Data Questions in GeoComputation

Brian G Lees

Department of Geography,
Australian National University,
ACT 0200,

An Air New Zealand Guest Keynote Speaker

Presented at the second annual conference of GeoComputation ‘97 & SIRC ‘97, University of Otago, New Zealand, 26-29 August 1997

Abstract

The new computational solutions which have developed in parallel with the widespread uptake of computer power in geography tend to represent a distinct move away from the more traditional, parametric statistical, methods introduced to the discipline during the so-called quantita-tive revolution. The implications of adopting these new methods have not yet been fully appreciated by many researchers, indeed the retreat by large sections of the discipline from any serious engagement with quantitative approaches to geographical problems means that there is a growing potential for misuse and abuse of these solutions as they become more accessible.

It is well worth remembering that our new computational solutions comprise several critical components. Firstly, the new hardware configurations without which they could not be implemented. Secondly, the new algorithms themselves. Thirdly, the data, and fourthly, the problem. Perhaps this fourth element could be better described as the problem statement. Success depends on the adequacy of all of these, and on their correct integration. The data, the problem statement and the links between the algorithm and the data, between the data and the problem statement, and the problem statement and the algorithm have all received much less attention than the matching of algorithms and hardware. As we become increasingly engaged in harnessing our new computer power to geocomputation it is perhaps worth casting our eyes over these other, equally important, facets of quantitative investigation, analysis and prediction in the geosciences.

In this discussion I will refer to only two of the large suite of data-driven modelling techniques rather than range across the field. There is simply neither time, nor space, to do otherwise. Most of the points I want to make are relevant to a much wider group of techniques and these are merely convenient examples for discussion. The two I will use are decision trees and so-called back-propagation. Both travel under fancier names on occasion, but these labels are broadly understood and will suffice.