when the large image is loaded into the data base. 
This would reduce the time of the image overview 
into seconds. 
Table 2 gives the result when the query is applied 
to a 2000 by 2000 pixels 24-bit image and an 4-bit 
overview image of 512 by 512. The overview image 
is generated by an edge preserved reduction 
technique. From this result, it has shown about 10 
times time saving. 
b) Query Two: Zoom/pan an Image 
Using Binary Large Object without image tilling 
and reordering, we need to seek through the whole 
image data, read the needed part of data, and then 
transform the image data using a required 
zoom/pan parameter. 
With proposed techniques such as tilling and data 
reordering, we tile the image first and reordering 
the tiles according to their spatial relationship. In 
case of zoom/pan, we always load 8 nearest tiles of 
the current tile into memory and then zoom/pan is 
a matter of deciding which pixels are "visible" for 
current query. 
Using an 1-bit image with size of 8000 by 8000 
pixels and a tile size of 512 by 512, the query cost 
using new method is shown in Table 2 and is about 
30 times fast. 
c) Query Three: Thematic Selection 
Without any DAD stored in the database, we need 
to retrieve the image using the image name. Then 
we read all image data into memory and use a rule 
to derive the "water coverage." Last, we calculate 
the size of the "water coverage" by counting the 
total pixel numbers. This is a very slow and tedious 
query processing. 
With our proposed method, we derive all thematic 
data that include "cloud coverage" theme using 
image classification and store them as DAD in the 
data base at the time when the image is loaded 
into the database. In this case, we store the theme 
"cloud coverage" as a polygon map with attributes 
associated with each individual polygon. When a 
thematic query is issued, it is just a matter of 
retrieving the right polygon from the database. 
Table 2 gives the result when the query is posed on 
our test database against a satellite image with 
the size of 1024 by 1024 pixels and the quantization 
level of 24 bits. From the table, the time difference 
between two methods is 2:15. 
d) Query Four: Range Query 
Using Binary Large Object without any data 
abstraction, we need to do the following three 
steps: first, perform spatial searching to get all 
image objects in the polygon ABCD; second, 
retrieve all image objects into memory using the 
image ids; third clip each image object against the 
polygon and return the clipped data to the 
application. 
In the proposed method, we tile all image files 
into small files and build a Minimum Enclosing 
Rectangle of each tile into DAD in the database. 
Then any range query is just a query that retrieves 
those tiles that intersect with the polygon ABCD 
and mosaic them together. This not only 
significantly reduces the amount of data involved, 
but also improves the processing speed. 
Table 2 shows a result using a scanned 1-bit map 
with size of 8000 by 8000 and a polygon size of 256 
by 256 pixels. It can be seen that the query cost is 1: 
81. 
e) Query Five: Automatic Analysis 
Without abstraction, we need to execute a function 
that can detect the existence of edges on the 
original image and display the result. This may 
take several minutes to finish. 
Instead of doing this we execute the function 
against the abstracted image(4 times resolution 
reduction) in the database. We can get almost the 
same result, but with significant performance 
improvement. 
Using a 24-bit 2000 by 2000 scanned map as an 
example, the query cost using two different 
methods is shown in Table 2. It is easy to see that 
the suggested method is about 15 times fast than 
the old method. 
From these examples we have concluded that the 
suggested techniques can improve the GIS query on 
large objects significantly. The initial cost to set up 
the DAD and DRAD in GIS data base is relatively 
small when we consider that GIS database can be 
used by multi users at various times. 
7. CONCLUSIONS AND RECOMMENDATIONS 
This research has shown that various techniques 
can be implemented in GISs to improve system 
performance on large spatial objects handling. Our 
data model distinguishes DAD, DRAD and Row 
Data in large spatial objects and is implemented 
218 
  
using ar 
query's. 
and pro 
compre: 
browsin 
large ot 
have 
signific: 
especial 
The futi 
can aut 
DAD ii 
system 
more te: 
GISs usi 
Special 
througt 
useful f 
time to 
Stoneb