European Floods in Summer 2021: the Extreme rainfall of 14-15 July

Summary

• The focus of this item is the northwest Europe flooding events that hit the headlines in July 2021; specifically the wettest day during that period: 24h to 06UTC 15th.
• Extreme rainfall on this particular day, following previous wet weather, is believed to have triggered the most deadly flash flood events to have affected Germany and Belgium.
• The synoptic drivers of the rainfall were a quasi-stationary cyclone, slow moving fronts, and related to those a plume of unusually warm, moist lower tropospheric flow curving west then south from the southern Baltic and areas beyond.
• 24h rainfall totals widely exceeded 75mm, with >150mm recorded in Cologne, and in the High Fens - Eifel area that spans the Belgian-German border, with the latter peak due in part to orographic enhancement over relatively steep north-facing slopes
• Rapid draining away, to the west and east, of the High Fens rainfall was probably the main cause of the flash floods and numerous fatalities in riverside towns such as Pepinster and Ahrweiler
• Operational IFS forecasts (cycle 47r2) gave reasonable broadscale guidance regarding the said 24h rainfall event, for lead times up to about 4 days in advance
• Whilst extreme rainfall details were generally not captured by the IFS, ENS and HRES did show signs of higher totals and probabilities in the High Fens area from 2-3 days in advance
• Whether IFS forecasts on their own (and a related hydrological model response) could have been used to anticipate and mitigate against extreme flood impacts is debatable, unless use of a low probability threshold such as 10% is deemed acceptable
• Reforecasts of the event, using cycle 47r3 with its new moist physics, had been expected to give a different flavour to the rainfall patterns: some features were as expected whilst others (e.g. lesser peaks) were not.
• Overall HRES forecasts from 47r3 were worse than those of 47r2, suffering in particular from a misplacement of rainfall to the east, although ENS forecasts from the two systems were of comparable quality.
• The shortest range 47r2 forecasts (for T+6-30h) were degraded relative to previous data times, showing less rain overall - the reason is not clear.
• Areally-integrated totals in short range forecasts from both IFS cycles were only slightly below observed values, so the major under-prediction errors one sometimes sees for big flood events were reassuringly absent.
• Analysis of 14,000 reforecasts per month, across an annual cycle, clearly highlights that in the affected areas a 24h rainfall event of the observed extreme magnitude was, climatogically speaking, far more likely to happen in summer than in winter.
• CAVEATS: this item has not directly examined LAM output, or hydrological model output, or flash flood products, or ecPoint forecasts, or sub-daily rainfall behaviour / totals

Background

Extreme rainfall between 12th and 15th July 2021 lead to flash floods and loss of life in Germany, Belgium, the Netherlands and Luxembourg, in what was one of Europe's most devastating weather-related disasters of recent decades. Radar-derived totals for western parts of Germany for the 3-day period 06UTC 12th to 06UTC 15th (Figure 1) suggest that the main event came at the end of the period (i.e. 24h ending 06UTC 15th), although rainfall in the preceding days undoubtedly "primed" the land surface for flash floods, by increasing soil water content and river levels. 

Figure 1. Radar-derived rainfall totals from DWD for western parts of Germany for 72h to ~06UTC 15 July 2021 (left panel), with a breakdown into consecutive 24h periods to the right.

Rainfall, floods and geographical considerations

This page focusses on a valid period of 06UTC 14th to 06UTC 15th, because it was the wettest of the three 24h periods, and because it was during this time that flash floods lead to the greatest loss of life. The highest death toll registered in Belgium was in the town of Pepinster, whilst in Germany it was in Ahrweiler. Their locations are marked on some of the following maps with white crosshairs (see Figure 2). Pepinster lies at the confluence of the rivers Vesdre and Hoëgne, whilst Ahrweiler lies on the river Ahr. All three rivers have their sources in the Eifel - High Fens Nature park area (called High Fens heareafter) which rises to a peak altitude of 694m. Vesdre and Hoëgne drain westwards, the Ahr drains eastwards. It was the dramatic overtopping of these rivers that lead to flash flooding and loss of life.

Figure 2. HRES topography (m) centred on the worst affected areas, with sites discussed in the text labelled.
Figure 3. 24h rainfall totals to 06UTC 15 July - peak values were 164mm in Belgium, and 153mm in Germany (two dark red spots). Data cropped to appear only in the following lat-long boxes: [N:53.0, W:4.8, S:51.0, E:9.5] plus [N: 51.0, W: 3.0, S:47.6, E:9.5]. The provision of high density rainfall observations over Belgium by RMI (Royal Meteorological Institute of Belgium), through Thomas Vanhamel at RMI, is gratefully acknowledged.

On the 24h rainfall totals map in Figure 3 the two wettest places (dark red spots) were Jalhay in Belgium, in the High Fens (164mm) and Cologne in Germany (153mm). Very high values (>125mm) surround Jalhay. Although observation density near Cologne is less, making comparison with Jalhay difficult, on the basis of the corresponding plot on Figure 1 it seems that the Cologne peak was more localised.

Synoptic setting

From a synoptic perspective low pressure, warm moist flow and slow moving frontal systems were the main drivers of the extreme rainfall (Figures 4 and 5). Note the tongue of exceptionally high theta-w air on Figure 5 extending from the southern Baltic, and beyond, down to the afflicted area. On the southeastern flank of this a slack low pressure centre moved along a cyclonically curved arc trajectory, broadly from north to south, passing just to the east of the High Fens, during the examined 24h rainfall event. This placed the High Fens in a northerly low level flow that contained the high theta-w airmass. It seems very likely that when this moist flow impinged on the relatively steep upslopes on the northern flank of the High Fens, orographic enhancement of rainfall occurred, which would help to explain the extreme totals in and around Jalhay, and lower values just south of there. The steepness of the northern slopes of the High Fens is more apparent in the 4km orography than in the HRES 9km equivalent - see figures 6 and 7 - with topographically forced ascent of >400m implied, and maybe as much as 600m. It is more difficult to explain the extreme rainfall in Cologne, a relatively low lying area, but perhaps this was more random, and related to convective activity embedded in the frontal rain.

Figure 4. Met Office surface analysis for 18UTC 14 Jul 2021. Orange dot signifies the Eifel - High Fens Nature Park. Red line shows the track of a low pressure centre between 06UTC 14th and 06UTC 15th.
Figure 5. Wet bulb potential temperature (theta-w) at 850mb from HRES (T+6) valid for same time as synoptic chart

Figure 6: HRES 9km orography versus 4km orography (m)
Figure 7: 4km orography with a low level orographic enhancement airmass pathway, that could explain the rainfall peaks in and around Jalhay, marked in red.

ECMWF forecasts

In assessing the utility of ECMWF IFS forecasts with regard to predictions of the extreme rain and its impacts (notably the aforementioned deadly flash flood events), we examine two aspects:

1. Ability of the IFS to predict the general envelope of very heavy rain during the cited 24h period - which we will loosely define as the area of >75mm/24h (magenta colouring and above on observation and quantitative forecast products).
2. Ability of the IFS to foresee that the High Fens area would see some of the highest totals.

For both of the above we examine the performance of the then operational ECMWF model cycle 47r2, and also the performance of the then experimental cycle 47r3 (that went operational in October 2021), with its attendant 'new Moist Physics' package. The new Moist Physics package is a particularly important aspect because it is known to alter rainfall forecast characteristics. It should also be stressed, however, that forecasters at the time would only have had access to the 47r2 runs.

1. Deterministic Forecasts

The following figures (8-11) show 24h rainfall total forecasts from the ECMWF 9km resolution model, along with an expanded-coverage observations chart for comparison. Colour schemes are the same.

Figure 8: Sequence of 47r2 HRES 24h precipitation forecasts for VT 06UTC 14th to 06UTC 15th Jul 2021, for lead time of 4 days (=T+78 to T+102), 3.5 days, ... 1 day (= T+6 to T+30).
Figure 9: Same as Figure 9 but for 47r3 HRES (esuite).
Figure 10: 24h rainfall observations as on Figure 3 for comparison, but larger area with no cropping.
Figure 11: All panels from Figures 8 and 9, in static format, longest leads at top to shortest at the bottom (very large figure - helps to zoom in a lot!).

Overall the operational 47r2 HRES gives quite good guidance on the magnitude and extent of the event, from lead times of 4 days down to 1 day (i.e. all 7 lead times shown). Values greater than 75mm appear in every forecast, in roughly the right general area, and there are even hints that totals > 125mm are possible on some forecasts. Virtually every forecast hints at localised orographic enhancement over the High Fens, and whilst mostly the totals there are in the 50-75mm category, which probably would not have been sufficient to trigger devastation on anything like the scales seen, the 1.5 day forecast (the penultimate one) was remarkably accurate in the said area. Looking then to the new 47r3 cycle we see it performs less well overall, with the 6 forecasts from 4 day down to 1.5 day leads all looking worse than their 47r2 counterparts. Often the main signal for large totals is too far east. The picture of high totals is more 'blobby'; 47r2 provided a 'smoother look' which is conceptually more appealing though may or may not be more correct in structural terms. This type of difference in appearance is a known characteristic of 47r3 runs. Another advertised characteristic of 47r3 is it's proclivity to deliver higher spot totals than 47r2 (which is believed to be better). However in this case study we see the opposite; 47r2 generally has the highest spot totals, and this looks like better overall guidance even if they are rarely in the right place. Regarding orographic enhancement over the High Fens only the last 3 forecasts in 47r3 really highlight this - so this aspect is also handled less well by 47r3. Although the picture is not clearcut there have been suggestions that 47r3 should handle orographic enhancement slightly better than 47r2 - so again this seems to be a counter-example. The very last forecast in the sequence is better in 47r3 than in 47r2 - 47r3 shows more accurate extent of high totals over Belgium, France and Luxembourg.

2. Ensemble Forecasts

The 9-up panels below show raw ENS probabilities of totals>75mm, from 47r2 (Figure 12) and 47r3 (Figure 13).

Figure 12: Probabilities of >75mm/24h for 06UTC 14th to 06UTC 15th July 2021, from consecutive 47r2 ENS forecasts at leads of 5 days (top left) across and then down to 1 day (bottom right). So the first two panels show forecasts not depicted for HRES in Figures 8, 9 and 11 above.
Figure 13: Same as Figure 12, but for cycle 47r3. the blank panels signify that no retrospective ENS forecasts were created for 12UTC data times.

The first two forecasts (from data times of 00 and 12UTC on 10 July 2021) provide little in the way of useful guidance for the areas affected, although they do show that extreme totals were possible somewhere in this part of western Europe. Thereafter, for leads of 4 days and less, the signals from both cycles do look useful for highlighting the at-risk areas, and there is also more run-to-run consistency than one sees in the HRES forecasts above, as one would expect from an ensemble. Nonetheless the probabilities do look less than one might have hoped - with peak values around 10% at leads of 4 and 3 days, rising to about 30% for 2 and 1 day leads. Also bear in mind that for totals larger than 75mm, which may well have been required to see such high impact flash floods, the probabilities would have been lower still. One could then ask how actionable is, say, a probability of 10%, and what level of action would be required/acceptable/justified? One also needs to remember that other model guidance, at much higher resolution, is also available to forecasters. For the High Fens area probabilities are elevated, compared to regions to the north and south, in both cycles, for day 3, day 2 and day 1 leads. This probably signifies the capturing of an orographic influence, which is useful, although on the other hand this is not the area with the highest probabilities. One should also remember that the ENS topography has a smoother representation of this and other orography than HRES, so we should be careful to not expect too much here.

With regard to cycle differences for ENS overall, these appear to not be that significant, and it is difficult to choose an overall 'winner' between 47r2 and 47r3. This differs from the HRES picture where 47r2 was the clear winner. For the shortest lead ENS forecasts however (lower right panels, for T+6-30) we see higher probabilities from 47r3, and better extent over Belgium, France and Luxembourg, which does mirror what we saw in HRES forecasts.

Moisture Budget

It is interesting to consider whether the total amount of rain created by the different runs, at short leads, mirrors what was observed. This is because in other cases, such as the devastating floods of 20 July 2021 around Zhengzhou in China, the areally-integrated rainfall totals from the IFS fall well short of observations (e.g. by 75% !) for reasons that are not clear, and such shortfalls would massively compromise the ability of hydrological models to create a meaningful response. To assess this for the current case we consider the T+6-T+30 forecasts from various IFS components, in cycles 47r2 and 47r3 (Figure 14), to compare with Figure 3. 

Figure 14: rainfall totals from different IFS components from 47r2 (top row) and 47r3 (bottom row) for VT 06UTC 14th to 06UTC 15th July 2021, from data time 00UTC 14th (which here we have been calling day 1). The scale is the same as for the other rainfall plots shown above.

On these maps we see slightly larger amounts in 47r3, as discussed above, with very reasonable envelopes of large totals in both cycles, but with a shortfall in the peaks, substantially so in some areas (e.g. Cologne). Nonetheless when we compute area averages over the two 'joined boxes' shown on Figure 3 (see Table 1), we see reasonable agreement between forecast and observed, with only a slight shortfall. This comparison assumes evenly spaced observational (gauge) data, which is not the case, although the impact of that incorrect assumption is not believed to be large. It looks as though 47r3 is delivering more rainfall, although as hinted at above that may well be a feature peculiar to this data time; for longer leads it seems the opposite may have been true. Another interesting aspect is that the ENS mean is less than the Control in both cycles - we have seen this for other extreme-rainfall-saturated-airmass cases and it may relate to the stochastic perturbations we are using in the perturbed members.

Table 1: Domain average rainfall totals for the joined boxes as on Figure 3, in mm, for 06UTC 14th to 06UTC 15th July 2021, for forecast lead times of 6 to 30h (forecasts are on Figure 14, observations on Figure 3).

Climatological Context

Using the re-forecasts that ECMWF routinely produces one can, with some assumptions, identify what a 'truly extreme' event, in a current climatological context, might look like. The plots below illustrate this for 24h rainfall (00-24UTC), in a framework where we have considered each month separately. Figures 15 and 16 show the picture of model-derived 500-year return extreme 24h rainfall events for mid winter (Jan) and for mid summer (Jul). Noise on the plots that is unrelated to topography should be disregarded as being a sampling issue.

Figure 15: Extreme 24h January rainfall at ENS gridscale, corresponding to a January return period of ~500 years, as derived, with assumptions, from ECMWF re-forecasts for January.
Figure 16: Same as Figure 15 but for July.

Examining a month-by-month sequence of plots like these we see clear-cut annual cycles. The main feature of these cycles, over the plot domain shown, is that in central and N Europe the truly extreme daily totals are much greater in summer than they are in winter, as illustrated on Figures 15 and 16 (although admittedly this picture tends to be reversed in west-facing upland coastal areas of N and W Europe, and indeed the cycle is rather different around the Mediterranean with autumn maxima there). Looking to where the floods under consideration occurred we see that the mid-winter extreme is about half of the mid-summer extreme (about 50 versus 100mm/24h). And clearly our observed values are close to or in excess of the July re-forecast extremes. If short period extreme rainfall (lets say 24h) is the main driver of flash floods like those that occurred, then it seems that in the areas affected these are much more likely to occur in summer than at any other time of year. Therefore it is not surprising that an event like this (if it was ever going to happen) actually occurred in summer. Physical considerations could also lead one to a similar conclusion - on Figure 5 we noted that the 850mb theta-w impinging on the north-facing upslopes of the High Fens was about 18C, whereas in winter-time an airmass from this "optimal" direction cannot really have 850mb theta-w above about 10C. In saturated conditions the former airmass can hold 90% more water vapour than the latter, which, if other things were equal, would mean rainfall about 90% greater also.