基于自监督学习的高光谱图像分类方法  

A Hyperspectral Image Classification Method Based on Self-Supervised Learning

作　　者：郑思睿 王俊伟 肖宏[3] 董怡卓 丁晨 张磊[4] ZHENG Sirui;WANG Junwei;XIAO Hong;DONG Yizhuo;DING Chen;ZHANG Lei(School of Computer Science,Xi'an University of Posts and Telecommunications,Xi'an,Shaanxi 710121;Beijing Special Engineering Design Institute,Beijing 100028;The 27th Research Institute of China Electronics Technology Group Corporation,Zhengzhou,Henan 450047;School of Computer Science,Northwestern Polytechnical University,Xi'an,Shaanxi 710072)

机构地区：[1]西安邮电大学计算机学院,陕西西安710121 [2]北京特种工程设计研究院,北京100028 [3]中国电子科技集团公司第二十七研究所,河南郑州450047 [4]西北工业大学计算机学院,陕西西安710072

出　　处：《航天技术与工程学报》2024年第3期13-25,共13页

基　　金：国家自然科学基金(62472350,62372379)。

摘　　要：高光谱图像是一种通过高光谱成像传感器在多个光谱波段捕获地物信息的图像,高光谱图像分类任务是为图像中的每个像素点确定所属类别。然而,现有高光谱图像分类任务中存在标记样本数量有限,大量未标记样本未被合理利用于训练网络模型与增强模型分类效果的问题。针对上述问题,本文提出了一种基于自监督学习网络的小样本高光谱图像分类方法。首先,通过自监督学习网络对大量未标记样本进行预训练,使得模型能够在无需标签的情况下对样本本身的内在结构或属性进行学习;其次,结合主动学习采样策略,对未标记样本进行采样及人工标注,扩充有限的数据集;最后,对分类模型进行微调,使模型能够根据已有知识对训练集中的新样本进行更准确、更稳定的分类判断,进一步提升其泛化能力。实验结果证明,该方法在增强模型泛化能力和提高分类准确性方面都表现出显著优势,在两个公开的高光谱数据集上都取得了良好结果。Hyperspectral images capture terrestrial information across multiple spectral bands using hyperspectral imaging sensors.The task of hyperspectral image classification is to determine the category of each pixel in the image.However,current hyperspectral image classification methods face challenges such as a limited number of labeled samples and the underutilization of a large number of unlabeled samples in training network models to enhance classification performance.To address these issues,we propose a few-shot hyperspectral image classification method based on a self-supervised learning network.Firstly,our approach pretrains the self-supervised learning network on a large number of unlabeled samples,enabling the model to learn the intrinsic structure or attributes of the samples without needing labels.Additionally,we incorporate an active learning sampling strategy to sample and manually annotate the unlabeled samples,thereby expanding the limited dataset.Finally,we fine-tune the classification model,allowing it to make more accurate and stable classification judgments on new samples in the training set based on existed knowledge,further improving its generalization ability.Experimental results demonstrate that our method shows significant advantages on model generalization and classification accuracy improvement,achieving excellent results on two public hyperspectral datasets.

关 键 词：高光谱图像分类 小样本学习 自监督学习 主动学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]