基于改进超像素分割算法的高光谱图像分类方法  

A hyperspectral image classification method based on improved superpixel segmentation algorithm

作　　者：孙中皋[1] 艾香辰 SUN Zhonggao;AI Xiangchen(School of Physics and Electronic Technology,Liaoning Normal University,Dalian 116029,China)

机构地区：[1]辽宁师范大学物理与电子技术学院,辽宁大连116029

出　　处：《辽宁师范大学学报(自然科学版)》2025年第1期95-105,共11页Journal of Liaoning Normal University:Natural Science Edition

基　　金：辽宁省教育厅科学研究服务地方项目(LF2020003)。

摘　　要：基于超像素分割的高光谱图像分类方法在显著降低数据复杂度的同时可以获得较高的分类精度.现有高光谱图像超像素分割算法未充分利用高维度纹理信息,为此,提出一种改进的流形-简单线性迭代聚类分割算法.改进算法在迭代聚类时采用组合值度量像素间距,组合值由高光谱图像全光谱维度表征的颜色和空间距离以及应用多主成分灰度共生矩阵的特征量表征的纹理距离构成,该方法充分利用了高光谱图像的高维度信息,改善了超像素分割效果.提取分割后超像素的光谱均值和加权光谱均值特征,采用图分类器对高光谱图像分类,在公开的高光谱数据集上进行实验验证,均取得了较高的分类精度,表明了改进分割算法的有效性.The hyperspectral image classification method based on superpixel segmentation can significantly reduce the data complexity and obtain high classification accuracy. The existing hyperspectral image superpixel segmentation algorithms do not make full use of high-dimensional texture information, so an improved manifold-simple linear iterative clustering segmentation algorithm is proposed. The improved algorithm uses the combined pixel distance which is composed of the color and spatial distance of the full spectral dimension and the texture distance represented by the feature value of the gray level co-occurrence matrix with multiple principal components, which makes full use of the high-dimensional information of the hyperspectral image and improves the superpixel segmentation effect. The average spectral value and weighted average spectral value of the segmented superpixel are extracted, and the graph classifier is used to classify the hyperspectral images. The experiment is verified on the public hyperspectral dataset, and the high classification accuracy is obtained, which shows the effectiveness of the method.

关 键 词：高光谱图像 超像素分割 流形-简单线性迭代聚类 图分类器 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]