Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Titre : Evaluating techniques for soil erosion modelling : a role for artificial intelligence ?

Auteur : Ellis, Fiona G

Université de soutenance : Australian National University

Grade : Doctor of Philosophy (1997)

Résumé partiel
This thesis evaluates two Artificial Intelligence techniques, Decision Tree Analysis (OTA) and Neural Networks (NN), in the modelling of soil erosion in two study areas in NSW, Australia. Soil erosion is a wide-spread problem in Australia. Insufficient knowledge of the location and severity of soil erosion is seen as a factor limiting its effective prevention. Several techniques are currently used to map and predict soil erosion, including : ground based surveys ; the interpretation of remote images to map and monitor soil erosion ; and process models to determine the distribution of eroded land and to predict the effect of land use changes. There are limitations associated with these techniques : ground surveys are labour intensive and hence very costly, and soil erosion cannot easily be measured over complex regions using either traditional process models or remote sensing. A different approach to these existing techniques is to model areas prone to soil erosion using multi-source data sets and data driven models. Al techniques such as OTA and NN are suitable for the analysis of multi-source datasets, as they are distribution free. In addition, OTA and NN are reputedly efficient in terms of data requirements. Whilst these properties make Al techniques attractive, it is somewhat unclear how to implement these types of inferential models, largely due to uncertainties associated with the effect of training set selection on the accuracy of these techniques. OTA and NN have not previously been used for modelling soil erosion. It was thought that these techniques would provide a way of predicting soil erosion using readily available data sets. In order for modelling to be a useful technique, maps of the necessary independent variables or surrogates must be available for modelling, or at least be easier to map than the phenomena under question. The two study areas chosen, suited the criteria for having data available. Both areas are located in the Southern Tablelands of south-eastern Australia. Michelago-Colinton was selected for the numerous erosion gullies, and previous work conducted in the region. The second study site, Dicks Creek, contains rill and gully erosion, and large areas of sheet erosion.

Présentation

Version intégrale (272 Mb)