期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2008
卷号：XXXVII Part B7
页码：601-606
出版社：Copernicus Publications
摘要：Classification methods are often employed to derive land cover information from satellite images. Although a variety of classifiers have been developed, some primary issues remain to be further investigated. Among others, two of them are: a) the determination of the number of distinct clusters, and b) obtaining optimal classification result utilizing the available classifiers. To investigate the first issue, a skewness measurement and a separation-cohesion index (SCI) are used to describe some characteristics of a clustering result. By plotting the two indices against the number of clusters, kinks and slope changes may emerge in the curves. With consideration of the spatial resolution of the imagery and the context of an application, an optimal number of clusters can then be determined. For most cases, the clusters in remote sensing images are different from the land cover types of interest. We need either to merge some clusters in order to form a land cover type or to split a cluster into more than one land cover types of interest. Because the statistical model of a land cover type does not always follow a distribution pattern, it may contain multiple models, or has no noticeable patterns, a collection of classifiers are proposed to accommodate any scenarios. This study used three classifiers: the maximum likelihood method for parametric model based approaches, the Kohonen's self-organizing map (SOM) for neural networks, and a classification tree method. Through case studies, practical procedures are proposed: 1) identify the number of clusters within an application context; 2) associate clusters with land cover types; 3) classify images using the three classifiers, and assess their accuracies; 4) accept the result of classes from any one of the three classifiers, and process the remaining classes in the next iteration, and each class is then analyzed independently. The proposed procedures were shown to be efficient using case studies with three imagery data sets
关键词：Remote sensing; Clustering; Classification; Cluster number identification; Classification optimization
