Use this URL to cite or link to this record in EThOS:
Title: Optimisation of pedotransfer function models for soil hydraulic properties using an artificial neural network ensemble method
Author: Baker, Lee
Awarding Body: University of Abertay Dundee
Current Institution: Abertay University
Date of Award: 2005
Availability of Full Text:
Access from EThOS:
Access from Institution:
Soil hydraulic properties are very time-consuming and expensive to measure directly. Conversely, routinely collected soil survey parameters (e.g. soil texture, dry bulk density and organic matter content) are relatively cheap and easy to collect. Fortunately, mathematical regression models called Pedotransfer Functions (PTFs) allow the transfer of data we have (soil survey parameters) into the data we need (soil hydraulic properties), and this thesis focusses on the potential and modelling methodologies of Artificial Neural Network (ANN) ensembles as a means by which to construct PTFs. The individual processes of the ensemble modelling procedure are examined herein, and suggestions as to how they may be improved upon, by mathematical, conjectural and empirical means, are discussed. Ultimately, it will show that more accurate, robust and efficient PTFs may be constructed by incorporating these into such models. An ensemble is a collection of individual ANNs, each of which provides an independent solution to the same problem. Thus, the combined strengths of the individual models are augmented, whilst their weaknesses are diminished. For any modelling methodology, a trade-off exists between the bias and variance components of the model error - decreasing the bias results in higher variance, and vice versa. However, for the ensemble method, combining ensemble members results in the reduction or elimination of the variance, whilst leaving the bias unaltered. Thus, the aim of the ensemble modeller is to determine an optimum balance between achieving low bias (regardless o f the consequences to the variance) and conserving data - a trade-off between the amount of data used and the value of that data. It has been empirically demonstrated here that by training ANNs using data selected such that all parts of the dataspace are equally represented, bias in ANN-PTF ensembles may be reduced to negligible levels. In addition, results showed that, using one-third or less the amount of data, the ensemble method yields results that are at least as accurate as single ANN methods.
Supervisor: Not available Sponsor: Carnegie Trust
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available