Go to Paper
Return to GeoComputation 99 Index
Dataspaces As An Organisational Concept For The Neural Classification Of Geographic Datasets
GAHEGAN, Mark (firstname.lastname@example.org) and TAKATSUKA, Masahiro, Pennsylvania State University, Department of Geography, 302 Walker Building, University Park, PA 16802
Key Words: classification, neural networks, self organising maps, dataspaces
Traditional geography and environmental analyses use the idea of a data space in order to add structure to the understanding of complex natural systems. A dataspace is some subset of the overall attribute or feature space where some geographical theme is represented. Examples are spectral space, geographic space, and environmental space (Aspinall and Lees, 1994). It has been suggested that the relationships within these spaces are not the same as those between them, due to the inherently different structure of each separate domain.
Many types of statistical classifiers (such as maximum likelihood) and inductive learning techniques (such as decision trees and neural networks) treat the entire attribute space as comprising a single descriptive vector. In geographic terms this is equivalent to the construction of a single and all-inclusive dataspace. While this approach has obviously met with a good deal of success, it seems likely that the mixing of concepts has two distinct disadvantages, namely:1. The classification tool has to effectively learn to disentangle data used in training. This increases the complexity of the learning task, which, in turn, may reduce accuracy. 2. The functioning of the classifier is difficult to explain in any language that is geographically meaningful.
This paper will investigate the use of a neural classification technique known as Self Organising Maps (Kohonen, 1995) to provide a structure to the classification problem based around the dataspace. The Self-Organising Map (SOM) is a set of artificial neurons, each of which is an ordered classifier in feature space. A 2-D SOM is commonly used to find (construct) classifiers, hence, provide a continuous topological mapping between the dataspace and the 2-D space. It is known that the SOM may develop a poor mapping if elements of an input vector have different scales (Kohonen, 1995). This situation arises when an input vector consists of signals from different domains (Takatsuka, 1996). In this study, separate SOMs are used for classifying and discovering underlying structures from different domains. These are then combined to produce an overall solution.
The following hypothesis is tested: that classification based on dataspaces can outperform traditional approaches in terms of accuracy, computing efficiency, and understandability, by providing much-needed structure to the problem. Results from a landcover classification exercise will be presented and compared to established neural and statistical classifier benchmarks (Gahegan and West, 1998).
Aspinall, R.J. and B.G. Lees (1994) Sampling and Analysis of Spatial Environmental Data. Proc. 6th International Symposium on Spatial Data Handling, Ed: Waugh, T.C. and R.G. Healey, Edinburgh, Scotland, pp. 1086-1098.
Gahegan, M. and G. West (1998) The Classification of Complex Geographic Datasets: An Operational Comparison of Artificial Neural Network and Decision Tree Classifiers, Proc. GeoComputation98, 3rd International Conference on GeoComputation. http://www.cs.curtin.edu.au/~mark/geocomp98/geocom61.html.
Kohonen, T. (1995) Self-Organizing Maps. Springer-Verlag, Berlin, Germany.
Takatsuka, M. (1996) Free-Form Three-Dimensional Object Recognition Using Artificial Neural Networks. Ph.D thesis, Dept. of Electrical and Computer Systems Eng., Monash University, Clayton 3168 VIC, Australia.