ARTMIP-early start comparison of atmospheric river detection tools: how many atmospheric rivers hit northern California’s Russian River watershed?

Abstract

Many atmospheric river detection tools (ARDTs) have now been developed. However, their relative performance is not well documented. This paper compares a diverse set of ARDTs by applying them to a single location where a unique 12-year-long time-series from an atmospheric river observatory at Bodega Bay, California is available. The study quantifies the sensitivity of the diagnosed number, duration, and intensity of ARs at this location to the choice of ARDT, and to the choice of reanalysis data set. The ARDTs compared here represent a range of methods that vary in their use of different variables, fixed vs. percentile-based thresholds, geometric shape requirements, Eulerian vs. Lagrangian approaches, and reanalyses. The ARDTs were evaluated first using the datasets documented in their initial publication, which found an average annual count of 19 ± 7. Applying the ARDTs to the same reanalysis dataset yields an average annual count of 19 ± 4. Applying a single ARDT to three reanalyses of varying grid sizes (0.5°, 1.0°–2.5°) showed little sensitivity to the choice of reanalysis. While the annual average AR event count varied by about a factor of two (10–25 per year) depending on the ARDT, average AR duration and maximum intensity varied by less than ± 10%, i.e., 24 ± 2 h duration; 458 ± 44 kg m^− 1 s^− 1 maximum IVT. ARDTs that use a much higher threshold for integrated vapor transport were compared separately, and yielded just 1–2 ARs annually on average. Generally, ARDTs that include either more stringent geometric criteria or higher thresholds identified the fewest AR events.

Appendices

Appendix 1: Estimating the precipitation contribution from ARs identified with different ARDTs

To quantify the impact of AR activity detected using different AR tracking schemes on precipitation regime at BBY, the contribution of AR-related precipitation to total annual precipitation accumulated at the area (certain grid cell) during the water years of 2005–2013 (Table 6) was estimated using Livneh’s (2013) precipitation dataset. Precipitation during AR days, defined as days with at least one 3-h time step associated with AR conditions, and the day after an AR day are counted. The set of ARDTs with the least strict criteria such as RSR2014, GSR2017 and GW2015 ranged from 55 to 60% of AR contribution per year, while WNR2013-IVT250 and MBM2016 ranged from 45 to 53% per year. The ARDTs focused on much stronger ARs, Ralphetal2013-OBS47, SGS2013 and WNR2013-IVT500 contribute roughly 10% of AR-related precipitation per year. The annual behavior of AR-related precipitation contribution is illustrated in Figure AI-1. In particular, during wet years such as 2006 the contribution of AR-related precipitation was as much as 70% for ARDT outputs with the least strict criteria, whereas strict AR detection schemes account for up to 30% of the contribution. During dry years both AR activity (Fig. 7 from the main text) and AR precipitation contribution (Fig. 11) are about 25% lower.

Proving the statement on the connection of AR strength and precipitation intensity (Sect. 6) we estimated the contribution of AR precipitation to all precipitation summed in the different percentile categories (Fig. 12). The results show that in general moderate to extreme precipitation accumulations are most likely to be associated with AR events. Namely, more than 40% of heavy precipitation and 80% extreme precipitation events are associated with ARs. The catalogs focused on much stronger ARs (SGS2013, WNR2013-IVT500 and Ralphetal2013-OBS47) tend to catch predominantly heavy and extreme precipitation cases. The set of ARDT outputs based on simpler (or no) geometric characteristics (GSR2017, RSR2014, GW2015) cover a wider spectrum of precipitation events (Table 6; Fig. 11, 12).

Table 6 Annual average contribution of AR-related precipitation to all precipitation

Fig. 11

Annual average contribution of precipitation associated with AR events counted by each MERRA2-based AR catalog at the grid cell containing BBY during water years 2005–2013

Fig. 12

Contribution of precipitation associated with AR days in different daily precipitation percentile categories counted by each MERRA2-based AR catalog at the BBY grid cell during water years 2005–2013

Appendix 2: Quantification of the difference between choice of ARDT and choice of reanalysis

Sensitivity of AR frequency, duration and intensity to the detection methodology (the reanalysis datasets) is quantified by the amount of shared variance in AR catalogs obtained from applying different (the same) detection algorithms to the same (different) reanalysis dataset. The percentage of shared variance represented by square of average correlation coefficient between pairs of AR catalogs shows the amount overlap variation of those catalogs. Two sets of catalogs are considered: six MERRA2-based AR catalogs developed using GSR2017, GW2015, RSR2014, MBM2016, WNR2013-IVT and WNR2013-IWV with solid/percentile-based IVT/IWV thresholds and with/without geometry characteristics at AR detection schemes (see Sect. 4), and three AR catalogs obtained from applying RSR2014 algorithm to NCEP/NCAR, ERA-Interim and MERRA-2 reanalysis datasets (see Sect. 5) with different spatial and temporal resolutions. The number of AR events, their average duration and IVT intensity were computed from November through April during 2005–2010 water years according to data availability in considered AR catalogs. The results (Table 6) show that the AR catalogs based on different ARDTs applied to the same reanalysis share 70% of interannual AR variability, whereas AR catalogs based on the same detection method applied to different reanalyses share 84% of AR variability. This illustrates that the choice of reanalysis has about 14% less of an effect on AR frequency than does the choice of ARDT. Shared variations in average duration and IVT intensity of different reanalysis based ARs are 14% and 20% higher, respectively (Table 7).

Table 7 Shared variance in AR catalogs obtained from applying GSR2017, GW2015, RSR2014, MBM2016 and WNR2013-IVT AR detection algorithms to MERRA2 Reanalysis dataset and RSR2014 algorithm to NCEP/NCAR, ERA-Interim and MERRA-2 reanalysis datasets