difficult to retrieve an image by its name or 
number(or other identifiers), but it may be for more 
difficult to do it through selection criteria that 
would be contained in the image. This may lead to 
very long research, or even sometimes to tedious 
image processing (e.g. for similarity retrieval). 
This observation leads us to consider a more 
powerful data model and processing methods that 
can be used to handle large objects. 
3. LARGE SPATIAL OBJECT QUERY PATTERNS 
Before we discuss new data model and processing 
methods that can be used to improve the system 
performance on large objects handling, we need to 
examine how users will make use of large spatial 
objects in GISs. First, many queries are concerned 
about the general information of large objects and 
these can be classified as requests of meta-data, 
i.e. information about large objects. The common 
characteristics of this type of queries are that 
queries touch a high volume of data, but the 
generated answers are tiny by comparison. With 
our model, it is not difficult to figure out that this 
type of queries can be easily answered using DAD 
or DRAD. 
The second type of query is a kind of data mining. 
Users try to query over a large set of objects in order 
to find a small set of specific data. For example, a 
user may want to find an image that has a certain 
percentage of the snow coverage. This type of 
queries does not require the exact and detailed 
information, but instead requires the "overview" of 
the special nature of a large object. In this case, low 
resolution images with certain rules specified in a 
query language to derive the required results are 
possible solution for this purpose. 
The third type of query displays a large spatial 
object on a computer screen in order to do image 
interpretation or image analyses. Due to the small 
size of computer screen, only a very tiny part of the 
large object can be displayed on the screen at a time 
with the original data resolution. In this case we 
can significantly improve the system performance 
by windowing the image first and then only render 
the small part of data on the screen. By buffering 
the nearby part of the image, we can also perform 
Zoom/pan in a very fast fashion. This can be very 
significant when we compare to the case where the 
whole image data needs to be read from a data 
base. 
The fourth type of query is data browsing. Users 
wish to perform data browsing because they are not 
sure which type of data they are dealing with. 
Class hierarchy or data base schema browsing can 
be very helpful to the new users. Useful summary 
information includes histograms showing data 
statistical feature, images outline showing the 
data extent and well-designed icons showing data 
characteristics. For GIS applications it is our belief 
that all these techniques are critical to the success 
of GISs when they are used to handle large spatial 
objects. 
4. OBJECT ORIENTED DATA MODEL 
Before we present our new data model, let us 
consider a simple example of a large image object. 
The image is a Landsat image with DRAD such as 
image dimension, the geographical location of the 
image origin, the band number, the pixel precision, 
the acquisition time, the control points used for 
image geometric rectification, the parameters used 
for radiation correction, the data quality, etc.. The 
DAD data for the Landsat image may be its 
histogram, classification result(a set of polygons 
with category numbers), a two dimensional array 
for the classification precision, a set of rules to 
derive a certain theme (for example, a rule like: 
"if a pixel value is large than 10 and less than 20, 
then the pixel can be classified as water".) It is 
difficult if not possible for the relational data 
model to handle this non primitive data (e.g. 
point, polygon, array), rules and data operation 
methods. However, they can be well handle by an 
object oriented data model [Zhou and Wilkinson, 
1993]. In the above example, we could define an 
object class Landsat image, with various 
attributes to store meta-data, with methods to 
present the correction models, and with class rules 
to present the data classification rules[Hughes, 
1993]. 
Two most often used methods to handle large 
objects in object oriented models are: first, the 
extended relational DBMSs use Abstract Data 
Types(ADTs) to support large and unstructured 
data such as images[Stonebraker, 1993]. It allows 
users to define their own functions to manipulate 
the ADTs. This method allows the system to be 
extended, but leaves the extensions to its users. In 
addition, the raw data is treated as unformatted 
byte strings and very little semantic information is 
maintained in the system. The second method 
stores large objects in user defined classes, and 
allows the data base administrator to define 
methods and other attributes for classes[Deux et 
al., 1991]. Usually both methods store raw data 
using Binary Large OBjects(BLOB), and provide 
the capability to store and retrieve them through 
a query language. 
In both methods the row data is treated as a whole 
214 
  
and no 
users. ] 
generall 
extende 
Therefor 
users mi 
and clus 
manage 
or opera 
To mak 
environ 
compre 
redunda 
investig; 
our dal 
patterns 
method 
strategi 
Using t 
implem 
help us 
schema] 
the data 
5.L 
Many ci 
to effect 
attribut: 
as remo 
DTM ai 
with v 
charact: 
often re 
answer 
inform: 
queries 
concerr 
objects r 
In orde: 
and ef 
applica 
Spatial 
devices 
traditio: 
imagen 
out lir 
dimens 
be opti 
object - 
current 
1000 by 
map) a 
only a