面向像素级的遥感图像云分割算法

Pixel-level cloud segmentation algorithms for remote sensing images

云遮挡对于非气象类遥感卫星的效能发挥具有重要影响,会导致卫星拍摄的遥感图像中无法显示有价值的地面信息,同时,大量无价值的数据下传还会极大的浪费卫星有限的数传资源。针对上述情况,结合星上在轨应用的相关要求,提出了一种基于低秩及稀疏约束制图的图像分割算法,将局部云纹理分布特征与低秩稀疏关联制图约束相结合。最后,通过对不同算法、不同地区、不同云覆盖程度的遥感图像块中的云区域进行分割实验与对比分析,结果表明算法可以在纹理特征空间中,有效捕获不同厚度云纹理的局部及全局结构化差异分布,从而产生更理想的云分割结果。而且在云分布更复杂的遥感图像中,还能有效分辨出厚云、薄云、无云3类像元,可为后续进一步对图像中云的处理提供支撑。

Cloud occlusion plays an important role in the effectiveness of non-meteorological remote sensing satellites. It will lead to the inability of displaying valuable ground information in remote sensing images taken by satellites. At the same time, a large number of worthless data downloading will greatly waste the limited data transmission resources of satellites. In this paper, an image segmentation algorithm based on low rank and sparse constraint mapping is proposed, which combines local cloud texture distribution characteristics with low rank sparse correlation mapping constraints. Finally, experiments and comparative analysisfor different algorithms, regions and cloud coverage levels show that the proposed algorithm can effectively capture the local and global structured difference distribution of cloud textures with different thickness in the texture feature space, thus producing better cloud segmentation results. Moreover, in remote sensing images with more complex cloud distribution, it can effectively distinguished three types of pixels:thick cloud, thin cloud and cloudless, which can support for further cloud processing.