PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yaoming Cai, Xiaobo Liu, Zhihua Cai

arXiv:1904.08269v1cs.CVcs.LG

TL;DR

HSI band selection must reduce redundancy without adequately capturing the nonlinear, global interactions among highly correlated spectral bands. BS-Net addresses this by learning band weights through attention and reconstructing the HSI from selected information; its FC and convolutional variants report better classification performance than competing methods across the evaluated images.

  • Problem

    Existing band-selection methods often estimate each band independently and do not fully model nonlinear, global spectral interactions, despite redundancy and Hughes-phenomenon challenges in HSI.

  • Method

    BS-Net treats selection as sparse spectral reconstruction, using a Band Attention Module to learn nonlinear band interdependencies and a Reconstruction Network to restore the original HSI.

  • Results

    BS-Net-FC and BS-Net-Conv outperform the compared methods in OA, AA, and Kappa, with BS-Net-Conv improving OA by 6.59% over BS-Net-FC.

  • Takeaways & Limitations

    The framework adaptively produces sparse band weights and provides flexible spectral-based and spectral-spatial implementations for HSI band selection.

Abstract

from arXiv · show

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.

I. INTRODUCTION

HSI’s many highly correlated spectral bands increase dimensionality, computational complexity, and Hughes-phenomenon risk. The paper frames band selection as sparse reconstruction and proposes deep networks to model nonlinear global spectral relationships.

  • Hundreds of narrow, highly correlated HSI bands increase time and space complexity and contribute to the Hughes phenomenon.
  • Band selection reduces redundancy by retaining an informative subset while preserving HSI physical properties more directly than feature extraction.
  • Many existing methods evaluate bands independently or use linear transformations, limiting their treatment of nonlinear and global spectral interrelationships.
  • BS-Net treats band selection as sparse spectral reconstruction, assuming the HSI can be reconstructed from a few informative bands.
  • The proposed framework uses deep neural networks to learn nonlinear spectral interdependencies and select significant bands from complete HSI cubes.

B. Convolutional Neural Networks

The paper introduces CNNs as nonlinear tensor-to-output functions trained by feedforward computation and gradient-based backpropagation. It then motivates band attention as channel-wise weighting of informative HSI bands.

  • B. Convolutional Neural Networks: A classical CNN applies multiple convolutional layers with nonlinear activations to map an input tensor x to an output y.
  • B. Convolutional Neural Networks: CNN training alternates feedforward feature computation with backpropagation that updates trainable parameters by gradient descent to minimize a cost function.
  • C. Attention Mechanism: Attention assigns importance weights through a gating function and nonlinear feature transformations.
  • C. Attention Mechanism: Visual attention may model spatial, channel-wise, or combined spatial-channel relationships.
  • C. Attention Mechanism: BS-Nets use channel-attention-like band attention to emphasize informative HSI bands and reduce trivial-band influence.

III. BS-NETS

BS-Nets convert band selection into sparse reconstruction: BAM learns band weights, BRW reweights inputs, and RecNet reconstructs the original HSI. Sparse regularization and averaged weights identify the informative top-k bands.

  • Architecture: BS-Nets comprise a band attention module, band re-weighting operation, and reconstruction network.BAM learns weights, BRW applies band-wise multiplication, and RecNet predicts the original input.
  • Band attention module: BAM maps an HSI cube to non-negative band weights while modeling interdependencies between spectral bands.A Sigmoid output activation guarantees non-negative weights.
  • Band re-weighting: BRW multiplies each input band by its learned weight to produce a re-weighted tensor for reconstruction.The re-weighted tensor is passed to RecNet, which outputs a prediction of the original input.
  • Training objective: Mean-square reconstruction error measures recovery quality, while an L1 constraint encourages sparse, interpretable band weights.The regularization coefficient λ balances reconstruction error against the regularization term, and optimization can use SGD or Adam.
  • Band selection: Bands with larger average weights are treated as more significant, and the top k bands form the selected subset.Average weights are computed across all training samples before selection.

B. BS-Net Based on Fully Connected Networks (BS-Net-FC)

BS-Net-FC models nonlinear spectral relationships with fully connected BAM and RecNet components, using individual spectral vectors as training samples.

  • Architecture: BS-Net-FC implements both BAM and RecNet with fully connected networks to model nonlinear spectral relationships.Its BAM uses a bottleneck structure with multiple fully connected layers and ReLU hidden activations.
  • Inputs and re-weighting: BS-Net-FC represents its training set as spectral-vector samples in a 4-D tensor X ∈R^S×1×1×b.Here S = M × N, and BRW performs element-wise multiplication Z = X ⊗W.
  • Reconstruction: The RecNet in BS-Net-FC uses a simple multi-layer perceptron to reconstruct the re-weighted spectral inputs.The excerpt states that the MLP uses the same number of hidden neurons with ReLU activations.

C. BS-Net Based on Convolutional Networks (BS-Net-Conv)

BS-Net-Conv extends BS-Net-FC by incorporating spectral-spatial information through convolutional processing, patch-based samples, and a convolutional-deconvolutional reconstruction network.

  • Motivation: BS-Net-Conv addresses BS-Net-FC’s omission of spatial information by using convolutional networks with spectral-spatial inputs.It is introduced to enhance use of the spectral-spatial information contained in HSI.
  • Band attention module: Its BAM uses 2-D convolutions, global pooling, and fully connected layers to generate band weights from spectral-spatial features.The convolutional layers extract spectral and spatial information simultaneously before spatial reduction.
  • Reconstruction: BS-Net-Conv uses a convolutional-deconvolutional network as RecNet to reconstruct the original HSI.The architecture is described as analogous to a classical auto-encoder.
  • Training samples: BS-Net-Conv produces fewer training samples than BS-Net-FC because it uses sliding HSI patches rather than individual pixels.The passage explicitly contrasts the sample counts of the two implementations.

D. Remarks on BS-Net framework

BS-Net uses learned band weights within a flexible framework evaluated across three real HSI datasets and multiple quantitative criteria.

  • D. Remarks on BS-Net framework: BS-Net explicitly learns spectral-band weights with deep neural networks and supports flexible implementations across network types.The framework is end-to-end trainable and can adaptively extract spectral and spatial information while modeling nonlinear band relationships.
  • D. Remarks on BS-Net framework: The experiments evaluate Indian Pines, Pavia University, and Salinas using classification performance against multiple existing band-selection methods.Evaluation includes comparisons with ISSC, SpaBS, MVPCA, SNMF, MOBS, and OPBS, plus an all-bands reference.
  • D. Remarks on BS-Net framework: SVM classification uses 5% labeled samples for training, the remainder for testing, and 20 independent runs to compute OA, AA, and Kappa.The classifier uses a radial basis function kernel, and training and testing samples are randomly selected for each run.
  • D. Remarks on BS-Net framework: Entropy measures band information content, whereas mean spectral divergence measures redundancy among selected bands.Higher entropy indicates more image detail, while higher MSD indicates less redundancy among selected bands.
  • D. Remarks on BS-Net framework: BS-Nets are accelerated on a NVIDIA TITAN Xp GPU, while baseline methods use an Intel Xeon CPU environment.HSI pixel values are scaled to [0, 1], and implementations use Python, Matlab, or TensorFlow-GPU according to the method.

B. Results on Indian Pines Data Set

Indian Pines is a 145×145-pixel AVIRIS scene with 224 spectral bands, reduced to 200 after removing water-absorption regions.

  • B. Results on Indian Pines Data Set: The scene includes agriculture, forest, natural vegetation, highways, a rail line, housing, structures, and roads.Its ground truth contains sixteen classes that are not all mutually exclusive.
  • B. Results on Indian Pines Data Set: Figure 4 presents convergence analyses for BS-Net-FC and BS-Net-Conv on Indian Pines.Panels show loss versus accuracy across iterations and average band weights under varying iterations.

2) Analysis of Convergence of BS-Nets :

On Indian Pines, BS-Nets converge quickly, improve classification with few bands, and select distributed informative bands while highlighting limitations of MSD as a standalone criterion.

  • 2) Analysis of Convergence of BS-Nets :: 21% and 12% classification-accuracy improvements are obtained by BS-Net-FC and BS-Net-Conv, respectively, during convergence analysis.BS-Net-FC rises from 42% to 63%, while BS-Net-Conv rises from 52% to 64%; losses approach convergence rapidly.
  • 2) Analysis of Convergence of BS-Nets :: Band-weight distributions become progressively sparser, making informative bands easier to distinguish as trivial bands receive very small weights.Average weights are visualized across band number and training iterations.
  • 3) Performance Comparison:: BS-Net-Conv achieves the best OA, AA, and Kappa when the selected subset exceeds five bands, while BS-Net-FC is comparable to MOBS and exceeds five other competitors.BS-Net-Conv needs 5, 17, and 17 bands to outperform all-bands performance on OA, AA, and Kappa, respectively.
  • 4) Analysis of the Selected Bands:: Both BS-Nets select relatively uniformly distributed, high-entropy bands while avoiding identified low-entropy noisy regions.Their selected subsets contain fewer continuous adjacent bands, which is associated with reduced redundancy.
  • 4) Analysis of the Selected Bands:: BS-Net-Conv has the best classification performance without the best MSD, because noisy bands can increase MSD despite reducing classification performance.BS-Net-FC has MSD comparable to OPBS and better than most competitors; example subsets [104, 144] and [104, 25] have MSDs 106.64 and 51.49, respectively.

C. Results on Pavia University Data Set

On Pavia University, BS-Nets converged to sparse band weights and achieved strong classification performance with small informative subsets. BS-Net-FC was strongest for larger subsets, while both methods remained competitive using 25 bands.

  • Both BS-Nets converged after several iterations, with losses tending to zero.
  • 78% to 83% and 72% to 83%: best-five-band accuracy increased for BS-Net-FC and BS-Net-Conv, respectively.
  • Increasing iterations produced sparse, distinguishable band weights and highlighted only a few reconstruction-useful bands.
  • Below 17 bands, both BS-Nets significantly outperformed other methods on OA, AA, and Kappa.
  • Using only 25 bands, both BS-Nets matched all-band performance without an obvious Hughes phenomenon.
  • With 19 bands, both BS-Nets achieved the best OA, AA, and Kappa among competitors, while BS-Net-FC won in 7 classes.

4) Analysis of the Selected Bands:

Across Pavia University and Salinas, BS-Nets selected less redundant, more distributed bands and avoided low-entropy or noisy regions. Their convergence and classification behavior supported sparse informative subsets.

  • Pavia University: On Pavia University, BS-Net subsets contained fewer continuous bands and concentrated more around higher-entropy positions than competitor selections.
  • Pavia University: Both Pavia University BS-Nets had MSD values comparable with OPBS and better than ISSC, SpaBS, MVPCA, and MOBS.
  • Salinas: After about 20 iterations on Salinas, BS-Nets’ loss and accuracy converged, while learned band weights became sparse.
  • Salinas: On Salinas, BS-Net-Conv avoided sharply decreasing low-entropy regions, including bands [106, 107] and [146, 147], while selected bands were relatively uniform and less adjacent.
  • Salinas: SNMF and ISSC achieved better MSDs than BS-Net-FC but selected noisy bands that reduced classification performance.

4) Analysis of the Selected Bands:

On Salinas and across runtime comparisons, BS-Nets combined strong classification with competitive computational cost. Their selected bands avoided noise-related trade-offs, and the framework is positioned for future integration with deep classifiers.

  • Salinas performance: When subsets exceeded 19 bands on Salinas, both BS-Nets significantly outperformed competing methods and all bands on OA, AA, and Kappa.
  • Salinas performance: With 19 bands on Salinas, both BS-Nets were generally superior on most classes.
  • Selected-band analysis: BS-Net-FC avoided noisy bands and achieved better classification performance than other methods despite having relatively lower MSD.
  • Computational cost: BS-Net runtimes were comparable with MOBS and significantly faster than SpaBS and SNMF.
  • Conclusion: The proposed BS-Nets balanced classification performance and running time across the evaluated comparisons.
  • Conclusion: The framework may be combined with deep-learning classification methods in future work to reduce computational complexity and enhance classification performance.
Loading 1904.08269v1…