ABSTRACT

Meteorology satellite visible-light images are critical for meteorologists. However, there are no satellite visible-light channels data at night, so we propose a method based on deep learning to create synthetic satellite visible-light images during night. Specifically, to produce realistic-looking products, we trained a generative adversarial network (GAN) model. The model can generate satellite visible-light images from corresponding satellite infrared (IR) channels data and numerical weather prediction (NWP) products. Considering to explicitly evaluating the contributions of different satellite IR channels and NWP products elements, we suggest using a channel-wise attention mechanic, e.g. a ‘Squeeze and Extraction Block’ (SEBlock) to quantitatively weigh the importance of different input data channels. The experiments based on the NWP products and the meteorology satellite data show that the proposed method is effective to create realistic synthetic satellite visible-light images during night.

摘要

气象卫星可见光图像对气象学家至关重要。然而，夜间没有卫星可见光通道数据，因此我们提出了一种基于深度学习的方法来创建夜间合成卫星可见光图像。具体来说，为了生产逼真的产品，我们训练了一个生成对抗网络 (GAN) 模型。该模型可以从相应的卫星红外（IR）通道数据和数值天气预报（NWP）产品中生成卫星可见光图像。考虑到明确评估不同卫星红外通道和 NWP 产品元素的贡献，我们建议使用通道注意机制，例如“挤压和提取块”（SEBlock）来定量权衡不同输入数据通道的重要性。