Precipitation from atmospheric rivers (ARs) presents both opportunity and challenge to human and natural systems near midlatitude western coasts: opportunity in that AR-related precipitation often satisfies a large portion of local water budgets in these coastal regions, and challenge in that the associated precipitation is often more extreme than that from other types of storm systems. Despite this importance, and despite a reasonably well-understood qualitative definition of ARs, the quantitative detection of ARs has remained surprisingly challenging. Results from the Atmospheric River Tracking Method Intercomparison Project (ARTMIP) indicate that there are a broad range of plausible AR detectors, and that scientific results can depend on the algorithm used. It is therefore imperative that AR-related science explicitly account for this uncertainty in the quantitative definition of ARs.

We present results from novel statistical and machine learning approaches that detect ARs in atmospheric data and explicitly account for uncertainty in the detection. The first method utilizes convolutional neural networks to estimate the probability that each location in a global dataset is associated with an AR. The second method utilizes a Bayesian approach to sample the posterior distribution of parameters in a simple, heuristic AR detector; the essential goal of the method is to probabilistically select heuristic AR detectors that get the 'right' counts for ARs relative to a dataset of synoptic fields in which experts counted ARs. We compare our results to other detection methods within ARTMIP, and we apply our probabilistic AR detection methods to a variety of climate model scenarios to explore the interactions among AR detection uncertainty, natural variability, and external forcing.