Convective Cloud Detection From Himawari-8 Advanced Himawari Imager Data Using a Dual-Branch Deformable Convolutional Network

Abstract

Himawari-8 satellite, equipped with an advanced Himawari imager (AHI), has been widely employed for cloud detection tasks due to its high-spatiotemporal resolution. In this article, we propose a deep learning model named dual-branch deformable convolutional network (DBDCN) to detect convective cloud from AHI data. Specifically, we first choose some infrared channels from AHI to compute their difference. These brightness temperature difference (BTD) data are sensitive to cloud detection, thus providing complementary information for AHI. Then, both AHI and BTD are used as inputs of DBDCN, which adopts a dual-branch structure to effectively combine them together. Considering that convective clouds often exist in variable shapes, we use deformable convolutions to extract convective clouds' features in each branch. In addition, in order to fuse the complementary features from both branches, we propose a crossbranch fusion block. To evaluate the performance of our proposed DBDCN, we collect a convective cloud dataset, covering most of regions in southern China (i.e., 97$^\circ$E to 122$^\circ$E, 20$^\circ$N to 35$^\circ$N). Experimental results show that DBDCN is able to achieve better detection performance than the widely used threshold methods and existing convolutional neural networks in most cases.

Keywords

• Brightness temperature difference (BTD)
• convective cloud detection
• crossbranch fusion
• deformable convolution