Satellite imagery analysis plays a vital role in remote sensing, but the information loss caused by cloud cover seriously hinders its application. This study presents a high-performance cloud removal architecture called Progressive Multi-scale Attention Autoencoder (PMAA), which simultaneously leverages global and local information. It mainly consists of a cloud detection backbone and a cloud removal module. The cloud detection backbone uses cloud masks to reinforce cloudy areas to prompt the cloud removal module. The cloud removal module mainly comprises a novel Multi-scale Attention Module (MAM) and a Local Interaction Module (LIM). PMAA establishes the long-range dependency of multi-scale features using MAM and modulates the reconstruction of the fine-grained details using LIM, allowing for the simultaneous representation of fine- and coarse-grained features at the same level. With the help of diverse and multi-scale feature representation, PMAA outperforms the previous state-of-the-art model CTGAN consistently on the Sen2_MTC_Old and Sen2_MTC_New datasets. Furthermore, PMAA has a considerable efficiency advantage, with only 0.5% and 14.6% of the parameters and computational complexity of CTGAN, respectively. These extensive results highlight the potential of PMAA as a lightweight cloud removal network suitable for deployment on edge devices. We will release the code and trained models to facilitate the study in this direction.

翻译：卫星影像分析在遥感领域发挥着重要作用，但云层覆盖造成的信息损失严重阻碍了其应用。本研究提出了一种名为渐进式多尺度注意力自编码器（PMAA）的高性能去云架构，该架构同时利用全局与局部信息。该架构主要由云检测主干网络和去云模块组成。云检测主干网络利用云掩膜增强云覆盖区域，以引导去云模块进行修复。去云模块主要包含新型多尺度注意力模块（MAM）和局部交互模块（LIM）。PMAA通过MAM建立多尺度特征的长程依赖关系，并利用LIM调节细粒度细节的重建过程，从而在同一层级同时表征细粒度和粗粒度特征。借助多样化的多尺度特征表征能力，PMAA在Sen2_MTC_Old和Sen2_MTC_New数据集上持续优于先前最先进的CTGAN模型。此外，PMAA在效率上具有显著优势，其参数量和计算复杂度分别仅为CTGAN的0.5%和14.6%。这些广泛的结果凸显了PMAA作为适合部署于边缘设备的轻量级去云网络的潜力。我们将公开代码和预训练模型以促进该方向的研究。