Exploring the Integration of Spatial Information into a Spectral Clustering Approach to Unsupervised Band Selection

Band selection methods are crucial for mitigating the challenge of high dimensionality in hyperspectral image data. Several unsupervised band selection techniques have already shown success, including the Particle Swarm Optimisation with Fuzzy Clustering Means (PSO-FCM). However, such an approach does not exploit the available spatial information. The work herein thus proposes Superpixel-enhanced PSO-FCM (SPPF), that exploits both the spectral and spatial redundancy in hyperspectral image data. To ensure more optimal exploitation of the local spatial context of the Superpixel map, various distance measures, together with various cluster-representative selection techniques are investigated in the light of different datasets. Comparison of the proposed SPPF with the original PSO-FCM shows the promise of this spatial-spectral BS solution.