基于全局注意力的自监督对比学习遥感图像场景识别方法

Self-supervised Contrastive Learning Based on Global Attention for Remote Sensing Scene Classification

摘要:

针对遥感图像的类间同质和类内多模态问题以及现有基于全监督范式的遥感图像场景识别方法的对数据持高度依赖性问题。本文提出了基于全局注意力的自监督对比学习遥感图像场景识别方法(Self-supervised contrastive learning based on global attention for remote sensing scene classification，GACL)。首先，提出并构建了基于时空不变的数据增强模块，用于来学习遥感图像在不同时空中的一致性特征；然后，为了充分挖掘并建立图像内部之间的空间上下文关系，构建了基于残差全局注意力的特征提取模块；最后，为了充分学习多层特征中的不变信息，降低样本不平衡度对识别精度所带来的干扰，基于焦点损失和多层对比损失构建了复合对比损失函数。在NWPU、UCM和MLRSNet数据集上的实验分别达到了79.83%、83.01%、94.46%的精度。验证了GACL的有效性与优越性。

Abstract:

Aiming at the inter-class homogeneity and intra-class multimodality problems of remote sensing images and the problem of high dependence on data holding of existing remote sensing image scene recognition methods based on fully supervised paradigm. In this paper, a self-supervised contrastive learning based on global attention for remote sensing scene classification(GACL), is proposed. First, a spatio-temporal invariant-based data enhancement module is proposed and constructed to learn the consistent features of remote sensing images in different spatio-temporal spaces; then, in order to fully explore and establish the spatial contextual relationships between images, a residual global attention-based feature extraction module is constructed; finally, in order to fully learn the invariant information in multilayered features, and to reduce the interference of the sample imbalance on the recognition accuracy, a focus loss and multilayered features-based approach is proposed and constructed. interference, a composite contrast loss function is constructed based on focus loss and multilayer contrast loss. The experiments on the NWPU, UCM and MLRSNet datasets achieve an accuracy of 79.83%, 83.01% and 94.46%, respectively. The effectiveness and superiority of GACL are verified.