The nature and acquisition of expert knowledge to be used in spatial expert systems for classifying remotely sensed images

Knowledge engineering is the process of acquiring expert knowledge from human domain experts. In this thesis the emphasis is on the acquisition of geographic or spatial knowledge from experts involved in interpreting multi-spectral satellite images. This thesis argues that spatial knowledge is primarily visual, hence tools to acquire it also need to be visual. Currently there is no methodology, other than ad hoc interview and protocol analysis, for acquiring expert knowledge of interpretation of satellite images. As a result, there cannot be an integrated knowledge acquisition toolkit, since this must be based on a formal methodology. This thesis offers a methodology to overcome this shortcoming and presents a series of tools to implement the methodology. In the first part of the thesis the nature of geographic knowledge is investigated. A geographic knowledge classification scheme is presented as the basis of the work in the rest of the thesis. It is shown that geographic knowledge can be divided into a six level hierarchy: ‚Äö Primitive knowledge about point, line and areal objects, ‚Äö Relationship knowledge about the relationships between primitive objects, ‚Äö Assembly knowledge about related collections of primitive objects, ‚Äö Non-Visual knowledge of expert heuristics (knowledge of short cuts acquired by experience), ‚Äö Consolidation knowledge of how to resolve and evaluate conflicting information and ‚Äö Interpretation knowledge of how to combine the other knowledge types to produce a classified image. This six level hierarchical classification of geographic knowledge forms the basis of the KAGES (Knowledge Acquisition for Geographic Expert Systems) methodology. Traditional knowledge acquisition procedures are studied and their relevance to a geographic domain discussed. This includes both human interaction techniques such as interviewing and automated knowledge acquisition methods such as neural networks and machine learning. It will be shown that although automated pattern recognition techniques are important, there is still a need to include knowledge acquired by human image interpreters in an automated image interpretation system. There is a theoretical discussion of new techniques to acquire visual knowledge of the types identified in the KAGES methodology. It is shown that these methods can be combined into an integrated knowledge engineering toolkit to acquire geographic knowledge from satellite image interpreters. Not all geographic knowledge is visual however. Three types of non-visual knowledge, algorithmic, heuristic and temporal, are identified and investigated. The first two are implemented in the knowledge engineering toolkit described in this thesis. It is shown that if there are multiple domain experts and multiple knowledge acquisition sessions multiple knowledge-bases will be produced. Techniques for the consolidation of these knowledge-bases is presented. The final section of the thesis involves evaluation of KAGES. This is done in two ways: user evaluation and application of the methodology in two domains. The user evaluation of the KAGES methodology and toolkit involved a number of image interpretation experts from a variety of domains and currently using a variety of tools. They were questioned about the usefulness and useability of the KAGES toolkit. The results of using the tools in the toolkit are evaluated by generating rules for two scenarios, one for sea ice identification and the other for crop recognition. The rules produced using the toolkit are compared with rules produced using other techniques. The effect of applying rules generated by the toolkit to classify images is compared with the results from other image classification methods.