基于卷积神经网络的场景级云分类算法

Scene level cloud classification algorithm based on convolutional neural network

遥感图像受云层覆盖的影响,使得光学卫星拍摄的遥感图像中的大量地表或地物被云层遮挡,大量下传此类云覆盖图像则严重浪费卫星对地数传资源。面向遥感图像的在轨云判应用,根据星上对不同含云量场景级遥感图像块的不同处理策略,提出一种新的云覆盖度等级场景分类准则,利用该准则进一步提出基于深度卷积网络的云图像场景分类算法,将传统图像分割算法只计算整幅图像云占比来进行云判的方式,精细到局部场景的不同级别云判,为卫星数据下传提供更精细的指示信息,更有效的利用在轨拍摄资源。通过多组实验分析,同时考虑到星上计算资源的限制,确定了合适的训练样本数量和深度卷积网络,最终证明提出的算法可以实现局部场景不同级别的精准云判。

Affected by cloud cover, a large number of earth surface in remote sensing images taken by optical satellite are sheltered. It is a serious waste of satellite data transmission resources if these cloud-covered images are transmitted downward to ground. Applied to the in-orbit cloud judgment, this paper proposed a new classification criteria of cloud cover level based on different in-orbit processing strategies for different cloud-covered content, then this paper further proposed classification algorithm for scene-level remote sensing image based on deep convolution network. Traditional segmentation algorithms only calculated the cloud proportion of entire image. Algorithms in this paper can finely judges the cloud-coverd level of scene-level remote sensing images and using the limited hardware resources in-orbit more efficiently. Through multiple groups of experimental analysis, considering into the limited in-orbit computing resources, choose the appropriate number of training samples and deep convolution networks. Finally, the proposed algorithm could achieve a better accuracy of cloud judgments.