Floods in the Alpine region can be destructive and cause large economic losses. Many rivers and lakes in Switzerland are regulated and flood damage can be mitigated through an optimal management of lake levels and runoff. This requires high-quality forecasts of atmospheric flood precursors extending beyond short-range (forecast days 1–5) predictions. In several places around the world atmospheric rivers or extreme integrated vapor transport (IVT) are causally related to flood events. Also in Switzerland, extreme IVT oriented perpendicular to the main orography heralds extreme flood events. This relationship is exploited in an operational flood warning system on the medium-range (here forecast days 6–10) time scale based on probabilistic medium-range forecasts of IVT and precipitation over Switzerland provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS). This entails first a comprehensive probabilistic verification of the direction and magnitude of (extreme) IVT and second the development of compact visualizations for the operational use by hydrologists. Based on 20 years of probabilistic reforecasts, we show that both regular and extreme IVT has a better predictability than precipitation and IVT is predictable out to day 8. As the direction of IVT is of central importance for flood risk in Switzerland, we develop a visualization that summarizes probabilistic information on both the direction and magnitude of the IVT together with users of the product. The result is an operational flood warning system based solely on atmospheric flood precursors to extend flood warning information beyond the range of high-resolution deterministic weather forecasts.
Floods in Switzerland are a natural hazard causing substantial economic losses (Hegg et al. 2000) and worldwide (e.g., Munich Re).1 Preparedness and preventive measures are key to limit these losses. Weather and hydrological forecasts provide important information to support decisions in this regard (e.g., Moore et al. 2005; Roberts et al. 2009) and to implement short-term preventive measures. Short-term preventive measures include the management of lake levels and runoff in regulated catchments (e.g., Romang et al. 2011; Balbi et al. 2016), person and building content evacuation (Hubbard et al. 2014), mobile flood barriers, flood gates, or simple sand bags. To implement mobile flood barriers civil protection authorities need to mobilize and prepare for floods (Pappenberger et al. 2015).
In Switzerland numerical discharge forecasts are produced by the Federal Office for the Environment using the WASIM hydrological model (http://www.wasim.ch/de/) driven by high-resolution (2 km) numerical weather forecasts from the COSMO-E model (Klasa et al. 2018) and are currently only available out to day five. However, flood warning systems providing information about potential flood conditions beyond this lead time may provide useful information to improve flood preparedness (Pappenberger et al. 2015). The existing literature gives hope that useful and potentially predictable flood precursors exist for Switzerland and other regions globally with strong topographical forcing, which can be used as a basis to build an early flood alert system.
In Switzerland, a relationship exists between extreme vertically integrated water vapor transport [IVT; see e.g., Lavers et al. (2012) for the computation of IVT] and major flood events (Martius et al. 2006; Froidevaux and Martius 2016, hereafter FM16). Exceptionally high IVT directed perpendicularly to the main mountain chains was present for 10 out of 14 major flood events in Switzerland in the past 40 years (FM16). FM16 distinguish three regions in Switzerland characterized by three different flow regimes conducive to floods. In northeastern Switzerland high IVT air reaching the Alps from the northeast driven by upper-level cutoff lows over eastern Europe was associated with major flood events. In southern Switzerland moist air reaching the Alps from the south driven by upper-level troughs located over western Europe was associated with major flood events (see also Massacand et al. 1998). In northwestern Switzerland high IVT air reaching Switzerland from the northwest along the eastern flank of a strong ridge over western Europe was associated with major flood events. FM16 identified region-specific thresholds of direction-dependent IVT that are related to the occurrence of major floods in Switzerland. In essence, high IVT episodes directed perpendicular to the Alps can serve as indicators of flood potential on a regional scale. Indeed, the strong link between very high direction-dependent IVT and floods in Switzerland provides a surprisingly simple atmospheric regional flood potential predictor that allows circumventing the complex process chains involved in flood generation (e.g., spatiotemporal precipitation characteristics, snowmelt, soil moisture). However, IVT cannot be used to infer information on the exact location or timing of potential floods.
European-wide there is also evidence of the relationship between extreme winter precipitation/floods and high IVT values (Lavers et al. 2011; Lavers et al. 2012; Lavers and Villarini 2013) occurring in so-called atmospheric rivers (ARs) associated with extratropical cyclones (Ralph and Dettinger 2011). ARs are coherent long and narrow structures of high IVT that typically occur ahead of cold fronts of extratropical cyclones (e.g., Neiman et al. 2008; Cordeira et al. 2013; Ralph et al. 2017; American Meteorological Society 2019). Hence, while ARs are associated with high IVT, high IVT values can also occur in the absence of ARs. For example, coherent high IVT structures on the Swiss Alpine south side do often not classify as ARs because they are too short. There is an increasing body of literature expressing the emergence of ARs as a science and application focus (e.g., Ralph et al. 2017). For example, Lavers et al. (2011) identified a connection between ARs and winter flood events in Britain. Furthermore, Lavers et al. (2014) illustrated that IVT has a higher potential predictability in Europe than precipitation. This is mainly due to the fact that ARs are connected to large-scale atmospheric variability, which is better predicted than the subsynoptic variability which affects precipitation. Lavers et al. (2016) further find increased skill in extreme IVT than extreme precipitation. When applying the European Centre for Medium-Range Weather Forecasts (ECMWF) extreme forecast index (EFI) on IVT and precipitation, the predictability of IVT is higher in week two of the forecast. This is also the case for North America (Lavers et al. 2017). Earlier awareness of extreme precipitation over the Iberian Peninsula is possible based on the EFI of IVT compared to the EFI of precipitation (Lavers et al. 2018).
The strong link between direction-dependent extreme IVT and floods in Switzerland and the enhanced predictability of IVT provide the motivation for our study. The ultimate aim of this study is the realization of an automated flood awareness alert system with a focus on the time window from forecast day 6 to forecast day 10 based on medium-range forecasts of IVT by the ECMWF, thereby linking IVT directly to the flood potential.
We illustrate the usefulness of such an alert system in two steps: first, we present a general verification of probabilistic forecasts of IVT magnitude, precipitation, and IVT direction. Unfortunately, the number of major flooding events in the dataset is too limited for quantitative evaluation of direction-dependent IVT to forecast flood potential. Instead we rely on the well-established link between direction-dependent IVT and flooding (FM16) and the verification of IVT and precipitation forecasts. To further document the usefulness of IVT forecasts for early warning of potential flood events, we complement the general verification with an illustration of two flooding events related with extreme IVT. Second, we develop visualizations of the probabilistic forecasts to summarize magnitude and direction of IVT with a focus on conditions conducive to flooding events in close collaboration with an end user (hydrological forecasting service in Switzerland).
2. Data and methods
We use reforecast data from ECMWF’s operational ensemble forecasting system for the forecast verification. These reforecasts are produced in parallel to the real time extended range forecasts, which are initialized every Monday and Thursday at 0000 UTC and extend to 46 days lead time. The reforecasts start on the same day and month for each of the past 20 years. In operation reforecasts are generated two weeks ahead of the corresponding forecast and consist of a smaller ensemble (11 members) than the real time forecast (51 members). Specifically, we used the reforecasts of the ECMWF’s Integrated Forecasting System (IFS) produced between 8 March 2016 and 7 March 2017, resulting in a dataset including all four seasons and the years 1997–2016. As the operational IFS version changed on 22 November 2016 (from Cycle41r2 to Cycle 43r1), our dataset also includes this model version change. It included an upgrade of the ocean resolution and changes in the data assimilation procedure. (The full documentation of the model updates can be found at https://www.ecmwf.int/en/forecasts/documentation-and-support/changes-ecmwf-model.) We expect these updates to have a minor influence on our analysis as the resolution of the atmospheric model did not change.
We analyze the instantaneous values of the reforecasts every 24 h out to day 10. The model output (both IVT and precipitation) is compared to the reanalysis dataset ERA-Interim (Dee et al. 2011) from 1997 to 2016. Both datasets are interpolated to a common 1° × 1° grid. IVT is then calculated following Lavers et al. (2012):
(kg m−1 s−1), where g is the acceleration due to gravity, q is the specific humidity, u and υ are the zonal and meridional wind components, and p is the pressure. Both IVT from the reforecasts and ERA-Interim are calculated from 10 pressure levels at 1000, 925, 850, 700, 500, 400, 300, 200, 100, and 50 hPa.
The analysis of IVT and precipitation is carried out for the three regions in Switzerland defined by FM16 and characterized by three representative grid points (Fig. 1). The grid points are defined as follows:
East (E) at 48°N, 9°E: This grid point is located to the north of the eastern Swiss Alps and is representative to capture moist airflow reaching the Alps from the northeast.
Jura (J) at 47°N, 7°E: This grid point is located north of the western Swiss Alps and ideal to capture moist air reaching the Alps from the north.
Tessin (T) in the south of Switzerland at 46°N, 9°E: This grid point is located south of the Alps and representative to capture southerly moist flow reaching the Alps.
1) Extreme forecast index
To analyze the capability of the forecasting system to predict extreme IVT and precipitation episodes we make use of the extreme forecast index. The EFI is an indicator of how extreme an ensemble forecast is compared to the model climate (i.e., the reforecasts) (Lalaurette 2003; Zsoter 2006). It is defined as follows:
where F(p) denotes the proportion of ensemble members lying below the p quantile of the model climate. The EFI of a reforecast from a given year is computed by comparing its 11-member ensemble to the corresponding model climate derived from the remaining 19 years of reforecasts. As IVT and precipitation have seasonal cycles, the reforecasts of each season are split into two for sampling the model climate [i.e., all reforecasts within a 6-week window (13 initialization dates × 19 years × 11 members = 2717 reforecasts)] are used for computing the quantiles of the model climate.
The EFI values range from −1 to 1. Generally, values from 0.5 to 0.8 irrespective of the sign indicate unusual weather and values larger than 0.8 indicate an extreme weather situation. Here we are interested in high IVT and precipitation values and therefore focus only on positive values of EFI. Note also that the thresholds for extremes depend on the lead time and we will discuss this point in more detail in section 3c. Since the EFI represents a single-valued ensemble summary it can be evaluated using the ROC as suggested by Lavers et al. (2016).
2) Probabilistic verification scores
We evaluate ECMWF IFS’s ensemble forecasting ability to predict the magnitude of IVT and precipitation at each of the three grid cells representing the three regions using four commonly used measures of forecast quality for probabilistic predictions (see Table 1). The continuous ranked probability score (CRPS) assesses both forecast calibration and sharpness. That is, if two competing forecasts are well calibrated or reliable (both forecast mean and forecast spread are free of errors), the sharper forecast results in better (lower) CRPS scores. The constant climatological reference forecast based on the verifying observation for example is well calibrated by design, but exhibits minimal sharpness. The CRPS does not focus on rare events in particular, but assesses the average accuracy of all forecasts.
To focus on relatively rare events, we use the Brier score (BS; Table 1). The BS measures the accuracy of binary probability forecasts (i.e., yes–no forecasts). Here we define the event to be assessed as IVT and precipitation exceeding the 95th percentile of the forecast and the reanalysis climatology, respectively. The 95th percentile captures moderate extremes; if we were to focus solely on very extreme values the verification sample would become too small and the results would become difficult to interpret. In addition, we also use the weighted CRPS (Lerch et al. 2017) to assess accuracy of forecasts of rare events. The weighted CRPS uses all data points but allows to specify a weight function to base the evaluation on the performance in the tail of the distribution of events. Here we use the weighted CRPS to solely focus on forecasts and reanalysis values exceeding the 95th percentile of the forecast and reanalysis climatological distribution, respectively. Both the BS and the weighted CRPS give an indication of how well the forecasting model is able to predict events at the upper end of the distribution. The BS indicates how well the occurrence of such events is forecast, and the weighted CRPS also indicates how well the intensity of rare events is forecast.
We express the above scores as skill scores with respect to an uninformed reference forecast and refer to these as the continuous ranked probability skill score (CRPSS) and Brier skill score (BSS) from here on. Positive values of the skill scores indicate that the forecast outperforms the reference, zero indicates that the forecast is as uninformative as the reference. Skill scores are typically calculated using either climatology or persistence as the reference forecast. Here we show skill scores against climatology only. After the first few days, climatology outperforms persistence and therefore a constant climatological forecast provides a more rigorous benchmark for the time scale of interest. The probabilistic climatological forecast consists of all ERA-Interim reanalysis of the same calendar day within the 20-yr reforecast period, that is, an ensemble forecast with 20 members.
In addition, two discrimination scores are used to understand how well the model distinguishes events from nonevents. Again we chose one to analyze the whole distribution, the generalized discrimination score [ “two alternatives forced-choice score” (2AFC)], and one only for the upper end of the distribution, the area under the receiver operating characteristics curve (ROCA) for binary forecasts of events exceeding the 95th percentile. Both the 2AFC and ROCA take values in the interval from 0 to 1, where 1 corresponds to a forecast with perfect discrimination and 0.5 corresponds to a forecast with no discrimination. To render these discrimination scores directly comparable with the skill scores defined above, we map the scores to the interval from −1 to 1 and refer to the skill scores as 2AFC skill score (2AFCSS) and ROCA skill score [ROCASS; see Wilks’s (2011) Eq. (8.46) or Kunii et al. 2011]. The 1 indicates a forecast with perfect discrimination, 0 indicates a forecast without discrimination, and −1 denotes the pathological case of a forecast with perfect discrimination but always giving the wrong response.
All analyses are carried out for the four seasons separately. The results are illustrated primarily for the summer season [June–August (JJA)] throughout the paper as this is a flood-prone season in Switzerland and generally the results are very similar for all four seasons. The case studies illustrate a summer and a fall flood case. Substantial deviations from the JJA results for the other seasons or regions are discussed in the text, otherwise the results are valid for all four seasons.
a. Forecast verification
1) IVT magnitude and precipitation
The verification scores for IVT as well as for daily precipitation at the three grid points indicated in Fig. 1 for JJA are shown in Fig. 2. Please note that only the amplitude of IVT was considered and not the direction of the flux, which is addressed below. Also, the verification of both IVT and precipitation is carried out on a relatively coarse spatial grid resolution of 1°. This resolution is well suited to capture the regional to synoptic-scale high IVT structures. Precipitation, however, typically exhibits smaller-scale structures (e.g., those associated with the local topography). The verification is thus limited to larger- (regional-) scale precipitation patterns. In addition, the precipitation that triggered floods in the past was typically located closer to the Alps than the location of the three “IVT grid points.”
We present a range of verification measures for IVT and precipitation forecasts for the three regions. These verification measures address different aspects of forecast quality. The CRPS (top row in Fig. 2) assesses the accuracy of IVT and precipitation forecasts. It is sensitive to forecast calibration and can thus be used to detect systematic biases. The CRPS as used here, however, is not specific for rare events, but rather characterizes the overall quality of precipitation and IVT forecasts. In contrast, the BSS (second row in Fig. 2) also measures forecast accuracy but focuses on rare events (above the 95th percentile) and is less sensitive to biases since the percentile values are derived from the forecast and the reanalysis climatology. The 2AFC and ROCASS on the other hand assess forecast discrimination in general (2AFC) and for forecasts of rare events (ROCASS). Forecast discrimination is insensitive to systematic biases and thereby provides an indication of potential predictability.
Generally, the IVT forecasts are slightly more skillful than the precipitation forecasts. The CRPS indicates that IVT magnitude can be skillfully forecasted at least to day 8 (day 6 in Tessin), with potential predictability as indicated the 2AFC likely extending beyond the forecast horizon analyzed here. Furthermore, also the upper end of the distribution can be forecasted accurately for at least 5 days with potential predictability extending at least to day 10 (Fig. 2, second and last row), which illustrates the potential to use IVT and/or precipitation forecasts for early warning of conditions conducive to flooding.
For the Tessin grid point the scores are generally slightly lower than for the other two grid points. Daily precipitation forecasts are also skillful, although in most cases the skill is slightly lower than for IVT. The negative CRPSS for precipitation in the Tessin is due to considerable systematic biases as the data is not debiased. The results for the other seasons do not differ in terms of skillful lead times from the JJA results shown in Fig. 2 (not shown). Hence, overall the skill of the forecasts is high, for accuracy as well as discrimination, only for the BS and CRPSS of daily precipitation the skill scores are rather low.
The forecast skill for both precipitation and IVT are evaluated at the grid points indicated in Fig. 1. When we compare forecasts of IVT with forecasts of precipitation at the same grid point where the IVT is analyzed, IVT is a less skillful predictand for heavy precipitation than precipitation itself (not shown). However, this comparison is relevant for the flood forecasting application, because heavy precipitation associated with high IVT often occurs downstream of the analysis grid points, which were on purpose selected to be located upstream of the orography and hence at some distance from the heavy precipitation that typically occurs along the orography.
2) IVT direction
In a next step, the representation of the direction of the IVT is assessed as it plays a crucial role for extreme events. As shown by FM16, IVT directed perpendicular to the mountains (Jura and Alps) creates the largest impacts in terms of precipitation and flooding and it is thus a critical aspect to incorporate. IVT thresholds indicating flooding potential including information on the relevant directional sector identified by FM16 are summarized in Table 2. Note that the thresholds as well as the flux direction sectors should be used as a guidance rather than absolute thresholds as they are based on a few cases only.
We analyzed the difference in the forecasted direction and the observed direction for IVT values exceeding the local 95th percentile. Only results for days 5–10 are shown in Fig. 3 as differences for the days before are very small. For lead times longer than 6 days the forecasted IVT direction may be off considerably. However, for a lead time of 7 days, still about 80% (50%) of all forecasts are within ±30° (±15°) of the analysis direction (see Fig. 3 for the east point). This fraction decreases to about 60% (30%) for a lead time of 10 days (third row for the example shown in Fig. 3). Hence, especially for the lead times of greatest interest (days 6–10), the forecasted IVT direction needs to be treated carefully and we will come back to this point in section 3c (IVT plots).
To analyze IVT magnitude and direction skill we proceed as follows: Depending on the direction of the IVT, the IVT is weighted. Within the sectors defined in Table 2, IVT magnitude is used directly without any weighting. Outside the sectors, the projection of IVT on the closest sector boundary is used, thus IVT further than 90° from the sector is set to zero. The CRPSS of this weighted IVT is shown in Fig. 4. For the Tessin the two sectors (see Table 2) are combined. For this direction-dependent measure skill decreases quickly over the first few days of lead time (Fig. 4). For the time window of interest, 6–10 days, there is still some skill, but it is rather low. Interestingly the skill for the Tessin is comparable to the skill at the two grid points on the north side of the Alps.
3) Verification of extreme IVT
As described earlier, extreme IVT values are needed to trigger a flooding event (FM16). By definition, the extreme events of interest are very rare. Verification of such events in a probabilistic setting has posed great difficulties in the past (Ferro and Stephenson 2011). Here, we take two different approaches to tackle this issue. The first approach is using the weighted CRPSS (Lerch et al. 2017, see section 2b for details). Figure 5 illustrates the outcome of this method for JJA. Based on this analysis, the magnitude of such extreme events can skillfully be forecast up to about six days ahead. Please note that the weighted CRPS evaluates not only threshold exceedances such as the ROCASS presented in Fig. 2, but also accounts for the errors in magnitude of forecasts of rare events. The results illustrate that while threshold exceedances may be skillfully forecast until day 10 (and probably beyond, see Fig. 2), the magnitude of IVT forecasts at longer lead times should not be over interpreted as these are not skillful beyond day 6 (Fig. 5).
The second verification approach is based on the EFI. The EFI is a good indicator of how extreme a specific ensemble forecast is and as a scalar ensemble summary the EFI can easily be visualized. It compares the forecast distribution with the distribution of the model climatology and is thus largely insensitive to model biases. As the EFI has no observational counterpart, direct verification is not possible. Its predictive performance can, however, be verified using scores for probability forecasts of threshold exceedances such as the ROC score (section 2).
The ROC curves in Fig. 6 describe how skillfully the EFI captures extreme IVT events above the 95th percentile. Hence, we compared the EFI value and whether at this value an extreme event occurred. According to the ROC analysis the EFI is skillful up to 10 days lead time, this is in agreement with the ROCASS for the 95th percentile of IVT shown in Fig. 2 (bottom row, left panel). The weighted CRPSS indicates predictability only to shorter lead times. The weighted CRPSS is stricter compared to the ROC curves as it uses here a percentile of the reanalysis climatology and is thus sensitive to model biases, while the EFI is based on a percentile of the reforecast climatology. It appears that occurrence of extreme IVT (as measured by the ROC of EFI and the BSS for the 95th percentile) can be detected quite early, but signals on the intensity (as indicated by the weighted CRPSS) can only be trusted at shorter lead times. The decreasing maximum EFI values in Fig. 6 suggest that with increasing lead time, lower EFI values can skillfully indicate an extreme event. We will discuss this in more detail below.
The maximum values of the EFI decrease with increasing lead time. The EFI rarely exceeds 0.7 in the 10-day forecasts (Fig. 6). This is due to the increasing spread of the probabilistic forecast with increasing lead time. Thus at long lead times, EFI values already lower than 0.8 can be associated with extremes. This phenomenon is discussed in detail in Pantillon et al. (2017). They analyze the time dependence of EFI and present a methodology to determine the optimal EFI threshold for each lead time. By optimizing the hit rate, the false alarm rate, and the Heidke skill score (HSS), the optimal EFI by lead time can be determined. In essence, this approach boils down to the question of whether you tolerate a higher false alarm rate to avoid misses, or whether you tolerate some misses to reduce the false alarm rate. For this step the users of our product were asked to give some indication of their tolerance. In the discussion it was decided that until some experience with the product could be gained an upper and lower bound of the EFI will be used as a guidance. The upper bound minimizes the false alarm rate (FAR), while the lower bound maximizes the probability of detection (POD). In both cases, it is the intention to maximize the HSS.
Figure 7 displays the upper and lower bound of the EFI values for lead times from 0 to 10 days. The red area in Fig. 7 depicts the range of minimum EFI values that indicate increased chances of an extreme event, defined as IVT exceeding the 99th percentile, as a function of lead time. Upper and lower bounds vary strongly with lead time due to the irregular sampling (forecasts issued twice a week) and small number of events analyzed for such extreme cases. Also, the increase in EFI values at higher lead times seems counterintuitive and is most likely related to sample size problems for such very rare events. With increasing lead time, there is not one optimal score but a range depending on how much POD, FAR, and HSS is desirable. To reduce the jumpiness that is likely related to the limited sample size, we applied a spline function to smooth the data. The shaded red area is based on the smoothed data, whereas the red diamonds indicate the underlying data. We have also analyzed optimal thresholds for moderately rare events characterized by the exceedance of the 95th percentile as used elsewhere in the study. Optimal threshold analysis for such moderately is qualitatively comparable to the results shown in Fig. 7, but overall less conclusive due to the even larger range at higher lead times (not shown).
b. Case studies
We have chosen two case studies to illustrate the potential and limitations of the EFI metric: one where the EFI metric worked very well and one where this was not the case.
1) October 2011
From 6 to 10 October 2011, locally 70–120 mm of precipitation fell on the north side of the Alps in Switzerland. This led to floods, landslides, and damages including interrupted rail roads and roads and damages at buildings (http://www.planat.ch/de/bilder-detailansicht/datum/2011/10/17/hochwasserereignis-vom-1011-oktober-2011/). The event was associated with an AR over the Atlantic and high IVT air reaching Switzerland from the northwest (Rossler et al. 2014; Piaget et al. 2015, Fig. 8). The high IVT air reached Switzerland from the Atlantic Ocean along the northern edge of a strong surface anticyclone. A detailed discussion of the moisture sources for the precipitation that fell over Switzerland and the local flow can be found in Piaget et al. (2015). Similar synoptic situations resulted in two other major flood events in northwestern Switzerland in the past 30 years (FM16).
The 6-hourly IVT values at the Jura grid point during this case were the highest IVT values in a direction-dependent 33-yr ERA-Interim climatology for this grid point (FM16) but they were not the highest values when ignoring the direction and looking only at the IVT magnitude. This is a key point for the interpretation of the EFI results. For floods in Switzerland the direction-dependent IVT (i.e., the magnitude of the IVT in the direction perpendicular to the local topography is key) (FM16). However, the IVT EFI is not direction dependent; it illustrates the “extremeness” of IVT forecasts relative to the local climatology based on the absolute IVT magnitude only. In the climatology the highest IVT values at the Jura grid point occur during southwesterly flow and maximum IVT values from direction northwest are slightly lower (Fig. 15). Hence, the highest IVT EFI values do not necessarily correspond to the most dangerous flow situations as they could arise during southwesterly flow. However, the critical direction-dependent IVT (see Table 2) is still high enough to result in high IVT EFI values.
Reforecasts with three different lead times are available for 0000 UTC 10 October, reforecasts started seven, four, and zero days earlier; day zero corresponds to the analysis. Figure 9 shows the EFI for IVT and daily precipitation for all available lead times. Please note that for precipitation—an accumulated quantity—we use the forecast integrating rainfall from 0000 UTC 10 October to 0000 UTC 11 October. Already seven days ahead of the flood event, the EFI of the reforecasts indicates the extreme event for both IVT and daily precipitation with EFI values exceeding 0.3–0.4. With decreasing lead time these values increase to about 0.9 for the analysis and the 24-h forecast, respectively (Fig. 9). The EFI of the analysis should be viewed as a perfect forecast which reflects the reality closely. Hence, this is an example showing that the early awareness system caught the signal and would have been able to raise early awareness to the flood event.
2) August 2005
On 22 August 2005, the most expensive flood event since 1910 happened in Switzerland (e.g., Stucki et al. 2012). Over the course of 21 and 22 August 2005 large areas along the northern Swiss pre-Alps received more than 220 mm of precipitation. This led to severe flooding and large damages to infrastructure and private properties. The synoptic situation was characterized by an upper-level cutoff low located over eastern Europe at the time of the flooding bringing moist air from the northeast toward eastern and central Switzerland (e.g., Stucki et al. 2012; FM16). This event was associated with moderately high IVT values compared to the total local climatology but with the highest IVT value in ERA-Interim relative to the local direction-dependent climatology from 1979 to 2014, FM16) reaching northern Switzerland from the north to northeast (Fig. 10).
This historical event is not captured by the EFI of IVT but quite well in the precipitation EFI (Fig. 11). This is due to the direction dependence of the high impact IVT. Climatologically the highest IVT values reach northeastern Switzerland from the southwest (FM16) and these cases are captured by high EFI values. However, the high IVT air coming from the southwest can pass through Switzerland without impinging on a major orographic obstacle and thus does do not produce high amounts of precipitation. IVT from the northeast on the other hand is directed perpendicular to the orography and can produce substantial precipitation, however, the absolute IVT values are lower and therefore not detected by the EFI. A direction-dependent EFI would be required to amend this issue. Indeed, when analyzing all 11 members of the reforecasts, it becomes evident that some forecast members were able to forecast the correct weather situation (not shown). These findings lead to the conclusion that the EFI of the daily precipitation as well as some indication whether single members forecast a dangerous situation need to be part of the final suite of visualization products. We present a visualization to deal with such cases in the following section. More historical flood events were analyzed as part of this study. In all cases, when taking into account the information of the EFI maps of IVT and daily precipitation and IVT of all individual members, an early warning could have been raised at least six days in advance.
c. Visualizations for the end users
In collaboration with operational hydrological forecasters (the end users of the forecasts) we developed a suite of graphics and data products to be used as guidance for them when issuing awareness alerts. To minimize the number of graphics to be inspected for routine operational use when no warnings need to be issued, they requested simple overview and summary information as on most of the days, a brief inspection of the overviews will be sufficient. For potential alert situations more extensive and detailed guidance with specific information is provided to allow users to inspect the situation in detail. Note that the weather forecast will consists of 51 members, contrary to the 11 members available in the reforecasts. Here we describe the guidance material and visualizations in more detail. All plots are introduced using an example of a high IVT episode that occurred in January 2018. All plots are designed to answer one or several specific questions by the forecasters.
IVT overview map: The first question that the hydrologists would like to have answered is what is the upstream IVT distribution over the Atlantic? To address this question an overview map showing the IVT situation over the Atlantic sector is produced at 12-h time steps from day 0 to day 10 into the future based on the control run (see Fig. 12).
Overview table and forecast consistency over time: The next questions are how consistent is the forecast over time? And how many ensemble members exceed the critical IVT thresholds? To address these points overview summaries in tabular and graphic form are created. An overview table (Fig. 13) shows forecasts issued at different lead times (rows) for a range of different verification times in the future (columns). This visualization offers the possibility to trace the evolution of forecast and potential warning situations over time, allowing to identify tendencies for potential alert situations. A color code is used to make it visually easier to spot whether members exceed the critical thresholds. When none of the ensemble members exceeds the thresholds, the box in the table is green, when 1–9 members exceed the threshold the box is yellow, and for 10 and more members the box is red. Because the thresholds are empirical and based on few historical cases and IVT values very close to the thresholds but not exceeding them might lead to critical situations as well, the full IVT distribution information is therefore provided as well (Fig. 14). While we have selected an interesting episode for illustration purposes, Fig. 13 also illustrates the limitations of such an early warning system. A considerable fraction of members indicates a strong IVT situation on 16 January, but none occurs in the final analysis; whereas there is early indication for an event on 18 January, but the forecast evolution is not very conclusive.
IVT diagram: The next question pertains to the variability among the ensemble members with respect to both the amplitude and the direction. This is addressed with IVT diagrams that allow a very compact climatological classification of the forecasts while also providing information both on the direction and the magnitude of the IVT for all 51 ensemble members (Fig. 15). Figure 15 shows the IVT diagram for a forecast issued at 0000 UTC 18 January 2018 for 1800 UTC 21 January 2018. The gray dots illustrate the ERA-Interim IVT climatology of about 37 years. A comparison between the forecast and ERA-Interim is informative as no major biases between the two datasets were found (not shown). The red dots indicate the IVT magnitude and direction of 51 forecast members for 1800 UTC 21 January 2018. The red line indicates the threshold for IVT proposed by FM16. In the example one ensemble member reaches the critical threshold. In the ERA-Interim climatology from 1979 to 2016 IVT was larger from this particular direction only for a handful of 6-h time steps. This indicates that the situation has the potential to be exceptional.
Plots of all ensemble members indicating IVT values above the threshold: The next question is how does the two-dimensional IVT field look for members that exceed the critical thresholds? In case of a flood awareness alert in the above overview products, the hydrologists can consult the European map of the members that show a dangerous situation for IVT and daily accumulated precipitation (not shown). This map plot shows a limited area over central Europe for every member. Members indicating IVT above the critical thresholds coming from the specific sectors are highlighted. The snowfall height (m) is also indicated for each member as this is a crucial information for the hydrologists to assess whether the precipitation will primarily be liquid or solid. These maps are generated on a 12-hourly basis.
EFI maps of IVT and precipitation and IVT and precipitation maps of all members: Along with the information above EFI maps for Europe will be generated for IVT and daily precipitation accumulation data. In addition, maps of the IVT and precipitation for an area zoomed to Switzerland for all members will be generated. However, these maps are not intended for daily use. They will be generated once a day, showing daily maximum IVT values and daily precipitation accumulations.
High IVT in a flow oriented perpendicular to the Alps or the Jura Mountains can cause devastating floods in Switzerland. An early flood awareness alert system can help society to be better prepared in case of a flood event. Here we show an example for an operational early flood awareness system based on medium-range IVT and precipitation forecasts. Verification analyses show that medium-range IVT forecasts are skillful, including for extreme values of IVT, even in the complex orography of Switzerland. However, the skill of the IVT forecasts depends on the skill score and the evaluated variable (i.e., the IVT magnitude or direction). Note that the verification is based on coarse resolution precipitation and IVT data (1° × 1°) from ERA-Interim, the reforecast climatology, and forecast data. This implies that the results are valid for regional-scale precipitation and IVT signatures and do not capture and represent complex local-scale precipitation patterns.
Based on these results we developed a flood awareness alert product together with the Swiss hydrological warning service consisting of an overview table and a number of forecast visualizations. These are produced operationally once every 24 h, a few even every 6 h. A requirement for the product suite was to build it in a way that minimizes the amount of information that hydrologists must screen on a daily basis. The first part of the product suite is intended to give a quick overview and provide information as to whether a dangerous situation might arise or not. Dangerous situations are rare and hence, on most of the days, the hydrologists do not have to look any further than at the overview information. In case of a signal, additional detailed information is provided for the hydrologists to analyze the situation such as the EFI maps and the individual members of the forecast that show a dangerous situation as well as an IVT diagram. In the complex topography of Switzerland, the direction of the flow is crucial and a direction-dependent EFI of IVT could capture this information. The development of the latter will be the next step in the development of the alert system.
While we have tried to provide some guidance on critical EFI thresholds with longer lead times, this guidance is of a rather qualitative nature due to the sampling uncertainty and balance between probability of detection and false alarm rate for very rare events. As a consequence, we would like to stress that gaining practical experience with these novel forecast products is crucial to develop recommendations on objective warning thresholds. Many studies have shown strong linkages between atmospheric rivers and flooding for other areas around the world [e.g., along the west coast of North America (e.g., Ralph et al. 2006), the United Kingdom (Lavers et al. 2011), South America (Viale and Nuñez 2011) and many other areas (see Gimeno et al. 2014 their Table 3 for a comprehensive overview)]. Therefore, there might be potential for similar early awareness systems in other regions of the world.
We thank three anonymous reviewers for thoughtful and thorough reviews that improved the quality of the paper substantially. We thank Johanna Ziegel, David Lavers, Sebastian Lerch, and Florian Pantillon for their helpful comments and advice. We further thank Swiss Federal Office (FOEN) for the Environment for the funding of this study. Therefore, parts of this study are additionally published in the form of a FOEN project report.