An unsupervised classifier for remote-sensing imagery based on improved cellular automata

Abstract

Traditional unsupervised classification algorithms for remote-sensing images, such as k-means (KM), have been widely used for massive data sets due to their simplicity and high efficiency. However, they do not usually take the interaction between neighbouring pixels into account, but only take individual pixels as the elements for clustering and classification. According to Tobler’s first law of geography, everything is related to everything else, but near things are more related than distant things. To make use of the spatial interaction between pixels, the cellular automata method can be employed to improve the accuracy of image classification. In cellular automata theory, the state of a cell at the next moment is determined by its current state and that of its neighbours. In traditional cellular automata methods, which are based on a standard neighbour configuration, even if the influence of neighbouring cells on the central cell is measured, the weights of these influences are the same. Hence, this article proposes an improved cellular automata method for image classification by allowing the cellular automata to diffuse in a geometrical circle, and by measuring the influence of the neighbouring cells using a fuzzy membership function. The proposed classifier was tested with typical Landsat Enhanced Thematic Mapper Plus (ETM+) and high-resolution images. The experiments reveal that the new classifier can achieve better results, in terms of overall accuracy and kappa coefficient, than cellular automata classifier based on Moore type (CAS), KM, and fuzzy c-means.

Acknowledgements

The authors gratefully acknowledge Professor Hu Xiangyun (Wuhan University, China) for his suggestions on algorithms and structure. This study was supported by funding from the National Natural Science Foundation of China (NSFC: 40871179).