首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 213 全文下载次数: 455
引用本文:

DOI:

10.11834/jrs.20210359

收稿日期:

2020-08-18

修改日期:

2021-04-02

PDF Free   EndNote   BibTeX
一种基于卷积核哈希学习的高光谱图像分类方法
薛朝辉, 张瑜娟
河海大学
摘要:

高光谱遥感可同步获取地表覆盖空间影像和连续且精细的光谱数据,能够实现对地物的精细分类与识别。然而,高光谱图像的高维特性对分类带来巨大挑战。为此,本文探讨了一种基于卷积核哈希学习的高光谱图像分类方法。哈希学习可以将高维信息表达为低维哈希编码,通过计算哈希编码内积并借助最小汉明距离实现分类。为了有效表达非线性数据,又发展了核哈希学习方法。然而,直接应用核哈希学习进行高光谱图像分类存在运行速度慢和未考虑空间邻域信息的不足。为此,本文在核哈希学习中引入径向基函数(Radial basis function,RBF)作为损失函数以提高运行效率;同时,借助四维卷积操作充分表达空间邻域信息,提出了基于卷积核哈希学习的高光谱图像分类方法(Supervised Hashing with RBF Kernel and Convolution,CKSH),同时探讨了该方法在仅利用光谱特征和光谱-空间联合特征上的分类效果。在国际通用测试数据Indian Pines和University of Pavia上进行了实验,结果表明:本文提出的CKSH方法优于传统分类方法(支持向量机、随机子空间)和其它哈希学习方法(如谱哈希、球哈希、监督离散哈希、潜在因子哈希等),同时在不同训练样本数量条件下均取得了较高的分类精度,达到96.12%(Indian Pines,10%)和98.00%(University of Pavia,5%),从而验证了该方法的有效性。

Hyperspectral image classification method based on Supervised Hashing with RBF Kernel and Convolution
Abstract:

Abstract: Objective:Hyperspectral remote sensing can simultaneously acquire spatial imagery of the land cover and continuous spectral data, which can realize the fine classification and identification of ground objects. However, the high-dimensional characteristics of hyperspectral images pose great challenges to classification. For this reason, this paper discusses a hyperspectral image classification method based on hash learning of convolution kernel. Hash learning can express high-dimensional information as low-dimensional hash codes, and realize classification by calculating the inner product of the hash codes and using the minimum Hamming distance. Method:In order to effectively express nonlinear data, the kernel hash learning method has been developed. However, the direct application of kernel hash learning for hyperspectral image classification has the disadvantages of slow running speed and lack of consideration of spatial neighborhood information. For this reason, this paper introduces RBF (Radial basis function) as the loss function in the kernel hash learning to improve the operating efficiency; at the same time, it uses the four-dimensional convolution operation to fully express the spatial neighborhood information, and proposes a high level of Supervised Hashing with RBF Kernel and Convolution (CKSH). The classification effect of this method in using only spectral features and spectral-space joint features is discussed. Results:Experiments were conducted on the international general test data Indian Pines and University of Pavia, and the extended morphological profile and the extended morphological attribute profile were used to extract the combined spectral-spatial features. Experimental results show that the CKSH method proposed in this paper is superior to traditional classification methods (Support Vector Machines, random subspace) and other hash learning methods (such as Spectral Hashing, Spherical Hashing, Supervised Discrete Hashing, Latent Factor Hash, etc.) , Under the condition of different training sample percentages, high classification accuracy was achieved, reaching 96.12% (Indian Pines, 10%) and 98.00% (University of Pavia, 5%), verifying the effectiveness of the method. Conclusions:Aiming at the two problems of kernel hash (KSH) loss function using l2 norm, which results in slow running speed and not considering spatial information, this paper uses four-dimensional convolution to introduce spatial information on the basis of kernel hash, while using RBF core instead of l2 norm. As the loss function, a hyperspectral image classification method based on Convolution Kernel Hash (CKSH) is proposed. Experiments on the general test data set confirmed the superiority of the CKSH method in terms of classification accuracy and running time. On the one hand, because CKSH uses four-dimensional convolution to mine the underlying structure information, the obtained hash code is more discriminative, which is conducive to improving the classification performance of the algorithm; on the other hand, the use of the RBF core as the loss function significantly reduces the running time, which is for subsequent hashing. The study of the learned loss function provides ideas. Keywords: hyperspectral image; hash learning; RBF; four-dimensional convolution; feature extraction; spectral-spatial classification

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮