The level l read on the scale of a fixed-cistern barometer is underestimated for high values (l>l0, where l0 is the zero level, i.e. the level where no correction is needed) and overestimated for low values (l<l0) due to the change in level in the cistern. Therefore, the following correction formula must be applied to the raw observations: L=l+d2D2-d2(l-l0), where L is the corrected level, d is the inside diameter of the tube and D is that of the cistern (assuming a circular section).

For the large majority of the early instrumental records, d, D and l0 are unknown. Even if we knew them, we could not say for sure whether or not the correction was applied before recording the observations or whether the correction was necessary at all. Most commercial barometers (including those intended for scientific use) were actually sold without the indication of l0 . In our metadata the observer clearly stated only in one case, for Cambridge (Harvard College), that “the barometer is provided with a floating gauge and scale of correction”.

We can try to quantify the maximum error that can arise from uncorrected observations. One case where the cistern level correction could be applied in the literature is the series from Stockholm: estimated a correction of 1 % to l-l0. This means that even for extreme high- or low-pressure values the error is less than 0.5 hPa. Using the metadata for the observatory in London suggests that any correction there would be even smaller, since the cistern / tube ratio was slightly larger than in Stockholm. A similar ratio is found for the barometer in Zwanenburg . We can expect smaller cisterns by some amateur observers; however, the errors introduced by the missing corrections are unlikely to be larger than 1 hPa.

In all mercury barometers, but in particular in fixed-cistern and Fortin barometers, too thin a tube can lead to underestimations in the readings due to capillarity. This error becomes larger than 1 hPa for d<8mm . The barometers in Stockholm and London had a tube with an internal diameter of only 3 and 6 mm, respectively; therefore, they were probably affected by a substantial error. Capillarity was indeed the largest source of error in barometers and could be fully bypassed only in the second half of the 19th century with the adoption of reference primary barometers . Nevertheless, correction tables had been around since at least 1776 , although their use is never mentioned in the metadata in our possession. The error introduced by capillarity can be assumed to be constant over a period of a few years, with the exception of siphon barometers, in which the tube is exposed to air (and thus to humidity and dust).

The scale was often prone to physical changes, such as mechanical drifts or irregular changes due to thermal expansion or to the humidity's effect on the wood of the support. The latter was estimated in as negligible; however, it depends on the individual instrument. Other significant errors and drifts can arise from the quality of the mercury or from bubbles of air that enter the tube. In general, most barometers probably had a drift of some kind and were less reliable after a few decades of use.

In 1815 only France had officially adopted the metric system; elsewhere, metric units were rarely used. The English inch (=25.40 mm) was the standard length unit in the English-speaking world. In the rest of the world, the most common unit for barometer scales was the Paris inch (=27.07 mm). We encountered four other non-metric units, which were used only in specific countries: the Swedish inch (=29.69 mm) in Sweden, the Vienna inch (=26.34 mm) in Austria, the Rijnland inch (=26.15 mm) in the Netherlands and the Castilian inch (=23.22 mm) in Spain. The English and the Swedish inch had decimal subunits (the resolution was usually 1/100 of an inch); the others were divided into 12 “lines”, which were in turn divided into 4 to 16 “points”.

The temperature was measured using either the Fahrenheit or the Réaumur scale. The only exceptions were in France and in Sweden, where the Celsius scale had already been adopted. We converted all temperature observations to ∘C.

Observation times are available in various formats in the original records. Usually the observations were fixed at specific hours, but for some series they were indicated only qualitatively (e.g. “morning”), and in some others one of the observations was made at sunrise or sunset, whose time varies during the year. In 1815 all the countries of the observatories in the collection had already adopted the Gregorian calendar.

We assumed all times to refer to local solar time, since official standardised times did not exist. This also includes observations from ships, which were usually made at local noon together with the calculation of the geographical coordinates. For qualitative observation times, we applied the following fixed conversions when we did not have any information from the available metadata: morning – 08:00 LT; noon — 12:00 LT; afternoon – 16:00 LT; evening – 20:00 LT. However, when quantitative observation times are indicated only at the beginning of a manuscript (e.g. only on the first page of a meteorological register), we assume that they hold for the whole manuscript or the whole series of manuscripts (e.g. if there is one volume per year and quantitative observation times are indicated only for the first year).

In cases for which observation times are noted as “sunrise” and “sunset”, the local sunrise and sunset is computed based on the date and latitude of the station using the following equation: Hsun=arccos⁡(-tan⁡ϕ⋅tan⁡δ)⋅242π, where Hsun is the half-day length in hours, ϕ the latitude of the station and δ the declination of the sun, computed applying the algorithms described in . The local sunrise (sunset) time is 0.5Hsun before (after) local noon.

If observation times for single observations are missing but observations were taken at regular intervals, we replaced the missing observation times with the most frequent observation time for this interval (e.g. 21:00 LT for evening observations if 21:00 LT is the most frequent known time for evening observations at one specific observatory).

We finally translated local observation dates and times to UTC. For this we used a simple equation based on the longitude of the station: tUTC=tloc-λ⋅24360, where λ is the longitude of the station in degrees east, tloc is the local time and tUTC is the UTC time.

About half of the observatories in our data set recorded the temperature of the barometer. It was in fact common to have a mercury thermometer fixed on the same support as the barometer. Since the mercury expands and shrinks depending on the temperature, observations made with a mercury barometer must be corrected accordingly: L0=(1-γT)Lmm, where γ is the thermal expansion coefficient of mercury at 0 ∘C (1.82×10-4K-1), T is the temperature of the barometer in ∘C, Lmm is the original observation in millimetres of mercury and L0 is the observation reduced to 0 ∘C. Today, the “neutral” temperature of 0 ∘C is dictated by international standards; this was partially the case already in the early 19th century. Note that some observers used to reduce their observations to other temperatures (10 ∘R being the most common).

When the temperature of the barometer was not available, we used outside air temperature for the reduction. In many cases this is a good approximation because often the barometer was located in an unheated room or in a meteorological window and was fairly close to the “outside” thermometer. At some observatories, however, the barometer hung in a heated room, in which case we have an unknown error, usually with some seasonal cycle. Note that we rarely know the location of the barometer from metadata. When outside temperature observations were also missing, we used the closest (in space and time) 30-year climatology of 2 m air temperature from the Twentieth Century Reanalysis at 3-hourly resolution. This reanalysis has a spatial resolution of 2 ∘ for both latitude and longitude. As the base period for the climatologies, we chose 1871–1900 to minimise the difference with early 19th century temperatures. To reduce variability, we applied an 11-day moving mean per time step, so that the climatology for temperature on 6 January, 12:00 UTC, is the average of temperature on 1–11 January, 12:00 UTC, in the years 1871–1900. The use of climatologies was necessary for four stations only – one in Europe (Paris in 1815) and three in North America (see Table ) – and for occasional gaps in the other series. In one case (Brunswick), metadata indicate that the barometer was in a heated room; therefore, we preferred to use an arbitrary constant temperature of 18 ∘C for the correction.

To evaluate the errors introduced by the use of outside temperatures or climatologies, we made use of the stations where the temperature of the barometer was measured by correcting their pressure observations using either outside temperature or climatology and analysing the differences with the “right” correction.

The errors in the mean (Fig. 3a and b) have, as expected, a seasonal cycle. In summer, differences between inside and outside temperatures are on average very small for all stations, but in winter the barometers located in heated rooms are 5 to 17 ∘C warmer than the outside air (corresponding to average errors of 1 to 3 hPa when using outside temperature for the pressure reduction). We obtained similar results using observations from a certain part of the day (e.g. only morning or afternoon observations); in particular, the average errors in summer always remain within ±1hPa (not shown). Climatologies from the reanalysis introduce errors similar to those introduced by outside temperatures; these are slightly larger when the barometer is not in a heated room.

In Sect.  we try to correct these errors using a statistical method. However, much larger errors (>5hPa) are possible for single sub-daily values in continental climates, specifically in New England and Fennoscandia, when large deviations from climatology occur.

Temperature has, in general, a vertical gradient along the barometer, meaning that the observed temperature of the barometer is actually the temperature of only one part of it (depending on where the thermometer is attached). Therefore, the correction can introduce errors of the order of some tenths of hPa even when the temperature of the barometer is available.

Compared to the mean, the variance is more strongly affected when using climatologies (Fig. 3c and d). Using outside temperature introduces a random error in the variance that does not depend on the season and is usually smaller than 5 % for all stations but one: in Natchez (Mississippi) there is a systematic overestimation of the variance of about 10 %, which could be due to the subtropical climate of this station (i.e. a smaller pressure variance than any other station in the collection) or simply on the quality of the temperature observations (e.g. unshielded thermometer). Climatologies introduce a seasonal cycle in the variance error for some stations, with an underestimation (overestimation) of the variance in winter (summer).

We did not apply corrections for the thermal expansion of other parts of the barometer (cistern, tube, scale), which are usually 1 order of magnitude smaller than the correction for mercury and depend on the material used to build the barometer.

We also used Eq. () to rebase to 0 ∘C pressure observations that had been reduced to some other temperature at the time of the readings. This results in a small inconsistency because the correction tables in use at the time were purely empirical, γ not being known with sufficient precision. Therefore, the original corrections do not correspond exactly to those resulting from Eq. ().

The series from Milan, Salem, Stockholm and Turin had already been reduced to 0 ∘C in previous works by data contributors (see the respective references for more details). In Exeter, the observer started to register the temperature at the barometer only in 1817; outside air temperature was used before that year (absolute differences between temperature at the barometer and outside temperature were on average smaller than 2.5 K during 1817).

The conversion of pressure readings from millimetres to hectopascal follows from the hydrostatic equation: Pn=ρgφ,hL0×10-5, where Pn is the absolute pressure in hectopascal reduced to normal gravity, ρ=1.35951×104kgm-3 is the density of mercury at 0 ∘C, gφ,h is the local gravity (see below) and L0 is the barometric reading in millimetres (corrected for temperature). This is equivalent to the usual procedure of first converting pressure readings from millimetres to hectopascal by using normal gravity acceleration in Eq. () and then correcting for local gravity by using Pn=gφ,hgnP0, where P0 is the absolute pressure not reduced to normal gravity and gn=9.80665ms-2 is the normal gravity acceleration.

We estimated the local gravity gφ,h from the latitude φ and elevation h (in m a.s.l.), assuming flat terrain around the station see: gφ,h=9.80620⋅1-0.0026442⋅cos⁡2φ-0.0000058⋅cos⁡22φ-0.000003086⋅hms-2.

Since all land stations in the data set are in the midlatitudes and at relatively low elevations, the gravity correction is on average small (ca. 0.5 hPa; positive for high latitudes and negative for low latitudes).

To use the pressure observations for synoptic analysis, we reduced P to sea level: P0=P⋅exp⁡gφ,hR⋅hTS+a⋅h2, where R=287.05Jkg-1K-1 is the gas constant for dry air, a=6.5×10-3Km-1 is the standard lapse rate of the fictitious air column below the station and TS is the outside temperature at the station in K.

We did not apply further corrections described in , since the uncertainty in our data set is much higher than that required for modern barometers (i.e. ±0.1hPa).

Similarly to the reduction in pressure readings to 0 ∘C (Sect. ), we used in situ air temperature observations where available and resorted to climatological temperatures from the Twentieth Century Reanalysis (1871–1900) otherwise. We did not use the temperature of the barometer to reduce pressure readings to sea level.

The series from Stockholm and Turin had already been reduced to sea level by the respective data contributors.

We inspected visually each sea level pressure (SLP) series (and differences with nearby stations) to flag erroneous outliers and clear inhomogeneities in the period 1815–1817. Nearly all outliers derive from mistakes in the digitisation or in the transcriptions by the observer. When possible (i.e. when the original sources were readily available) we corrected them; otherwise, we flagged them as erroneous and excluded them from the analysis.

The total number of pressure observations flagged after the quality control is 4657, corresponding to 4.1 % of the 1815–1817 data set. Most of the flagged observations correspond to long periods in a few series where we detected large inhomogeneities: Madrid (whole year 1815 flagged), New Haven (most of autumn and winter of 1815/16), Valencia (all summer observations), Växjö (whole 1817) and Ylitornio (11 months in 1817). The number of flagged observations for each series is indicated in Table .

Another requirement for a synoptic analysis is that observations must be simultaneous. To achieve this, we linearly interpolated all pressure observations to four daily, equally spaced time steps: 00:00, 06:00, 12:00 and 18:00 UTC. If no observations of a certain station were available within ±6 h from a certain time step, then we did not interpolate and considered the station to have no data for that specific time step. In Europe (on which our analysis will focus), most observations were made very close to 06:00, 12:00 and 18:00 UTC; interpolated values for 00:00 UTC are in general less reliable and will not be analysed. Across all stations, the mean absolute differences between the interpolated values and the closest observations are 0.9 hPa for 00:00 UTC, 0.5 hPa for 06:00 UTC, 0.4 hPa for 12:00 UTC and 0.8 hPa for 18:00 UTC. By using a linear interpolation, we did not account for the daily cycle of pressure; this choice does not significantly affect the results because the amplitude of the daily cycle is much smaller than the day-to-day variability that we want to study.

We did not interpolate outside temperature observations because of their larger daily cycle and its strong dependance on other meteorological variables such as cloud cover and wind.

Even though the results of the previous section demonstrate a good consistency among the variability in most of the series, the lack of metadata for many of them causes large systematic errors in the mean values. A statistical approach is the only viable option to obtain absolute SLP values accurate enough for a synoptic analysis; thus, we use the reconstruction by as a reference to correct the land series in Europe.

It is important to mention that the reconstruction is not independent; in fact, the monthly means of 16 series in our collection were used as input for the reconstruction. However they were all homogenised by ; therefore, we are confident that the reconstruction offers the best possible estimation of mean SLP and that the application of the corrections guarantees a better reliability of synoptic weather maps.

Using the original SLP observations, we calculated seasonal means for each series in the period 1815–1817 and then applied a constant offset necessary to match the 1815–1817 seasonal means of the nearest grid point in the reconstruction. This was possible only if enough data were available: we calculated the offset using only the years with at least 90 % of the days in the target season having at least one observation available. When a series does not have enough data in any year for a certain season, we used the average of the offsets from the available seasons. The seasonal offsets were then applied directly to the interpolated SLP values described in Sect. .

If the data are insufficient in every season, the series is not used in this section. This was the case for Derby, which has only 1 month of data. Moreover, we excluded the series from Valencia and Växjö, which showed low reliability in the previous section. We did not correct the already homogenised series from Milan and Stockholm.

Since the reconstruction is based on monthly means, in turn calculated from daily means, we must apply a further correction to the offsets to take into account that our data are instantaneous observations rather than daily means. For this, we estimated the mean daily cycle of SLP for each season from the 1981–2010 climatology of the closest grid point of the MERRA (Modern-Era Retrospective analysis for Research and Applications) reanalysis . MERRA offers the advantage of an hourly resolution and a higher spatial resolution (1/2∘ latitude × 2/3∘ longitude) than the Twentieth Century Reanalysis. For stations with variable observation times, we used for the calculation the observation times (rounded to the hour) adopted most frequently at the target station in the target season. The resulting corrections are very small for all series and smaller than 1 hPa even for stations with only one observation per day.

On the other hand, the total statistical corrections are in some cases larger than 10 hPa (Fig. 8), while their root mean square is 4.4 hPa. The average correction (thick line in Fig. 8) has a seasonal cycle with a peak-to-peak amplitude of 1.9 hPa, indicating overestimated values in winter relative to summer. This is nothing more than what we expected because of incorrect temperature corrections for barometers in heated rooms (Fig. 3a and b).

The largest corrections are usually related to amateur observatories with scarce metadata and with the temperature of the barometer missing, but in some cases (e.g. Prague) they are also related to official observatories with high scientific standards. An important source of systematic errors is the uncertainty of the barometer elevation: according to Eq. (8), considering a standard atmosphere, an uncertainty of 20 m (which applies to most stations) results in an uncertainty in SLP of about 2.5 hPa near the sea level or less for higher elevations. Moreover, the statistical correction can also take into account capillarity (see Sect. ), which is probably the reason why the majority (about two thirds) of the applied offsets are positive (capillarity always causes an underestimation in mercury barometers) and represent the main contributor to the large corrections needed in some of the official observatories.

As already mentioned, the winter of 1815/16 was not a typical post-volcanic winter in terms of temperatures, being colder than usual in most of Europe. From our temperature data we detected two severe cold spells that hit in quick succession between the end of January and the first half of February, which significantly contributed to the cold anomaly. We use these two cold spells as a case study to evaluate the quality of the corrected SLP data set.

Figure 9 shows four SLP synoptic maps corresponding to the initial phase of the two cold spells. We plotted the 06:00 UTC time step because more temperature observations (also shown in the maps) are available near that time.

A common SLP pattern is evident for the two cold spells, although the one in February, the most severe, is characterised by much lower pressure values. In both cases there is a low-pressure system over southern Europe and a high pressure area over northern Europe (note that the position of the centre of cyclones and anticyclones drawn by the isobars in the maps is often an artefact due to the lack of observations near the borders, in particular in the Mediterranean). This pattern represents a typical blocking situation and drives a westward flow of cold continental air towards western Europe e.g., consistent with a severe cold outbreak.

A curious anecdote is related to the cold spell of February 1816. Samuel Parkes, a contemporary British chemist, exploited the unusual cold for an experiment on the freezing point of wine. His results were published as a short article in the first issue of The Quarterly Journal of Science, Literature and the Arts, where he reported that the temperature on the morning of 9 February (probably near his house in London) was “22∘ below the freezing point”, which, assuming a Fahrenheit scale, corresponds to -12 ∘C. On the following night another London chemist, Luke Howard, made several observations with different thermometers in Tottenham (Annals of Philosophy, vol. 7). On 10 February at 07:30 LT he measured a temperature of -19 ∘C, ca. 2 m (8 feet) above the ground. According to Howard, this was the lowest value measured in London since 1797.

According to our data set, temperatures were particularly low in Sweden, reaching -38.5 ∘C in Umeå and -37 ∘C in Härnösand, while the absolute minimum was measured in Ylitornio (Finland) with -40 ∘C (not shown). The homogenised series from Stockholm allows a comparison with modern data: in the reference period 1961–1990 there were three cold spells that were more severe than that of February 1816. However, cold spells of this magnitude were probably not as rare in the early instrumental period. In fact, and found a particularly high frequency of very cold winter days in Stockholm and Uppsala in the late 18th and early 19th century, although they warned that the data might still be affected by inhomogeneities. Daily temperatures lower than those of February 1816 were registered, for example, in January 1814 and again in December 1817. Similar results (not shown) are obtained by analysing the temperature series from St. Petersburg, in north-western Russia .

We analyse here one case at the beginning of July, one of the coldest periods of the summer of 1816 in central Europe i.e. about 7 ∘C colder than usual in Geneva; see. There was abundant rainfall in the Alps, where, in the night between 3 and 4 July, a huge landslide, about 300 m wide, killed at least 14 people near the town of Uznach in eastern Switzerland (Erdrutsch in der Au (Goldingertal): Situationsplan, State Archives St. Gallen, Ref. KPG 1/65.1, 7/16).

A shallow low-pressure system crossed the Alps between 1 and 2 July (Fig. 10), which is consistent with cold and rainy weather, and then probably moved to south-eastern Europe on 3 July, when easterly winds were observed at the stations in eastern Europe (not shown). Afterwards, the Alps remained under the influence of unstable air coming from the Atlantic for several days. The weather diary kept for Aarau, in northern Switzerland, reports precipitation every day until 19 July, always accompanied by westerly winds except on 1 day.

An area of high pressure was present over north-eastern Europe during the whole period shown in Fig. 10, suggesting fair weather there (confirmed by temperatures). In particular, in the north-eastern corner the maps show temperatures registered in Ylitornio at 14:00 LT which are remarkably high for that latitude (the maximum temperature is reached on 5 July at 31 ∘C, not shown). The quality of the measurements is questionable see; however, it is interesting to note that the average 14:00 LT temperature measured in Ylitornio in the first week of July 1816 is 9 ∘C higher than the average 14:00 LT temperature of the whole summer 1816. Therefore, our data suggest the occurrence of a heatwave in north-eastern Europe in conjunction with the cold period in western Europe. Again, the daily temperature series from St. Petersburg supports our conclusion, indicating 13 consecutive days (6–18 July) with a mean daily temperature of >20∘C, the longest such series in the years 1815–1817. The month of July as a whole had, nevertheless, slightly negative temperature anomalies in that region , showing how even relatively long-lasting events can be overlooked when considering monthly means only. Note also that the SLP values in Ylitornio are clearly underestimated in the analysed period, and in general they are not very reliable because of the continental climate of the region (see Sect. ).

After a relatively mild winter, the spring of 1817 struck a serious blow to Europe. In particular, as already mentioned, the month of April was extremely cold (see Fig. 4).

To gather more information on the most important weather events that distinguished this month, we examined contemporary newspapers and other historical sources. The worst affected area was probably the northern slope of the Alps. Exceptional snowfalls and avalanches were often reported in that month, especially in Austria: in Innsbruck (574 m a.s.l.), for instance, snow fell on 18 out of 30 days , while over 2 m of snow were reported in Annaberg (976 m), near Vienna, after 16 consecutive days of snowfall (Lemberger Zeitung, 9 May 1817). At Buchlovice (234 m, south-east Moravia), a priest, Šimon Hausner, recorded snowfall on 11–14, 19–26 and 28 April, i.e. on 13 days (with another 2 days with sleet). Permanent frosts were also typical in this month. Hausner concluded that “no previous April has been this bad” (Tägliche Witterungs-Beobachtungen des Buchlowitzer Pfarrer Simon Hausner von Jahren 1803 bis 1831 excl., Moravský zemský archiv Brno, fond G 138 Rodinný archiv Berchtoldů (1202) 1494–1945, inv. č. 851).

One episode in particular attracted the attention of newspapers. In 1817 the Austrian foreign minister, the influential Prince von Metternich, had organised an ambitious scientific expedition to Brazil, the first major overseas mission ever undertaken by the Austrian navy. On 10 April two frigates, the Austria and the Augusta (Fig. 11), weighed anchor from the port of Trieste, in today's north-eastern Italy, and headed to Rio de Janeiro. On the morning of the second day of navigation, near the coast of Istria, the ships were surprised by a violent storm and suffered heavy damage. The Austria was able to dock in Pula (today's Croatia) and could resume the journey after only 1 week. The Augusta was shorter on luck, losing all its masts, sails and boats, and reached the port of Chioggia, near Venice, with great difficulties, 4 days after the storm hit. The repair of the ship took about 7 weeks.

Many renowned scientists and intellectuals were on board the two frigates. Among them were two members of the Bavarian Academy of Sciences: Johannes Baptist von Spix (biologist) and Carl Friedrich Philipp von Martius (botanist), who were on the Austria. Their detailed account of the expedition gives us a description of the storm:

“The night passed over quietly; but in the morning we were all awakened from our sleep by an uncommonly violent motion of the ship. Those whom sea-sickness had not rendered insensible, readily perceived […] that we were in a great storm.

The Bora, a cold, very violent north-east wind, which, especially in spring, frequently blows from the Istrian mountains, and prevails in the northern part of the Adriatic sea, had suddenly assailed the two ships. A black cloud, hanging very low, was the only indication that the officer on duty had of the approach of the gale; so that there was scarcely time to take in the sails. In a few minutes we lost sight of the Augusta, which hitherto had kept at a small distance from us. A thick fog enveloped our ship; a cold rain, mixed with hailstones, which the storm furiously drove before it, covered the deck with pieces of ice of considerable size, and almost froze the crew. The ship was tossed violently; the yards and tackle were torn and broken: the waves rushed through the window into the forecastle, partly filled the hold with water; and at last, when the storm was at its height, the bowsprit broke short off. The hurricane raged with the utmost fury till noon, when the sea grew calmer, and the bleak Bora being succeeded by a mild east wind, we cast anchor in the middle of the sea, about three miles to the west of Rovigno.”

As suggested by the two German scientists, who demonstrate a remarkable knowledge of climatology, the storm was related to a severe bora wind event . The event was also felt in most of the Po Valley (northern Italy), where four of the stations in our data set are located. The observatories of Padua and Bologna, which are close to the Adriatic coast, reported thunderstorms, very strong wind from the north-east and snow flakes on that day. Newspapers reported heavy snowfall in the eastern Alps during the same event; in particular, about 50 cm of snow were measured in northern Slovenia and 10 cm in the city of Klagenfurt (Gazzetta di Milano, 8 May 1817).

In Fig. 12 we show the synoptic maps for 10 and 11 April. The position of the Austrian frigates in the morning of 11 April is marked by a star in the third map. Again, we are dealing with a blocking pattern, characterized by high pressure over north-western Europe and low pressure over Fennoscandia. This configuration represents the negative phase of the so-called Scandinavian pattern e.g., which is normally completed by a third pole (in this case, an area of high pressure) over central Siberia. This pattern stayed in place for most of the month of April 1817, continuously pumping Arctic cold air towards central and southern Europe. singled out this configuration as the feature most commonly responsible for widespread exceptional snowfall in Italy.

A cold air outbreak is the condition necessary for a severe Adriatic bora storm, and the synoptic pattern of 11 April (third panel in Fig. 12) is in fact a typical pattern for severe bora events . The maps show clearly the build-up of a pressure gradient between the northern and the southern slope of the Alps caused by the interaction of the cold air with the orographic barrier. It is likely that an orographically induced cyclone formed in the Mediterranean as a typical “cut-off” and that it remained there for several days e.g.. wrote that when the Austria was near the coast of southern Italy on 22 April, they could see the Gargano promontory, which reaches a maximum elevation of 1065 m, “covered with snow very low down”. They also repeatedly reported stormy weather in the following days, culminating in another violent storm on 27 April that forced the frigate to seek shelter in Malta. They wrote

“On the following morning we were already forty-two leagues to the west of Malta, when the wind suddenly settled in the N.N.W. It soon increased, and the waves ran so high, that it was impossible to keep the course to the south-west. The frigate rolled so violently, that in a short time the tackling was materially damaged; every thing movable was thrown backwards and forwards; and it seemed dangerous longer to expose the ship to the fury of the waves.”

The direction of the wind suggests that a cyclone was centred close to eastern Sicily. Note that the position reported in the official translation is probably affected by a mistake in the unit conversion: the original German version reports “vierzig Seemeilen”, literally 40 nautical miles (74 km). The word Seemeile was, however, commonly used also to indicate the “league” (i.e. 3 nautical miles), although it is not clear why the translator converted it to 42 leagues instead of 40. In any case, we think that the literal translation gives a more realistic position for the ship, which would have been only a few kilometres from the coast of Africa otherwise.

In fact, between 26 and 27 April another cold outbreak affected southern Europe. This time the snow fell abundantly even in the Po Valley. In Bologna, about 15 cm (“half shoe”) of snow were reported (Osservazioni meteorologiche 1817, Historical Archives of the Astronomical Department, University of Bologna), again accompanied by a strong north-easterly wind. The wind was also responsible for the spread of a fire which destroyed the Hungarian town of Szombathely, near the Austrian border, where 250 houses were reportedly burned to the ground in the night between 26 and 27 April (Corriere di Milano, 17 May 1817). Snow was observed even in Rome, but the exact date is unknown ; in any case it would be the latest snowfall ever recorded in Rome.

The synoptic pattern underlying this event is shown in Fig. 13. Again, a star marks the position of the frigate Austria when it was hit by the second storm. The large-scale SLP pattern had not changed from 11 April, but the temperatures registered were even lower in some places, despite the season being advanced, and stayed low for days. On 29 April, -4 ∘C were measured in Geneva, the lowest temperature of the whole spring there. At the beginning of May, in the Austrian Alps at 600 m a.s.l., the snow was still “higher than fences” . When temperatures finally returned to more usual levels, the enormous amount of snow in the mountains melted rapidly (including the snow that had not melted in the previous summer), causing widespread floods in the Alps and surrounding regions e.g..