Automatic clustering of multispectral imagery by maximization of the graph modularity

Abstract

Automatic clustering of spectral image data is a common problem with a diverse set of desired and potential solutions. While typical clustering techniques use first order statistics and Gaussian models, the method described in this paper utilizes the spectral data structure to generate a graph representation of the image and then clusters the data by applying the method of optimal modularity for finding communities within the graph. After defining and identifying pixel adjacencies to represent an image as an adjacency matrix, a recursive splitting is performed to group spectrally similar pixels using the method of modularity maximization. The careful selection of pixel adjacencies determines the success of this spectral clustering technique. The modularity maximization process uses the eigenvector of the modularity matrix with the largest positive eigenvalue to split groups of pixels with non-linear decision surfaces and uses the modularity measure to help estimate the optimal number of clusters to best characterize the data. Using information from each recursion, the end result is a variable level of detail cluster map that is more visually useful than previous methods. Additionally, this method outperforms many typical automatic clustering methods such k-means, especially in highly cluttered urban scenes. The optimal modularity technique hierarchically clusters spectral image data and produces results that more reliably characterize the number of clusters in the data than common automatic spectral image clustering techniques.