The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1316
In order to assess the area measurement accuracy, reference
parcels with a known area and perimeter have to be selected
from available sources or acquired using independent tools,
here precise orthophoto. This aerial orthophoto with 0.5m of
GSD was acquired in 14 th of May 2005 as multispectral image
RGB using UltraCamD digital camera. Pixel level accuracy was
determined in a separate experiment (Spruyt, pers. comm.).
2.2 Methods
The methodology of area measurement evaluation is based on
statistical analysis of discrepancies between the measured and
reference areas (Pluto-Kossakowska et al. 2007). In order to
derive the tolerance above which an inspector will reject the
area claimed by the farmer with a risk of a=5%, an initial
verification that the distribution of the buffer is normal must be
made. To obtain the final tolerance for the measurements, a
repeatability limit of the buffer was applied.
The scheme of validation procedure is proposed as followed:
1. Data processing and acquiring:
- images orthorectification - needed to make cartographic
product to measure
- Acquiring the reference parcels - from digital orthophoto
- Area measurement of the selected parcels on the images
- Buffer calculation based on measurements and reference
data
2. Statistical analysis of the buffer value
- Anomalous measurements detection and elimination
- Normality test and analysis of variance (SLS, ANOVA)
- Determination of tolerance for the measurements as
reproducibility limits.
It is practical to model the maximum acceptable discrepancy
between the measured area and the claimed area, i.e. the
tolerance, as the parcel perimeter multiplied by a width. This
width, also called buffer width (or simply buffer) around the
parcel perimeter, is expected to vary as a function of the
measurement tool, whether it is an image or a GPS-device. For
a given parcel, the knowledge of its reference (i.e. true) area
and reference perimeter allow the transformation of the area
error (measured area - reference area) into this buffer width
using:
B = (a¡_JV) (eq . l)
iV
where B, = buffer width for measurement i
aj = measured area for measurement i
a,. e f = reference area of the parcel
p ref = reference perimeter of the parcel
Using the buffer values from different observations we can
determine the tolerance between two independent
measurements under the specified condition (the same parcels,
same image, and independent operators). The simplest way is to
verify whether the distribution of the buffers follows a normal
law using different tests.
Detection of outliers is recommended prior to verifying the
normality of the buffer widths. According to ISO 5725 (1994),
the detection of anomalous measurements may be made using
different tests. Outlier detection was performed here using the
Jacknife distance test in JMP 6.0 (SAS Institute). A Standard
Least Square (SLS) procedure was then performed to identify
the factors (and 2 nd /3 rd order interactions) significantly
explaining the observed variability of the buffer. Table 1
presents the list of factors and related modalities. “Shape”
factor was distinguished on to three modalities: simple i.e.
rectangular alike shape, medium i.e. rectangular shape with
little changes and complex i.e. shapeless. The visibility depends
on parcel itself, parcel surrounding and image properties: good
visibility i.e. all parcel borders are easy to recognise; poor
visibility - part of the border is difficult to recognise and must
be deduced. “Operator” presented two different modalities
(skilled vs. unskilled) based on the level of “experience” of
each photointerpreter had at the beginning of the survey.
Finally, the assumption of normal distribution of the buffer
values leads to the derivation of a tolerance (at a=5%), above
which an inspector would reject the area claimed by the farmer.
For the needs of our survey and following the ISO 5725 (1994),
the tolerance can be interpreted as reproducibility limit (eq. 2).
Reproducibility refers to the ability of the measurement to be
accurately reproduced by someone else working independently,
i.e. is a value less than or equal to which the absolute difference
between two results obtained under reproducibility conditions
may be expected to be with a probability of x%.
R = f*cr R * yfn (eq. 2)
Where a R = standard deviation under reproducibility condition
(for the method of calculation refer to ISO 5725, 1994)
f = multiplication factor of standard deviation to
determine the confidence interval on specified level of
probability (here 95%)
n = number of test results to be compared, here n=2
For normal distribution at 95% probability level, f is 1.96 and
f*V2 then is 2.77. The simple “rule of thumb” R=2.8o r is
applied instead of equation (2) (ISO 5725, 1994).
Factors
Modalities
Operator (n=5)
Skilled (n=3)
Unskilled (n=2)
Image (n=3)
Orthophoto
Cartosat-1 Aft
Cartosat-1 Fore
Image visibility (n=4)
Good on all images
Good on ortho,
poor on cartosat
Poor on ortho,
good on cartosat
Poor on all images
Parcel shape (n=3)
Simple
Medium
Complex
Parcel size (n=3)
Small (< 2ha)
Medium (2ha> >8ha)
Large (>8ha)
Land cover type (n=7)
Bare soil
Green cover
Marsh
Olive trees
Orchard
Pasture
Vineyard
Table 1. List of factors tested and related modalities