A multi-scale convolutional neural network for automatic cloud and cloud shadow detection from Gaofen-1 images

Abstract

The reconstruction of the information contaminated by cloud and cloud shadow is an important step in preprocessing of high-resolution satellite images. The cloud and cloud shadow automatic segmentation could be the first step in the process of reconstructing the information contaminated by cloud and cloud shadow. This stage is a remarkable challenge due to the relatively inefficient performance of the current methods in complex scenes in multispectral high-resolution satellite images. In recent years, using deep convolutional neural networks has largely improved the performance of cloud and cloud shadow segmentation. Increasing the generalization capability of cloud and cloud shadow segmentation is one of the problems of deep convolutional neural networks. In this paper, we focus on tackling the poor generalization performance of automatic cloud and cloud shadow segmentation in Gaofen-1 (GF-1) images. In this regard, we propose a deep learning multiscale method, founded on multi-dimension filters, for accurate segmentation of cloud/cloud shadow in single date GF-1 images which is based on a new multi-scale deep residual-convolutional neural network called MultiCloud-Net. The cloud/cloud shadow masks are extracted based on a new loss function to generate the final cloud/cloud shadow masks. The MultiCloud-Net was implemented in the Google Colab and was validated using 12 globally distributed GF-1 images. The quantitative assessments of test images show that the average F1 score, the average Jaccard Similarity Index (JSI), and the Kappa coefficient for cloud (cloud shadow) segmentation are about 97 (95.5), 96 (94.5), and 0.98, respectively. The experimental results using the GF-1 images demonstrate a more reasonable accuracy and efficient computational cost achievement of the proposed method compared to the automatic cloud/cloud shadow segmentation performance of two advanced deep learning and statistical methods.