Earlier awareness of extreme winter precipitation across the western Iberian Peninsula

Extreme winter precipitation across the western Iberian Peninsula is frequently connected to atmospheric rivers, which are bands of intense water vapour transport (integrated vapour transport, IVT). Recent research over Western Europe has shown that by using the IVT forecasts in the form of the extreme forecast index (EFI), a tool that identifies anomalous conditions, it is possible to provide earlier awareness of extreme precipitation events than by using the precipitation EFI directly. The aim of the present paper is to assess these findings further by identifying the regions of Iberia where such IVT forecasts are skilful and for what lead times. Employing the EFI on the European Centre for Medium‐Range Weather Forecasts (ECMWF) Integrated Forecasting System ensemble for winters 2015/2016 and 2016/2017 and using high‐density daily surface precipitation observations, the IVT EFI is shown to have slightly more skill (than the precipitation EFI) in discriminating extreme precipitation anomalies across the western Iberian Peninsula (Portugal and northwestern Spain) from forecast day 11 onwards. The reasoning is that the higher IVT predictability means that the IVT EFI is more able to detect the approximate location of extreme events at earlier lead times than the precipitation EFI. In contrast, for shorter forecast lead times, the precipitation EFI is more skilful. Also, when considering the entire Iberian Peninsula, the precipitation EFI is more skilful and more appropriate to monitor for potential extremes even at longer lead times.

Extreme winter precipitation across the western Iberian Peninsula is frequently connected to atmospheric rivers, which are bands of intense water vapour transport (integrated vapour transport, IVT). Recent research over Western Europe has shown that by using the IVT forecasts in the form of the extreme forecast index (EFI), a tool that identifies anomalous conditions, it is possible to provide earlier awareness of extreme precipitation events than by using the precipitation EFI directly. The aim of the present paper is to assess these findings further by identifying the regions of Iberia where such IVT forecasts are skilful and for what lead times. Employing the EFI on the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System ensemble for winters 2015/2016 and 2016/2017 and using high-density daily surface precipitation observations, the IVT EFI is shown to have slightly more skill (than the precipitation EFI) in discriminating extreme precipitation anomalies across the western Iberian Peninsula (Portugal and northwestern Spain) from forecast day 11 onwards. The reasoning is that the higher IVT predictability means that the IVT EFI is more able to detect the approximate location of extreme events at earlier lead times than the precipitation EFI. In contrast, for shorter forecast lead times, the precipitation EFI is more skilful. Also, when considering the entire Iberian Peninsula, the precipitation EFI is more skilful and more appropriate to monitor for potential extremes even at longer lead times.

K E Y W O R D S
atmospheric rivers, extreme forecast index (EFI), extreme precipitation, Iberian Peninsula, water vapour transport The European Centre for Medium-Range Weather Forecasts (ECMWF) uses the extreme forecast index (EFI) (Lalaurette, 2003;Zsoter, 2006;Zsoter et al., 2014) to identify regions that experience anomalous weather conditions in a forecast. It is calculated by comparing the ECMWF forecasts and the model climate, and this method is applied to many variables, for example, 2-m temperature, precipitation, wind speed and convective parameters. Recent research has investigated the possibility of also employing the EFI on vertically integrated horizontal water vapour transport (integrated vapour transport, IVT) (Lavers et al., 2016;Lavers et al., 2017). The IVT, particularly in the form of an atmospheric river within an extra-tropical cyclone, is a key contributor to extreme winter precipitation and flooding in western North America (e.g. Ralph et al., 2006), the United Kingdom (Lavers et al., 2011) and the Iberian Peninsula (Ramos et al., 2015). Results have shown that the IVT EFI can provide earlier awareness of extreme precipitation (compared with using the precipitation EFI) in Western Europe (Lavers et al., 2016(Lavers et al., , 2017 and western North America (Lavers et al., 2017) from about forecast day 10 onwards. It is the large-scale nature and, thus, higher predictability of the IVT field that leads to the IVT being more skilful at these lead times.
One region of Western Europe with a strong link between the IVT and extreme precipitation is the Iberian Peninsula, and this has been shown by Ramos et al. (2015) for the large international river basins located in western Iberia (i.e. the Minho, Tagus and Duero river basins). In previous research using the IVT EFI (Lavers et al., 2017), it was hypothesized that the skill obtained by using the IVT EFI could be increased by only considering an area with a significant link between the IVT and precipitation (compared with using a pan-European domain). As these areas with the strongest relationship between the IVT and precipitation are smaller in size, the availability of a high-resolution data set is essential to account for the reduced sample. The aim of this study is to use highdensity gauged precipitation observations across the Iberian Peninsula to evaluate the forecast skill of the IVT and precipitation EFI in capturing extreme precipitation. This will show whether the IVT EFI skill can be raised by only considering a region where the IVT is a main driver of extreme events.

| DATA
This section gives a brief overview of the data sets and methods, as further details can be found in Lavers et al. (2017). The 51 ensemble members (with forecasts out to 15 days) from the ECMWF Integrated Forecasting System (IFS) ensemble (ENS) were retrieved for the 0000 UTC initialization for two winter seasons, November 1, 2015, to February 29, 2016andOctober 20, 2016, to February 28, 2017 (253 forecasts). The dates differed in the two seasons because the experimental model version was used from October 20 to November 21, 2016 (this became the operational model on November 22, 2016), which increased the sample size and provided a constant model version for winter 2016/2017. Daily total surface precipitation was retrieved, and the magnitude of the water vapour transport (IVT) was calculated (e.g. Neiman et al., 2008) using the specific humidity and zonal and meridional winds at 300, 400, 500, 700, 850, 925 and 1,000 hPa.
The ECMWF also produces re-forecasts with the same horizontal resolution as the ENS to generate a model climate; this climate can then be compared with the ENS. The re-forecasts have 11 ensemble members and have been run twice a week for the past 20 years. As with the ENS, the daily total surface precipitation was retrieved and the IVT was calculated. The model climate was constructed using nine re-forecast dates centred on the forecast of interest, which provides a climate of 1980 members (9 dates × 20 years × 11 members).
Daily gauged precipitation data for Spain were obtained from the National Spanish Meteorological Service (AEMET) archives for the 2015/2016 and 2016/2017 winters (Vicente-Serrano et al., 2017). These data correspond to 24 hr precipitation totals measured at 0700 UTC. For Portugal, the daily precipitation data were obtained from the Portuguese Meteorological and Oceanic Institute (IPMA) for the same two winters and the daily precipitation was accumulated between 0000 UTC and 2359 UTC. Note that this analysis did not change the Spanish observations to a 0000 UTC-2359 UTC period; the day with the longest time coverage was considered as the day when the precipitation fell. The average number of gauges available across the Iberian Peninsula on each day was 2,146. Extreme observed precipitation events were identified as exceedences over the 99th percentile of the model climate precipitation at the closest grid point to the gauge (as in Lavers et al., 2017). This approach guards against the loss of stations due to them not having enough data to generate a gauge climatology. The model climate precipitation used was taken from forecast day 7 from the model climate nearest the middle of the month.

| METHODS
Using the ECMWF forecasts and model climate, the EFI (Lalaurette, 2003;Zsoter, 2006;Dutra et al., 2013;Zsoter et al., 2014) was used to determine how extreme a forecast was by comparing the probability distribution of the ENS and the model climate.
The EFI was calculated as: where F(p) is the proportion of ensemble members that lie below the pth percentile of the model climate. The EFIs ranged from -1 to 1, where -1 are extremely low and 1 are extremely high values compared with the model climate.
The EFI was calculated for the November 1, 2015, to February 29, 2016 and for the October 20, 2016, to February 28, 2017 (253 dates) forecasts for each forecast day (1-15) for the IVT and precipitation. The EFI forecasts' ability to discriminate extreme precipitation (> 99th percentile) across the Iberian Peninsula was assessed using the hit and false-alarm rates, which are defined as the fraction of forecast hits with respect to the total observed events, and as the number of false alarms to total non-events respectively. These were investigated using relative operating characteristic (ROC) curves for the EFI thresholds ranging from 0 to 1 (Wilks, 2006). The ROC area or score was also calculated; it ranged from 0 to 1, with areas > 0.5 representing a skilful forecast.
A bootstrap procedure was employed to assess the statistical significance of the differences in ROC areas between the IVT and precipitation. This process resampled with replacement the 253 forecasts (for a particular forecast day) and calculated the ROC area, and then repeated the process 1,000 times (as in Lavers et al., 2017).

| EFI verification
This section concentrates on two sub-regions of the Iberian Peninsula to determine how the verification varies with region, thus showing the importance of regional aspects. These areas are shown by the boxes in Figure 1: (1) a subregion of the Iberian Peninsula from the Atlantic coastline to 7 W (hereafter referred to as "7 W"), an area known to be affected by atmospheric rivers (Lavers and Villarini, 2013;Ramos et al., 2015); and (2) the Iberian Peninsula from the Atlantic coastline to 1 W (hereafter referred to as "1 W"), which includes the sub-region to 7 W. Figure 1 also displays the recording gauges (black triangles) and gauges where the precipitation exceeded the 99th model climate percentile (magenta triangles) on February 12, 2016 (this event is discussed further in Section 4.2). The ROC areas and bootstrap distributions for each forecast day calculated over the two winters studied for the IVT and precipitation EFI for the 7 and 1 W regions are shown in Figure 2a, b respectively. It is evident that the precipitation EFI is more useful (i.e. has more discriminatory power) up to about forecast day 10 for 7 W and right through the medium-range forecast horizon for 1 W. From forecast day 11, the results show that more often than not there is a slight advantage in using the IVT EFI (compared with the precipitation EFI) to discriminate extreme events in the western part of the Iberia Peninsula. However, there is overlap in the bootstrapped distributions of the ROC area differences (precipitation-IVT) from forecast day 11 (Figure 2c), which suggests some level of equality between the IVT and precipitation EFI after forecast day 11 (except on forecast day 14). In general, this confirms previous findings for Western Europe and shows larger ROC area differences (between the IVT and precipitation), and, hence, skill than those shown for Western Europe in Lavers et al. (2017). Therefore, the IVT EFI has potential usefulness for early awareness of extreme precipitation in this particular subregion of Europe. For the 1 W domain (Figure 2d), the IVT EFI generally shows little added usefulness over the precipitation EFI.
The ROC curves on forecast day 12 are given for the two study regions in Figure 2e,f. At this lead time, the slightly larger ROC area for the IVT EFI for the 7 W region indicates that the IVT EFI is more skilful in capturing extreme precipitation events than the precipitation EFI. It is thought that the large-scale characteristics of the IVT, which  (e, f ) The ROC curves for the IVT (black lines) and precipitation (grey lines) for all 253 forecasts on forecast day 12 for 7 and 1 W respectively. The ROC areas are provided in the legends and the number of extreme precipitation events is also given results in higher predictability, means it is more likely to detect the location of extreme events. In contrast, the precipitation is linked to more local-scale processes, such as those associated with the land surface topography, and is therefore less able to discriminate extremes at these longer lead times. For the 1 W domain the precipitation EFI is more skilful as seen by the precipitation ROC curve generally being closer to the top left corner of the plot. This suggests that when considering all gauges up to 1 W, there are processes other than the IVT that are important to the generation of extreme precipitation. Note that other domains were considered across the Iberian Peninsula, and as the domain was extended to the east, the advantage of using the IVT EFI was gradually reduced. This is related to the fact that western Iberian precipitation is more closely associated with synoptic-scale low-pressure systems that come from the North Atlantic than the eastern sector (Cortesi et al., 2013;Cortesi et al., 2014) and, in particular, the impact of atmospheric rivers in extreme precipitation is very small in Spanish provinces close to the Mediterranean (Lavers and Villarini, 2013;Ramos et al., 2015).

| Example of the EFI over the Iberian Peninsula on February 12, 2016
On February 12, 2016, an atmospheric river affected the Iberian Peninsula resulting in extreme precipitation and flooding. A large number of stations recorded precipitation in excess of the 99th percentile of the model climate (their location is depicted by lighter/magenta triangles in Figure 1). The EFI and mean sea level pressure (MSLP) ensemble mean forecasts for this event are shown on  forecast day 12 (Figure 3a,b), forecast day 7 (Figure 3c,d), and forecast day 1 (Figure 3e,f ). Visual inspection of Figure 3a,b indicates that the IVT EFI (mean EFI = 0.29; maximum = 0.35) has a stronger and, hence, earlier signal across the region for this event compared with the precipitation EFI (mean EFI = 0.15; maximum = 0.32). The IVT EFI also captures the large-scale flow across the North Atlantic (as shown by the IVT EFI pattern around the ensemble mean MSLP) due to its relationship with larger synoptic-scale processes. By forecast day 7, the signature of an atmospheric river is striking (Figure 3c), and although the signal is clearer than that shown for the precipitation EFI (Figure 3d), this comes at the expense of the IVT EFI having more false alarms over the Mediterranean. In turn, this means that the interpretation of the IVT EFI to forewarn of extreme precipitation requires regional knowledge, such as about the local topography that is known to impose a very distinct contribution, during intense IVT events, to precipitation totals in the western and eastern Iberian river basins (Ramos et al., 2015). This corroborates the results presented in Figure 2a,b, which showed that on forecast day 7 the precipitation EFI is more able to discriminate which gauges will experience extreme precipitation. Similar findings occur on forecast day 1 when the large-scale IVT field not only captures the extreme events but also many surrounding gauges, thus resulting in a higher false-alarm rate and less discriminating power than the precipitation EFI. This is the reason behind the IVT EFI being less skilful up to about forecast day 10, as presented in Figure 2a.

| CONCLUSIONS
The aim of this study was to assess whether the integrated vapour transport (IVT) extreme forecast index (EFI) shows more usefulness (compared with the precipitation EFI) in discriminating extreme precipitation events over the Iberian Peninsula, a region with a strong connection between IVT and extreme precipitation. Using the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble forecasts from two winters and high-density gauged precipitation observations across Spain and Portugal, the results corroborate the findings of Lavers et al. (2017) and show that the IVT EFI is slightly more skilful in detecting extreme precipitation in Portugal and northwestern Spain (7 W region) from forecast days 11-15. By focusing on a region of Western Europe with a strong association between the IVT and precipitation, the difference in the relative operating characteristic (ROC) score (between the IVT and precipitation) was increased. At earlier lead times, however, the precipitation EFI is more skilful at identifying extreme events because the larger-scale IVT EFI fields tend to cover a broader area resulting in a higher false-alarm rate. This implies that knowledge of local topography, for example, is required for the IVT EFI to provide earlier awareness of extreme precipitation events. When considering a broader region of the Iberian Peninsula (1 W domain), the precipitation EFI was generally more skilful than the IVT EFI throughout the 15-day forecast horizon, suggesting that the IVT has less influence on extreme precipitation events farther east in Spain. These results are in agreement with previous findings that atmospheric rivers do not explain extreme precipitation over eastern regions of the Iberian Peninsula, as shown for the Ebro River basin by Ramos et al. (2015). The findings herein add to the body of literature showing the potential usefulness of using the IVT EFI in the mid-latitudes during the late medium-range forecast horizon, and the most appropriate approach for implementing it operationally is being discussed.