The ART of BARD: The Role of Domain Selection and the Background Field on Atmospheric River Tracking (ART) of BARD over the pan-Atlantic

Over the past decade, the number of methods for Atmospheric River (AR) detection has increased, highlighting the growing understanding that uncertainty in detection may affect scientific knowledge. This study evaluates and validates the regional scale implementation of Bayesian AR Detector (BARD), a statistical machine learning model developed to reduce the uncertainty in AR tracking (ART), using three different horizontal and vertical domains of background integrated water-vapour transport (IVT) field and focusing on the pan-Atlantic region during 1940-2022 using ERA5 data. The consistency in seasonal AR Probability (ARP) and IVT differences across 3 model runs indicates that all configurations capture the general seasonal cycle of ARs, with enhanced activity and moisture transport in the midlatitudes during winter. However, discrepancies in selected IVT backgrounds and domains led to anomalies’ magnitude and spatial distribution, particularly in AR detection probability and AR IVT over Western and Northern Europe. These discrepancies among model runs are large over the ocean where ARs take shape and are consistent in climate modes such as a strong positive El Niño-Southern Oscillation (ENSO+) of 2015-2016. This reflects the robust and inherent differences in how each configuration maps AR dimensions and their associated transport processes. Further, these biases in AR mapping across model runs led to higher differences in AR-induced precipitation, wind speed, and temperature in Northern Europe and Scandinavia. This comparison underscores the importance of evaluating model configurations to assess uncertainties in AR representation under varying IVT background fields across regional domains and climate conditions.