Quantifying Uncertainty in Precipitation Type Detection with Bayesian Deep Learning
Abstract
Understanding Eraths water and energy cycles is tightly coupled with the understanding of precipitation. Ability to accurately capture spatio-temporal distribution of precipitation is limited. Currently, satellite-borne Passive Microwave (PMW) radiometers provide precipitation retrievals which aid in understanding of global total precipitation. These precipitation retrievals are prone to large systematic errors because of the inability to adequately relate changes in the precipitation intensity to the change in radiometric signatures driven by the variability in the cloud structure. Understanding the type of precipitation, i.e. convective vs stratiform, could significantly improve the estimation of the precipitation. To address this challenge, we present a study on the Bayesian Deep Learning to accurately classify precipitation type from PMW radiometer inputs while providing uncertainty in the classification. In this novel approach to precipitation classification, we adopt a variational inference framework to develop a Bayesian form of the Residual Neural Networks (ResNet) architecture to classify these structural cloud differences while quantifying uncertainty in the classification. We benchmark several ResNet architectures in both deterministic and probabilistic configurations to achieve accuracies of above 86% for deterministic and above 90% for probabilistic configuration. Unique to probabilistic approach is that for the Bayesian Deep Learning model configurations each classification output is accompanied with the uncertainty estimate. We demonstrate that by utilizing uncertainty one can further improve model performance by filtering out least certain predictions. Moreover, this is one of the largest datasets used for studying the utility of Bayesian Deep Learning given that we utilize more than one year worth of satellite data resulting in more than 14 million input feature vectors used for model development.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H21E..03O