摘要
针对目前高光谱图像分类数据冗余度高,计算效率低下,且易丢失光谱信息等问题,该文提出一种可以有效地利用光谱信息通过多尺度样本熵提取图像特征的方法。先描述多尺度样本熵计算过程,并对参数进行分析,选取最优参数。在此基础上,分析多尺度样本熵曲线变化规律,设计最优多尺度样本熵特征选择方法。将选取的最优多尺度样本熵特征矢量代入支持向量机分类器(SVM),实现高光谱图像分类。将该文算法与深度特征融合网络(DFFN)算法和基于自适应波段选择(ABS)算法在PaviaU图像和Indian Pines图像上进行对比实验,并对其结果进行定量精度评价。实验结果表明,对于两组高光谱图像,该文算法在总体分类精度上分别达到了98.64%和96.49%,明显高于两种对比算法,同时在时间效率上也有了显著提升。
In order to address the current problems of high redundancy of hyperspectral image classification data,low computational efficiency,and easy loss of spectral information,this paper proposes a method that can effectively use spectral information to extract image features by multi-scale sample entropy.Firstly,the calculation process of the multi-scale sample entropy is described,and the parameters are analyzed to select the optimal parameters.On this basis,the change rule of multi-scale sample entropy curve is analyzed,and the optimal multi-scale sample entropy feature selection method is designed.Finally,the selected optimal multi-scale sample entropy feature vector is substituted into the Support Vector Machine(SVM)classifier to achieve hyperspectral image classification.The algorithm in this paper is compared with Deep Feature Fusion Network(DFFN)algorithm and Adaptive Band Selection(ABS)based algorithm on PaviaU images and Indian Pines images for experiments,and the results are evaluated for quantitative accuracy.The experimental results show that for the two sets of hyperspectral images,the algorithms in this paper achieve 98.64%and 96.49%in overall classification accuracy,which is significantly higher than the two comparison algorithms,and also have a significant improvement in time efficiency.
作者
赵泉华
张杰
李玉
ZHAO Quanhua;ZHANG Jie;LI Yu(Institute of Remote Sensing Science and Application,School of Surveying,Mapping and Geographical Sciences,Liaoning Technical University,Fuxin,Liaoning 123000,China)
出处
《测绘科学》
CSCD
北大核心
2023年第1期117-126,共10页
Science of Surveying and Mapping
基金
国家自然科学基金青年基金项目(42001286)
辽宁省教育厅项目重点攻关项目(LJ2020ZD003)
关键词
高光谱图像分类
多尺度样本熵
支持向量机
深入特征融合网络
自适应波段选择
hyperspectral image classification
multiscale entropy
support vector machine
deep feature fusion network
adaptive band selection
