The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1331
measurements and singular measurement, instead of mean, the
true value should be used, if it is accessible. Alternative method,
to RMS, could be also involved to the accuracy assessment,
especially if the experiment is properly prepared and
measurements are repeated by a few operators. In this case, the
experiment is similar to laboratory measurements usually made
in chemical research, precisely determined by ISO standard.
According ISO 5725 “Accuracy (trueness and precision) of
measurement methods and results”, among others, the following
notation is used: sample, operator, laboratory, equipment,
repeatability and reproducibility. As a sample, in our research
we understand digitised impervious surface on IKONOS, on
each of test area. In the experiment, 6 operators have been
taking part using 2 equipments (IKONOS PAN and RGB). Each
operator is understood as a laboratory, so we have 6
laboratories. Each test area has, as a reference, the impervious
surface area, digitised on the aerial ortophoto. Six operators
digitised PAN/RGB 3 times, so we have 18 measurements for
each test area. Finally, 108 measurements were analysed
because of 6 test area and 18 measurements of one test area. At
the beginning, outliers were found for PAN and RGB images.
According ISO 5725 outliers are found using graphical or
numeric method, Corchan/Grubbs tests. In our research group,
there was no statistician so we identified outliers in traditional
way, as used usually in surveying. Relative differences between
reference area and each measured area were calculated and
histogram of the error was prepared for PAN and RGB (Figure
4). Outliers were defined as 5% of the external measurements
on the histogram. If at least one of measurement is marked as
an outlier, all group of measurements (3 repetitions) are
discarded. After outliers removing, accuracy analysis was
performed.
Assuming “y”, as a result of measurement of the impervious
surface area for each test area for each measurement, the
following can be written:
y = m + B+ e
where:
m- expected value (reference impervious area),
B -bias in repeatability condition (difference between measured
area and reference area),
e - random error in repeatability condition.
Variance of B describes between laboratories variance:
var (B) = CT 2 group
where:
cfgroup - standard deviation (between laboratories).
Variance in repeatability condition for one laboratory is defined
as following:
var (e) = a 2 i
CTi - standard deviation within group (laboratory), calculated for
the test area digitized 3 times by one operator.
Each operator (laboratory) is described by standard deviation a\.
For all operators average variance is calculated, called variance
of repeatability:
CT 2 re P et=var(e) = ct 2 , .
Finally, accuracy is defined as standard deviation of
reproducibility and it is the sum of the between groups variance
and the within groups variance:
_2 2 , 2
° reprod 0 group ' M- repet
where:
Ogroup- standard deviation in group (one for one test area),
CT rep et- average standard deviation of repetibility (average of six
standard deviations for each operator).
Results of photointerpretation were validated using ISO
standard and in traditional way basing on RMS. Besides
absolute RMS [sq m], relative area error (RRMS - Relative
Root Mean Square = RMS/reference area with value: 0-100%
or 0-1) was analysed.
4.2 Part 2 - influence of the photointerpretation accuracy
on the imperviousness factor calculated in simulated
Landsat pixel
Landsat classification (e.g. unmixing) requires training area,
obtained from field surveying or from photointerpretation of
VHR images (e.g. IKONOS). In our research, we ask the
question: how does the photointerpretation accuracy influence
on the percent of impervious surface in Landsat pixel?
For each test area, grid of 30m cell size was simulated. Then the
impervious surface area was calculated in the grid. The
procedure was performed for all measurements (18 observations
for each test area) and for the reference. In each simulated
Landsat pixel, area of impervious surface and percent of
impervious surface area in the pixel was calculated for each
measurement and for the reference. The percentage of
impervious surface area in the pixel is called: imperviousness
factor. Reference area and measured area were analysed in 100
pixels for the test area. Analysis was made in two aspects:
comparison to the relative area error (RRMS) calculated
generally for all test areas (data processing described in 4.1 Part
1) and evaluation of the absolute and relative error of
imperviousness factor. Firstly, RRMS of the impervious area
was calculated for each pixel. Then the differences between
percent of impervious cover of the pixel basing on observations
and on reference were calculated. Then, absolute RMS of
imperviousness factor was calculated for each pixel. Finally,
RMMS of impervious surface area was calculated for all pixels.
Accuracy analysis was performed on the Landsat pixel level in
10 groups of imperviousness factor varies from: from 0% to
100% by each 10%.
Figure 3 Part of test area no. 1 with overlaid all digitised on
Pansharp colour IKONOS area; each pixel is labelled by
percent of impervious surface [0,1].
5. RESULTS
Initially, all observations for PAN and RGB were statistical
analysed. Histogram of relative area error for PAN and RGB is
on the figure (Figure 4) presented. Mean of the difference
between observed area and reference area (bias of the method)