结合多尺度空间滤波和层级网络的基于结构保持的高光谱特征选择

发布时间：2018-01-16 11:20

本文关键词：结合多尺度空间滤波和层级网络的基于结构保持的高光谱特征选择　出处：《光子学报》2017年05期 　论文类型：期刊论文

【摘要】：为了充分利用高光谱图像蕴含的丰富的光谱信息和空间信息,提出了结合多尺度空间滤波和层级网络的基于结构保持的高光谱特征选择算法.算法利用基于l2,1范数的数学模型,选出同时保存全局相似性结构和局部流形结构的特征子集;在多个尺度的窗口中使用双边滤波,自适应计算滤波核,自动在光谱数据中融入空间信息,增强了类内相似性和类间相异性,避免了参量选择;引入层级结构实现空间信息和光谱信息的深入融合,提高了分类准确度;讨论了层级数目和窗口尺度个数对分类准确度的影响.在Indian Pines和PaviaU两个数据集的实验表明,该算法在大部分地物种类上的分类准确度都有较大幅度的提升,总体分类准确度分别达到90.98%和94.20%,相比其他方法明显提高了地物分类准确度.
[Abstract]:In order to make full use of the rich spectral information and spatial information contained in hyperspectral images. A hyperspectral feature selection algorithm combining multi-scale spatial filtering and hierarchical network is proposed. The algorithm uses a mathematical model based on l2m-1 norm. The feature subsets of both global similarity structure and local manifold structure are selected. Using bilateral filtering in multi-scale windows, adaptive computing filter kernel, automatically incorporating spatial information into spectral data, enhancing intra-class similarity and inter-class differences, avoiding parameter selection; The hierarchical structure is introduced to realize the deep fusion of spatial information and spectral information, and the classification accuracy is improved. The effects of the number of levels and the number of window scales on the classification accuracy are discussed, and the experimental results of Indian Pines and PaviaU are given. The classification accuracy of the algorithm has been greatly improved in most kinds of ground objects, and the overall classification accuracy has reached 90.98% and 94.20%, respectively. Compared with other methods, the classification accuracy of ground objects is improved obviously.
【作者单位】：火箭军工程大学信息工程系;中国科学院西安光学精密机械研究所;西安交通大学电信学院;
【基金】：国家自然科学基金(No.61401471) 中国博士后基金(No.2014M562636)资助~~
【分类号】：TP751
【正文快照】： 1-3000150conventional methods.0引言近年来高光谱遥感技术发展迅速,更高的光谱和空间分辨率能捕获地物精细的光谱响应及空间细节特征,带来了更精细的光谱波段和更丰富的地物信息,对分类任务提出了更高要求.高光谱图像数据集(图像立方体)具有图谱合一的特点[1-5].虽然丰富的

【相似文献】