BS-Nets: An End-to-End Framework For Band Selection of Hyperspectral Image (1904.08269v1)

Abstract: Hyperspectral image (HSI) consists of hundreds of continuous narrow bands with high spectral correlation, which would lead to the so-called Hughes phenomenon and the high computational cost in processing. Band selection has been proven effective in avoiding such problems by removing the redundant bands. However, many of existing band selection methods separately estimate the significance for every single band and cannot fully consider the nonlinear and global interaction between spectral bands. In this paper, by assuming that a complete HSI can be reconstructed from its few informative bands, we propose a general band selection framework, Band Selection Network (termed as BS-Net). The framework consists of a band attention module (BAM), which aims to explicitly model the nonlinear inter-dependencies between spectral bands, and a reconstruction network (RecNet), which is used to restore the original HSI cube from the learned informative bands, resulting in a flexible architecture. The resulting framework is end-to-end trainable, making it easier to train from scratch and to combine with existing networks. We implement two BS-Nets respectively using fully connected networks (BS-Net-FC) and convolutional neural networks (BS-Net-Conv), and compare the results with many existing band selection approaches for three real hyperspectral images, demonstrating that the proposed BS-Nets can accurately select informative band subset with less redundancy and achieve significantly better classification performance with an acceptable time cost.

• The paper introduces BS-Nets, an end-to-end deep learning framework with Band Attention and Reconstruction modules for efficient hyperspectral image band selection by reconstructing the original image.
• Two BS-Net variations are proposed: BS-Net-FC focuses on spectral data, while BS-Net-Conv incorporates spectral-spatial features for enhanced selection accuracy and context.
• Experiments show BS-Nets, especially BS-Net-Conv, outperform traditional methods on standard datasets, achieving higher classification accuracy with fewer bands.

Overview of BS-Nets: An End-to-End Framework for Hyperspectral Image Band Selection

The paper presents an innovative framework, termed Band Selection Network (BS-Net), for hyperspectral image (HSI) band selection aimed at addressing challenges related to the Hughes phenomenon and high computational costs in HSI processing. The pivotal idea is the assumption that a complete HSI can be effectively reconstructed from a few informative bands, thereby making it possible to perform efficient spectral and spatial feature extraction.

Methodology

The BS-Net framework is structured into two core modules: the Band Attention Module (BAM) and the Reconstruction Network (RecNet). The BAM is integral in modeling the nonlinear interdependencies across spectral bands, while RecNet aims to restore the original HSI cube from the selected informative bands. The end-to-end architecture supports seamless training and integration with existing networks, ensuring a flexible design adaptable to different implementation scenarios.

Two variations of the BS-Nets have been proposed:

• BS-Net-FC: Employs fully connected networks, focusing on spectral information.
• BS-Net-Conv: Utilizes convolutional neural networks to incorporate both spectral and spatial data, enhancing selection accuracy by contextual feature extraction.

Experimental Evaluation

The authors conducted experiments on three widely-acknowledged HSI datasets—Indian Pines, Pavia University, and Salinas. The comparison of BS-Nets-FC and BS-Nets-Conv implementations against six traditional band selection methods demonstrated notable improvements in classification accuracy and a reduction in redundancy and noise among the selected bands. Particularly, BS-Net-Conv consistently outperformed others in overall accuracy (OA), average accuracy (AA), and Kappa across varying band subset sizes. The evaluations employed SVM classifiers to assess the selected band subsets’ efficacy, revealing BS-Net’s robustness against the dimensionality curse through achieving higher classification performance with fewer bands than traditional methods.

Implications and Future Work

The proposed BS-Nets present a promising approach to HSI band selection, emphasizing the importance of leveraging deep neural networks to capture complex spectral-spatial relationships. The implementation flexibility allows for customization based on specific application needs, potentially leading to improved performance in HSI-driven domains like agriculture, land management, and medical imaging.

Future work could explore deeper integration of BS-Nets with more specialized classification networks to further boost performance while minimizing computational load. Additionally, exploring hybridization with other deep learning paradigms such as generative adversarial models or recurrent neural networks could open new possibilities for optimizing information extraction from HSIs, enhancing both capability and resilience in various practical contexts.