Efficient Unsupervised Band Selection for Hyperspectral Imagery with Mamba-based Classifier – An In-Depth Comparative Analysis

TMLR Paper6888 Authors

Abstract: Band selection is a critical step in processing hyperspectral imagery (HSI), reducing input dimensionality to mitigate redundancy, enhance computational efficiency and improve learning accuracy. Efficient unsupervised deep-learning-based band selection methods have recently garnered significant attention due to their strong feature representation capabilities. In existing literature, we observe that there is a broader and more general line of research regarding feature selection, which some recent deep learning-based HSI band selection methods have drawn inspiration from. This work concentrates on efficient unsupervised deep-learning-based band selection methods from the standpoint of unifying two research lines: the more general feature selection and the more specific HSI band selection. Specifically, we conduct an in-depth comparative analysis in terms of downstream classification performance and computation cost, on six state-of-the-art efficient unsupervised HSI band selection methods, of which one does not involve deep learning and the other five do. Classification experiments are carried out using three publicly available remote sensing benchmark datasets, where we incorporate a recent Mamba-based classifier that outperforms the typical SVM substantially in classification accuracy by a ∼10-20% margin. To our best knowledge, this is the first work that puts together and compares the aforementioned efficient unsupervised methods in the context of HSI band selection and employs a Mamba-based classifier in the analysis.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Patrick_Flaherty1

Submission Number: 6888