There currently exist many different atmospheric river (AR) detection algorithms. The AR Tracking Method Intercomparison Project (ARTMIP) compares more than a dozen different threshold-based AR detection and tracking algorithms (heuristics). These heuristics typically rely on the researcher’s expertise to set various thresholding criteria, either absolute or relative, on variables such as integrated vapor transport or integrated water vapor, and geometric thresholds on the length or width of an AR. Differences in these heuristics can lead to large discrepancies in the output.
To overcome the need for engineered features, we use a state-of-the-art deep learning (DL) model trained on the ClimateNet dataset – an open, community-sourced human expert-labeled curated dataset – for AR detection. Previous work (Kashinath, et al., 2020) has shown that this approach is not only accurate but also very efficient for detection across large datasets and generalizes well to climate change scenarios that the DL model has not been trained on. Here we present results from the DL model applied to coupled historical simulations in the Energy Exascale Earth System Model (E3SM) and two modern reanalysis products (ERA5 and MERRA2) and provide a comparison to two heuristics - tARget v3 (Guan and Waliser, 2019) and TECA-BARD (O’Brien et al., 2020). We show that the DL model corrects for known biases in heuristic-based detection algorithms, such as an underestimation of ARs compared to human controls. In conclusion, the DL model better emulates expert understanding of ARs, thus providing a potential for more accurate predictions of the impacts of ARs in a changing climate.