Satellite images are indispensable for weather monitoring and forecasting. Visible and infrared satellite images are commonly used for such purpose. Image of the former type has higher resolution which enables more detailed analysis of weather systems’ features, but it is generated only during daytime when there is sunlight shining on the Earth. The latter type of image is available round-the-clock despite having a lower resolution.

With the advance of artificial intelligence technology, the Observatory has developed a deep learning model for generating virtual nighttime visible satellite images to facilitate weather forecasters in determining the centre of a tropical cyclone during nighttime. It is a conditional generative adversarial networks (CGAN) model comprising a generator and a discriminator (Figure 1). The model has been trained with past satellite images from different channels of the Himawari-8 (H-8) satellite for optimizing the performance of the generator and the discriminator. After training, real-time satellite images could be fed into the CGAN model for producing virtual visible images. Useful results were obtained as exemplified in the case of tropical cyclone Rai in December 2021 (Figure 2).