Assessment of AWS Constellation on potential Nowcasting impact at high latitudes
Title: Assessment of AWS Constellation on potential Nowcasting impact at high latitudes
SNIC Project: Berzelius-2022-221
Project Type: LiU Berzelius
Principal Investigator: Bengt Rydberg <bengt.rydberg@smhi.se>
Affiliation: SMHI
Duration: 2022-10-31 – 2023-05-01
Classification: 10508
Homepage: https://www.smhi.se/forskning/forskningsenheter/meteorologi/bengt-rydberg-1.184885
Keywords:

Abstract

The Arctic Weather Satellite (AWS) is a new weather satellite developed by OHB Sweden and the European Space Agency (ESA). A single unit of AWS is planned to be launched into a polar orbit in 2024, and the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) is currently investigation the benefit and possibility for a future constellation of several AWS units, for improving the sampling frequency at high latitudes of the existing weather satellite system. A better sampling frequency is anticipated to lead to better nowcasting and forecasting products of, for example, precipitation. The current weather satellite system only provides a frequent sampling at high latitudes during certain time of the day, and the aim of the current project is to estimate the benefit of adding a constellation of AWS units into the exiting weather satellite system. For this reason we intend to use machine learning techniques, and use data from sensors onboard current weather satellites together with data from a network of ground-based precipitation monitoring radars, in order to develop a data driven now-casting model to predict precipitation for the coming nearest hours. We anticipate that the forecasting skill of the data driven model will vary for forecasts issued during various time of the day since new input satellite data is not availabe during some hours of the day. Since the the aim of the AWS constellation is to fill in the gaps, we can then roughly estimate the impact of adding AWS into the system. Both the input and output of the data-driven model can be thought of as images, and U-Net convolutional neural networks, originally developed for biomedical image segmentation, has been proven to fit with the problem at hand. For this we reason we apply for GPU resources to train such models.