Classification by progressive generalization: a new automated methodology for remote sensing multichannel data

A new procedure for digital image classification is described. The procedure, labelled Classification by Progressive Generalization (CPG), was developed to avoid drawbacks associated with most supervised and unsupervised classifications. Using lessons from visual image interpretation and map making, non-recursive CPG aims to identify all significant spectral clusters within the scene to be classified. The basic principles are: (i) initial data compression using spectral and spatial techniques; (ii) identification of all potentially significant spectral clusters in the scene to be classified; (iii) minimum distance classification; and (iv) the use of spectral, spatial and large-scale pattern information in the progressive merging of the increasingly dissimilar clusters. The procedure was tested with high- (Landsat Thematic Mapper (TM)) and medium- (Advanced Very High Resolution Radiometer (AVHRR) 1 km composites) resolution data. It was found that the CPG yields classification accuracies comparable to, or better than, current unsupervised classification methods, is less sensitive to control parameters than a commonly used unsupervised classifier, and works well with both TM and AVHRR data. The CPG requires only three parameters to be specified at the outset, all specifying sizes of clusters that can be neglected at certain stages in the process. Although the procedure can be run automatically until the desired number of classes is reached, it has been designed to provide information to the analyst at the last stage so that final cluster merging decisions can be made with the analyst's input. It is concluded that the strategy on which the CPG is based provides an effective approach to the classification of remote sensing data. The CPG also appears to have a considerable capacity for data compression.