The use of a newly developed algorithm of divisive hierarchical clustering for remote sensing image analysis

Abstract

The ISODATA (or K-means) clustering algorithm has been widely used to perform unsupervised classification and to aid in collecting training samples for supervised classification. However, it requires the analyst to have some a priori knowledge on data so that several input parameters can be specified manually, particularly the initial number of clusters and their locations. When processing large data sets, it is very slow because of its iterative process. The objective of this study was to develop a divisive hierarchical clustering (DHC) algorithm to avoid the before-mentioned limitations and to serve as an alternative to ISODATA. The DHC algorithm was developed based on the concepts of the hyperplane and the binary tree. This study shows that an inappropriate assignment of initial seeds for ISODATA would reduce final classification accuracies significantly. In contrast, DHC requires a minimum of user input with only three parameters and has the ability to determine these parameters automatically from the data set itself, thereby avoiding the adverse effects associated with ISODATA. Because DHC passes through the data set only once, it is fast when processing large sets of remotely sensed data; however, the overall accuracy of DHC is slightly lower than that of ISODATA. Although the algorithm lost a bit in accuracy, it gained much in speed. In summary, this study concludes that DHC is well suited to serve as a surrogate for ISODATA when applied to remote sensing image analysis with large data sets.