The occurrence of overshooting cloud tops can cause extremely severe weather such as strong winds and heavy rainfalls. The traditional overshooting cloud top (OT) detection methods not only need to choose a reasonable threshold, it is also very hard to make full advantage of the multispectral information of cloud images. These make small-scale OT detection very difficult with poor accuracy of OT boundary determination. In order to utilize the multispectral information of Himawari-8 satellite cloud images, in this article, we propose a method for detecting OT based on the dual channel multiscale deep network (DCMSDN). The brightness temperature of infrared window and the difference of brightness temperature between the infrared window and water vapor window are used as dual channel inputs, respectively. Then, DCMSDN introduces a multiscale prediction module to improve the accuracy of small target detection, which makes the network more suitable for the detection of the OT with small spatial scale. Experimental results indicate that the proposed method provides competitive performance with acceptable computational efficiency. Specifically, for the quantitative indicators of OTs detection, our approach achieves the accuracy of 89.36%, the precision of 95.63%, the recall of 88.90%, and the F1-measure of 91.61% for the test cloud images, which outperforms that of comparative methods.

中文翻译：

---

使用双通道多尺度深度网络通过 Himawari-8 图像检测过冲云顶

出现过冲云顶会导致极端恶劣的天气，如强风和暴雨。传统的超调云顶（OT）检测方法不仅需要选择合理的阈值，而且很难充分利用云图像的多光谱信息。这些使得小规模 OT 检测非常困难，OT 边界确定的准确性较差。为了利用Himawari-8卫星云图像的多光谱信息，在本文中，我们提出了一种基于双通道多尺度深度网络（DCMSDN）的OT检测方法。红外窗的亮温和红外窗与水汽窗的亮温差分别作为双通道输入。然后，DCMSDN 引入了多尺度预测模块来提高小目标检测的准确性，使网络更适合小空间尺度的 OT 检测。实验结果表明，所提出的方法以可接受的计算效率提供了具有竞争力的性能。具体来说，对于OTs检测的量化指标，我们的方法实现了89.36%的准确率、95.63%的准确率、88.90%的召回率和91.61%的测试云图像的F1-measure，优于对比方法。实验结果表明，所提出的方法以可接受的计算效率提供了具有竞争力的性能。具体来说，对于OTs检测的量化指标，我们的方法实现了89.36%的准确率、95.63%的准确率、88.90%的召回率和91.61%的测试云图像的F1-measure，优于对比方法。实验结果表明，所提出的方法以可接受的计算效率提供了具有竞争力的性能。具体来说，对于OTs检测的量化指标，我们的方法实现了89.36%的准确率、95.63%的准确率、88.90%的召回率和91.61%的测试云图像的F1-measure，优于对比方法。