Convolutional neural networks (CNNs), a form of machine learning, are a novel and powerful technique for detecting weather events in atmospheric datasets. By training CNNs on datasets of pre-identified weather events, CNNs can learn how to recognize these events. Atmospheric rivers (ARs) are a particularly challenging class of weather event, since there is no single community-accepted AR identification algorithm. This results in a fundamental uncertainty that must be represented when training CNNs. To represent the uncertainty expressed by contemporary, state-of-the-science AR tracking methods, we create probabilistic AR detection fields from 14 algorithms submitted to the Atmospheric River Tracking Method Intercomparison Project (ARTMIP). Each algorithm identifies grid cells associated with ARs in over 30 years of 3-hourly data from the MERRA reanalysis. At a given time, we estimate each grid cell’s probability of AR detection as the proportion of ARTMIP algorithms that identify an AR in that grid cell. We refer to this probabilistic identification as the ‘ARTMIP mean’. With a deep-learning-based segmentation model, we generate probabilistic AR identifications that are quite close to the ARTMIP mean, with an average RMSE of 0.03.

Additionally, the ARTMIP algorithms largely rely on integrated vapor transport (IVT; the mass-weighted vertical integral of moisture times wind) for heuristic identification of ARs. This quantity yields superior AR detection results, with ARs localized closer to areas of intense precipitation. However, calculating IVT requires three 3-D, subdaily fields: zonal and meridional wind and specific humidity. Because these fields need a large amount of storage, IVT is a somewhat rare field in climate model output. We demonstrate CNNs’ ability to produce AR probabilistic masks using the more widely available ‘precipitable water’ field (the mass-weighted vertical integral of moisture). We demonstrate that our probabilistic deep-learning-based method can readily be applied to arbitrary atmospheric model output, and we explore the viability of this method as a general-purpose probabilistic AR detection method. Finally, we discuss methodological advances that yield a probabilistic, CNN-based AR detection method which emulates expert judgment in AR detection.