潜在特征融合和最优聚类的高光谱图像降维  

Hyperspectral image dimensionality reduction via latent features fusion and optimal clustering

作　　者：聂萍 汪国强[1] NIE Ping;WANG Guo-Qiang(School of Electronic Engineering,Heilongjiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080

出　　处：《黑龙江大学工程学报》2023年第1期51-59,共9页Journal of Engineering of Heilongjiang University

基　　金：国家自然科学基金项目(51607059);黑龙江省自然科学基金项目(QC2017059)。

摘　　要：高光谱图像(Hyperspectral Images, HSI)可提供几十到数百个连续的光谱波段,但这些波段导致数据处理的复杂性增加,并且相邻波段的冗余度较大。为了解决这些问题,提出了一种潜在特征融合和最优聚类的高光谱图像降维方法(Latent Features Fusion and Optimal Clustering Framework, LFFOCF)。该方法使用超像素分割将HSI分割为多个区域,以便充分保留HSI的空间信息。通过构造相应的拉普拉斯矩阵获取先验信息,生成一组低维潜在特征,进一步增强不同波段之间的可分性;通过融合区域感知的潜在特征,获得HSI的共享潜在特征表示,以有效捕获HSI的频带冗余;通过最优聚类框架搜索HSI中的最优聚类结构,在一种排序策略的基础上获得最优聚类结果,生成相关性较低且具有更多鉴别信息的波段子集。该方法充分利用了光谱和空间特性,在两个公共数据集上的大量实验表明,与Optimal Neighborhood Reconstruction(ONR)、Optimal Clustering Framework(OCF)和Region-aware Latent Features Fusion based Clustering(RLFFC)方法相比,所提出的方法在OA、MA和Kappa系数3个指标上都优于其他算法。Hyperspectral images(HSI)can provide dozens to hundreds of continuous spectral bands.However,these bands lead to increased complexity in data processing,and the redundancy of adjacent bands is large.To tackle these issues,a hyperspectral image dimensionality reduction method via latent features fusion and optimal clustering framework(LFFOCF)is proposed.The method employs superpixel segmentation to segment HSIs into multiple regions so that the spatial information of HSIs can be fully preserved.The prior information is obtained by constructing the corresponding Laplacian matrix,from which a set of low-dimensional latent features are generated to further enhance the separability between different bands.Then,a shared latent feature representation of HSIs is obtained by fusing region-aware latent features to effectively capture the band redundancy of HSIs.Finally,the optimal clustering structure in HSI is searched by the optimal clustering framework.Based on a ranking strategy,the optimal clustering result is obtained and generates a band subset with lower correlation and more discriminative information.The method makes full use of spectral and spatial properties,extensive experiments on two public datasets show that the proposed method is superior to other algorithms in OA,MA,and Kappa coefficients when compared with other state-of-the-art methods such as Optimal Neighborhood Reconstruction(ONR)、Optimal Clustering Framework(OCF)and Region-aware Latent Features Fusion based Clustering(RLFFC).

关 键 词：潜在特征学习 降维 最优聚类 高光谱图像 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]