International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
previous methods. This case demonstrate that in this method 
samples have better distribution in image, then with less 
samples can achieve appropriate results. Also there was not 
overestimating tendency problem in these four study areas. This 
subject also is a confirmation about efficiency of this sampling 
schema in all of these images. In view of graphs of means and 
standard deviations in Figure 5 this subject is delineated that in 
images with large image size i.e. image #3 and image #4, means 
are closer to real values (Figure 5(a)) whereas standard 
deviations are appropriate (Figure 5(b) and Figure 5(c)), then it 
scems that this method is more efficient in large study areas. 
If results of image #3 and image #4 in Figure 5 are compared 
together, it is seen that image #4 with larger field sizes has 
better results ie. smaller difference of means (Figure 5(a)) and 
smaller standard deviations (Figure 5(b) and Figure 5(c)), then 
in addition to large image size, large object size is a factor for 
achievement of good results. In the case of real image, this 
matter is confirmed: this image with smaller image size than 
image #3 and because of larger field size has results better than 
image #3. 
vverall accuracy 
  
differenee et 
  
Figure 5. The difference of average overall accuracies after 
stability with actual overall accuracies (a) and standard 
deviations from means (b), and standard deviations from real 
values (c), using SYSTEM method (each sampling schema for 
each sample size has been repeated 30 times and the results 
have been averaged) 
With comparing Figure 5(a) with same graphs in previous 
methods (Figure 2(a) and Figure 4(a)), it is obvious that the 
range of values in this graph is lower than the others. Then this 
method is more appropriate from other methods. The results of 
computing overall accuracies in real image show that the results 
go towards stability after almost 50 samples for each class. 
4.4 Experiment #4: Investigation of Stratified Systematic 
Unaligned Sampling (SSUS) schema 
In this sampling method the same result as STRAT sampling 
about the stability of the means of overall accuracies was 
achieved. In other words with 50 sample for each class and 
without considering the size of images, the stability of results 
was acquired. The reason for this matter is that SSUS method 
has either random or systematic characteristics. Then results of 
SSUS method not as SYSTEM method with small sample size 
for all images and not as SRS method with large sample size for 
large images, but with almost 50 samples for each class go 
towards stability. 
There is not overestimating problem in graphs of this method. 
This subject is also visible in comparing of graphs in Figure 6, 
because standard deviations from means (Figure 6(b)) are close 
to standard deviations from real overall accuracies (Figure 6(c)) 
in all of images. This subject is a reason for efficiency of this 
sampling schema in all of these images. 
The best results have been produced in image #3 and image #4 
that are images with large image size. These results are the best 
because of smallness of differences from real overall accuracies 
(Figure 6(a)) and standard deviations (Figure 6(b) and Figure 
6(c)). 
     
É tT Tr Mp pU TTI 
2 \ 3 ; A 
\ B / / v. 
\ fed fc 3 { 
t PA 
x / \ / f \ / 
\ \ / 
\ US 
z elt Y y Lo Lt A 
a ! » 1 i 3 1 4 35th 
i 
a awe urn be: b ' e age manber 
Figure 6. The difference of average overall accuracies after 
stability with actual overall accuracies (a) and standard 
deviations from means (b), and standard deviations from real 
values (c), using SSUS method (each sampling schema for each 
sample size has been repeated 30 times and the results have 
been averaged) 
The results of computing of overall accuracies with various 
sample sizes with stratified systematic unaligned sampling 
schema for the real TM image showed that the results go 
towards stability after 50 samples for each class. 
4.5 Experiment#5: Investigation of Cluster Sampling (CS) 
Schema 
In this experiment CS schema with cluster shape 3 by 3 1s 
investigated. In this sampling method in all of images the means 
of overall accuracies go to stability after 60 or 70 samples for 
each class. These values for sample size are the largest values in 
comparing with the other methods. This case shows that in this 
method samples have not suitable distribution in image then 
with further samples can achieve the some better results. 
In graphs of image #1 the overestimating of results are 
observed. In the case of image #4 although the overestimating 
tendency don't exist, this image has the maximum value of 
standard deviation (Figure 7(b) and Figure 7(c)) even with 
respect to graphs of the other methods. Then the results of this 
image are not appropriate too. Totally neither of these images 
has preferable and suitable results because if averages of 
calculations (Figure 7(a)) are small value, standard deviation of 
that (Figure 7(b) and Figure 7(c)) is large and vice versa. Then 
this method is not a suitable sampling schema in neither of 
images in this paper. The reason for this is that with cluster 
sampling, the distribution of samples in images is not suitable 
and the samples don’t represent the population properly. 
S sigmOV 
  
    
image number page number 
Figure 7. The difference of average overall accuracies after 
stability with actual overall accuracies (a) and standard 
deviations from means (b), and standard deviations from real 
values (c), using CS method (each sampling schema for each 
sample size has been repeated 30 times and the results have 
been averaged) 
The results of the real TM image with this sampling schema | 
showed that the results go towards stability after 70 samples for 
each class. 
1040 
In 
gr 
sh 
im 
thi 
C: 
nu 
all 
mé 
(F 
de 
sci 
  
de 
of 
On 
me 
san 
Fig 
and 
ach 
diff 
(Fi 
  
  
de 
Imi 
In à 
Tesu 
size 
calcı 
prod 
grap 
It is 
calci 
achi 
unre 
elim