Fast Blind Hyperspectral Unmixing Based On Graph Laplacian
Fast Blind Hyperspectral Unmixing Based On Graph Laplacian
复制标题
DOI：
10.1109/whispers.2019.8921375
复制
发表时间：
2019-09
期刊：
影响因子：
--
通讯作者：
Jing Qin;Harlin Lee;Jocelyn T. Chi;Y. Lou;J. Chanussot;A. Bertozzi
中科院分区：
文献类型：
--
作者：
Jing Qin;Harlin Lee;Jocelyn T. Chi;Y. Lou;J. Chanussot;A. Bertozzi
高光谱盲解混是遥感中的一个具有挑战性的问题，其目的是从给定的高光谱数据中推断物质的光谱和丰度。许多传统的方法遭受材料的差的识别和/或昂贵的计算成本，这可以通过用效率来交换准确性而部分地得到缓解。在这项工作中，我们提出了一个快速的基于图的盲解混方法。特别是，我们应用Nyström方法来有效地近似对应于归一化图拉普拉斯算子的矩阵的特征值和特征向量。然后，交替方向乘子法（ADMM）产生一个快速的数值算法。在真实的数据集上的实验表明，该方法在准确性和效率方面具有很大的潜力。
Blind hyperspectral unmixing is a challenging problem in remote sensing, which aims to infer material spectra and abundances from the given hyperspectral data. Many traditional methods suffer from poor identification of materials and/or expensive computational costs, which can be partially eased by trading the accuracy with efficiency. In this work, we propose a fast graph-based blind unmixing approach. In particular, we apply the Nyström method to efficiently approximate eigenvalues and eigenvectors of a matrix corresponding to a normalized graph Laplacian. Then the alternating direction method of multipliers (ADMM) yields a fast numerical algorithm. Experiments on a real dataset illustrate great potential of the proposed method in terms of accuracy and efficiency.