Title

A neural network approach to solving areal interpolation problems

Date of Completion

January 2004

Keywords

Geography|Artificial Intelligence

Degree

Ph.D.

Abstract

A problem often encountered in spatial analysis is the unavailability of aggregated data for areal units that are not the desired units. To overcome this problem, areal interpolation methods are used to estimate unknown data values for a set of areal units based on the known data values for another set of areal units. A number of areal interpolation methods exist and there is no single method that is considered superior. Choosing the appropriate method to use often depends on the type of problem to be solved and the desired level of accuracy. ^ This study proposes an areal interpolation method by means of artificial neural networks. Artificial neural networks are computation models inspired by the biological neural network of the human brain. These computational models have been widely used in spatial analysis including applications within the sub-fields of remote sensing, spatial interaction modeling, and spatial interpolation. One advantage of using neural networks for spatial analysis problems is that there are no underlying assumptions that need to be made about the data being interpolated. ^ In this study, neural networks are applied to two areal interpolation scenarios; the “missing data” problem and the “alternative geography” problem. In both instances, the data to be estimated by neural networks are total population values for census block groups in Hartford County, Connecticut. A number of neural network models are generated containing various combinations of input data, which are considered to be spatial and demographic in nature. The accuracy of the population estimations derived from neural networks is measured by comparing them to the estimations from two existing areal interpolation methods; areal weighting and urban weighting. The results from this study indicate that neural networks achieve optimal estimations when applied to the “missing data” problem. When applied to the “alternative geography” problem, neural networks also produce better results when compared to areal weighting and urban weighting. ^