小样本下基于空谱特征增强的高光谱图像分类  

Classification of hyperspectral images with small-sized samples based on spatial-spectral feature enhancement

作　　者：陆瑶 王立国[1,2] 石瑶 LU Yao;WANG Liguo;SHI Yao(College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;College of Information and Communication Engineering, Dalian Minzu University, Dalian 116600, China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]大连民族大学信息与通信工程学院,辽宁大连116600

出　　处：《哈尔滨工程大学学报》2022年第3期436-443,共8页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(62071084).

摘　　要：针对小样本下基于深度学习的高光谱图像分类方法精度不高的问题,本文提出一种基于空谱特征增强残差网络的高光谱图像分类方法。该方法利用三维卷积核同时提取高光谱图像的空间和光谱特征。同时,在普通三维卷积神经网络结构前加入空间和光谱特征增强的残差网络模块,通过恒等映射保留高光谱原始信息的同时用较少的参数增强网络的信息流,并且在构造更深层网络的同时避免了梯度消失的问题。实验表明:本文方法能充分利用高光谱图像的空间和光谱特征,在印第安纳森林数据集、帕维亚大学数据集上分别能够取得99.29%、99.74%的总体分类精确度,提高了小样本下的分类精度。To address the low accuracy of the hyperspectral image classification method based on deep learning with small-sized samples,a hyperspectral image classification method based on a spatial-spectral feature enhancement residual network is proposed in this paper.This method uses a three-dimensional convolution kernel to extract the spatial and spectral features of hyperspectral images.At the same time,a residual network module of spatial and spectral feature enhancement is added in front of the network to retain the original hyperspectral information through identity mapping while enhancing the information flow of the network with fewer parameters.Then,it avoids the gradient disappearance problem while constructing a deep network.Experiments show that the method in this paper can make full use of the spatial-spectral feature of hyperspectral images and can achieve 99.29%and 99.74%overall accuracy on the Indian Pines dataset and University of Pavia dataset,respectively.The proposed method improves the classification accuracy with small-sized samples.

关 键 词：遥感 高光谱 分类 小样本 深度学习 卷积神经网络 残差学习 特征提取 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]