-
- A11A-05: Moving beyond parameterized precipitation processes using holistic machine learning and satellite observations
-
NOLA CC
Author(s):Generic 'disconnected' Message
Raul Moreno, University of Washington Seattle Campus (First Author, Presenting Author)
Dale Durran, University of Washington Seattle Campus
Nathaniel Cresswell-Clay, University of Washington Seattle Campus
Accurately predicting rainfall is one of the hardest problems in weather forecasting. Traditional models rely on approximations of complex processes to compute the precipitation variable, and these approximations result in large errors, especially for light and extreme precipitation. We train a neural network on satellite observations to estimate precipitation from easily observable fields. This allows us to bypass the typical approximations and assumptions that traditional models must make while also making use of observations. We find that the neural network produces more accurate precipitation estimates than ERA5 reanalysis, especially in capturing heavy rainfall events. Without additional training, it also performs well when using input from weather forecasts. These results emphasize the potential to use machine learning and observations to improve precipitation modeling.
Scientific DisciplineSuggested ItinerariesNeighborhoodType
Enter Note
Go to previous page in this tab
Session
