Precipitation Type from PMW Observations: Constellation Retrieval with Uncertainty Estimates
Abstract
Satellite-borne passive microwave (PMW) radiometers and the associated research on their precipitation retrievals deliver a highly accurate understanding of global total precipitation. At regional scales, however, PMW precipitation retrievals are prone to large systematic errors due to their limited ability to accurately relate changes in surface precipitation to the changes in radiometric signatures induced by varying cloud microphysics. Accurate information on precipitation system microphysical properties is expected to greatly mitigate this problem. A Machine Learning model is developed to retrieve precipitation type -convective vs. stratiform- using PMW observations of GMP mission constellation. The approach relies on Bayesian Deep Learning to deliver accurate classification of precipitation type and its uncertainty. By adopting Bayesian form of Residual Networks (ResNet) architectures, we extract the information from PMW observation vectors to identify the structural differences of cloud systems while providing, per pixel, classification uncertainty estimates. Additionally, the model offers quantitative insight on the origin of the output uncertainty for each of GPM radiometers, reporting accuracy of over 90%.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H35M1296P