Atmospheric rivers do not explain UK summer extreme rainfall

Extreme rainfall events continue to be one of the largest natural hazards in the UK. In winter, heavy precipitation and floods have been linked with intense moisture transport events associated with atmospheric rivers (ARs), yet no large‐scale atmospheric precursors have been linked to summer flooding in the UK. This study investigates the link between ARs and extreme rainfall from two perspectives: (1) Given an extreme rainfall event, is there an associated AR? (2) Given an AR, is there an associated extreme rainfall event? We identify extreme rainfall events using the UK Met Office daily rain gauge data set and link these to ARs using two different horizontal resolution atmospheric data sets (ERA‐Interim and Twentieth Century Reanalysis). The results show that less than 35% of winter ARs and less than 15% of summer ARs are associated with an extreme rainfall event. Consistent with previous studies, at least 50% of extreme winter rainfall events are associated with an AR. However, less than 20% of the identified summer extreme rainfall events are associated with an AR. The dependence of the water vapor transport intensity threshold used to define an AR on the years included in the study, and on the length of the season, is also examined. Including a longer period (1900–2012) compared to previous studies (1979–2005) reduces the water vapor transport intensity threshold used to define an AR.


Introduction
Extreme rainfall continues to be one of the greatest natural hazards in the UK. A number of recent extreme rainfall events have led to widespread flooding highlighting the vulnerability of the UK to this hydrohazard. Aside from flooding, extreme rainfall can lead to hazardous driving conditions and disruptions to other travel. With extreme rainfall events expected to change in intensity and frequency with climate change [Gregerson et al., 2013;Kendon et al., 2014], understanding the causes of these extreme rainfall events is of particular importance.
Previous studies have shown a strong relationship between winter flooding over the UK and atmospheric rivers (ARs) [Lavers et al., 2012], synoptic features which have also been associated with winter flooding in the remainder of Europe [Lu et al., 2013;Lavers and Villarini, 2013a], the U.S. [Lavers and Villarini, 2013b;Ralph et al., 2006;Neiman et al., 2011], and South America [Viale and Nuñez, 2011]. Dacre et al. [2015] show that this enhanced moisture transport is formed by the cold front of an extratropical cyclone sweeping up water vapor in the warm sector as it catches up with the warm front. This results in a narrow band of high water vapor content forming ahead of the cold front at the base of the warm conveyor belt airflow. The UK study of Lavers et al. [2011] used river flow-gauge data and was focused on basins in the west of the country. The present study exploits rain gauge data for the whole country to identify events that may potentially lead to flooding in other regions.
Previous studies have shown that 70% of the winter precipitation in the UK can be attributed to extratropical cyclones [Hawcroft et al., 2012]. While there can be small-scale processes that lead to an increase in the precipitation intensity (e.g., orographic enhancement), the scale of the conditions that lead to winter precipitation, e.g., fronts, is of the scale of thousands of kilometers and can last several days. Hence, winter precipitation is typically widespread and for a prolonged period [Hand et al., 2004]. Therefore, it can be expected that extreme winter events are also associated with a large-scale process associated with cyclones. For example, an AR could be providing the "feeder" moisture for heavy precipitation associated with a seeder-feeder mechanism [Bergeron, 1965;Browning, 1974]. This is where upper level precipitation (seeder) falls through lower level orographic precipitation (feeder), here an AR, causing an increase in the precipitation intensity. There have been 10.1002/2014JD022863 a large number of studies that have shown the significant role orography has on the enhancement of precipitation associated with ARs in the U.S. [Neiman et al., 2013;Smith et al., 2010;Ralph et al., 2006;Dettinger et al., 2004] and elsewhere in the world [Viale and Nuñez, 2011].
Summer precipitation, however, is generally dominated by small-scale (less than tens of kilometers), short-lived (a few hours) precipitation [Hand et al., 2004], with flash floods (those lasting only a few hours) dominating the flood record. This type of precipitation is more commonly associated with convective events which can lead to extremely high rainfall rates. There are a number of dynamical processes that lead to convective precipitation [Bennett et al., 2006] which typically occur on relatively small spatial scales. Therefore, it is not clear whether ARs are associated with such events, since ARs are defined as a large-scale region of atmospheric convergence, while the processes that cause the summer events are typically much smaller in scale.
If the atmospheric precursors are better known, this could contribute toward better prediction of flash floods, thus reducing their potential impact on the UK. The aims of this paper are to investigate whether ARs can be used to explain summer extreme rainfall events over the UK and to examine how the threshold for defining an AR depends on the season and number of years used in the study.

Atmospheric Data sets
Two atmospheric data sets were used in this study, ERA-Interim [Dee et al., 2011] and the ensemble mean of the twentieth Century Reanalysis (20CR) [Compo et al., 2011]. ERA-Interim is a high-resolution (0.7 ∘ latitude by 0.7 ∘ longitude) reanalysis product from the European Centre for Medium-Range Weather Forecasts (ECMWF) that spans from 1979 until the present day; the period up until the end of 2013 was used in this study. The 20CR is a lower resolution (2.0 ∘ latitude by 2.0 ∘ longitude) reanalysis ensemble product from the National Oceanic and Atmospheric Administration (NOAA) covering a much longer period, 1871-2012; only the period 1900-2012 was used in this study. The year 1900 was chosen as the start year due to the availability of the rain gauge data (discussed later).
The higher resolution of the ERA-Interim data increases the likelihood that fields with small spatial scales associated with ARs, e.g., specific humidity, are identified; the longer period of the 20CR allows for trends in AR occurrence to be investigated. It should be noted that 20CR only assimilates surface observations of synoptic pressure, monthly sea surface temperature, and sea ice distribution [Compo et al., 2011]; therefore, the synoptic details are quite likely less accurate although the temporal continuity is greater, compared to the ERA-Interim. The number of ARs identified in each data set, and their association with known flooding events, are presented in the next section.

Atmospheric Rivers Identification
We apply a comparable method to Lavers et al. [2013] to detect ARs. Vertically integrated horizontal water vapor transport (IVT) is calculated by integrating the zonal and meridional moisture fluxes through each atmospheric layer between 1000 hPa and 300 hPa, as described in equation (1), where u and v represent the wind field in the zonal and meridional directions, respectively, q is the specific humidity, and p is the pressure at different pressure levels. The integral was calculated using two layers, one between 1000 hPa and 750 hPa and another between 750 hPa and 300 hPa. For ERA-Interim this involved integrating 11 vertical levels between 1000 hPa and 750 hPa (at 25 hPa intervals) and 10 vertical levels between 750 hPa and 300 hPa (at 50 hPa intervals). For 20CR, where the output has a constant vertical interval of 50 hPa, there were seven vertical levels between 1000 hPa and 750 hPa and nine vertical levels between 750 hPa and 300 hPa. This was calculated as a global field for every model time step. This does not take into account any contributions below 1000 hPa, which could be addressed by using surface pressure instead of 1000 hPa. For comparison to previous studies 1000 hPa was used.
The IVT values (kg m −1 s −1 ) between 50 ∘ N and 60 ∘ N at 4 ∘ W are examined, and the 85th percentile of the daily maxima at 1200 UTC is taken as the threshold for defining an AR, as used in previous studies [Lavers et al., 2013]. The limits of 50 ∘ N and 60 ∘ N were chosen for direct comparison to Lavers et al. [2013]. While this may result in some filaments of enhanced IVT that occur on the southern coast of the UK being missed, the limits were retained for comparison to previous studies. It was also found that between 20% and 25% of identified ARs occur north of 58 ∘ N, where there is very little landmass and very few rain gauges. This may result in a number of ARs being identified that do not have an effect on the precipitation over land. However, as before, the limits were retained for comparison to previous studies.
When the globally precalculated IVT field has exceeded the 85th percentile threshold between the limits of 50 ∘ N and 60 ∘ N at 4 ∘ W, the IVT field is tracked back to see whether this threshold is met for at least 20 ∘ in longitude and must be persistent for 18 h, three time steps, to meet the criteria for the persistent AR definition. Only a 4.5 ∘ latitude displacement (between time steps) to the north or south of the initial IVT maximum at 4 ∘ W was allowed. By assuming that the maximum IVT represents the midpoint of the AR (at 4 ∘ W), and that ARs have been considered to be on the order of 1000 km wide [Neiman et al., 2008], a 4.5 ∘ movement (which is approximately equal to 500 km) means that even if the central location of the AR moves by 4.5 ∘ , the AR may still be present over a specific location.
Finally, we ensure that each AR is unique by ensuring there are four time steps (1 day) in between identified IVT exceedances. A region of moisture convergence that fits all these criteria is then labeled as an independent AR. An example of a region of IVT that meets the AR definition for the UK is shown in Figure 1 from the ERA-Interim data set, along with the fields that are used to calculate IVT: specific humidity (q), zonal wind (u), and meridional wind (v) (all shown at 850 hPa).
Brönnimann et al. [2012] note that individual members of the 20CR are better able to represent extremes compared to the ensemble mean which is used here. This caveat is therefore considered when interpreting differences between the reanalysis products. Brönnimann et al. [2012] also show that the ensemble variance of extreme winds decreases after 1950. This suggests that the variance of the IVT values selected would similarly decline; this is not considered in the present study since trends are not assessed.

Extreme Rainfall Identification
To determine whether ARs over the UK can be linked to extreme rainfall events, daily rain gauge totals from the Natural Environment Research Council (NERC) Centre for Environmental Data Archival [Met Office, 2012] were used to identify extreme rainfall events. The locations of all the gauges included in this study are shown in Figure 2 (top). While it is known that intense summer rainfall is typically convective [Hand et al., 2004], occurring on time scales shorter than a day (typically hours), daily data were used due to a lack of a nationwide hourly rain gauge data.
To identify an event as extreme, a peak over threshold (POT) method was used. The threshold used was the top decile (90th percentile) of the maximum rainfall observed over the UK for all days for the period of the data set (1979-2013 for ERA-Interim and 1900-2012 for 20CR), for each season considered. In addition, only days when precipitation was observed are included, using a value of 2.54 mm as the lower daily threshold. This method is similar to the one used by Rutz et al. [2014], with their lower threshold being based on the resolution of their data set. The intensity resolution of the data set used here is 0.2 mm. However, this resulted in a very low top decile, and it was decided to use a similar threshold as used by Rutz et al. [2014]. The top decile calculated for each season is shown in Table 1. Other methods [Davison and Smith, 1990] have suggested selecting only a certain number of events for the period of study, such as the POT method. The POT method identifies a set number of events for the period, i.e., using POT1 an average of one event per year would be extracted; however, multiple events could come from the same year. Using POT2 and POT3, an average of two and three events per year, respectively, would be selected. These are applied separately to winter and summer. These thresholds are shown alongside the thresholds calculated earlier in Figure 3. This method was considered suitable for data sets where the number of events of interest is low (e.g., when river flow-gauge data exceed the banks and therefore causes flooding). However, in this case where the number of events of interest is higher it was decided that these thresholds identified too few extreme events. Hence, the top decile method was used to identify extreme events as being of interest.
The rain gauge data set had already undergone a number of quality control checks by the UK Met Office. However, it was necessary for further quality checks to be made to remove further problems identified in the data set. The number of rain gauges included in this study varies significantly for different periods. For the period 1900-1960, the number of operational rain gauges (post quality control) were relatively few (typically 100-300). After 1960 there was a large increase in the number of operational rain gauges to over 4000. For the period 2001-2011 there were a number of data quality issues due to changes to the data quality flags, used to identify data as potentially incorrect. This could not be resolved in this study; thus, the number of rain gauges included for this study during this period was around 1000.
These variations in the number of rain gauges included may result in the number, or intensity, of the extreme events identified changing. This was investigated, and it was found not to have a significant effect; the most extreme events in a year were still being recorded by a rain gauge somewhere in the UK. In the analysis, when multiple rain gauges recorded an extreme event on the same day, it was only classed as a single event. Therefore, the variation in the number of rain gauges is not considered to have a significant effect on the results. Studies have also shown that percentile indices potentially introduce inhomogeneities into the time series [Zhang et al., 2005], although this is more important for climate change detection and monitoring which is not the focus of the present study.
An AR was considered to be associated with an extreme rainfall event if the AR was first detected up to a day before the extreme rainfall event. Lavers et al. [2012], who used river flow-gauge data, allowed 3 days between an AR and flood event. This is to allow for the catchment response, time the time it takes for rain falling across the catchment to feed into the rivers and potentially result in flooding. The 1 day period used here is also the same used by Rutz et al. [2014], who also used rain gauge data in their study.
The sensitivity of the results to the length of this period was tested, up to 5 days, and is discussed in more detail alongside the main results in section 3.2. On average, for each day the period was extended by, an extra 1% of AR events had an extreme rain event associated with them. Given an extreme rain event, an additional 2% per day for which the period was extended by, had an AR associated with them. This study also uses the end date of the identified AR, allowing an extreme rainfall event to be associated with an AR until the final day it is identified. The sensitivity of the results to the end date of an identified AR was not tested. This study was not able to take into account any positional errors in the locations of the ARs in the reanalysis.

Threshold Selection
In previous studies [Lavers et al., 2013], the threshold to define an IVT as an AR was chosen as the 85th percentile of IVT identified at 4 ∘ W for the period 1979-2005 at 1200 for October to March (ONDJFM). In this study, more recent years were available (up until 2013 for ERA-Interim and 2012 for 20CR). The extra years available were used to test the dependence of the threshold value on the period over which it was calculated. It is known that there is variability between seasons and years on the prevalence of extreme rainfall [Burt and Howden, 2013;Jones et al., 2013]. This suggests that there will be variability in the IVT intensities depending on the years chosen to calculate the threshold and on the season being investigated. The dependence of the threshold on the length of the season used, i.e., an extended winter/summer (ONDJFM/AMJJAS) or a shortened winter/summer (DJF/JJA), was also investigated.
Initially, the study looked at the winter period to compare to known previous thresholds [Lavers et al., 2013] who used the period 1979-2005 and for an extended winter.  The results for summer, Table 3, highlight a number of differences to the results seen for winter (Table 2). In contrast to the winter results, the threshold value of IVT is similar for ERA-Interim and 20CR. For winter, the 20CR thresholds were around 10% lower than the ERA-Interim values. For summer the thresholds calculated never differ by more than 2%, a similar magnitude of difference as seen in the winter investigation into the effect of the length of the season and the years included. The reason for these differences could not be determined in this study.
This effect is also highlighted when comparing the summer IVT thresholds to the winter thresholds. For ERA-Interim there is typically a reduction of around 5% in the summer threshold compared to the winter threshold (with the exception of the 1979-2013 JJA value). The 20CR results, however, show an increase in the threshold value for summer compared to winter, differing by up to 10%. The dependency of the results on model resolution are similar to Hagos et al. [2015] who found a decrease in the frequency of AR events with model resolution. The causes of the differences observed here are outside the scope of the present study but merits further investigation. The threshold investigation has shown that for winter, by including the whole period available , there is a reduction in the threshold value of IVT. This is not seen for summer. As stated earlier, it is predicted that the variance of the IVT threshold decreases during this period, particularly after 1950. A study into whether this is the case would be an interesting extension to this result. The threshold investigation has also shown that the two different resolution data sets have similar threshold values of IVT for summer. For winter around a 10% difference is seen. It is worth remembering the results from Brönnimann et al. [2012] that the ensemble mean of 20CR (used here) does not represent extremes as well as individual members. Thus, the average IVT threshold may be artificially lowered as the most extreme IVT values are not captured,

ERA-Interim 20CR
Given an AR is there an extreme rain event? Figure 4. The percentage (to the nearest integer) of (left) extreme rainfall observations that have an AR associated with them and (right) the percentage of ARs that have an extreme rainfall observation associated with them for both ERA-Interim (black) and 20CR (grey) for both winter and summer, shortened and extended. compared to ERA-Interim. Any implications these results may have in determining the main cause for moisture availability over the UK in the summer is discussed in the next section. Lavers et al. [2012] linked AR events to winter flooding using a POT1 method on flow-gauge data in nine river basins to identify flooding events. The present study focuses on extreme rainfall as observed from rain gauges to capture events that might not be captured by flow-gauge data, e.g., overland flow. Not all of these events will lead to flooding nor will all flooding events be captured due to a number of events being over short time scales (hours) and very localized, such as summer flash floods caused by convective events. This study will also not look at events with less extreme rainfall intensities but which last for a number of consecutive days that may also lead to flooding. However, the use of rain gauges can provide useful insights into the causes of extreme rainfall in the UK and potentially identifying overland flow.

Relevance to UK Extreme Rainfall
The results are analyzed from two perspectives. The first is as follows: Given an extreme rainfall event, is there an associated AR? These results are shown in Figure 4 (left). For a flood event to be attributed to an AR the criteria used by Lavers et al. [2012] was to allow an AR to be present up to 3 days prior to the flood. In this study, due to the rain gauge response being almost instantaneous, the approach taken by Rutz et al. [2014] is used where the AR must be present on the same day or a maximum of 1 day previously, although this only reduces the number of associated events by up to 4%. The second perspective examines how many extreme rainfall events occur given there is an AR, answering the question as to whether an AR can be used as an indicator of an extreme rainfall event occurring; these results are also shown in Figure 4 (right). The location of the rain gauges that had events linked to ARs is shown in Figure 2 (bottom), for both winter (left) and summer (right). The locations are the same for both perspectives considered.
The results for the first perspective (given there is an extreme rainfall event, is there an associated AR?, Figure 4, left) show that for a short winter (DJF), 65% of the extreme rainfall events identified had an AR associated with them. These values are very similar to those found by Lavers et al. [2012], although that was for an extended winter and for a smaller sample of nine river basins. The results for an extended winter show a smaller percentage, around 50%. The results between ERA-Interim and 20CR are very similar, an interesting result given the longer period the 20CR data are examined over, 1900 to 2012. This is also similar if different periods of the 20CR data are chosen, e.g., 1979-2012. The results for summer show a very different result, with fewer than 20% of the extreme rainfall events identified, regardless of season length of data set, being associated with an AR. While a reduction in the connection was expected due to the more convective, and therefore localized, nature of extreme rainfall events in summer (hence, the further reduction in the short summer), this shows a very large reduction in the connection.

10.1002/2014JD022863
The thresholds for defining both an extreme rainfall event and an AR are both lower in the summer, although the relation is considerably weaker.
This suggests that ARs are not a good indicator of summer extreme rainfall possibly due to the relatively small scale summer convection compared to the large-scale nature of ARs. The sensitivity of the results to the length of time between the identified AR and an associated extreme rainfall event showed that for every day the period was extended by (up to 5 days tested), the number of extreme events that had an AR associated with it went up by 2-3%. This result was independent of the season, or the length of the season.
The second perspective taken in this study is as follows: Given an AR, is there an extreme rainfall event? (Figure 4, right). It is immediately clear that for summer, and winter, the percentages are much lower than the results for the prior perspective. These results suggest that while extreme rainfall events, in winter, can be associated with ARs, there are far more ARs (using the definition of Lavers et al. [2012]) than extreme rainfall events, where an extreme rainfall event is defined by being in the top decile of all rainfall events over the whole of the UK. Further results (not shown) indicate that it is also not necessarily the most extreme summer rainfall events that are identified by ARs, suggesting that there are other factors that are more important in determining the severity of summer rainfall events. The same sensitivity test to the period between an AR and an extreme event was performed and showed that for each day the period was extended and the number of events went up by only 1-2% (up to 5 days tested).
The location of the rain gauges which had events associated with ARs identified in ERA-Interim ( Figure 2 for both winter (bottom left) and summer (bottom right)) also highlight two conclusions that agree with previous studies. The first is that almost all of the rain gauges are located on the west coast of England, Scotland, and Wales with only two rain gauges being identified in the east of the UK. However, there is little variation between the north and the south of the UK. The second is that they are also predominantly associated with areas of high orography. This agrees with previous work on linking ARs to flooding in river basins in the northwest of Britain [Lavers et al., 2012] and the importance of orographic forcing on the enhancement of precipitation associated with landfalling ARs [Neiman et al., 2013].

Conclusions
This study has examined the link between extreme daily rainfall in the UK and intense moisture transport episodes associated with ARs during the winter and summer seasons. The study has produced similar results to the flow-gauge study of Lavers et al. [2012], demonstrating a connection between ARs and winter extreme rain gauge events, although the percentages are reduced for an extended winter. However, the results have shown that ARs cannot be used to explain summer extreme rainfall events, with less than 20% of summer extreme rainfall events having an AR associated with them. The results have also highlighted the importance of orography in the connection between ARs and extreme rainfall events. This study examined whether ARs could be defined as a precursor to extreme rainfall events. The results showed that many more ARs are identified than are associated with extreme rainfall events, with less than 10% of ARs in the short summer being associated with an extreme rainfall event.
While these results suggest that ARs are not particularly useful for the summer, there are a number of factors that ought to be taken into account. The first is the short time scale over which summer extreme rainfall events typically occur, which is hours rather than days. Therefore, the rain gauge data set used in this study, a daily data set, may not capture many of the extreme rainfall events, and it could be that many of the ARs identified do cause extreme rainfall, however, on much shorter time scales. The second factor is that due to quality control issues the number of rain gauges between 2002 and 2008 were significantly reduced, compared to surrounding years. Also, for the case of 20CR, the number of gauges before 1960 is significantly smaller.
The results also show that for both seasons (and shortened and extended), more of the extreme rainfall events were associated with ARs identified in 20CR than in ERA-Interim. The periods over which the relationship was tested were very different, 35 years (1979 to 2013) for ERA-Interim and 113 years (1900 to 2012) for 20CR. This suggests that the correlation may be affected by natural variability, i.e., the effect of wet seasons and dry seasons; however, this is not tested in this study. This result could also be strongly affected by the quality control issue of the rain gauge data set mentioned earlier.
The results for winter showed a reduction (5%) in the threshold value of IVT when a much longer period of 1900-2012 compared to a period of 1979-2005, or 1979-2012, was used in the 20CR data. This reduction was not seen for the summer. The difference between the periods 1979-2005and 1979(for ERA-Interim, 1979) was very small. The effect of the resolution of the data set on the threshold value of IVT was shown to be more pronounced in winter than in summer. In winter, the threshold values of IVT were around 10% lower in 20CR than in ERA-Interim. For summer, the threshold values of IVT were very similar. It was also seen that by using an extended winter caused an increase in the threshold value of IVT, whereas a reduction was seen for summer. The summer thresholds were also around 5% lower than in winter. Taking into account the representation of extremes in the 20CR ensemble mean [Brönnimann et al., 2012], this implies that different processes are operating depending on the season. It also highlights the potential variability in the threshold value of IVT depending on the season, and years included, in the study.
This preliminary study considered the connection between atmospheric rivers and ungridded daily rain gauge data observations. The limitations to this work include the use of daily rain gauge data to try to observe hourly rainfall events in the summer months; the choice of the threshold value to define an event as extreme and that ARs are defined at 4 ∘ west, and the rain gauges are taken over the whole country. However, the results show an agreement to Lavers et al. [2012] as to how many winter events are associated with an AR and also show that far fewer summer events are associated with an AR. The results also highlight that a lot more ARs are identified than extreme rainfall events. Future work will be to use an hourly rain gauge data set to identify more convective events in the summer months and to use alternative moisture identification methods, such as the ones used by Lavers and Villarini [2013a] and Rutz et al. [2014]. comments from all the reviewers in improving this paper. This research is part of the SINATRA project which is supported by the United Kingdom NERC Flooding From Intense Rainfall program (grant NE/K00896X/1). The ERA-Interim data set was made available from the European Center for Medium-Range Weather Forecast (ECMWF) via the ECMWF Data Server (http://www.ecmwf.int). The Twentieth Century Reanalysis V2 data are provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/. Support for the Twentieth Century Reanalysis Project data set is provided by the U.S. Department of Energy, Office of Science Innovative and Novel Computational Impact on Theory and Experiment (DOE INCITE) program, and Office of Biological and Environmental Research (BER), and by the National Oceanic and Atmospheric Administration Climate Program Office. The rain gauge data set was downloaded from the NERC Centre for Environmental Data Archival (http://www.ceda.ac.uk).