTS
4.1 Outliers
The workflow included a statistical detection of outliers, a great
variety of which are available from the literature and which
could be used in this experiment. Using the Jacknife distances
test, 81 observations out of 4995 (1.6%) were identified as
significant. SLS procedures allowed the identification of the
factors and related interactions responsible of the outliers’
population distinction ((F(6;80)=30.51 p-value<0.0001,
^=0.71). The maximum value of the outliers was 61,2m while
the minimum value was -79,2m.
The majority of the outlying observations occurred once per
parcel. Less frequent were the parcels several times with
multiple outliers: these concerned essentially the three
replicates from the same operator (and the same image). Even if
existing, it was very rare that the same image was identified as
an outlier by several operators from several images. To
illustrate this specific case, an example is given in Figure 1.
Figure 1.Example of outliers: 1 on Cartosat Aft, 6 on Cartosat
Fore, none on orthophoto
The two main groups of factors explaining outliers were (i) the
“parcel#image properties” defined as the visual representation