Cloud Detection Using Fully Convolutional Network with Zynq SoC for Spaceborne Application

Cloud detection is an important step to avoid the interference of contaminated areas in the remote sensing image. At present, the onboard cloud detection using deep learning is an attractive idea to provide the solution for detecting cloud contaminated region with high accuracy in real-time. However, the method based on deep learning has a large amount of model parameters and requires high computation resources, which is difficult for deployment in the onboard scenario. To address this issue, the cloud detection using the fully convolutional network with Zynq SoC is proposed in this article. Multiple convolution layers in a fully convolutional network are used to extract deep semantic features to improve the accuracy of cloud detection in different scenarios. And a custom computing architecture with full-precision parameters is conducted, which utilizes the loop tiling for feature maps and general matrix multiplication with parallel computing for convolution. The proposed network is deployed under the limited hardware resource. Experimental results indicate that the mean intersection over union of the proposed method is 90.39%, and the pixel accuracy reaches 95.79%. Compared with the implementation on ARM, the proposed method can achieve about 18.84 times speedup.