PaperlayerView exact revision on arXiv

Source-linked AI summary

Hyperspectral Image Denoising Employing a Spatial-Spectral Deep Residual Convolutional Neural Network

Qiangqiang Yuan, Qiang Zhang, Jie Li, Huanfeng Shen, Liangpei Zhang

arXiv:1806.00183v3cs.CV

TL;DR

HSI noise harms subsequent interpretation, while existing approaches face spectral, tuning, and generality challenges. The paper proposes HSID-CNN, a combined spatial-spectral residual CNN with multi-scale extraction and multi-level representation. Simulated and real-data experiments report stronger denoising evaluation, visual quality, and classification accuracy than mainstream methods.

  • Problem

    HSI noise affects subsequent interpretation, while existing methods can cause spectral distortion, require per-image parameter tuning, and fail to model HSI spectral redundancy.

  • Method

    HSID-CNN learns a nonlinear end-to-end noisy-to-clean mapping using combined spatial-spectral convolutions, multi-scale feature extraction, residual learning, and multi-level feature representation.

  • Results

    HSID-CNN outperforms many mainstream methods in evaluation indexes, visual effect, and classification accuracy across simulated and real-data denoising experiments.

  • Takeaways & Limitations

    Joint spatial-spectral learning with multi-scale and multi-level features provides the paper’s supported framework for recovering noise-free HSIs.

Abstract

from arXiv · show

Hyperspectral image (HSI) denoising is a crucial preprocessing procedure to improve the performance of the subsequent HSI interpretation and applications. In this paper, a novel deep learning-based method for this task is proposed, by learning a non-linear end-to-end mapping between the noisy and clean HSIs with a combined spatial-spectral deep convolutional neural network (HSID-CNN). Both the spatial and spectral information are simultaneously assigned to the proposed network. In addition, multi-scale feature extraction and multi-level feature representation are respectively employed to capture both the multi-scale spatial-spectral feature and fuse the feature representations with different levels for the final restoration. The simulated and real-data experiments demonstrate that the proposed HSID-CNN outperforms many of the mainstream methods in both the quantitative evaluation indexes, visual effects, and HSI classification accuracy.

I. INTRODUCTION

HSI noise degrades later interpretation, while existing denoising methods struggle with spectral distortion, per-image parameter tuning, and complex spatial-spectral noise. The paper addresses these limitations with HSID-CNN, combining spatial-spectral learning and deep end-to-end restoration.

  • HSIs contain spatial and spectral information but are affected by random noise, stripe noise, and dead pixels.
  • Noise substantially affects subsequent HSI information interpretation and understanding, making denoising important before analysis.
  • Band-by-band denoising methods often cause larger spectral distortion because they do not jointly preserve HSI spectral correlations.
  • Spatial-spectral methods can improve results but require precise parameter tuning for each HSI, making processing unintelligent and time-consuming.
  • HSI denoising needs a method combining spatial-spectral constraints with deep learning because generic image denoisers do not model spectral redundancy.
  • HSID-CNN learns a nonlinear end-to-end mapping from noisy to clean HSIs using combined 2D spatial and 3D spatial-spectral convolutions.

A. Hyperspectral Noise Degradation Model

The section contrasts transform- and spatial-domain HSI denoising approaches with the need for a faster, more efficient, and more universal framework across diverse HSI data.

  • HSI denoising methods are broadly organized into transform-domain and spatial-domain approaches.
  • Transform-domain methods separate clear signals from noisy data using principal component, Fourier, or wavelet transformations.
  • Spatial-domain methods use priors such as total variation, non-locality, sparse representation, and low-rank models to preserve spatial and spectral characteristics.
  • Existing methods may require precise parameter tuning for each HSI, making them unintelligent and time-consuming across different data.
  • The proposed framework targets fast, efficient, and universal adaptation to HSIs with different situations.

III. METHODOLOGY

HSID-CNN combines noisy-band spatial data with adjacent-band spectral data in a residual deep CNN, adding multi-scale extraction and multi-level feature fusion for restoration.

  • The Proposed Framework Description: HSID-CNN combines a single noisy spatial band with its adjacent spectral bands to learn a nonlinear end-to-end denoising mapping.
  • The Proposed Framework Description: The architecture accepts W × H spatial input and W × H × K spectral-cube input, enabling denoising one spatial band at a time regardless of total band count.
  • Joint Spatial-Spectral Multi-Scale Feature Extraction: A combined 2D and 3D CNN exploits single-band characteristics together with correlations and complementarity among adjacent bands.
  • Joint Spatial-Spectral Multi-Scale Feature Extraction: Multi-scale convolutional kernels capture spatial and spectral features across diverse receptive-field sizes for multi-context noise removal.
  • Deep CNN with Residual Learning Strategy: Residual learning represents the reconstructed output through residual noise, improving training stability and efficiency while reducing degradation in deeper networks.
  • Multi-Level Feature Representation for Restoration: The multi-level representation unit concatenates feature maps from different-depth layers and uses multiple skip-connections to transfer comprehensive features for final restoration.

IV. EXPERIMENTAL RESULTS AND DISCUSSION

The experiments evaluate HSID-CNN against mainstream HSI denoising methods on simulated and real data using normalized imagery and multiple quality measures. The design uses fixed adjacent-band context and datasets from Washington DC Mall, Indian Pines, and University of Pavia.

  • Experimental Setup: HSID-CNN is compared with HSSNR, LRTA, BM4D, and LRMR using MPSNR, MSSIM, and MSA in simulated experiments.
  • Parameter Setting and Network Training: The adjacent spectral-band number K was fixed at 24 during both simulated and real-data training procedures.
  • Data and Training: The Washington DC Mall image was divided into 200 × 200 × 191 testing data and 1080 × 303 × 191 training data before 20 × 20 patch cropping.
  • Test Data Sets: The experiments used Washington DC Mall, Indian Pines, and University of Pavia datasets after removing specified water-absorption or atmospheric-disturbed bands.

A. Simulated-Data Experiments

Table II presents the simulated-experiment results.

  • Table II reports results for the simulated experiments.
  • The table is part of the simulated-data experiments section.
  • The supplied table passage identifies the results table but provides no individual values.

QUANTITATIVE EVALUATION OF THE DENOISING RESULTS OF THE SIMULATED EXPERIMENTS

The simulated experiments evaluate denoising under multiple noise settings using quantitative, visual, spectral, and spectral-difference analyses. HSID-CNN achieves the strongest reported overall performance and preserves spectral features more reliably than the compared methods.

  • Quantitative evaluation: HSID-CNN achieves the highest MPSNR and MSSIM and the lowest MSA across all simulated noise levels.The comparison covers the noise conditions summarized in Table II.
  • Visual comparison: BM4D suppresses noise under uniform and non-uniform intensities but can over-smooth small texture features.LRTA can generate artifacts, while HSSNR leaves residual noise under strong noise.
  • Spectral analysis: For pixel (83, 175), the proposed method’s spectral curve is closest to the ground truth among the compared methods.The spectral plots use DN values across band numbers.
  • Spectral analysis: The proposed method produces smoother spectral-difference curves with residual values closer to zero for roof, grass, and road classes.The curves compare restoration results with the noise-free HSI across spectral bands.

B. Real-Data Experiments

Real-world HSI experiments compare denoising methods through supervised classification using a common SVM evaluation setup. Overall accuracy and the kappa coefficient are the reported evaluation indexes.

  • Real-data evaluation: Two real-world HSI data sets are used to evaluate the denoising methods.
  • Classification evaluation: SVM classification is conducted under the same environment for all restoration results.
  • Classification evaluation: Overall accuracy and the kappa coefficient serve as the classification evaluation indexes.

1) AVIRIS Indian Pines Data Set:

On the AVIRIS Indian Pines data set, HSID-CNN removes noise and stripes while retaining local detail and structure. It also achieves the best reported SVM classification performance after denoising.

  • Denoising results: HSID-CNN effectively removes noise and stripes while preserving local details and structural information.The comparison includes HSSNR, LRTA, BM4D, and LRMR restoration results.
  • Classification results: The Indian Pines experiment evaluates 16 ground-truth classes using supervised SVM classification.Training uses 10% of the test samples randomly generated from each class.
  • Classification results: 85.65% OA and 0.8338 kappa are the highest reported classification values for HSID-CNN.Before denoising, OA and kappa are 75.96% and 0.7220, respectively.

2) ROSIS University of Pavia Data Set:

On Pavia University data, HSID-CNN produced the strongest reported denoising and classification results among the compared methods, while reducing fragmentary effects.

  • HSSNR and LRMR reduced some noise, but non-uniform noise remained in the restored Pavia University results.
  • HSID-CNN achieved the highest Pavia University classification accuracy, with OA of 86.99% and kappa coefficient of 0.8319.
  • The classification experiment used the first 20 spectral bands because Pavia University noise was mainly concentrated in some first bands.
  • The proposed method reduced the fragmentary effect better than HSSNR, LRTA, BM4D, MH, and LRMR.

C. Further Discussion

Further experiments examine spectral-band selection, feature modules, deep-learning baselines, and runtime. They identify K = 24 as optimal in the reported experiment and describe HSID-CNN as computationally efficient in GPU mode.

  • Adjacent spectral band number K: The spatial-spectral input uses the current band together with K adjacent spectral bands to exploit redundant correlated information.
  • Adjacent spectral band number K: HSID-CNN performance first increased with adjacent-band number K and reached its highest MPSNR at K = 24 before declining.
  • Multi-scale feature extraction: Multi-scale convolution extracts features from contextual information at different scales for HSI recovery.
  • Multi-level feature representation: Multi-level feature representation merges feature information from different-depth layers for final restoration.
  • Comparisons with DL-based Denoising Methods: The study compares HSID-CNN with DnCNN and nine-layer 3D-DnCNN using quantitative indexes and detailed denoising results.
  • Runtime Comparisons: In GPU mode, HSID-CNN had the lowest runtime complexity among the compared HSI denoising algorithms.

V. CONCLUSION

The paper concludes that HSID-CNN combines spatial-spectral deep learning with multi-scale feature extraction for HSI denoising. Simulated and real-data experiments reported advantages across evaluation indexes, visual quality, and classification accuracy, while future work targets mixed noise and spectral distortion.

  • HSID-CNN learns a non-linear end-to-end mapping from noisy to clean HSIs with a combined spatial-spectral convolutional network.
  • The network jointly uses spatial information and adjacent correlated bands, with multi-scale extraction for spatial and spectral features.
  • Experiments on simulated and real data reported better evaluation indexes, visual effects, and classification accuracy than many mainstream methods.
  • Future work will investigate more efficient structures for mixed noise, including stripe noise, impulse noise, and dead lines.
  • Future work will also consider priori constraints or structures to reduce spectral distortion and improve texture details.
Loading 1806.00183v3…