In the study area a random sample consisting of 
1454 pixels yields 
  
  
  
là | [0.647 0.265 0.364 0.096 0.183 0.310 | 
fa 0.052 0.122 0.068 0.04 0.083 0.092 
ps | _ | 0.162 0.132 0.296 0.079 0.117 0.117 
pa | | | 0.07 0.291 0.148 0.581 0.383 0.315 
bs 0.011 0.032 0.04 0.119 0.150 0.075 
[js | | 0.059 0.159 0.085 0.083 0.083 0.092 | 
| 0.1940 | 
0.1467 
0.1229 
0.1838 (5) 
0.0419 
| 0.3110. 
  
  
and the estimated proportions are 
Far here] 042206 | 
f» 0.07340 
p - | | _ | 01401 
b. 0.3181 
fs 0.0614 
| ps | | 0.0003 | 
  
  
  
  
The estimates deviate from the true proportions 
known from the field survey only by a small margin. 
The reason for is the large size of the sample and samp- 
ling with the sampling unit pixel. But, the coefficient 
of reliability K is very low. In case of ML classification 
of the raw SAR data both classifiers are not correlated. 
The sampling accuracy depends almost entirely on the 
true classifications. As K approaches 0 the variance 
depends only on the size of the sample n (Equation 7). 
If K rises, the left term of Equation 7 becomes smal- 
ler with (1 — K). But the right term must be added. 
This term is, however, very small because of large N. 
Thus, some expensive true observations are replaced 
by a large number of misclassified observations. 
Larger K values are found in the results of the ma- 
Jority filtering. In that case the results of the ML clas- 
sification of the entire imagery improve the accuracy 
of the estimates from the small sample with true clas- 
sifications. But the majority filtering rises the entire 
cost of the sampling. 
In the sampling scheme described above a full image 
classification was used. The full frame classification 
can be replaced by a sample with N pixels, where N >> 
n. In that case the n members of the sample are called 
subsample. With a given coefficient of reliability K 
and the probability of misclassification 0, both n and 
International Archives of Photogrammetry and 
N can be optimized for given accuracy requirements 
Tenebein (1972) describes the procedure. 
3. CONCLUSIONS 
From the results of this study can be concluded, 
that the double sampling has the capability to cor. 
rect misclassification errors of the ML classification, 
However, the coefficient of reliability is with ERS.1. 
SAR-data very low. This means that a comparatively 
large and expensive sample with true classifications is 
needed to get sufficient accurate estimates of the e. 
ror probability matrix and other parameters. The re. 
sults apply to the sampling unit pixel. The sampling 
performance can be improved by a carefully trained 
classifier, use of systematical sampling and exclusion 
of non-agricultural areas from the sampling. 
  
The author thanks ESA for providing the ERS-1 SAR 
imagery 
References 
Bross, I. (1954): Misclassification in 2*2 tables. Biometrics 
(10), 478-486. 
Card, D. H. (1982): Using known map categorical mar- 
ginal frequencies to improve estimates of thematical 
map accuracy. Photogrammetric Engineering and Re- 
mote Sensing A8, 431—439. 
Congalton, R. G. (1991): A rewiew of assessing the accura- 
cy of classifications of remotely sensed data. Remote 
Sensing of Environment 37, 35-46. 
Czaplewski, R. L. and Catts, G. P. (1990): Calibrating 
area estimates for classification error using confusion 
matrix. Technical Papers 1990 ACSM-ASPRS Annu- 
al Convention 4, 431-440. 
Czaplewski, R. L. and Catts, G. P. (1992): Calibration 
of remotely sensed proportion or area estimates for 
misclassification error. Remote Sensing of Environ. 
ment 39, 29-43. 
Smiatek, G. (1993): Erfassung der Flächennutzung mit 
Hilfe von LANDSAT/TM-Stichproben. Dissertation 
Institut für Navigation der Universität Stuttgart 
Stuttgart. 156 S. 
Smiatek, G. (1995): Sampling thematic mapper inagery 
for land use data. Remote Sensing of Environment 
52, 116-121. 
Tenebein, A. (1970): A double sampling scheme for esti 
mating from binomial data with misclassifications. 
JASA 65(331), 1350-1361. 
Tenebein, A. (1972): A double sampling scheme for 
estimating from misclassified multinomial data with 
applications to sampling inspection. Technometrics 
14, 187-202. 
626 
Remote Sensing. Vol. XXXI, Part B7. Vienna 1996