! 2004 
pp. 
DTM 
test. 
^ CDD. 
acy of 
ection 
mated 
rected 
m on 
1989. 
using 
ology 
1978. 
fodels 
1978. 
fodels 
ations 
metry, 
SPOT 
laska's 
emote 
Stereo 
cience 
ual of 
FACTORS CAUSING UNCERTAINTIES IN SPATIAL DATA MINING 
Hanning YUAN* Shuliang WANG® 
*School of Remote Sensing Information Engineering, Wuhan University, Wuhan 430079, China 
? International School of Software, Wuhan University, Wuhan 430079, China 
E-mail: hnyuanslwang@yahoo.com, 
Commission IV, WG IV/3 
KEY WORDS: Factors, Uncertainties, Spatial data mining 
ABSTRACT: 
Spatial data mining is to extract the unknown knowledge from a large-amount of existing spatial data repositories areas (Ester et al., 
2000). The spatial data are to represent the spatial existence of an object in the infinitely complex world. They may be incomplete, 
noisy, fuzzy, random, and practical because the computerized entities are different from what they are in the real spatiotemporal 
space, i.e., observed data different from true data. 
For it works with the spatial database as a surrogate for the real entities in the 
spatial world, spatial data mining is unable to avoid the uncertainties. If the uncertainties are made appropriate use of, it may be able 
to avoid the mistaken knowledge discovered from the mistaken spatial data. The uncertainty parameters, such as, supportable level, 
confident level and interesting level, may further decrease the complexity of spatial data mining. Otherwise, it is unable to discover 
suitable knowledge from spatial databases via taking the place of both certainties and uncertainties with only certainties. Based on 
the unsuitable even mistaken knowledge, the spatial decision may be made incorrectly. The uncertainties mainly arise from the 
complexity of the real world, the limitation of human recognition, the weakness of computerized machine, or the shortcomings of 
techniques and methods. Their current constraints might further propagate even enlarge the uncertainty during the mining process. 
1. OBJECTIVE REALITY 
The world is an infinitely complex system that is large, 
changeable, nonlinear, and multi-parameter, about 80% 
information of which is spatial-referenced (Wang, 2002). In 
the spatial world, there are more inexact entities with 
indeterminacy or inhomogeneity than the exact ones. The 
spatial entity in the world includes historical information, 
current status, and future trend. At any moment, it receives the 
information from other entity, and it also eradiates its own 
information. The information of different entities may be 
overlapped, mixed, or deformed. Two entities of the same 
classification may eradiate different spectrum information, 
while two entities that eradiate the same spectrum information 
may belong to different classifications. As a result, it is 
confused to correctly classify the pixels with the same gray 
degrees in the boundary area where two different 
classifications overlap. In the real world, the information 
cannot be incarnated if it is not sensed by the observation of a 
certain instrument. Remote sensing captures spatial data via 
detecting the spectrum with sensors. Traditionally, it was 
presumed that the spatial world stored in spatial database was 
crisply defined, precisely described and accurately measured in 
computerized databases (Burrough, Frank, 1996). For instance, 
an object model assumes that the spatial entities may be 
precisely described via points with exactly known coordinates, 
lines linking a series of crisply known points, and areas 
bounded by sharply defined lines. However, these cases 
seldom happened in the real world, and in many cases, there do 
not exist the pure points, lines, and polygons with geometric 
definitions (Wang, Shi, 2002). 
Some true spatial values are even inexact or inaccessible. The 
true values of spatial data are the actual characteristics of the 
spatial entity reality. Some true spatial values exist but are 
impossible to obtain. One is unobservable for they are spatial 
261 
data with long history, the other is impractical to observe 
because they are too complex, difficult or expensive for human 
to get in the constraint contexts of current cognition, 
instruments and techniques, times and capitals. As to some 
spatial values, there are further no true values at all in the real 
world. Some spatial entities have no sharp boundaries or 
cannot be precisely determined. Take it for example that the 
spectrum of the spatial entity makes the image data uncertain. 
It is a fundamental function to determinate whether or not the 
spatial element belongs to the predefined entity, and the 
classification determination is performed on the accessible 
spatial values that are measured by sensors. The overlapped or 
mixed pixel of remote sensing images comprehensively reflects 
the classifications of different but neighbor objects on the 
ground. The additional but indispensable measurement step 
will further cause uncertainty because of the limitations in the 
process. Remote-sensing images of different objects may show 
the phenomena of spectral uncertainty created by spatial 
entities. One is that two objects belong to the same type or 
species but with different spectrums, which cannot be uniform 
as one spectral curve, but are composed of a series of different 
spectral curves, and cause a wide distribution. In a generalized 
category it also includes the multi-angular, multi-temporal and 
multi-scale effect, e.g., Rocks/Minerals, Vegetation. The other 
is that two objects belong to different classifications but with 
the similar or same spectral features in a certain wavelength 
range, e.g., the camouflage in military. New uncertainties may 
further be caused during the process of additional but 
indispensable measurements. 
The uncertainty is more popular in macro-world (e.g., astro- 
space) and micro-world (e.g., the space that electron, proton 
are moving), both of which are moving at a high speed 
(Duncan, 1994). The length of moving objects, and the 
distance between two objects, all have contractility. The