The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1333
Observations were analysed to identify and discard outliers:
PAN: 8 outliers (4 groups - 12 measurements),
RGB: 5 outliers (3 groups - 9 measurements)
After outliers discarding standard deviations received 0.20
(PAN) and 0.14 (RGB). Repeatability, reproducibility and RMS
for the test areas are for PAN in table (Table 1) presented and in
table (Table 2) for RGB. Comparison between: bias (average
observed area minus reference area), standard deviation (from
the average observed area), reproducibility and RMS is possible
in tables (Table 3, Table 4). Relationships between
reproducibility, RRMS and the test area are shown on the
diagrams (Figure 5 and Figure 6). Mean value of reproducibility
for PAN is equal: 13.3% and respectively RRMS: 21.6%. Better
results were obtained for RGB, mean reproducibility is equal:
8.9% and respectively RRMS: 12.7%. Photointerpretation of
PAN was biased more than RGB (see Table 3 and Table 4).
Mean bias on PAN for test areas 1,3,4,6 (digitized impervious
surface) was -18% and for test areas 2, 5 (digitized pervious
surface): +12.3% (average error of Ain Table 3 is 16%). It
means that in photointerpretation on PAN impervious surfaces
were underestimated in compare to the reference. This
relationships is not observed on the RGB, bias is smaller than
on PAN, varies from minus to plus values and mean of RRMS
is 2.9% (Table 4). Relationships between the RRM and test area
are presented for PAN and RGB on the diagrams, respectively
on Figure 5 and Figure 6. The bias, appearing in the PAN
observations, is possible to be seen on the figure Figure 5 as a
shift of RRMS up to the reproducibility (ISO). This
phenomenon is not, in this scale, observed in RGB observations
(compare Figure 5 and Figure 6). For example, RRMS for the
test area 3 and 4 is almost equal to the reproducibility, and the
bias is small.
Before statistical analysis, values of RMS for all test area and
all pixels, were combined into 10 groups depending of the mean
imperviousness factor from 0 to 100% by each 10%. Bias, and
RMS of imperviousness factor are presented in 10 groups for
PAN (Figure 8) and RGB (Figure 9) observations. Maximum
bias in PAN measurements was slightly more then -30%, it
means that operators underestimated impervious areas.
Absolute RMS increases with increasing of imperviousness
factor even more then 35% on IKONOS PAN. IKONOS
PANSHARP (RGB) allowed obtaining much better results
(Figure 9), small, neglected bias and RMS in some cases only
slightly more then 10%.
Finally, the RRMS of impervious surface area was calculated in
each pixels for comparison to the RRMS of impervious surface
area calculated for all test area (compare Figure 5, 6 and Figure
10,11). In the analysis the value of the impervious area should
be take into consideration, Fig, 5, 6: 10 000 - 60 000 sq m, and
Figure 10, 11:0- 900 sq m.
The figures: Figure 7 - Figure 11 show the accuracy analysis of
data, in simulated Landsat pixels (30m). Relation between the
RRMS, relative area error of impervious surface in the pixel,
and reference area of impervious surface is presented on
diagram (Figure 7). Usually, the absolute area error increases
with increasing of the measured area, but simultaneously
RRMS is decreasing. The relationship is valid for the area from
digitalisation of remote sensing images or surveying. It is easy
to be observed in the case of cadastre parcels. Our object of
interest was however other. There are in fact many polygons,
analysed totally in test areas (Figure 5 and Figure 6). Therefore,
the tendency of decreasing of RRMS with the area is slightly to
be noticed on the diagram (Figure 5) and better in the RGB
observations (Figure 6). The total measured impervious areas in
test areas are varying between 10 000 - 60 000 sq m, and the
RRMS varies from 6%, in RGB, to even 25 % in PAN
observations.
Analysing the same relationship in grid of 30m, we could obtain
even huge error for small areas heading to zero, because in this
case RRMS heads to infinity. Maximum area in grid 30m is
9000 sq m, so in Landsat pixel impervious area varies from 0
to 900m, and RRMS varies in range of zero to infinity. But the
decreasing tendency with area is noticeable on Figure 7 , even
in the “cloud of points”. On the other hand we use results of
photointerpretation as a test area in Landsat image classification.
Therefore, we compared imperviousness factor calculated for
pixel on the basis on all observations and on the reference.
Usually we assume result of one photointerpretation as a
reference. In our research we have 18 measurements because 6
operators digitised the area 3 times.
Figure 8 Relation between the absolute errors of
imperviousness factor: RMS, bias and the imperviousness factor
combined into 10 groups, PAN
Figure 9 Relation between the absolute errors of
imperviousness factor: RMS, bias and the imperviousness factor
combined into 10 groups, RGB