Improved Classification Method Based on the Diverse Density and Sparse Representation Model for a Hyperspectral Image

Na Li, Ruihao Wang, Huijun Zhao, Mingcong Wang, Kewang Deng, Wei WeiSensors 2019 · DOI

TL;DR

This paper introduces a novel hyperspectral image classification method (NCM_DDSR) that effectively tackles the "small sample size" (SSS) problem. By combining a "diverse density" (DD) approach for robust dictionary learning with an improved "sparse representation" (SR) model, the method achieves high classification accuracy even with limited training data.

What Problem Does This Paper Address?

Hyperspectral imaging offers rich spectral information, enabling detailed land cover classification. However, a significant challenge arises from the "small sample size" (SSS) problem. This occurs when the number of available training samples is insufficient relative to the high dimensionality of hyperspectral data (many spectral bands). This imbalance can lead to poor classification accuracy, a phenomenon known as the Hughes phenomenon. Traditional solutions like dimensionality reduction can lose valuable spectral information, while methods like Support Vector Machines (SVMs) often require extensive parameter tuning based on user experience. The paper aims to provide a more robust and less parameter-dependent solution for hyperspectral classification under SSS conditions.

Methods

The proposed NCM_DDSR method is built upon two core components:

  • Diverse Density (DD) Algorithm for Dictionary Learning: This algorithm is used to learn a dictionary of spectral features. Instead of relying on noisy or mixed spectral signatures, DD identifies "pure" spectral atoms that best represent a ground cover class. It achieves this by finding feature vectors that are highly similar to positive training samples (representing the target class) and dissimilar to negative training samples (representing other classes). This process helps mitigate noise and spectral mixing issues.
  • Improved Sparse Representation (SR) Model: The paper utilizes an SR framework where each pixel's spectral signature is represented as a linear combination of a few dictionary atoms. An "improved matching pursuit algorithm" is employed to compute sparse coefficients. This algorithm is enhanced to ensure non-negative coefficients that sum to one, and it incorporates a relaxation constraint to handle mixed pixels more effectively. The sparse coefficients are then used for classification.

Datasets

The proposed method was evaluated on two publicly available hyperspectral datasets:

  • Push-broom Hyperspectral Imager (PHI) data: Collected over the Fanglu Tea plantation area in Jiangsu province, China. This dataset has 65 spectral channels and seven ground cover classes.
  • Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data: Specifically, the Salinas_A dataset from California, USA. This dataset has 166 spectral channels and six ground cover classes.

The experiments were conducted under small sample conditions, specifically using training sample sets that were 0.5, 1, and 1.5 times the number of spectral bands.

Key Insights and Ideas

  • DD for Robust Dictionary Atoms: The DD algorithm's ability to select dictionary atoms that are representative of pure spectral signatures is crucial for overcoming noise and spectral mixing, which are common issues in hyperspectral data.
  • SR for Dimensionality Reduction and Feature Extraction: Sparse representation naturally handles the high dimensionality of hyperspectral data by seeking compact representations, effectively reducing the reliance on dimensionality reduction techniques that might discard useful information.
  • Improved SR for Mixed Pixels: The relaxation constraint in the SR model is a key innovation for handling mixed pixels, where a single pixel contains signatures from multiple ground covers. This allows for a more nuanced representation.
  • Synergy of DD and SR: The combination of DD for dictionary learning and SR for representation and classification creates a powerful framework that is particularly effective under SSS conditions.

Key Results

The NCM_DDSR method demonstrated superior performance compared to traditional PCA-based minimum distance (PCA_MinD) and Support Vector Machine (SVM) classifiers, especially under small sample size conditions.

  • PHI Data: Achieved an overall accuracy of 91.59% and a Kappa coefficient of 0.897 when training samples were 0.5 times the number of bands.
  • AVIRIS Data: Achieved an overall accuracy of 92.83% and a Kappa coefficient of 0.91 when training samples were 0.5 times the number of bands.
  • The proposed method showed significantly higher and more stable classification accuracy as the number of training samples decreased, outperforming PCA_MinD and SVM, which experienced substantial drops in performance.

Limitations and Future Work

  • Parameter Sensitivity: While the paper claims reduced reliance on user experience, the DD algorithm itself might have parameters that require careful selection, although this is not explicitly detailed as a limitation.
  • Computational Complexity: Sparse representation algorithms, especially with improved matching pursuit, can be computationally intensive. The paper does not extensively discuss the computational cost.
  • Future Work: The authors suggest further exploration of the method's performance with different datasets and under varying sample sizes. Investigating more advanced dictionary learning techniques and sparse representation optimization strategies could also be avenues for future research.

Summary updated

Abstract

To solve the small sample size (SSS) problem in the classification of hyperspectral image, a novel classification method based on diverse density and sparse representation (NCM_DDSR) is proposed. In the proposed method, the dictionary atoms, which learned from the diverse density model, are used to solve the noise interference problems of spectral features, and an improved matching pursuit model is presented to obtain the sparse coefficients. Airborne hyperspectral data collected by the push-broom hyperspectral imager (PHI) and the airborne visible/infrared imaging spectrometer (AVIRIS) are applied to evaluate the performance of the proposed classification method. Results illuminate that the overall accuracies of the proposed model for classification of PHI and AVIRIS images are up to 91.59% and 92.83% respectively. In addition, the kappa coefficients are up to 0.897 and 0.91.
BibTeX
@inproceedings{li2019improved,
  author = {Na Li and Ruihao Wang and Huijun Zhao and Mingcong Wang and Kewang Deng and Wei Wei},
  title = {{Improved Classification Method Based on the Diverse Density and Sparse Representation Model for a Hyperspectral Image}},
  booktitle = {Sensors},
  year = {2019},
  doi = {10.3390/s19245559},
}
Figure