5.2 Ordering 
The second area that we want to look into is to 
explore the internal nature of the large spatial 
objects. For example, it has been demonstrated that 
remotely sensed imagery has very high spatial 
correlation among pixels in a neighbor and between 
various spectral bands. Partition and re-ordering 
the spatial data to take these features into 
consideration can speed up many GIS queries. 
The traditional method of storing large object data 
is linear allocation whereby the data is stored by 
some predetermined orders. Typical date storage 
orders may be one of the pixel interleave, line 
interleave or plane interleave. This order may be 
efficient for some data processing that involved all 
pixels, but will make other access such as point or 
window queries in-efficient. Optimizing the 
allocation of large object data becomes increasingly 
important as more and more multi-spectral, multi- 
temporal large objects are available today in GISs. 
In GIS application such as display, editing and 
analysis, only a small part of data is required at a 
time. Point or window based queries to extract parts 
of the image are most often performed. This makes 
data ordering particular important for large object 
handling. To improve the system performance, 
several strategies have been proposed, which 
include: 
a) Tiling: which divides an image into tiles that 
are stored and accessed independently. By 
combination with parameters such as optimism 
buffer size, display window size, these techniques 
can improve the system performance dramatically, 
especially for network based GISs. 
b) Reordering: which permutes the order of image 
pixels to reduce average seek distance. For 
example, in data display, we could anticipate that 
zoom/pan is frequently performed on the large 
objects. Therefore, instead of linear storing each 
tile, we use the Mount-Caro order[Samet, 1989] to 
organize the tile files. 
c) Redundancy: which stores redundant copies of 
the image. The redundant images are abstracted 
and organized differently to optimize different 
patterns of access. For example, for overview 
purpose, a low resolution image can be used to give 
users the rough idea on the object. Only when the 
user specifically zooms in the data, the original 
data is provided to give more details. This 
technique not only makes the system easy to use, 
but also makes most interactive applications 
possible. 
5.3 Abstraction 
The third area that needs to be investigated is 
object abstraction. The conventional DBMS 
paradigm handles well-formatted data and does 
not need to perform any abstraction on the data. In 
contrast, large spatial objects are intrinsically 
information intensive and semantics richer. 
Intelligence needs to be build into the system before 
the system can be used to effectively and 
efficiently answer user's queries. The DAD in our 
data model is mainly used for this purpose, because 
we believe that data abstraction can be built into 
DAD either through automatic image processing 
methods or through manually editing or indexing 
methods and can be used to answer many queries in 
GISs. 
Abstraction can be defined as the process of 
constructing new concepts from a given collection of 
concepts. Abstraction is a method of extraction of 
the essential part of the data set. It is created 
using some combination of statistical analyses, 
data classification and artificial intelligence. 
Derived Attribute Data is the result of abstraction. 
Classification in the object oriented modeling is 
the example of abstraction. Image classification is 
another example of abstraction. 
The key observation is that while a query may be 
posed to a large object, the answer to the query may 
be tiny. For example, a query such as "what is the 
size of Lake Ontario" can be answered very easily 
if the abstracted DAD data is available. While 
many techniques can be used to perform data 
abstraction[Sequio 2000 Report 11], we only limit 
ourselves to the construction of meta-data and the 
vector counterpart of large spatial objects. For 
example, we use statistical analyses to derive the 
mean, variance, histogram, and other parameters 
of image objects. We also use image classification 
techniques and rules to extract classes such as 
water, cloud, land use, etc. with manual editing 
techniques followed to derive vector representation 
of the image. These derived data is modeled as 
DAD data and is stored in GIS data base for 
various applications. 
5.4 Indexing 
The last method to be investigated is ability to 
provide value indexing. The conventional DBMS 
paradigm provides object indexing. Hence, one can 
designate one or more fields in a record as indexed, 
and DBMS will build the appropriate kind of 
index on the data in the required fields. Value 
indexing may be reasonable in traditional 
216 
applica 
large © 
queries 
conten 
Ontaric 
of a cl 
and no 
for im 
values 
examp 
return 
and the 
To be 
datab: 
param 
throug 
large « 
further 
when 
large c 
proper 
we nc 
conter 
rarely 
seconc 
type « 
search 
end-u: 
retriey 
Howe 
used 
huma 
[Wiec 
obser 
techni 
Altho 
our lc 
and « 
gener 
paran 
Befor 
consi 
bein: 
a)as 
W 
b) a 
al 
cal 
N 
With 
datat