基于对比自监督的遥感图像场景分类  

Remote sensing image scene classification based on contrastive self-supervised learning

作　　者：吴奕恒 陈知明 戴颖成 欧阳文欣 李建军[1] 刘峰 WU Yiheng;CHEN Zhiming;DAI Yingcheng;OUYANG Wenxin;LI Jianjun;LIU Feng(School of Computer Science and Information Engineering,Central South Univensity of Forestry and Technology,Changsha,Hunan 410000,China)

机构地区：[1]中南林业科技大学计算机与信息工程学院,湖南长沙410000

出　　处：《燕山大学学报》2023年第3期236-245,共10页Journal of Yanshan University

基　　金：国家自然科学基金资助项目(31570627);引进国际先进农业科学技术计划(948计划)资助项目(2015-4-17);湖南省自然科学基金资助项目(2020JJ4938);湖南省高等学校科学研究重点资助项目(20A506);智慧物流技术湖南省重点实验室资助项目(2019TP1015)。

摘　　要：基于视觉Transformer的自监督模型掩码自编码器因其优秀的全局特征捕捉能力,被广泛应用于遥感图像分类领域。但该模型存在图像重建训练时局部上下文语义信息易丢失从而限制其分类精度的进一步提升。针对以上问题,本文提出了融合掩码重建和对比学习的三阶段自监督遥感图像分类新模型——对比掩码自编码器。第一阶段进行掩码重建预训练,以提取遥感图像全局特征;第二阶段则通过对比学习中的正负样本补充第一阶段掩码建模过程中丢失的局部上下文信息;最后通过训练线性分类器完成特征分类。在公开遥感图像数据集AID和NWPU-RESISC45上将本文方法与主流自监督分类方法、监督分类方法进行对比实验。实验结果表明,该模型在两个数据集上分类精度分别达到95.37%和95.14%,性能优于DINO、MoCo、SSGANs等主流自监督模型,接近GLANet、CANet、MG-CAP(Sqrt-E)等主流监督模型,具有良好的应用价值。The elf-supervised model Masked Autoencoders based on vision Transformer is widely used in the field of remote sensing image classification because of its excellent global feature capture ability.However,the Masked Autoencoders has a tendency to lose local contextual semantie information during image reconstruetion training,which restricts its futher improvement of cassifcation accuracy.To solve the above problems,a new two-stage self-supervised remote sensing image cassification model,Masked Autoencoders with Contrast,is proposed,which oombines mask reconstruction and contrast leaning.In the first stage,the model performs mack rconstruction pre-training to extract global features of remote sensing images.In the second stage,the local context information lost in the first stage mask modeling process is supplemented by positive and negative samples in contrast leaning.Finally,the feature cassification is completed by training a linear cassifier.The method is compared with mainstream self-supervised classification methods and supervised classification methods in experiments on the publicly available remote sensing image datasets AID and NWPU-RESISC45.The experimental results:show that the proposed model achieves 95.37%and 95.14%classification accwracy on the two datasets,respctively,with better performance than mainstream self-superised models such as DINO,MoCo,SSCANs.It is close to the mainsteam supervised models such as GLANet,CANet,MG-CAP(Sqt-E),and has good application value.

关 键 词：自监督学习 对比学习 遥感图像 特征提取 无监督预训练 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]