Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986 
541 
An efficient classification scheme for verifying lack fidelity 
of existing county level findings to cultivated land cover areas 
Yang Kai, Lin Kaiyu, Chen Jun & Lu Jian 
Wuhan Technical University of Surveying and Mapping, China 
ABSTRACT: In order to verify the lack fidelity of existing findings to cultivated land 
cover areas in county level for the purpose of rural economical planning, a special com 
puter aidded classification scheme has been tested and is introduced in this paper. The 
schene consists of three main aspects, namely the preprocessing of IANDSAT MSS multitem 
poral image data, the selecting of optimum feature image set and the classification pro 
cessing with auxiliary ground height information. Based on this schene, the improvement of 
classification accuracy amongting to 8% was obtained, and the man-male mistakes in existing 
findings of testing county have been checked out. 
1. Introduction 
In order to meet the needs of planning for develo 
ping- rural economies, it is neccesory for the author 
ities of different level's goverments to know the 
actual situation of land cover types. But unfortu 
nately, the existing county level findings in ques 
tion in China are mostly not reliable enough for 
planning purpose. This is because not only of the 
technical reasons, but mainly of the historical and 
man-made reasons, for instance, the definition of 
distal ce or area measure for cultivated land was not 
corresponding to the couversion between Chinese rule 
and metric one; and the peasents used to report the 
cultivated land areas less than actual one for paying 
less agricultural tax. However, the uncorrectness of 
existing findings to cultivated land was rather 
systenatical, especially within a local regin. So 
it becomes possible to estimate the actual cultiva 
ted lend areas all over the regin, by determining 
the correction factor from some sampling area, e.g. 
a typical county. This is the main purpose of our 
subject. 
Under above description, the next question could 
be hov can detormine the actual land cover situation 
in county level. There are options, the one is based 
on cor ventional photogrammetry procedure which is 
accurate but time consuming; the other is based on 
satellite remote sensing techniques which is less 
accurate than the former but rather time saving. We 
have chosen the latter one. Of course, the problem 
following our choice is to design a scheme to impro 
ve the classification accuracy based on satellite 
images, which is the key point that our research 
subject was going to solve. 
Our classification scheme can be cha.racterrized 
by following key words: LANDSAT images; multitem 
poral data, image transformation in multispectral 
domair, feature image selection, and computer-aided 
classification with auxiliary ground height informa 
tion. The principle of the scheme and corresponding 
experimental results will be introduced as follows. 
2. Pieparation of basic image data 
The Xian-ning county of Hubei province in China was 
selected as sampling and testing area. Three types 
of basic image data, were neccesary for performing 
the classification scheme, namely: 
1. origional LANDSAT MSS images within the county 
bound ery. 
In this experiment two-temporal LANDSAT MSS images 
with total 8 bands were used, which were seperately 
imaged on Oct. 16, 1978 (digital image CCT) and on 
June 16, 1979 (negtives with scale 1:3*36 million). 
Three steps was needed to form the required image 
windows as shown in figure 1 (taking one band as 
example). 
(a) read (or scan) a rectangular image window covering 
testing county area from the raw image carrievs; 
(b) digitally rectify the image window based on 
selected control points; 
(c) digitally mask the rectified image and form the 
image windows within the county boundary. 
2. digital terrain model (DTM) image 
The DTM image in our exeriment was created by 
direct reading the height value of DTM grid with 
200m intervals on existing topographic map. Then 
it is densified in computer by interpola.tion method 
for each DTM pixels which have the same geometric 
resolusion as origional image ha.s. 
3* real ground feature encoding image in selected 
sampling areas 
vVithin the testing county area six sampling area.s 
selected, which were distributed at typical parts 
of the county, seperately with the height range; 
(1) 0 - 20m, (2) 21 - 50m, (3) 51 - 100m, (4) 
101 - 200m and (5) 201 - 900m. Each of them has 
60 x 60 pixels whose class attribute has been 
Figure 1. image window within county boundary.