The Application of Neural and Visual Techniques to the Analysis of Spatial Data

1. Improvements in computing performance and the development of massively parallel architectures have enabled previously intractable analysis problems to be addressed via deterministic means.

2. Progress in pattern recognition, classification and function approximation tools, originating from the artificial intelligence community (such as decision trees, neural networks and genetic algorithms) now provide sophisticated capabilities for tackling a range of non-deterministic problems.

3. Advances in graphical display technology provide the basis for data exploration using visualisation or virtual reality techniques.

Taken together, these newer computing tools show considerable promise in that they are capable (in theory at least) of managing large, many-layered and heterogeneous datasets. This is just as well, since spatial analysis, whether concerned with the built or natural environments, has to deal with an increasing volume and diversity of data. Furthermore, with the advent of concern over global environmental issues, the scale and complexity of the tasks to be conducted is set to rise. However, many problems remain before this new technology is effectively harnessed. Many of these are methodological; sophisticated tools require sophisticated setup and operation.

My main area of interest is in the development of suitable methods to make good use of these tools in a geographic setting, specifically for problems involving complex, high dimensionality datasets. Work to date has focussed on three different application areas: landcover classification (at the floristics level), spatial epidemiology and geological interpretation, and has led to the development of sophisticated visualisation tools (http://www.cs.curtin.edu.au/gis/visualisation/) and neural network-based classifiers (http://www.cs.curtin.edu.au/gis/donnet/). Both of these areas are described briefly below.

Visualisation for exploratory data analysis

It is important to note that the aim of exploratory visualisation is not to analyse the data per se, but rather to present the data to the user in a way that promotes the discovery of inherent structure and relationships (MacEachren & Ganter, 1990). In psychometric colloquialism this is known as inducing visual ‘pop out’ (Csinger, 1992). Thus, a collaborative mode of interaction is developed between the user and the machine, where the visualisation environment produces a stimulus via the visual encoding of the data which is then interpreted by the user, enabling full advantage to be taken of the unsurpassed abilities of humans to perceive complex structural relationships.

Methods to achieve effective visual encoding strategies for spatial data have been investigated (Gahegan, 1996; 1998) including the use of expert knowledge to mediate and navigate through a combinatorially explosive range of possible solutions (Gahegan and O’Brien, 1997).

 
 
Figure 1. Four scenes depicting a range of visualisation techniques applied to the same data (an environmental dataset of a coastal region of New South Wales, Australia). Top left, mark composition using arrows draped on an elevation model. Top right, ‘interactors’ describing relationships between different data layers. Bottom left, a scatterplot, enhanced with planar point icons to encode additional information. Bottom right, two different environmental surfaces are dynamically ‘intersected’. A comparison of these techniques can be found in Gahegan, 1998).

Neural networks and decision trees for classification

The kinds of classification problems that arise in geography or the wider earth sciences are often characterised by their complexity, both in terms of the classes and the datasets used. For example, classes may be difficult to define, may vary with location and over time, and their properties may overlap in attribute space. Datasets increasingly contain many descriptive variables (layers) and often contain a mix of statistical types; for example remotely sensed reflectance values (quantitative data) supplemented with nominal data such as soil type or geology and ordinal data such as slope or aspect. Data ‘saturation’ seems set to increase with the adoption of sensing devices of greater sophistication, resulting in a higher spatial resolution and many more channels. The complexity of the tasks to which these data are applied is also increasing; for example the classification of deep geological structure from 300 channel airborne electro-magnetics data, or socio-demographic indices from combinations of many indicator variables.

Addressing geographic classification problems successfully with tools based around inductive learning and search requires detailed attention to methodology and often also a good deal of further development and enhancement. To this end a black-box neural classifier has been specifically engineered for application to geographic problems (German and Gahegan, 1996). It is based on a feedforward, multi-layer perceptron, with various enhancements made (German et al., 1997; Gahegan et al., 1998) Importantly, it is designed to be self-configuring, requiring only the same setup as a standard Maximum Likelihood Classifier. Results so far show increased classification accuracy over established techniques, that is maintained as the number of data layers (attributes) increases to twenty or more.

References

Benediktsson, J. A., Swain, P. H. and Ersoy, O. K. (1993). Conjugate gradient neural networks in classification of multisource and very high dimensional remote sensing data. International Journal of Remote Sensing, Vol. 14, No. 15, pp. 2883-2903.

Civco, D. L. (1993). Artificial neural networks for landcover classification and mapping. International Journal of Geographical Information Systems, Vol. 7, No. 2, pp. 173-186.

Csinger, A. (1992). The psychology of visualisation. Technical Report Series, Department of Computer Science, University of British Columbia, Canada.

Eklund, P. W., Kirkby, S. D. and Salim, A. (1994). A framework for incremental knowledge update from additional data coverages. Proc. 7th Australasian Remote Sensing Conference, Melbourne, Australia, Remote Sensing and Photogrammetry Association of Australia, pp. 367-374.

Fitzgerald, R. W. and Lees, B. G. (1994). Assessing the classification accuracy of multisource remote sensing data. Remote Sensing of the Environment, Vol. 47, pp. 362-368.

Freidl, M. A. and Brodley, C. E. (1997). Decision tree classification of land cover from remotely sensed data. International Journal of Remote Sensing, Vol. 61, No, 4, pp. 399-409.

Gahegan, M. N. (1996a). Visualisation strategies for exploratory spatial analysis. Proc: Third International Conference on GIS and Environmental Modelling, NCGIA, Santa Barbara, USA.

Gahegan, M. (1998). Scatterplots and scenes: Visualisation techniques for exploratory spatial analysis, Computers, Environment and Urban Systems, Vol. 22, No. 1. (in press).

Gahegan, M., German, G. and West, G. (1998). Some solutions to neural network configuration problems for the classification of complex geographic datasets. To appear in Geographical Systems.

Gahegan, M. N. and O’Brien, D. L. (1997). A strategy and architecture for the visualisation of complex geographical datasets. International Journal of Pattern Recognition and Artificial Intelligence, Vol. 11, No. 2, pp. 239-261.

German, G. and Gahegan, M. (1996). Neural network architectures for the classification of temporal image sequences. Computers and Geosciences, Vol. 22, No. 9, pp. 969-979.

German, G., Gahegan, M. and West, G. (1997). Predictive assessment of neural network classifiers for applications in GIS. Second International Conference on GeoComputation, Otago, New Zealand, pp. 41-50.

Haslett, J., Bradley, R., Craig, P., Unwin, A. and Wills, G. (1991). Dynamic graphics for exploring spatial data with application to locating global and local anomalies. The American Statistician, Vol. 45, No. 3, pp. 234-242.

Hearnshaw, H. M and Unwin, D. (Eds.) (1994). Visualization in Geographical Information Systems, John Wiley & Sons, Chichester, England.

Hepner, G. F., Logan, T., Ritter, N. and Bryant, N. (1990) Artificial neural network classification using a minimal training set: comparison to conventional supervised classification. Photogrammetric Engineering and Remote Sensing, Vol. 56, No. 5, pp. 469-473.

Kamata, S. and Kawaguchi, E. (1993). A neural network classifier for multi-temporal Landsat images using spatial and spectral information. Proc. IEEE 1993 International Joint Conference on Neural Networks, Vol. 3, pp. 2199-2202.

Kanellopoulos, I. and Wilkinson, G. (1997). Strategies and best practice for neural network image classification. International Journal of Remote Sensing, Vol. 61, No, 4, pp. ???

Keller, P. R. and Keller, M. M. (1993). Visual Cues: Practical Data Visualization, IEEE Press, Los Alimatos, CA, USA.

Lees, B. G. and Ritman, K. (1991). Decision tree and rule induction approach to integration of remotely sensed and GIS data in mapping vegetation in disturbed or hilly environments. Environmental Management, Vol. 15, pp. 823-831.

MacEachren, A. M. and Ganter, J. H. (1990). A pattern identification approach to cartographic visualisation. Cartographica, Vol. 27, No. 2, pp. 64-81.

Paola, J. D. and Schowengerdt, R. A. (1995). A detailed comparison of backpropagation neural networks and maximum likelihood classifiers for urban landuse classification. IEEE transactions on Geoscience and Remote Sensing, Vol. 33, No. 4, pp. 981-996.

Rheingans, P. and Landreth, C. (1995). Perceptual principles for effective visualisations. In: Perceptual Issues in Visualisation, (Eds. Grinstein, G. and Levkowitz, H.), Springer-Verlag, Berlin, pp. 59-69.

He lectures in various aspects of GIS, including courses on Spatial Data Structures and Spatial Analysis. Research interests include: systems integration (specifically GIS and remote sensing), landcover classification techniques, cartographic and scientific visualisation, data translation, interoperability and semantics in GIS.

He is chair of ISPRS working group II/2, software and modelling aspects for integrated GIS (http://www.cs.curtin.edu.au/gis/isprs/), on the editorial board of the International Journal of Geographical Information Science, and on the international steering committee of the GeoComputation conference series (http://www.ashville.demon.co.uk/geocomp/).

Recent publications relevant to spatial analysis using ‘newer’ computational approaches

Journals

German, G. and Gahegan, M. (1996). Neural network architectures for the classification of temporal image sequences. Computers and Geosciences, Vol. 22, No. 9, pp. 969-979.

Stockwell, T., Daly, A., Phillips, M., Masters, L., Gahegan, M., Midford, R., Philp, A. (1996). Total versus Hazardous Per Capita Alcohol Consumption as Predictors of Acute and Chronic Alcohol Related Harm. Contemporary Drug Problems 23, pp. 441-464.

Gahegan, M. N. and O’Brien, D. L. (1997). A strategy and architecture for the visualisation of complex geographical datasets. International Journal of Pattern Recognition and Artificial Intelligence, Vol. 11, No. 2, pp. 239-261.

Gahegan, M., German, G. and West, G. (1998). Some solutions to neural network configuration problems for the classification of complex geographic datasets. Accepted for publication in Geographical Systems.

Gahegan, M. (1998). Scatterplots and scenes: Visualisation techniques for exploratory spatial analysis, Computers, Environment and Urban Systems, Vol. 22, No. 1. (in press).

Gahegan, M. N. (1998). Visualisation as a geocomputational tool. To appear in: GeoComputation (Eds. Openshaw, S. and Abrahart, B.), Harwood Academic, UK.
Conferences

O’Brien, D., Gahegan, M. N. and West, G. A. W. (1995). Information overload, the visualisation of multiple spatial datasets for mineral exploration. Proc. GISRUK '95, Newcastle, England.

Gilchrist, J. and Gahegan, M. N. (1995). Issues in the design of a conceptual data model for socio-demographic applications in GIS. Proc. International Conference Computers in Urban Planning and Urban Management, Melbourne, Australia.

Gahegan, M. N. (1996). Visualisation strategies for exploratory spatial analysis. Proc: Third International Conference on GIS and Environmental Modelling, Santa Fe, USA, NCGIA. (http://www.cs.curtin.edu.au/~mark/santafe.html)

Gahegan, M., German, G. and West, G. (1996). Automatic neural network configuration for the classification of complex geographic datasets. First International Conference on GeoComputation, University of Leeds, UK, pp. 343-358.

Midford, R., Stockwell, T., Phillips, M., Gahegan, M., Masters, L., Daly, A., Philp, A. (1996). Alcohol Consumption and Injury in Western Australia: A Spatial Correlation Analysis using GIS Technology. 3rd International Conference on Injury Prevention and Control, Melbourne, Australia.

German, G., Gahegan, M. and West, G. (1997). Predictive assessment of neural network classifiers for applications in GIS. Second International Conference on GeoComputation, Otago, New Zealand.

Gahegan, M. (1997). The visualisation of relationships between geographic datasets. Second International Conference on GeoComputation, Otago, New Zealand, pp. 335-343.

Gahegan, M. (1997). Accounting for the semantic differences between geographic information systems. NCGIA Conference on Interoperation of GIS.

Gahegan, M. (1998). Four barriers to the development of effective exploratory visualization tools for the geosciences. ICA Commission on Visualization, Poland. Also available via the internet from http://www.cs.curtin.edu.au/~mark/visworkshop/visproblems.html

German, G., Gahegan, M. and West, G. (1998). Improving the Learning Abilities of a Neural Network-based Geocomputational Classifier. Proc. ISPRS Symposium II, Cambridge, UK. ISPRS Publications.

Gahegan, M. and West, G. A. W., (1998). A theoretical approach to the use of artificial intelligence techniques in the classification of complex geographic datasets. (To appear in) Third International Conference on GeoComputation, Bristol, England. Also available via the internet from http://www.cs.curtin.edu.au/~mark/geocomp98/geocom61.html

Other publications

Go back to list