Developing Smart Spatial Systems Using CommonKADS

Abstract

The aim of this paper is to show how spatial reasoning systems in a geographic context can be developed using CommonKADS. KADS was initiated as a "Structured methodology for the development of knowledge based systems" (Motta. 1997). The limitations of production rules, combined with their inherent non-reusability contributed significantly to the impetus to develop methodologies like KADS. The two central principles that underlie the KADS approach are the introduction of multiple models as a means of coping with the complexity of the knowledge engineering process, and the use of knowledge-level descriptions as an intermediate model between expertise data and system design.Motta (1997) coins the term "knowledge modelling revolution", which refers to the paradigm switch from symbol level (rule based) approaches to knowledge level task centred analysis.This heralded the necessary decoupling of the task specification and the problem solving method.The CommonKADS methodology has been used extensively in a wide range of domains, but not that of spatial reasoning. The advantage of CommonKADS is that it provides a full model of an application, not just a model of the knowledgebase (Schriber et al 2000) .Visual geographic knowledge which can be extracted from satellite remote sensing images has characteristics which are not commonly found in nonvisual domains. Traditionally geographic expert systems have worked either at the pixel level of raster images or the object level of vector images. This has shortfalls when knowledge acquired from a human image interpreter has to be incorporated into an expert system to aid interpretation.

Urban expansion and the loss of agricultural lands and forest cover in Limbe, Cameroon

“…In addition, the training data were randomly selected throughout the entire image, and these were spectrally homogenous areas. The general rule is that the training data were being extracted from numbers of spectral bands (n) and then greater than 10n pixels of training data were collected for each class (Crowther et al,1999). Each class has minimum sample size of 10*n and thus the total minimum training class size: 10*n*c, where c is number of classes.…”