Quantifying Uncertainty in the Detection of Atmospheric Rivers

The Science

The Lawrence Berkeley National Laboratory (Berkeley Lab) Calibrated and Systematic Characterization, Attribution, and Detection of Extremes Scientific Focus Area (CASCADE SFA) has created a high performance–computing (HPC) tool for quantifying uncertainty in the detection of atmospheric rivers (ARs): Toolkit for Extreme Climate Analysis Bayesian AR Detector (TECA-BARD v1.0.1).

The Impact

This is the first time that a statistical machine-learning technique has been used to train a method (TECA-BARD v1.0.1) to simultaneously (a) emulate the behavior of experts when identifying ARs and (b) quantify uncertainty associated with the expert’s subjective opinions. This research enables investigation of AR research that explicitly quantifies a major source of uncertainty. It builds on an HPC code (TECA) that allows AR analyses to be done quickly and efficiently on HPC machines like those at the National Energy Research Supercomputing Center (NERSC).

Summary

The Berkeley Lab CASCADE SFA developed a novel AR detection, with the goal of “training” it to emulate how an atmospheric science expert would detect ARs. To accomplish this, eight members of the team counted ARs in meteorological maps. This dataset of expert AR counts was used within a statistical machine-learning (Bayesian) framework to determine optimal parameter settings for the novel AR detection algorithm. This machine-learning process resulted in a set of 128 separate AR detectors designed to emulate each of the eight experts, resulting in a total of 1,024 separate AR detectors. These 1,024 detectors were incorporated in TECA as the TECA-BARD v1.0.1.

The team used TECA-BARD v1.0.1 to investigate the question “How does El Niño affect the number of ARs?”; TECA- BARD v1.0.1 provided 1,024 different answers. Differences indicate that the answer to the question (whether there are more or fewer ARs during El Niño) depends on which expert the AR detector was trained, leading the authors of the paper to call for more research to constrain their definition of ARs.

TECA-BARD v1.0.1 is publicly available to the scientific research community as part of the TECA software suite.

Principal Investigator(s)

William D. Collins
Lawrence Berkeley National Laboratory
[email protected]

Funding

This work was done as part of the Regional & Global Climate Model (RGCM)-funded Calibrated and Systematic Characterization, Attribution, and Detection of Extremes Scientific Focus Area (CASCADE SFA), which aims to understand how energy use impacts the occurrence of rare and extreme weather.

References

O’Brien, T.A., Risser, M.D., Loring, B.  et al. “Detection of atmospheric rivers with inline uncertainty quantification: TECA-BARD v1.0.1.” Geoscientific Model Development 13(12), 6131–6148 (2020). [DOI:10.5194/gmd-13-6131-2020]