Hyperspectral band selection based on multi-feature metric and hierarchically optimized spectral clustering

Band selection is an effective method for dimensionality reduction of hyperspectral images, and the key challenge lies choosing the appropriate bands. This paper proposes an unsupervised band selection method based on multi-feature measure and hierarchically optimized spectral clustering (MHSC-BS). This method improves the traditional spectral clustering algorithm by combining multiple similarity metrics to form a new metric method to quantify the similarity between bands. Additionally, hierarchical clustering is used to optimize k-means clustering in the clustering stage, which captures the data’s hierarchical structure while improving clustering’s stability. Finally, the silhouette coefficient index is used to preliminarily evaluate the clustering effect to determine the appropriate number of clusters, and representative bands are selected from the clusters obtained by clustering. Experiments on two hyperspectral datasets verify the effectiveness of the proposed method.