›› 2013, Vol. 42 ›› Issue (3): 0-0.

• 学术论文 •    下一篇

应用稀疏非负矩阵分解聚类实现高光谱影像波段的优化选择

施蓓琦1,刘春2,孙伟伟2,陈能3   

  1. 1. 同济大学测量与国土信息工程系;上海师范大学城市信息研究中心
    2. 同济大学
    3. 上海师范大学
  • 收稿日期:2012-04-05 修回日期:2012-09-05 出版日期:2013-06-20 发布日期:2014-01-23
  • 通讯作者: 刘春 E-mail:liuchun@tongji.edu.cn
  • 基金资助:
    国家重点基础研究发展计划(973计划);上海市教育委员会科研创新项目

Sparse Nonnegative Matrix Factorization for Hyperspectral Band Selection

  • Received:2012-04-05 Revised:2012-09-05 Online:2013-06-20 Published:2014-01-23
  • Contact: Chun Liu E-mail:liuchun@tongji.edu.cn

摘要: 针对高光谱影像数据高维性、高度相关性和冗余性等特点,本文提出应用稀疏非负矩阵分解聚类实现高光谱影像波段的优化选择。论文通过稀疏非负矩阵分解方法对高光谱影像进行稀疏化表示,同时顾及其可聚类的特性,在保留所选波段物理意义的基础上,得到波段选择后的高光谱影像降维数据。此外,应用稀疏非负矩阵分解聚类的方法对PHI-3高光谱影像进行波段选择的实验分析,应用聚类特征有效性分析波段聚类结果,并采用波段子集的信息量、相关性和可分性三类评价指标来验证方法的有效性。最终,算法实用性评价从运行效率和分类精度两方面证明了基于无监督聚类的稀疏非负矩阵分解可有效地进行高光谱影像的波段选择。

关键词: 高光谱影像,波段选择,稀疏表示,非负矩阵分解,概率潜语义分析聚类

Abstract: Hyperspectral data provides imagery with hundreds of spectral bands, but many of them contain redundant information. Band selection is often applied to reduce the dimensionality of the data. This paper proposes a new technique for band selection through sparse nonnegative matrix factorization (SNMF-BS) which decomposes the image data into the multiplication of basic matrix and the coefficient matrix. Using sparse NMF, Hyperspectral images can be described as sparse representation. Since sparse NMF acts as co-clustering, it can be used for band clustering without considering the distance metric between different spectral bands. By imposing the sparsity constraint on the coefficient matrix, clustering assignments of bands can be easily indicated through the largest entry in each column of the matrix. Through band clustering, sub-bands are selected to serve the need of dimensionality reduction, while preserving the physical meanings of the selected bands. The PHI-3 real hyperspectral dataset is experimented for band selection. Clustering validity indexes, KLC, band correlation and separability are used to do the evaluation. The experimental results show that sparse NMF provides considerable insight into the unsupervised clustering-based band selection problem and also offer the proper band combinations with spectral meanings. And the results also demonstrate that the proposed method outperforms other popular methods.

Key words: Hyperspectral Imagery, Band Selection, Sparse Representation, Nonnegative Matrix Factorization, Probabilistic Latent Semantic Analysis

中图分类号: