Ground‐Based Cloud Classification Using Task‐Based Graph Convolutional Network. Issue 5 (2nd March 2020)

• Record Type: Journal Article
• Authors: Liu, Shuang; Li, Mei; Zhang, Zhong; Cao, Xiaozhong; Durrani, Tariq S.
• Is Part Of: Geophysical research letters. Volume 47:Issue 5(2020)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-03-02
• Subjects: Geophysics -- Periodicals; Planets -- Periodicals; Lunar geology -- Periodicals; 550
• DOI: 10.1029/2020GL087338
• Languages: English
• ISSNs: 0094-8276

Abstract

Abstract: Clouds play a significant role in weather forecasts, water cycle, and climate system. However, existing methods ignore the relations of ground‐based cloud images. In this letter, we propose a novel method named task‐based graph convolutional network (TGCN) for ground‐based cloud classification, which takes image relations into consideration. To this end, we construct the graph using convolutional neural network‐based features of ground‐based cloud images which are learned in a supervised manner, and incorporate the graph computation into TGCN. Given that existing ground‐based cloud databases are with limited labeled training images and categorized according to different classification criteria, we release the largest ground‐based remote sensing cloud database (GRSCD) to provide a comparative study for different methods and to further improve the study of regional sky conditions. The experimental results on GRSCD manifest the effectiveness of TGCN for ground‐based cloud classification. Plain Language Summary: Clouds are an important indicator of weather conditions. The investigation of ground‐based cloud classification is performed using hand‐crafted methods or convolutional neural networks (CNNs). Although CNNs have exhibited remarkable performance in ground‐based cloud classification, they are limited to fixed data structures, which results in inadequate learning for cloud representations. Specifically, CNNs take ground‐based cloud images and their labels as inputs, and therefore they cannot discover the intrinsic data structures without considering the relations among ground‐based cloud images. As a kind of natural texture, clouds possess large intraclass and small interclass variances. Furthermore, the effectiveness of deep learning is highly dependent on labeled training samples, which could not be satisfied by the existing public ground‐based cloud databases. In this letter, we release the largest ground‐based remote sensing cloud database (GRSCD) with 8, 000 cloud images and propose the task‐based graph convolutional network to explicitly construct the relations among ground‐based cloud images so as to mine their inherent structure information. As far as we know, we first apply graph convolutional network to model image relations in the deep network for ground‐based cloud classification. Extensive experiments validate that the proposed method can exceed the conventional and CNN‐based methods on GRSCD. Key Points: We first apply graph convolutional network to model image relations in the deep network for ground‐based cloud classification We release the largest ground‐based remote sensing cloud database with 8, 000 cloud images Extensive experiments validate that the proposed method is superior to other methods on the ground‐based remote sensing cloud database