la. A GIS file. Each 
field is digitized as 
a polygon with a unique 
polycentre identifier. 
Figure 1. A GIS file and 
1 
1 / 
— 
1 
1 / 
$ 
1 
_^1 _ 
! 
I 
w. 
1 / 
I 
— 
r>' 
I 
m 
/1 
! 
/ 1 
( 1 
; 
lb. The corresponding 
image file(s) and the 
corresponding pixel(s) 
to each polycentre in 
the GIS file. 
corresponding RSS file. 
is selected and the crop type of this sample 
is identified, by using remotely sensed data. 
3.2 Approaches to the GIPS/ALP concept 
3.2.1 In the GIS data base each polygon has a 
unique identifier, i.e. the polycentre. And, 
the coordinates and attributes (e.g. the crop 
type, the acreage, the crop yield...etc) of each 
polygon are stored at a high level of accuracy 
and up dated as required, which can be retrieved 
directly and easily if a proper data structure 
is used. 
Field 
Identifier 
Location of 
Polycentre 
Attributes 
1 
xl yl 
la lb lc 
2 
x2 y2 
2a 2b 2c 
3.2.2 If the 
position of 
each pixel of the 
image 
is known to 
a certain 
level of accuracy, 
the 
corresponding 
pixel to 
each polycentre in 
the 
GIS can be identified and extracted at the 
leve 1 
of accuracy. For a multiband or multidate images, 
registration techniques can be used to help define 
i:he fieldcentre pixels. It is also possible to 
estimate the purity of the polycentre pixel to 
see whether the extracted pixel is a pure pixel, 
composed of one major cover type, or a mixed 
pixel, composed of more than one major crop types. 
Figure 2. A flow chart showing the approaches to 
the implementation of classification in GIPS/ALP. 
containing this particular fieldcentre is identified 
i.e. the crop type of the fieldcentre pixel is 
assigned to the whole field. By comparing this 
newly derived crop type with that originally 
stored in the data base, the change, if any, 
of 
the 
cover type 
can be 
detected. The 
acreage 
of 
each 
crop type 
in 
the 
study area is 
derived 
or 
computed from 
the 
GIS 
data instead 
of the 
RSS 
data 
to achieve 
the 
accuracy at the GIS 
accuracy 
leve 1. 
Identified Change 
F.I. L.PC. A. L.PCP. P.PCP. New Cover Condition 
1 x 1 y 1 lalblc L1S1 P Cl N (no ch) 
2 x2y2 2a2b2c L2S2 M C2 C (chngd) 
3.2.5 The field check or ground truth verification 
if it is to be applied can be directed to those 
areas which appear as mixed pixels or are detected 
as changed cover. These location can be accurately 
identified from a GIS generated map, on which 
the suspected areas are marked automatically 
on a user specified base map. 
3.2.6 The resulted outputs can include (i) statistic 
tables and diagrams, (ii) maps, and (iii) images, 
whatever available functions in both the geographic 
information system and remote sensing system. 
Field Location of Location 
Identifiar Polycentre Attribute PC Pixel 
T' xl yl la lb lc LI SI 
2 x2 y2 2a 2b 2c L2 S2 
Purity 
PC Pixel 
P (pure) 
M (mixed) 
3.2.3 If the calibration data or training data 
collected in their primitive form and stored 
in GIS are structured properly, they can be 
retrieved for deriving particular data sets to 
meet a particular application. The derived data 
sets are used in turn to assist the classification 
of the extracted pixels along with some other 
ancillary data available in the data base. One 
of the derived data sets of particular importance 
is related to the a priori probability distribution 
of each existing cover type for a particular 
time in the study area. This can be computed 
from the archival land cover file in the data 
base. 
3.2.4 As soon as the crop tyr 2 of the field centre 
pixel is identified, the crop type of the field 
3.3 On the implementation of classification 
Figure 2 is an example to demonstrate the specific 
classification approach in GIPS/ALP. A brief 
discussion is given here. The geographic information 
system and remote sensing system are the two 
supporting pillars of GIPS/ALP. The geographic 
information system is for handling the fully 
structured data base; the remote sensing system, 
for collecting the most uptodate spatial data. 
At the training stage, four tasks are involved, 
namely (i) locating the fields with known land 
use type during the time images were taken, (ii) 
extracting image data corresponding the fields 
located in the previous step, (iii) extracting 
calibration data from GIS for modifying the image 
data, and (iv) obtaining the training statistics 
needed for a particular classification algorithm. 
The a priori prabability distribution of each 
cover type in study area can be estimated from 
the archival land cover file in GIS. 
The unknown image data, of which a value stands 
for a field centre pixel, are derived from RSS 
file and stored in a proper structure. 
The result 
the inputs tc 
the cover t] 
subsequent i 
cover data 
land cover fi 
4. UNIQUE ADV. 
This approac] 
volume which 
resolution i 
pixels of e; 
be kept cons 
for a define 
in a certaii 
polygon data 
that of the 
The classif 
improved due 
and the hig 
available in 
by this systei 
too. 
The geometr 
in the GIS ’ 
higher than 
sophisticate!} 
statistics r 
on basis of 
system. 
5. CONCLUSION 
Both geograp 
sensing systi 
spatial data, 
special featx 
image process 
They are the 
and geometric 
REFERENCES 
Allan, J.A. 
sensing. In 
Internations 
Remote Sens: 
Anderson, J.I 
Witmer 197 
classificat: 
data. Geo] 
964, usgs. ; 
Marble, D.F. 
information 
R.N. Coweli 
vol.I, p.92: 
Tseng, W.T. 
of computer 
and the st 
land covers 
& Service Ox