结合随机子空间与SSAE-LR的高光谱图像分类  

Hyperspectral Image Classification Combining Random Subspace and SSAE-LR

作　　者：程圆娥 吕志慧[1] 袁春琦 CHENG Yuan’e;LYU Zhihui;YUAN Chunqi

机构地区：[1]江苏省测绘研究所,江苏南京210013 [2]北方信息控制研究院集团有限公司,江苏南京211153

出　　处：《地理空间信息》2020年第9期96-101,I0007,共7页Geospatial Information

基　　金：江苏省科技厅资助项目(BM2018030)。

摘　　要：针对高光谱数据维度高、样本数有限对分类精度产生局限性的问题,提出了一种结合随机子空间与堆栈式稀疏自编码器深度网络模型(SSAE-LR)的高光谱图像分类新方法。首先基于像素邻域窗口进行像素邻域结构信息的空间特征提取,并将其与光谱信息进行串行拼接;然后利用随机子空间方法在得到的整体特征空间中随机选取若干大小相同的特征子集;最后在这些特征子集上利用SSAE-LR训练获得基分类器,并将所有基分类器的输出进行集成得到分类结果。通过两组高光谱图像数据的实验结果可知,该方法能有效提高传统方法的分类效果。Aiming at the limitation problems of classification accuracy caused by high-dimensional hyperspectral data and the limited sample size,we proposed a new hyperspectral image classification method combining random subspace and SSAE-LR in this paper.Firstly,we carried out the spatial feature extraction of pixel neighborhood structure information based on the pixel neighborhood window,and serial spliced with spectral information.And then,we used the random subspace method to randomly select some feature subsets of the same size from the obtained global feature space.Finally,we used the method of SSAE-LR to train the base classifier on these feature subsets,and integrated the output of all the base classifiers to obtain the classification results.The experimental results of two sets of hyperspectral image data show that this method can effectively improve the classification effect compared with the traditional method.

关 键 词：高光谱图像分类 SSAE-LR 特征提取 随机子空间方法 

分 类 号：P237[天文地球—摄影测量与遥感]