Goldsmith, Kevin (2005) Artificial intelligence techniques for soil erosion mapping and risk assessment in Almeria Province, Southeast Spain. (PhD thesis), Kingston University, .
Abstract
This thesis provides an alternative method to for mapping soil erosion. The method is conducted in a small study area of 40 km2 in the Sorbas Basin, Almería province, Southeast Spain. Soil erosion is one of the most destructive land degradation processes and can often lead to serious environmental problems. It is important to implement appropriate management strategies to meet these challenges at a range of scales. However, prior knowledge of erosion processes and the extent to which they operate spatially is often limited and, traditional methods of soil erosion mapping are often time and labour intensive. This thesis explores the use of two Artificial Intelligence (AI) techniques for soil erosion mapping; Artificial Neural Networks (ANNs) and Decision Tree Classifiers (DTCs). The opportunities for employing such methods relate in part to their non-linear capabilities, their ability to learn in an inductive manner and incorporate multi¬source data sets. AI training and test data were collected from 520 individually sampled locations within the study area. At each site the dependent variable erosion was estimated, as were a range of independent variables through field study. Two Digital Elevation Models were developed. Laboratory analysis was also undertaken to explore the physico-chemical processes relating to soil dispersion and to determine the applicability of a soil sodicity meter developed by the Co-operative Research Centre for Soil and Land Management in Adelaide, Australia. Results demonstrate that classification accuracy.and overall performance is strongly dependent on the independent and dependent variables used, with the more expensive field collected data providing improved variables to those extracted from the Digital Elevation Models. Discriminant Analysis (DA) classifications were also employed to provide a linear comparison to the AI techniques, and performed comparably well. In the Artificial Neural Network classifications the composition of the training set was seen to exert significant bias, leading to poor performance and often misleading results. Laboratory analysis highlights the complex physico-chemical relationships associated with soil dispersion. The findings also indicate that no discernible relationship exists between the sodicity meter and standard laboratory procedures employed to measure the sodic properties of a soil. The thesis demonstrates the potential for employing these methods for erosion risk analysis and the ability of inductive approaches to formulate rules that may enhance current levels of understanding associated with soil erosion processes. Mapped outputs produced by these methods may prove valuable in the management of landscapes susceptible to soil erosion.
Actions (Repository Editors)
 |
Item Control Page |