高光谱图像类别独立的域适应分类  

Class-independent domain adaptation for hyperspectral image classification

作　　者：余龙 李军 贺霖[3] 李云飞 YU Long;LI Jun;HE Lin;LI Yunfei(School of Geography and Planning,Sun Yat-sen University,Guangzhou 510006,China;Hubei Key Laboratory of Intelligent Geo-Information Processing,School of Computer Science,China University of Geosciences,Wuhan 430078,China;School of Automation Science and Engineering,South China University of Technology,Guangzhou 510640,China)

机构地区：[1]中山大学地理科学与规划学院,广州510006 [2]中国地质大学(武汉)计算机学院,智能地质信息处理湖北省重点实验室,武汉430078 [3]华南理工大学自动化科学与工程学院,广州510640

出　　处：《遥感学报》2024年第3期610-623,共14页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家自然科学基金(编号:62071184);国家杰出青年科学基金(编号:T2225019);广东省自然科学基金(编号:2022A1515011615);广州市科技计划项目(编号:202002030395)。

摘　　要：利用已有图像的标记样本对新的高光谱图像分类面临光谱偏移带来的分类性能差的问题。基于特征表示的域适应方法通过学习域不变特征来解决这个问题。然而现有方法难以同时将多个类的样本对齐,在对齐多类样本的同时又忽略了类间混合对可分性造成的影响。本文提出了一种类别独立的域适应分类方法。首先,为每个地物类别构造一个独立的降维子空间,在多个类别独立的子空间中对齐源域和目标域的样本。然后,在每个类别独立子空间中,利用对齐样本学习出目标域样本的后验概率。接着,融合所有类别独立子空间得到的后验概率得到分类标签,目的是增加后验概率的可信度。最后,利用空间先验平滑分类标签后将其作为伪标签用于迭代学习,更新类别独立子空间和目标域的分类结果。另外,本文还设计了代表性样本选择策略,有利于学习出更具共性的特征表达子空间。在两个真实的高光谱数据集上的实验结果表明,本文算法比原始的联合域适应算法的最近邻分类精度分别提升了9.56%和18.45%。Hyperspectral image supervised classification is a crucial and challenging task in remote sensing,as its performance depends heavily on the quantity and quality of labeled samples.However,labeling hyperspectral data is a difficult and time-consuming procedure.This problem results in a limited number of labeled samples in real-world scenarios,rendering the supervised classifiers vulnerable to the issue of overfitting.To address this problem,researchers have sought solutions in the field of unsupervised domain adaptation,utilizing labeled samples from previous images(source domain)to classify new hyperspectral data(target domain).Most existing domain adaptation methods strive to learn domain-invariant features in a new space,but many of them focus on aligning the overall statistics of the two domains without considering the spectral shifts in each class.Other methods attempt to align every class of the source and target domains simultaneously but often overlook the issues of mixture of samples across classes and incorrect sample selection.This may lead to a negative transfer and reduced separability of data.The significant discrepancies across domains will further compound the problem.In this paper,we propose a novel class-independent domain adaptation algorithm that addresses these issues in hyperspectral image classification.Our method first creates an independent subspace for each class and then aligns the corresponding single-class samples of the two domains in those subspaces.The posterior probabilities are learned independently through the aligned samples in each subspace.Then,the posterior probabilities obtained from multiple subspaces are fused to produce the final classification labels,aiming at increasing the confidence of results.Additionally,we use smoothed classification labels as pseudo labels for further iteration and incorporate a strategy for selecting representative samples to enhance subspace performance.Experimental results on two real hyperspectral datasets demonstrate the high classification p

关 键 词：遥感 高光谱图像 域适应 分类 类别独立子空间 资源一号02D 高分五号 

分 类 号：TP701[自动化与计算机技术—检测技术与自动化装置] P2[自动化与计算机技术—控制科学与工程]