Cluster-Based Spectral-Spatial Segmentation of Hyperspectral Imagery

In this paper, we present a multi-step method for performing unsupervised segmentation of hyperspectral imagery using modified clustering algorithms which incorporate both the spatial and spectral information present within the scene. The algorithm does not require apriori knowledge of the number of targets, spectral signatures, or any training data. Instead, the method divides the image into spatially connected regions based solely on an intrinsic notion of spectral similarity. As such, the method is expected to generalize well to many real-world analysis scenarios. The overall runtime complexity of the method is O(n
³
), but the method is amenable to GPU-based parallel computation and algorithmic optimizations which allow for considerable speedup in practice. Using a simplified, synthetic data generation model, we perform Monte Carlo simulations on various images to provide performance metrics in idealized scenarios. Finally, we provide anecdotal evidence of the viability of the method by segmenting several common hyperspectral images found in the literature and providing the graphical result.