The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
1319 
single factor influencing the buffer. All together, only 
“visibility” was significant and common to the three image 
datasets (F=3.875 p=0.009). Mean buffer value was generally 
significantly underestimated (-0.33m ± 1.75SD) for the third 
visibility modality “Poor on Ortho, good on Cartosat-1” 
whereas it was generally overestimated for the three other 
modalities (global mean of the three remaining modalities = 
0.17m ±2.04SD). 
With regard to the orthophoto and Cartosat-1 Fore images, 
significant factors and interactions were very similar. Only 
“land cover” had a significant effect (F=24.29 p=0.0064) on 
buffer for Cartosat-1 Fore. For these two images, variance of 
the buffer was high between operators. Unskilled operators 
generally delineated the parcel boundary well on the orthophoto, 
but largely underestimated the parcel area on Cartosat-1 Fore (- 
0.05m ±0.99SD and -1.92m ±2.26SD respectively). Conversely, 
skilled operators tended to overestimate the buffer with 
Cartosat-1 (+2.56m ±2.13SD) but to generally delineate parcels 
correctly from the orthophoto. Once again, “magnification 
effect”, “unconscious compensation effect” and/or loss of 
reference when passing from orthophoto to panchromatic could 
explain overestimation of parcel area by skilled operators. 
Generally the larger the parcel, the smaller the overestimation 
of the parcel area by skilled operators on Cartosat-1 Fore. On 
the orthophoto, the smaller the parcels, the higher the difference 
between the digitised area and the true parcel area. On the 
contrary, parcel size didn’t really influence the measurements 
for unskilled operators (F=0.93 p=0.39): parcel area was 
consistently underestimated when using Cartosat-1 Fore and 
relatively well measured with orthophoto, irrespective of the 
parcel size. Regarding the shape of the parcels, only a limited 
effect on the measurement of the parcel area (and subsequently 
of the buffer value) was observed. Parcel shape had a greater 
influence on buffer measurement when interacting with 
“image” and “parcel size”. This was especially true for skilled 
operators for whom a complex parcel shape led to significant 
overestimation of the parcel area and consequently to higher 
positive buffer values. With regards to Cartosat-1 Aft, for which 
“visibility” was the only significant factor, numerous 2 nd order 
interactions were significant. These principally concerned the 
shape and the size of the parcels, then the operators and finally 
the land cover. 
Firstly, from the previous results concerning image types, we 
showed that the characteristics of the parcel clearly influenced 
the precision of the parcel area delineation: shape and size of a 
parcel, either separately or combined, are interpreted differently 
depending on the operator’s experience. For experienced 
operators, large and/or complex parcels boundaries are 
generally smoothed because of the magnification effect or 
possibly as a consequence of productivity criteria (i.e. to cost- 
effectively process a maximum quantity of parcels a day), thus 
leading to overestimation of the parcel area. On the contrary, 
unskilled operators seemed to be less influenced by the parcel 
characteristics and constantly underestimated the parcel area. 
Secondly, image quality, as defined by the “visibility” factor, 
strongly influenced the final accuracy. Skilled operators used to 
working with orthophoto obtained relatively good results with 
orthophoto, but they tended to lose this advantage when 
switching to panchromatic images. Conversely, unskilled 
operators were frequently inaccurate with both orthophoto and 
with panchromatic images. Consequently, the use of one type of 
image cannot be decided without knowing the staff 
competences by assessing their abilities to transfer and use 
memorised CAPI experience. Therefore a first question could 
concern the best way to choose an operator according to the 
image type. When CAPI has to be performed on orthophoto or 
panchromatic images, a cost-effective solution would be to 
choose skilled operators, but with the risk that smoothing 
(complex/large) and compensation (complex/small) effects will 
generally lead to overestimation of the true area. On the other 
hand, if the strategy of the enterprise is to contract new photo 
interpreters, we suggest that one should assess their recognition 
capacity; this could be undertaken on true colour composite 
images regularly compared to their panchromatic equivalent. 
Image 
Carto 
Carto 
Ortho- 
Aft 
Fore 
photo 
Mean Value = bias [m] 
0,04 
0,52 
-0,06 
St. Dev. Repeatability [m] 
1,85 
2,22 
0,92 
Repeatability Limit [m] 
5,18 
6,23 
2,59 
St. Dev. Reproducibility [m] 
1,85 
3,13 
1,02 
Reproducibility Limit [m] 
5,17 
8,76 
2,86 
Critical difference to reference [m] 
1,65 
3,21 
0,96 
Table 4. Results from area measurement on orthophoto and 
Cartosat-1 
Intentionally, a last factor has not yet been discussed: land 
cover. The decision was made to discuss it separately so as not 
to risk overloading results or incorrectly classifying the main 
factors to consider from this survey. Indeed, land cover 
appeared to be significant only within 2 nd order interactions and 
never as a single significant factor, suggesting that land cover 
cannot be discussed independently of other factors. From the 
SLS results, we showed that land cover was mainly associated 
with “visibility”, “parcel size” and “parcel shape”; this 
indicated that land cover could be perceived as a characteristic 
of the object, i.e. the parcel, at the same level as “shape” and 
“size”. Whatever the operator and his level of experience, we 
showed that parcel area measurement was always more accurate 
and less variable when there was bare soil, annual crops or 
pastures. On the contrary, for orchards, vineyards or olive trees, 
parcel area was often overestimated and highly variable. This 
was true especially with Cartosat-1 images. For instance mean 
buffer values were 0.44m +1.27SD, 0.50m ±2.67SD, 0.61m 
±3.64SD for orchards and -0.13m ±0.68SD, -0.08m ±1.34SD, 
0.54m ±0.15SD for bare soil, respectively with orthophoto, 
Cartosat-1 Fore and Cartosat-1 Aft. This trend was maintained 
between operators, the sole difference being that unskilled 
operators continued to proportionally underestimate parcel area 
according to image type. When considering parcel size, larger 
parcels with bare soils, annual crops or pastures were often 
overestimated than small parcels. On the other hand, ligneous 
crops appeared to be the main source of underestimation of the 
area of small parcels. Finally, regarding parcel shape, the same 
results were obtained: greater difference from the true area and 
greater variability was evident for parcels with ligneous crops. 
From these results, land cover seemed to aid the operator in the 
correct identification of a parcel regarding its content; it 
allowed the operator to recognise more clearly the parcel but it 
remained relatively useless when delineating the parcel 
boundaries. Tree canopies extending outside of the parcel could 
lead to overestimation of the area because of the difficulty of 
clearly distinguishing the parcel area and surrounding natural 
vegetation; and crops boundaries were delineated more often