一种波段聚类和多尺度结构特征融合的高光谱图像分类模型  

A Hyperspectral Image Classification Model Based on Band Clustering and Multi-Scale Structure Feature Fusion

作　　者：王彩玲[1] 张静 王洪伟 宋晓楠 纪童 WANG Cai-ling;ZHANG Jing;WANG Hong-wei;SONG Xiao-nan;JI Tong(College of Computer Science,Xi'an Shiyou University,Xi'an 710065,China;School of Artificial Intelligence,Optics and Electronics,Northwestern Polytechnical University,Xi'an 710072,China;College of Grass Industry,Gansu Agricultural University,Lanzhou 730070,China)

机构地区：[1]西安石油大学计算机学院,陕西西安710065 [2]西北工业大学光电与智能研究院,陕西西安710072 [3]甘肃农业大学草业学院,甘肃兰州730070

出　　处：《光谱学与光谱分析》2024年第1期258-265,共8页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61401439,41301480,41301382);陕西省重点研发计划项目(2019GY-112);西安航空学院高教研究项目(2018GJ1005);陕西西安教育科学“十三五”规划2018年度课题(SGH18H435);西安石油大学研究生创新与实践能力培养计划项目(YCS21213257);陕西省自然科学基金面上项目(2022JM-396)资助。

摘　　要：高光谱图像包含丰富的地物信息,在农业、工业和军事等领域应用广泛。因此,高光谱图像的识别与分类是一项重要的研究课题。然而,高光谱图像存在光谱维度高、噪声大、标记样本有限等问题,并未取得很好的分类效果。针对以上问题,提出一种波段聚类和多尺度结构特征融合的高光谱图像分类模型(ASPS-MRTV)。该方法主要包括以下几个步骤,首先,对高光谱数据进行归一化处理,将归一化后的三维图像按光谱维等分为n个子空间;其次,采用粗细划分策略构造自适应子空间光谱特征提取框架,将每个空间波段拉伸为一维向量后用信息散度构造波段的相似性矩阵,按照聚类的思想对n个子空间进行自适应;然后,将每个自适应子空间的光谱波段平均值进行叠加,形成光谱特征;最后,对所得到的光谱特征数据利用多尺度相对全变分技术提取结构特征。为了增强样本的线性可分性,在数据堆叠之后进行核主成分分析,最终形成空谱特征。对比实验中统一使用惩罚参数C和核参数σ都为24.5的SVM进行分类。经测试,ASPS-MRTV网络模型在Indian Pines、 University of Pavia两个数据集上分别以5%, 1%的训练样本达到了97.06%、 98.98%的总体分类精度。实验结果表明,该模型与SVM、 ASPS(ED)、 ASPS(ID)、 ASPS-LBP、 ASPS-GlCM、 ASPS-BF模型相比,在分类性能和计算效率方面都取得了更优的效果,有效提高小样本下高光谱图像的分类精度。Hyperspectral images contain abundant ground object information and are widely used in agriculture,industry,military and other fields.Therefore,its identification and classification is an important research topic.However,hyperspectral images have problems,such as high spectral dimension,large noise and limited labeled samples,so they have not achieved good classification results.This paper proposes a hyperspectral image classification model based on band clustering and multi-scale structural feature fusion(ASPS-MRTV).The method mainly includes the following steps.First,hyperspectral data is normalized,and the normalized 3D image is divided into n subspaces according to spectral dimension.Secondly,an adaptive subspace spectral feature extraction framework is constructed using the coarse and thin division strategy.Each spatial band is stretched into a one-dimensional vector,and then the similarity matrix of the band is constructed by using the information divergence.Then,the spectral band averages of each adaptive subspace were superimposed to form spectral features.Finally,the multi-scale relative total variation technique extracts structural features from the obtained spectral feature data.In order to enhance the linear separability of the samples,kernel principal component analysis was performed after data stacking to form the null spectral features.In the comparison experiment,SVM with penalty parameter C and kernel parameter of 24.5 were uniformly used for classification.After testing,the ASPS-MRTV network model achieves the overall classification accuracy of 97.06%and 98.98%on Indian Pines and University of Pavia datasets with 5%and 1%training samples,respectively.Experimental results show that compared with SVM,ASPS(ED),ASPS(ID),ASPS-LBP,ASPS-GLCM and ASPS-BF models,the proposed model achieves better classification performance and computational efficiency and effectively improves the classification accuracy of hyperspectral images under small samples.

关 键 词：高光谱图像 多尺度结构特征 信息散度 核主成分分析 空谱特征 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]