Using Deep Learning for an Analysis of Atmospheric Rivers in a High‐Resolution Large Ensemble Climate Data Set

Abstract

Atmospheric rivers (ARs) are elongated corridors of water vapor in the lower troposphere that cause extreme precipitation over many coastal regions around the globe. They play a vital role in the water cycle in the western US, fueling the most extreme west coast precipitation and sometimes accounting for more than 50% of total annual west coast precipitation (Gershunov et al., 2017). Severe ARs are associated with extreme flooding and damages while weak ARs are typically more beneficial to our society as they bring much needed drought relief (Ralph et al., 2019). Future climate projections show an increase in US west coast precipitation variability caused by AR precipitation increasing and non-AR precipitation decreasing (Gershunov et al., 2019). From 2012 to 2016, California experienced a historic drought, which was followed by the state's second wettest year on record. 2020 and 2021 are two of the driest years on record over much of the western US (Williams et al., 2022). The extreme interannual variability in western US precipitation in recent years coinciding with climate change projections of increased precipitation variability is a serious cause for concern over how patterns may change in the coming decades (Polade et al., 2017;Shields & Kiehl, 2016).A necessary step in understanding changing patterns in ARs as a function of climate change is employing an AR detection method. There are a variety of different algorithms used to track ARs due to their relatively diverse definitions (Shields et al., 2018). The Atmospheric River Tracking Intercomparison Project (ARTMIP) organizes and provides information on all the widely accepted algorithms that exist. Nearly all the algorithms included in ARTMIP rely on absolute and relative numerical thresholds. Absolute thresholds are static constraints that are required for an AR to exist, typically coming in the form of length, width, minimum inte-Abstract There is currently large uncertainty over the impacts of climate change on precipitation trends over the US west coast. Atmospheric rivers (ARs) are a significant source of US west coast precipitation and trends in ARs can provide insight into future precipitation trends. There are already a variety of different methods used to identify ARs, but many are used in contexts that are often difficult to apply to large climate datasets due to their computational cost and requirement of integrated vapor transport as an input variable, which can be expensive to compute in climate models at high temporal frequencies. Using deep learning (DL) to track ARs is a unique approach that can alleviate some of the computational challenges that exist in more traditional methods. However, some questions still remain regarding its flexibility and robustness. This research investigates the consistency of a DL methodology of tracking ARs with more established algorithms to demonstrate its high-level performance for future studies.Plain Language Summary Atmospheric rivers (ARs) are long corridors of water vapor in the lower atmosphere that are...

Deep Learning Image Segmentation for Atmospheric Rivers

“…Besides detection, deep learning has also been employed to perform image segmentation for ARs. In this case, the desired Higgins et al (2023) Zonal wind at 850 mb AR IoU 5 20%-50% Meridional wind at 850 mb Surface pressure IWV Tian et al (2023) Zonal wind at 850 mb AR IoU 5 38.5% Meridional wind at 850 mb IWV output of the network is an array similar to that output by TECA or any other heuristic that shows whether each position in the image belongs to an AR. This is more useful than simple classification as the spatial attributes of features, ARs in this case, can be more easily studied.…”

“…Furthermore, a NEPAC-trained network obtained an all-class mIoU of 57% when trained and tested on NEPAC-only data, showing that a region-specific network might be useful. Higgins et al (2023) trained a network based on CGNet (Wu et al 2021) on data spanning different regions, different horizontal resolutions and different climate models. The authors argued that IVT was computationally expensive to calculate and that some model outputs do not have this field present.…”

East Asia Atmospheric River Forecast With a Deep Learning Method: GAN‐UNet

“…Nevertheless, DL has some applications focused on ARs, including AR detection (Higgins et al, 2023;Prabhat et al, 2021; and postprocessing of AR forecasting (Chapman et al, 2019(Chapman et al, , 2022. employed an ensemble of 20 different DL models to perform semantic segmentation for ARs.…”

“…Nevertheless, DL has some applications focused on ARs, including AR detection (Higgins et al., 2023; Prabhat et al., 2021; Tian et al., 2023) and postprocessing of AR forecasting (Chapman et al., 2019, 2022). Tian et al.…”