Optimal Clustering Framework for Hyperspectral Band Selection

Qi Wang, Fahong Zhang, Xuelong Li

Abstract

Band selection, by choosing a set of representative bands in hyperspectral image (HSI), is an effective method to reduce the redundant information without compromising the original contents. Recently, various unsupervised band selection methods have been proposed, but most of them are based on approximation algorithms which can only obtain suboptimal solutions toward a specific objective function. This paper focuses on clustering-based band selection, and proposes a new framework to solve the above dilemma, claiming the following contributions: 1) An optimal clustering framework (OCF), which can obtain the optimal clustering result for a particular form of objective function under a reasonable constraint. 2) A rank on clusters strategy (RCS), which provides an effective criterion to select bands on existing clustering structure. 3) An automatic method to determine the number of the required bands, which can better evaluate the distinctive information produced by certain number of bands. In experiments, the proposed algorithm is compared to some state-of-the-art competitors. According to the experimental results, the proposed algorithm is robust and significantly outperform the other methods on various data sets.

Introduction

Hyperspectral sensors capture hundreds of narrow spectral bands, providing a "data cube" rich in detail but plagued by massive redundancy. This leads to the "Hughes Phenomenon": as the number of features (bands) increases, the accuracy of a classifier actually decreases unless the amount of training data grows exponentially. To mitigate this, researchers use Band Selection to identify a small subset of representative bands. Most existing methods rely on "greedy" searches or "evolutionary" algorithms, which are prone to local optima. The challenge is the combinatorial explosion; clustering 200 bands into 15 groups involves roughly $10^{223}$ possibilities, making brute-force search impossible.

Methodology: OCF and RCS

The proposed approach moves from heuristic searches to mathematically optimal solutions through two primary components: the Optimal Clustering Framework (OCF) and the Rank on Clusters Strategy (RCS).

Overview of the proposed band selection framework consisting of OCF and RCS.

Optimal Clustering Framework (OCF)

The OCF leverages the Contiguous Band Indexes Constraint (CBIC). By observing that bands with similar wavelengths are naturally more correlated, the authors propose that an optimal cluster should only contain contiguous bands. This physical constraint drastically shrinks the search space and introduces an "optimal substructure." Using Dynamic Programming, the problem is broken into smaller sub-problems, allowing the global maximum of an objective function (such as Normalized Cut) to be found in $O(L^2K)$ time.

Rank on Clusters Strategy (RCS)

Once the optimal clusters are formed, the RCS provides a criterion for selecting the most representative band within each cluster. Instead of simply picking the center band, RCS ranks all bands based on importance metrics like variance or information entropy, selecting the highest-ranked band from each optimal partition.

Physical Insights and Global Optimality

The core innovation lies in the realization that the smooth curve of spectral reflectance justifies the contiguous constraint. This physical property makes the dynamic programming approach viable and guarantees a globally optimal result for the chosen objective function, unlike methods like k-means which are sensitive to initialization.

Visualization of spectral bands showing the smooth curve of spectral reflectance which justifies the contiguous constraint.

To enhance robustness, the authors introduced the Top-Rank Cut (TRC) objective function. TRC focuses on maximizing the contribution of the most discriminative bands, making the selection process more stable on noisy datasets compared to standard methods that might be misled by high-variance noise.

Experimental Results

The framework was evaluated on four classic hyperspectral benchmarks: Indian Pines, Pavia University, Salinas Scene, and the Kennedy Space Center (KSC). Across all datasets, the OCF-based methods consistently outperformed state-of-the-art competitors in classification accuracy. Furthermore, the method demonstrated high stability across different classifiers, including K-Nearest Neighbors (KNN), SVM, and Linear Discriminant Analysis (LDA). Despite its mathematical rigor, the $O(L^2K)$ complexity remains efficient, processing datasets in seconds on modern workstations.

Conclusion

This research presents a significant advancement in HSI processing by providing a framework for globally optimal band selection. While current efficiency is slightly limited by initial similarity matrix calculations, future work will explore incorporating manifold learning to better understand the underlying geometric structure of the data. The general nature of the framework also allows for the future integration of even more powerful objective functions.

Summary updated

Abstract

Band selection, by choosing a set of representative bands in hyperspectral image (HSI), is an effective method to reduce the redundant information without compromising the original contents. Recently, various unsupervised band selection methods have been proposed, but most of them are based on approximation algorithms which can only obtain suboptimal solutions toward a specific objective function. This paper focuses on clustering-based band selection, and proposes a new framework to solve the above dilemma, claiming the following contributions: 1) An optimal clustering framework (OCF), which can obtain the optimal clustering result for a particular form of objective function under a reasonable constraint. 2) A rank on clusters strategy (RCS), which provides an effective criterion to select bands on existing clustering structure. 3) An automatic method to determine the number of the required bands, which can better evaluate the distinctive information produced by certain number of bands. In experiments, the proposed algorithm is compared to some state-of-the-art competitors. According to the experimental results, the proposed algorithm is robust and significantly outperform the other methods on various data sets.
BibTeX
@article{wang2018optimal,
  author = {Qi Wang and Fahong Zhang and Xuelong Li},
  title = {{Optimal Clustering Framework for Hyperspectral Band Selection}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2018},
  url = {https://openreview.net/forum?id=zRVTYNOHIe},
}
Figure