Satellite video, as an emerging data source for Earth observation, enables dynamic monitoring and has wide-ranging applications in diverse fields. Nevertheless, cloud occlusion hinders the ability of satellite video to provide uninterrupted monitoring of the Earth’s surface. To mitigate the interference of clouds, cloud-free areas need to be selected before application, or an optimized solution like a cloud removal algorithm can be utilized to recover the occluded regions, both of which inherently demand the precise detection of clouds. However, no existing methods are capable of robust cloud detection in satellite videos. We propose the first sub-second-level satellite video cloud detection model VCDFormer to handle this problem. In VCDFormer, a spatial–temporal-enhanced transformer consisting of a local spatial–temporal reconfiguration block and a spatial-enhanced block is introduced to explore global spatial–temporal correspondence efficiently. Additionally, we construct WHU-VCD, the first sub-second-level synthetic dataset specifically designed to capture the more realistic motion characteristics of both thick and thin clouds in satellite videos. Compared to the state-of-the-art cloud detection methods, VCDFormer achieves an approximate 10%–15% improvement in the IoU metric and a 5%–8% increase in the F1-Score on the simulated test set. Experimental evaluations on Jilin-1 satellite videos, involving both synthetic and real-world scenarios, demonstrate that our proposed VCDFormer achieves superior performance in satellite video cloud detection tasks. The source code is available at https://github.com/XyJin99/VCDFormer.

中文翻译：

---

VCDFormer：研究亚秒级卫星视频中的云检测方法


卫星视频作为新兴的对地观测数据源，能够实现动态监测，并在各个领域有着广泛的应用。然而，云遮挡阻碍了卫星视频提供对地球表面的不间断监测的能力。为了减轻云的干扰，需要在应用前选择无云区域，或者可以使用云去除算法等优化解决方案来恢复被遮挡的区域，这两者都需要精确检测云。然而，现有的方法都无法在卫星视频中进行稳健的云检测。我们提出了第一个亚秒级卫星视频云检测模型 VCDFormer 来处理这个问题。在 VCDFormer 中，引入了一个由局部时空重构块和空间增强块组成的时空增强变换器，以有效地探索全局时空对应关系。此外，我们构建了 WHU-VCD，这是第一个亚秒级合成数据集，专门用于捕捉卫星视频中厚云和薄云的更真实运动特性。与最先进的云检测方法相比，VCDFormer 在模拟测试集上的 IoU 指标提高了大约 10%-15%，F1-Score 提高了 5%-8%。对吉林一号卫星视频的实验评估，涉及合成和真实场景，表明我们提出的 VCDFormer 在卫星视频云探测任务中取得了优异的性能。源代码可在 https://github.com/XyJin99/VCDFormer 上获得。