An Improvement on the Clustering of High-Resolution Satellite Images Using a Hybrid Algorithm

Abstract

The challenge of obtaining training data for supervised classifications of satellite images has led researchers to unsupervised algorithms, i.e. cluster analysis. Numerous researches have been conducted to improve quality and decrease uncertainty of results of this analysis. This study proposes a hybrid cost function as well as a hybrid clustering algorithm-Artificial Bee Colony optimization approach for the clustering of high-resolution satellite images. In order to evaluate viability of the proposed methodology, it is compared to some other classic clustering algorithms such as modified K-Means, K-Medoids, Fuzzy C-Means, and Kernel-based Fuzzy C-Means methods over three different study areas selected from a WorldView-2 satellite image. The Shannon entropy technique, Kappa coefficient, compactness, and separation criteria are used as quality and uncertainty indicators for the evaluation. The results of the study show that, compared to other methods, the hybrid algorithm obtained from the proposed cost function, Kernel-based Fuzzy C-Means method, and ABC algorithm provide clustering capabilities of higher quality and lower uncertainty levels.