ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS’’, Bangkok, May 23-25, 200t 
256 
direction is applicable. The direction is measured according to 
this established cardinal direction. (Frank 1992) 
3.3 Qualitative Distances 
To qualitatively represent distance, we have map distance to 
finite number of symbols. This representation should be 
Euclidean approximate enough in order to obtain meaningful 
result. Therefore, this process is application dependent. 
Basically, we may represent distance with two symbols, three 
symbols or more. (Frank 1992) The first and most obvious 
concept of distance is a set of two symbols for far and close. 
Medium could be introduced which leads to a system with 
three symbols: close, medium, far. These concepts can be 
generalized to create any finite sequence of symbols of 
increasing distance symbols. To keep our paper concise, we 
only adopt three symbols. 
N 
S 
Figure 3. Cone-shaped directions 
3.4 Brief Concluding Remarks 
We have discussed the most essential qualitative 
representations of spatial information upon which we can 
employ common data mining methods. Although we only 
present three kinds of qualitative representation, our method 
does not limit to these representation. Upon different 
circumstances, more complex and rich representation could be 
introduced. 
From above three kinds of representation, we obtain 19 
predicates: 
Topology: disjoint, contains, inside, equal, meet, covers, 
coveredBy, overlap 
Direction: north, northeast, east, southeast, south, southwest, 
west, northwest 
Distance: near, medium, far 
These predicates could be used to seek spatial association 
rules as Koperski did in (koperski 1999). But we won’t stop 
here. Comparing these predicates of different time, the 
changes could be detected. Then the sequential pattern can 
be found among these changes. 
4. DETECTING SPATIAL CHANGES 
FROM QUALITATIVE REPRESENTATION 
According to different domain of application, proper time 
interval can be selected. No matter how to incorporate 
temporal information into spatial data or what kind of temporal 
database is used, we may generate qualitative representation 
at different times. The method has been discussed in precious 
section. The change is computed in a way of comparing the 
difference between consecutive snapshots of the data 
represented. 
As illustrated in figure 1, at time ti and ti+1, the geospatial data 
is retrieved and represented in a qualitative way. Then the 
qualitatively representations of ti and ti+1 are compared to 
detect changes. Once the changes are found, they are 
recorded in a table with timestamp for the mining the 
sequential pattern from these changes. 
For example, a change could be presented as follows. 
Changel: near(A, B)~+far(A, B) (time T, Support X) p 
It means that at time T, the distance between spatial object A 
and B increases greatly. Since we are mining massive dataset 
to obtain general pattern, generally A and B refer to a class of 
spatial objects. Furthermore, spatial objects could be arranged e 
into a hierarchy. By this way, the change could be detected at 
multiple levels. We would not cover this topic in this paper. 
Support X is the count of this change, which depict how often 
this change occurs. Only the patterns whose support exceeds E 
given threshold is regarded as interesting and will be 
considered in next step. 
5. MINING SEQUENTIAL PATTERN 
OF SPATIAL CHANGES 
Once the changes are detected and presented by predicates, 
the using WINEPI to discover frequent episode is 
straightforward. Those changes correspond to events in 
(Mannila, Toivonen et al. 1997). Frequent episodes could be 
detected to present that relationship among changes. From 
these frequent episodes, we may also derive rules. The Rules 
should show the connection between events or changes more 
clearly than frequent episodes alone. 
For example, changes have been detected following the above 
steps. All the changes have associated time of occurrences. 
The WINEPI, which is based on the discovery of episodes by 
only considering an episode when all its subepisodes are 
frequent, and on incremental checking of whether an episode 
occurs in a window, could efficiently find the frequent 
episodes. Quite similar to rule generation from frequent 
itemset, rules could also be generate from frequent episode. 
6. CONCLUSION AND FUTURE RESEARCH ISSUES 
In this paper we present a method to discover sequential 
pattern from spatio-temporal data. Some data mining methods 
originated from transaction database and constructed from the 
perspective of machine learning. Therefore, generally 
information has to be represented in a symbolic way for 
convenience of automatic reasoning. However, most of 
geospatial information is gathered with reference to certain 
quantitative coordinate system. To introduce those data mining 
methods, geospatial information has to be transformed into to 
symbolic representation. Although qualitative spatial 
representation has been studied in the field of qualitative 
spatial reasoning, this process is not direct. Furthermore, 
beside frequent episode, more complex sequential pattern 
could be introduced into the field of GIS. Some future research 
issues are: 
• Symbolic representation of spatio-temporal 
information. Although some research of this area has 
already been done, it is basically for purpose of spatial 
reasoning. A systematical representation does not yet 
exist for our purpose. 
• Learning qualitative representation from geospatial 
dataset efficiently. It is computationally expensive to 
infer qualitative representation, especially when 
dataset is very huge. Efficient algorithm is needed. 
• Incorporate temporal semantics into sequential 
pattern. More expressive temporal relations would 
make the result more meaningful. (Rainsford 1999)