Research And Application On The Knowledge Auto-extraction In Bayesian Expert System Classifier

Image classification and geo-information extraction is a fundamental process during the application of remote sensing technology. Among many different classification methods, the expert system classifier is effective and can synthesize multi-sources of geo-data and carry out the computer-based automatic classification based on the expert knowledge. However, the bottleneck of knowledge acquirement in classification limits the application of expert system classifier and so has not been solved till now. The traditional processes of knowledge base construction, which are realized by field experts and knowledge engineers, would take a lot of time and energy.This study focused on solving the problem of expert knowledge acquirement and knowledge base construction for the Bayesian expert system classifier. Firstly, the bottleneck of actual Bayesian expert system classifier was identified according to the process analysis. Secondly, the expert knowledge auto-extraction method which was used to estimate the prior-probability and conditional-probability for each class was realized based on statistical analysis of reference samples. Finally, the Bayesian expert system classifier with knowledge auto-extraction model (BESCKAM) was designed and verified using simulated and controlled spatial data.After the development of BESCKAM, two applications of landscape classification were carried out in Foping Nature Reserve in the Qinling Mountains and the Beijing Yanfang Industrial Region. In these applications the knowledge was extracted from the field samples and expert system classification was achieved automatically. Besides, in this study the maximum likelihood classifier and decision tree learning classifier were also used to compare with integrated BESCKAM classifier.Results showed that the prior-probabilities and conditional-probabilities could beeffectively estimated based on the statistical analysis of the reference samples, which means that the knowledge could be extracted automatically for the Bayesian expert system classifier. The landscape classification using BESCKAM classifier had a higher accuracy than the classifications by using maximum likelihood classifier or decision tree learning classifier. According to the knowledge auto-extraction method developed in this study, the bottleneck of knowledge acquirement in Bayesian expert system classifier could be well solved.