tigated is 
al DBMS 
and does 
he data. In 
trinsically 
'S richer. 
tem before 
vely and 
AD in our 
e, because 
built into 
processing 
' indexing 
queries in 
rocess of 
1lection of 
raction of 
s created 
analyses, 
elligence. 
ostraction. 
)deling is 
fication is 
y may be 
juery may 
'hat is the 
ery easily 
le. While 
orm data 
only limit 
a and the 
jects. For 
lerive the 
arameters 
ssification 
; such as 
al editing 
>sentation 
odeled as 
base for 
ability to 
al DBMS 
>, one can 
; indexed, 
kind of 
is. Value 
aditional 
applications, but will not work for the type of 
large objects such as remotely sensed data. First, 
queries are issued to retrieve images by their 
content, e.g. to find all images that have Lake 
Ontario. This search requires indexes on the result 
of a classification function (on virtual attributes) 
and not on raw images. Second, indexing functions 
for images and text often return a collection of 
values for which efficient access is desired. For 
example, a keyword extraction function might 
return a set of relevant keywords for a document, 
and the user desires indexing on all keywords. 
To be able to retrieve large object stored in 
databases we must associate identifying 
parameters with each of the large objects. It is 
through these parameters that we can select stored 
large objects for display, comparison, queries and 
further analysis. Primary parameters are produced 
when large objects are obtained and describe the 
large object's acquisition procedures and it's related 
properties. For fast systematic analyses and access 
we need to retrieve images based upon their 
contents. The pixels that make the image are 
rarely suitable for direct search. If we can provide 
secondary parameters indicating the location and 
type of objects seen on the image, the database 
searches could be enabled that directly serve the 
end-user’ objectives. Modern GISs without such a 
retrieval capability will not be justifiable. 
However, the majority of secondary parameters 
used today, when available at all, is entered by 
humans after visually scanning an image 
[Wiederhol 1989 and Chang 1992], This 
observation motivates us to use abstraction 
techniques to generate the secondary parameters. 
Although a completely automated abstraction is 
our long term goal, only semi-automatic analysis 
and classification are used in this research to 
generate the DAD that is used as secondary 
parameters in the content based queries. 
6. PERFORMANCE EVALUATION 
Before we look into our examples, let us first 
consider how big some typical spatial objects may 
be in GISs: 
a) a scanned 8 bits map data 10, 000 by 10,000 pixels 
without compression is about 100 MB; 
b) a SPOT panchromatic image 6000 by 6000 is 
about 36 MB; 
¢) a Landsat Thematic frame(7 bands) is about 300 
MB; 
With more than 1000 large objects in a GIS 
database, we can imagine how important it is to 
effectively and efficiently store, transfer and 
process these large objects. The methods used to 
handle these large spatial objects will mainly 
decide the success of GISs. 
6.1 Application Examples 
Without practical examples with real GIS queries 
to large spatial objects, we can only speculate on 
the efficiency of the methods we presented. In the 
previous section, we have analyzed types of 
queries a user may actually run against large 
spatial objects and shown how often the abstracted 
data can be used to answer user's queries. In this 
section we will demonstrate the difference in 
system performance when our approaches are used. 
To show the performance improvement using our 
proposed methods, we use a data base that consists 
of ten large image objects(about 50 MB), 5 MB 
vector and attribute data(including DRAD and 
DAD). We consider this database very small 
compared to real GIS data bases. This database is 
stored in the disk on a SUN SPARC 2 station with 
32 MB memory. We post the following five 
different types of queries on large spatial objects in 
the database: 
1) Overview: display an image in a window; 
2) Zoom/pan: Zoom/pan an image. 
3) Thematic selections: Show me the percentage of 
"Water Coverage" in a large image object. 
4) Range Query: Show the image that is contained 
in a polygon ABCD. 
5) Automatic analysis: Detect all edges in a 
Landsat image using a predefined method" 
These selections are often followed by further 
processing steps in practical applications to refine 
the selections. This is an important feature in large 
object queries. We will not discuss this in this 
paper because this refinement is very specific for 
each particular application. 
6.2 Query Cost Comparisons 
a) Query One: Image Overview 
Using the traditional method, we would need to 
read all image data from the database into 
memory, zoom out the image to make it fit in the 
window and then display image data on the 
window. This will take several minutes to 
complete. 
In the proposed method, we just retrieve the 
overview image that is generated using lower data 
quantization and spatial resolution at the time 
217