ul 2004
——M—
curacies
] Figure
s after
dard
m real
for each
s have
| various
sampling
sults go
ing (CS)
qi.
by 3 is
he means
nples for
values in
iat in this
age then
sults are
stimating
value of
ven with
Its of this
se images
erages of
viation of
rsa. Then
neither of
th cluster
xt suitable
y.
image number
ies after
ndard
rom real
for each
ts have
chema
amples tor
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
After previous experiments, it is possible to draw some new
graphs using the same values. In Figure 8 these graphs has been
shown. This figure is consisting of five graphs for each type of
images used in this paper, in one diagram. In all of these graphs,
x-axis introduces sampling methods with numbers of 1 to 5 that
these are accordance to: SRS, STRAT, SYSTEM, SSUS and
CS, and y-axis in graphs 8(a) shows the values of difference of
means of overall accuracies with real values and in graphs 8(b)
is the values of standard deviations from means.lt is seen that in
all of images systematic sampling is clearly better than the other
methods because of minimum values of mean differences
(Figure 8(a)) and almost most uniform and small standard
deviations (Figure 8(b)). Then with assurance this sampling
schema produces suitable results in all of these images.
sampling schema
sampling schema
Figure 8. The difference of average overall accuracies after
stability with actual overall accuracies (a) and standard
deviations from means (b) using all of sampling methods for all
of images with 30 times repeating, each color shows results of
using all of methods in one image
On the other hand with graphically displaying the graphs of
mean differences and standard deviations related to all of the
sampling methods (graphs of Experiment#1 to 5) according to
Figure 9, it is obvious that in image #2 that has small image size
and small field size, in all of methods appropriate results are
achieved. This image has the least differences in resulted mean
differences (Figure 9(a)) and resulted standard deviations
(Figure 9(b)), using all of the sampling schemas.
03 fe
0.24
t
image number image number
Figure 9. The difference of average overall accuracies after
stability with actual overall accuracies (a) and standard
deviations from means (b) in all of images using 5 sampling
methods with 30 times repeating, each color shows results of
one sampling methods in all of images
In addition to previous experiments and for confirmation of
results, with sample size equal to 1000 as a confident sample
size in each of sampling schemas, with one , 10 and 30 times
caleulating of overall accuracies several graphs has been
produced that because of similarity of results with previous
graphs, it has been restrained from displaying of them. However
it is important to be realized that with one time sampling and
calculating of overall accuracy, the previous results are not
achieved, because the problem of the chance of obtaining an
unrepresentative sample which is always possible doesn’t
eliminate. Then for assurance, these computations have to be
1041
done in repetitive manner and average of values must be
considered as the final result (Hashemian. 2004).
^
5. CONCLUSIONS
With due attention to experiments in this paper, several results
are achieved. Here these results are summarized:
l. With 50 to 70 samples for each class in all of images
used in this paper, the results of overall accuracies go
towards stability and the best results. Then 100
samples for each class in image can be a reliable
sample size.
2. In images with large fields, SRS and STRAT methods
overestimate the overall accuracy, but with SYSTEM
and SSUS methods this problem doesn't exist.
3. SRS method has the best results for images with small
fields used in this paper.
4. STRAT method produces better results in smaller
study areas.
SYSTEM schema is the more efficient method in large
images with large fields.
6. SSUS schema is a more suitable method for accuracy
assessment in images with large size.
7. CS schema is not a suitable method in neither of the
images used in this paper, and don’t have the good
results on comparing with the other methods.
8. Totally, the sampling schemas with systematic basis
achieve rather more suitable results in all of images
used in this paper.
9. In image with small image size and small field size
(images with good distribution of classes), it is
expectable that the good results are produced from
each of these sampling schemas.
10. For achieving the best results, computations of
accuracy assessment have to be done in repctitive
manner and average of values be considered as final
result.
Cn
REFERENCES
Abkar, A.A. 1999. Likelihood-Based Segmentation and
Classification of Remotely Sensed Images, A Bayesian
Optimization Approach for Combining RS and GIS, PhD-
thesis, University of Twente, Enschede / ITC, Enschede, 132 p.
Congalton, R.G., 1988. A Comparison of Sampling Schemes
Used in Generation Error Matrices for Assessing the Accuracy
of Maps Generated from Remotely Sensed Data,
Photogrammetric Engineering & Remote Sensing, 54(5), pp.
593 — 600.
Congalton, R.G., 1991. A Review of Assessing the Accuracy of
Classification of Remotely Sensed Data, Remote Sensing
Environ, 37, pp. 35 - 46.
Fatemi, S.B., 2002. Model-based Image Analysis of Remotely
Sensed Images: An Agricultural Case Study, MSc Thesis, KN
Toosi University of Technology.
Hashemian, M.H., 2004. Study of accuracy assessment
techniques for classification of remote sensing data. MSc
Thesis, KN Toosi University of Technology.
