基于多特征随机最近正则化子空间的高光谱影像分类算法  

Hyperspectral Image Classification Algorithm Based on Multi-feature and Random Subspace Nearest Regularized Subspace

作　　者：李倩楠 彭菲 虞瑶 LI Qiannan;PENG Fei;YU Yao(Basic Geographic Information Center of Jiangsu Province,Nanjing 210013,China)

机构地区：[1]江苏省基础地理信息中心,江苏南京210013

出　　处：《地理空间信息》2025年第4期25-29,共5页Geospatial Information

基　　金：江苏自然资源智库2023年度开放合作项目(ZK23002)。

摘　　要：基于单一特征的随机子空间的最近正则化子空间集成算法(random eubspace based nearest regularized subspace,RSNRS)精度受限,因此设计一种基于多特征的随机最近正则化子空间(multi-meature random subspace-based nearest regularized Subspace,MFRSNRS)的高光谱遥感算法。提取高光谱遥感影像的Gabor特征、局部二值模式特征(local binary pattern,LBP)特征、拓展形态学多属性剖面(extended multi-attribute profiles,EMAPs)特征和光谱特征,利用随机子空间的集成方式构建特征子空间,基于特征子空间训练基分类器最近正则化子空间(nearest regularized subspace,NRS),并将基于分类器的输出结果进行集成得到分类结果。MFRSNRS算法在数据集Indian Pines总体精度达到92.88%,与RSNRS算法相比,总体精度提高了17.05%。在数据集Pavia University上总体精度达到78.88%,与RSNRS算法相比,总体精度提高7.73%。实验结果表明,多特征的使用具有一定的提升效果。The random subspaces nearest regularized subspace(RSNRS) based on a single feature has a limited accuracy,we proposed a new hyperspectral remote sensing algorithm named multi-feature random subspace nearest regularized subspace(MFRSNRS).Firstly,we extracted Gabor features,LBP features,EMAPs features and spectral features by hyperspectral remote sensing images.Then,we selected the nearest regularized subspace(NRS) as the base classifier,and used the random subspace method to construct the feature subspace.Finally,we trained the base classifiers based on the feature subspace,and integrated the results of base classifiers to obtain the final classification results.The OA of MFRSNRS algorithm reached 92.88% in the Indian Pines data set,which is 17.05% higher than RSNRS algorithm.The OA of MFRSNRS algorithm reached 78.88% in the Pavia University data set,which is 7.73% higher than RSNRS algorithm.The experimental result indicates that the use of multiple features has shown a certain improvement in performance.

关 键 词：多特征 最近规则子空间 高光谱影像分类 随机子空间 

分 类 号：P237[天文地球—摄影测量与遥感]