Enhancing Deep Learning Bias Correction of Climate Extremes with Uncertainty Quantification
Abstract
Prediction of key climate quantities are undermined by biases in the climate models due to their simplification of physical and dynamical processes in tandem with coarse spatial and temporal resolutions. The Regularized Adversarial Domain Adaptation (RADA) is a deep learning methodology that has been developed to correct biases within climate simulations. Deep learning models have increasingly become standard in many applications yet have an absence in proper interpretation and difficulty in quantifying uncertainty. The RADA system is expanded to incorporate epistemic and aleatoric uncertainty quantification with complementary performance metrics for bias correction of climate simulations for precipitation and temperature extremes. The improved forecasts of climate extremes provide the development of surrogate models for enhanced streamflow and wildfire fuel moisture predictions with deep learning. The improved accuracy of climate extremes, with integral quantification of uncertainties, facilitates enhanced planning and strategies to mitigate threats to infrastructure that are essential to national security.
Prepared by LLNL under Contract DE-AC52-07NA27344.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMGC16C..02L