结合波段选择和保边去噪滤波的高光谱遥感图像分类  

Hyperspectral Remote Sensing Imagery Classification Combining Band Selection and Edgepreserving Filtering

作　　者：王巧玉[1] 陈锻生[1] 

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021

出　　处：《小型微型计算机系统》2017年第5期1098-1102,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(61370006)资助;福建省科技计划项目(工业引导性)重点项目(2015H0025)资助

摘　　要：高光谱遥感图像分类是遥感图像在众多领域得以有效应用的基础.为了提高其分类精度,提出一种结合波段选择和保边去噪滤波的分类方法.通过波段选择对高光谱数据进行降维同时保留了光谱的物理信息;然后通过对各波段图像进行保边去噪滤波将空间信息和光谱信息结合起来;最后选取结合了空间信息的光谱曲线作为由堆栈降噪自编码器加Softmax分类器构成了深度学习网络的输入,对其进行特征提取及分类.波段选择去除了一些对分类不利的波段,保边去噪滤波将空间信息与光谱信息结合了起来,深度学习提取特征避免了手动选择特征的主观臆断,通过在Indian Pines和Pavia University数据集上的实验,结果表明本文算法分类精度高、稳定性好.Hyperspectral remote sensing image classification is the foundation of its use in many fields. In order to improve the classification accuracy, a method based on band selection and edge-preserving filtering is proposed. The dimension of the hyperspectral data was reduced by the band selection while the physical information of the spectrum was maintained. Then the edge-preserving filtering was performed on the selected bands, through which could combine the spatial and spectral information together. Finally, a network with stacked denoising autoencoder and Softmax classifer is used to obtain the useful high-level features and do the classification;the bands which are not useful for the classification were removed by the bands selection, and the spatial information was combined with the spectral information through the edge-preserving filter; and using deep learning to extract features avoids manually choose subjective. Experiment results on Indian Pines and Pavia University dataset demonstrated the proposed algorithm not only has high accuracy but also has good stability.

关 键 词：波段选择 保边去噪滤波 堆栈降噪自编码器 空间-光谱结合 高光谱图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]