The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1318
of parcel characteristics within an image and (ii) the level of
experience of one operator when interpreting image.
Concerning “parcel#image properties”, only “image” (F=6.11
p<0.0157) and “parcel size” (F=58.06 p<0.0001) were single
factors significantly affecting the area buffer. Then, 2 nd order
interactions between the shape of the parcel (“image* parcel
shape” F=33.79 p<0.0001), the size of the parcel (‘image *
parcel size” F=11.83 p=0.001) and “image” were significant.
From ANOVA, orthophoto counted for 23% of the outliers and
was significantly underestimated (mean value = -16.7m; F=6.50
p=0.0025) when Cartosat-1 Aft and Fore counted respectively
for 18% and 58% with mean values of -2.6m and +7.5m.
As discussed in the next paragraph, this result did not signify
that the orthophoto is a major source of underestimation (or
Cartosat-1 Fore as a source of overestimation). In fact, the
operator’s interpretation was mainly the source of
underestimation. Concerning parcel size, small and large
parcels counted respectively for 22% and 26% of the outliers
when medium parcels represented 56% of the outliers.
Difference of buffer was significant between parcel sizes
(F=26.2 p<0.0001) with mean buffer values of +12.6m, +8.5
and -28.1m respectively for small, medium and large parcels.
Independently of the image, of the parcel shape or of the
operator, small parcel areas were often overestimated and large
parcel areas were underestimated. One would suggest that this
effect was a consequence of the magnification function used
during the measurement: each parcel was enlarged and fitted to
the full screen size to facilitate boundary recognition. The
smaller the parcels, the higher the magnification was and the
higher the dilution of the contrast. On the other hand, one could
imagine a “compensation” effect from the operator which,
unconsciously, searched to not disadvantage small parcels and
tended to put boundaries off and inversely.
As explained previously, few outliers came from orthophoto
when most of them were due to panchromatic Cartosat-1
images: Cartosat-1 Aft and Fore images were respectively
responsible of 18.5 and 58% of the outliers detected. Clarity of
the object displayed on the screen and true colour (RGB)
composition seemed essential first for a good recognition of the
object, second for the better delineation of the parcel boundaries.
Thus, “operator’s recognition capacity” and “operator’s object
memorisation” could be of prime importance, at least to explain
outlier’s existence. They both compose “operator’s experience”
and condition the interpretation of the size and shape of the
parcel (and of possible contained object to be excluded). With
Cartosat-1 images, we assumed that parcel boundaries
recognition was more difficult and more deductive for operator.
We showed that “operator* image” effect on buffer
measurement was significant (F=3.6 p=0.018). Over the five
operators, four were responsible for the outliers; among those,
three were skilled. The unskilled operator was responsible of
most of the outliers (70%) and generally underestimated the
buffer (mean value = -6.2m) when others always overestimated
the buffer (from +12.2 to +14.6m). All together, the results
obtained from the analysis of the outliers’ population clearly
suggested a tripartite relationship existing between (i) the parcel
with its particular characteristics, (ii) the image as the
information vector conveying parcel characteristics and (iii) the
operator as the place of interpretation where his personal visual
recognition capacity and object memorisation interact and
determine the accuracy of the measurement. Here we showed
that the unskilled operator was mainly responsible of
underestimation of parcel area certainly because of a too limited
object recognition experience on image, this, whatever the type
of image considered. On the contrary, the three skilled
operators often overestimated parcel areas. They were
responsible for the major part of overestimation on Cartosat-1
image. This should be related to the fact that their experience
initially concerned true colour composition imagery.
Consequently, (i) effect of magnification, (ii) unconscious
parcel size related “compensation effect” and/or (iii) lost of
reference when passing from orthophoto to panchromatic could
explain outliers from skilled operators.
By analysing extreme area discrepancies, we showed that area
measurement accuracy is mainly conditioned by the
relationship between operator and the image properties. CAPI
training for panchromatic and true colour images should be
organised to reduce risk of wrong delineation of the agricultural
parcels.
4.2 Final buffer population
4.2.1 Normality test
A normality test (KSL/Kolmogorov-Smimov-Lilliefors) was
performed on each type of image to determine if observed
buffer values were normally distributed and consequently to
decide if an analysis of variance was relevant to identify the
main significant factors and differences between modalities.
Whatever the type of image, small p-values were obtained
(W=0.951 p-value<0.001; W=0.989 p-value<0.001; W=0.996
p-value=0.001 for Orthophoto, Cartosatl Fore and Aft
respectively) and the null hypothesis that observed buffer
values have a normal distribution were rejected. However,
buffer value distribution was very close to the normal
distribution, being relatively symmetrical around the mean.
Furthermore, the number of observations by type of image was
sufficiently high (Table 2) to allow for processing SLS
procedures and analysis of variance without restriction and to
limit the risk of misinterpretation of factors’ effects.
Buffer values obtained from the survey were significantly
different between image modalities; mean values and standard
deviation are given in table 2. Consequently, the SLS procedure
and analysis of variance have been conducted for each type of
image independently and will be discussed separately.
Parameters
orthophoto
Cartosat
Cartosat
[m]
Aft
Fore
Mean
-0.059
0.041
0.515
Std Dev
0.990
1.848
2.857
Std Err Mean
0.024
0.046
0.074
upper 95% Mean
-0.011
0.131
0.661
lower 95% Mean
-0.106
-0.049
0.370
Number of observations
1642
1613
1485
Table 2.Distribution parameters for each image
4.2.2 SLS and ANOVA: main significant effects
The SLS procedure applied separately to each image dataset
showed that the factors responsible for the observed variability
of the buffer were different between images. Whereas
“operator”, “shape” and “size” of the parcel or even “visibility”
were the single factors to be retained for the orthophoto, “land
cover” should also be considered of interest for Cartosat-1 Fore.
Concerning Cartosat-1 Aft, only “visibility” was retained as the