[Paper Review] Band Attention Convolutional Networks For Hyperspectral Image Classification

Hongwei Dong, Lamei Zhang+1 more|arXiv (Cornell University)|Jun 11, 2019

Remote-Sensing Image Classification25 references20 citations

TL;DR

This paper proposes a Band Attention Module (BAM) for hyperspectral image (HSI) classification that enables deep CNNs to adaptively weight or select informative spectral bands, reducing noise and redundancy. The BAM is a plug-and-play component that integrates into existing CNN architectures for end-to-end training, achieving state-of-the-art performance on benchmark datasets like Indian Pines and KSC.

ABSTRACT

Redundancy and noise exist in the bands of hyperspectral images (HSIs). Thus, it is a good property to be able to select suitable parts from hundreds of input bands for HSIs classification methods. In this letter, a band attention module (BAM) is proposed to implement the deep learning based HSIs classification with the capacity of band selection or weighting. The proposed BAM can be seen as a plug-and-play complementary component of the existing classification networks which fully considers the adverse effects caused by the redundancy of the bands when using convolutional neural networks (CNNs) for HSIs classification. Unlike most of deep learning methods used in HSIs, the band attention module which is customized according to the characteristics of hyperspectral images is embedded in the ordinary CNNs for better performance. At the same time, unlike classical band selection or weighting methods, the proposed method achieves the end-to-end training instead of the separated stages. Experiments are carried out on two HSI benchmark datasets. Compared to some classical and advanced deep learning methods, numerical simulations under different evaluation criteria show that the proposed method have good performance. Last but not least, some advanced CNNs are combined with the proposed BAM for better performance.

Motivation & Objective

• To address the challenge of spectral redundancy and noise in hyperspectral images (HSIs) that degrade CNN-based classification performance.
• To design a deep learning framework tailored to HSI characteristics, rather than adapting general-purpose CNNs.
• To enable end-to-end learning of band selection or weighting, avoiding separate, non-differentiable stages.
• To develop a plug-and-play module that enhances existing CNNs for HSI classification without architectural redesign.
• To validate the generalizability and effectiveness of the proposed BAM across diverse CNN backbones and benchmark datasets.

Proposed method

• The Band Attention Module (BAM) is designed to learn channel-wise attention weights for spectral bands using global average pooling and fully connected layers.
• The BAM processes the input HSI through a series of 1D convolutions and global pooling to generate a weight vector of size c×1 for c spectral bands.
• The learned weights are applied via channel-wise multiplication to re-weight the input bands, emphasizing informative bands and suppressing noisy or redundant ones.
• The BAM is integrated as a plug-in module into standard CNNs, enabling end-to-end training with joint optimization of band weighting and classification.
• The method uses a ratio r in 1D convolution layers to control information aggregation, with r=2 found optimal in ablation studies.
• The framework is evaluated on Indian Pines and KSC datasets using multiple backbones, including VGGNet, ResNet, DenseNet, and Xception.

Experimental results

Research questions

• RQ1 Can a learnable band attention mechanism improve HSI classification accuracy by selectively emphasizing informative spectral bands?
• RQ2 How does the proposed BAM compare to existing attention mechanisms like squeeze-and-excitation (SE) blocks in the context of HSI classification?
• RQ3 Does the BAM module generalize across different CNN architectures and achieve consistent performance gains?
• RQ4 What hyperparameters (e.g., training epochs, compression ratio r) significantly affect the performance of the BAM-based model?
• RQ5 Can the BAM be used as a plug-and-play component to enhance existing CNNs without architectural modification?

Key findings

• The proposed BAM achieved the highest overall accuracy (OA) of 95.06% on the KSC dataset, outperforming all compared methods including SE blocks and standard CNNs.
• On the Indian Pines dataset, the BAM-based model achieved an OA of 91.38%, with a 4–5% improvement over the baseline CM and at least a 2% improvement over SE blocks.
• The BAM improved classification accuracy by 4–5% across multiple evaluation metrics (OA, AA, Kappa) on the KSC dataset compared to the baseline CM.
• The BAM demonstrated strong generalization, consistently improving performance when combined with advanced backbones such as ResNet, DenseNet, Xception, and TSCNN.
• Hyperparameter analysis showed that r=2 and 1000 training epochs yielded optimal performance, with convergence observed after 500 epochs.
• The BAM module is effective as a plug-and-play component, enhancing performance across diverse CNN architectures without architectural redesign.