An expert system for land cover classification

A framework to represent a broad class of problems in the analysis of remote sensing imagery is proposed, and an inference engine to tackle such problems is derived. A simple model for spectral knowledge representation is used along with a method for quantification of knowledge through an evidential approach. An automatic knowledge extraction technique is also proposed to gather knowledge from training samples. The techniques of knowledge extraction, representation and inferencing have been used to do a land cover analysis on two data sets, and the results are compared with contemporary digital techniques. It is found that the proposed approach has the advantages of avoiding commission errors, and can incorporate non-spectral and collateral knowledge, while its accuracy using only spectral knowledge is comparable with standard digital methods. >