DeepCloud: An Investigation of Geostationary Satellite Imagery Frame Interpolation for Improved Temporal Resolution.

Satellite imagery is a crucial product for today’s weather forecasts. Worldwide, billions of dollars are invested annually on state-of-the-art satellites to produce imagery with better resolution and more spectral channels. This paper proposes a deep-learning methodology to enhance the temporal resolution of multi-spectral weather products from already deployed geostationary spacecraft (GOES or Himawari). Our method can dramatically improve the smoothness of an animation made with a series of pictures from an area of interest (Hurricanes, Tropical Storms, or Polar Jet Streams or Cold Fronts) while maintaining a good level of detail. To do so, we applied the transfer learning technique to fine-tune an already published frame interpolation neural network with real satellite weather imagery. By doing so, our technique can synthesize an intermediate frame of two adjacent frames scanned at the standard image interval. We also investigated the effects of using multiple combinations of spectral channels (e.g. Visible, True-Color, and Infrared) to test and train this network. The proposed temporal improvement can be helpful for weather forecasters to recognize atmospheric trends faster and more precisely. Moreover, our method does not require space hardware modifications thus being a safe and cost-effective solution.