【摘要】 深度卷积神经网络已被证实是高分辨率遥感影像场景分类中最有效的方法之一。过去的研究大多关注于单一光学遥感影像的场景级分类，并且多为单标签分类。然而，单一光学遥感影像容易受到天气条件的限制，并且单标签的标注难以全面描述复杂的图像内容。因此，本文利用欧洲空间局于2020年获取的SAR和光学遥感图像，构建了武汉市多模态多标签场景分类数据集SEN12-MLRS,并设计了一种基于并行双注意力融合网络(PDANet)的多标签场景分类方法。PDANet通过双分支特征提取、自适应特征融合及多级特征融合，实现了光学和SAR图像的多模态与多层级的特征融合。试验结果表明，在SEN12-MLRS数据集上，PDANet相较于多种先进模型取得了最佳性能，并通过消融试验进一步验证了本文方法的有效性。更多还原

【Abstract】 Deep convolutional neural networks have proven to be one of the most effective methods for scene classification of high-resolution remote sensing images. Most previous studies focus on scene-level classification of single optical remote sensing images and are primarily limited to single-label classification. However, single optical remote sensing images are often constrained by weather conditions, and single-label annotations cannot fully describe complex image contents. Therefore, in this paper, we constructed a multimodal, multi-label scene classification dataset called SEN12-MLRS, using SAR and optical remote sensing images acquired by the European Space Agency in 2020. We proposed a parallel dual attention fusion network(PDANet) for multi-label scene classification. PDANet achieves optical and SAR image feature extraction as well as multi-modal and multilevel feature fusion through two-branch feature extraction, adaptive feature fusion, and multilevel feature fusion. Experimental results demonstrate that PDANet achieves superior performance compared to many state-of-the-art models on the SEN12-MLRS dataset. The effectiveness of the proposed network and its modules is further validated through ablation experiments.更多还原