2004 
—— 
ess of 
e than 
range 
point 
. This 
‘anges 
'anner 
ed for 
patch 
The 
from 
of the 
The 
‘range 
when 
ore be 
nse of 
t of a 
which 
canner 
1 close 
)wman 
ent of 
vithout 
ipment 
unding 
e Rs ~ 
portant 
rveying 
jour of 
| Laser 
of the 
ng and 
Corfu, 
flecting 
Laser 
ospatial 
\nalysis 
rnal of 
)9. 
Colour 
ccuracy 
ospatial 
. 2003. 
per XIX 
October 
STUDY OF SAMPLING METHODS FOR ACCURACY ASSESSMENT OF CLASSIFIED 
REMOTELY SENSED DATA 
M.S.Hashemian ", A. A.Abkar", S.B.Fatemi ? 
* Department of image processing, National Cartographic Centre (NCC) of Iran, Meraj Av., Azadi Sq., Tehran, Iran, 
hashemian@ncc.neda.net.ir 
? Faculty of Geodesy and Geomatics Eng., KN Toosi University of Technology, Vali Asr Street, Mirdamad Cross, 
Tehran, Iran, abkar@itc.nl, sbfatemi@yahoo.com 
KEY WORDS: Remote Sensing, Land cover, Classification, Sampling, Accuracy 
ABSTRACT: 
Accuracy assessment is an important step in the process of analyzing remote sensing data. It determines the value of the resulting 
data to a particular user, i.e. the information value. Remote sensing products can serve as the basis for political as well as economical 
decisions. Users with a variety of applications should be able to evaluate whether the accuracy of the map suits their objectives or 
not. In the conventional accuracy assessment an error matrix and some accuracy measures derived from it are used. An error matrix 
is established using some known reference data and corresponding classified data. There are various factors that affect the 
performance of the accuracy assessment by influencing the error matrix through out the ground truth data collection. In practice, the 
techniques are of little value if these effective factors are not considered. In this paper the necessity considerations for accuracy 
assessment including the sampling schemas and the sample size for these sampling methods are studied. Also the factors that affect 
selecting and applying appropriate sampling schemas and sample size are investigated. For this study numbers of synthetic images 
and one real image and some reference data are used. Sensitivity of the various sampling schemas has been investigated using the 
synthetic images and using the real image the obtained results have been confirmed. The results represent that depend on specific 
conditions such as type and size of the study region and object characteristics, different sampling methods and sample sizes are 
preferred. 
1. INTRODUCTION simulated datasets is one means of overcoming this problem. 
One real image is used for confirmation of the obtained results 
The value of the classified map is clearly a function of the from synthetic images. The synthetic images are: 
accuracy of the classification. A responsible use of the stored * Image number ! is a quadrangle area with three 
geodata is only possible if the quality of these data is known. generated bands and 200 pixels in rows and 200 pixels 
Then accuracy assessment is an important step in the process of in columns and has 10 classes and large fields. There 
analyzing remote sensing data. In the conventional accuracy are 10 fields in the image totally and the average area of 
assessments an error matrix is established using some known the fields is 4000 pixels. 
reference data and corresponding classified data and some e Image number 2 is a quadrangle area with three 
accuracy measures derived from it are used for accuracy generated bands and 200 pixels in rows and 200 pixels 
assessment. Assessing the accuracy of land cover map in columns and has 10 classes and small fields with 
generated from remotely sensed data is expensive in both time good distribution in total of image. The average area of 
and money. Obviously, a total enumeration of the mapped areas the fields is 99 pixels. 
for verification is impossible. Sampling, therefore, becomes the * Image number 3 is a quadrangle area with three 
means by which the accuracy of the land cover map can be generated bands and 512 pixels in rows and 512 pixels 
derived. Using the improper sampling approach can be costly in columns and has 10 classes and small fields with 
and yield poor results. To wit, poor choice in sampling scheme good distribution in total of image. The average area of 
can result in significant biases being introduced into the error the fields is 119 pixels 
matrix, which may over or under estimate the true accuracy. In e Image number 4 is an image with three bands (R, G, B) 
this paper the suitability of five sampling methods including and 768 pixels in rows and 576 pixels in columns. This 
simple random sampling, stratified random sampling, image is a RGB-CCD image of a model of agricultural 
Systematic sampling, stratified systematic unaligned sampling fields: thaicineans: a color CCD: video camera has 
and cluster sampling are investigated with some synthetic produced three bands (R,G,B) of it, this image has 9 
images and one real i image. classes consist of two types of roads, five different crop 
fields, farmhouses and forest (Abkar, 1999). The 
average area of the fields is 29491 pixels. 
2. DATA USED 
For generation of first three synthetic images firstly it has been 
In this paper, in order to have a controllable process and reliable assumed that our data has a known normal distribution. Then by 
investigation, some experiments are performed using generated — having this assumption, some values for mean vectors and 
Synthetic images. Also, for comparison of the different covariance matrices have been considered and using them the 
üpproaches that have been hindered by the scarcity of satellite pixel values, on the basis of normal distribution have been 
imagery for which full reference data are available, use of generated. The values of means and standard deviations of each 
1037