Using fuzzy sets to improve cluster labelling in unsupervised classification

摘要

This paper presents a fuzzy set approach to improving the cluster labelling process for unsupervised classification. Assuming normality for spectral data distribution, this approach represents both spectral clusters and the spectral signatures of information classes as fuzzy sets, with their fuzzy membership functions being defined using spectral means and variances. A fuzzy similarity index, defined through a fuzzy binary relation between each cluster fuzzy set and its corresponding class fuzzy set, is derived by applying a max-min fuzzy composition to the paired fuzzy sets. A second max-min fuzzy composition was performed to determine the final labelling of individual clusters based on these indices. Using a Thematic Mapper (TM) image of Imperial Valley, California as the test dataset, the approach was compared with both a human interpretation and a supervised classification resulting from the maximum-likelihood classifier. Both the process and the results indicated that the fuzzy approach surpassed the human interpretation in speed and equated the supervised classification in accuracy. In comparison to the Bayesian maximum-likelihood classifier, this approach supports a quick identification of mixed clusters and missing classes and incomplete classification. In addition, the fuzzy labelling approach provides a great potential for rapid a posteriori cluster validation.