Aerosol optical thickness (AOT) data over the study period are estimated from both regional AERONET station data and Collection 6 (C6) Terra MODIS Dark Target (DT) aerosol products (Levy et al., 2013). AERONET AOTs are derived from the measured solar energy at seven wavelengths including 340, 380, 440, 500, 675, 870, and 1020 nm (Holben et al., 1998). For the study period, quality-assured Level 2.0 AERONET data are not available, and thus the cloud-screened Level 1.5 AERONET data are used in this study. To derive fine-mode AOT associated with smoke and to help remove any thin cirrus contamination that may be a residual in the level 1.5 data, the Spectral Deconvolution Algorithm as described by O'Neill et al. (2003) and verified by Chew et al. (2011) and Kaku et al. (2014), is utilized. Retrievals of several aerosol-related parameters, including effective radius, spectral single scattering albedo and upwelling and downwelling aerosol forcing efficiencies are also obtained from the AERONET inversion products (Dubovik and King, 2000).

No AERONET data are available at the 550 nm spectral channel. To be consistent with the MODIS AOT data, AERONET τ550 values are derived by interpolating AERONET AOTs reported at the 500 and 675 µm channels using a method described in Shi et al. (2011). While there are a number of AERONET sites installed in mid-to eastern United States, four observed the nature of the plume particularly well: Grand Forks, North Dakota, (47.91∘ N, 97.33∘ W); Sioux Fall, South Dakota (43.74∘ N, 96.63∘ W); Ames, Iowa (42.02∘ N, 93.77∘ W), and Bondville, Illinois (40.05∘ N, 88.37∘ W). These are labeled in Figs. 1 and 3a, c and e, with 500 nm fine-mode AOTs listed in Fig. 2e.

Over land, MODIS DT aerosol data are available over dark surfaces such as non-desert regions (Levy et al., 2013), and in this study, the Terra MODIS nadir 10 km resolution τ550 retrievals are used, which best correspond to the midday 12:00 LST/18:00 Z forecast period evaluated. The accuracy of C6 MODIS AOT is reported to be on the order of 0.05 + 15 % × AOT (Levy et al., 2013), although individual retrieval uncertainties may be higher (e.g., Shi et al., 2011). As verification, Terra MODIS retrievals were compared to AERONET sites listed above for the period of 29 June through 4 July 2015, with five data points available at Grand Forks having τ550 spanning from 0.88 to 3.7, three at Sioux Falls spanning 0.12 to 3.98, and one at Ames with a τ550 of 0.58. Regression showed MODIS having a slight 10–20 % high bias, and outstanding regression coefficients (r2=0.98). However, AOT retrievals failed for τ550 above ∼ 4 due to saturation of the aerosol signal.

The hypotheses developed for this effort originated from observations of significant temperature forecast errors in the Dakotas in association with the central Canadian smoke plume. Thus a key comparison for forecasted and observed daily maximum temperatures is performed between Grand Forks (47.95∘ N, 97.18∘ W), in the center of the plume, and Bismarck (46.77∘ N, 100.75∘ W), 300 km to the west and outside of the plume. These sites are marked in Fig. 3a, c. Official forecast data were obtained from the National Weather Service-issued text weather reports (Point Forecast Matrices and Climate Reports) from the Grand Forks and Bismarck, ND, stations respectively. The NWS Point Forecast Matrices include forecasted daily maximum near-surface air temperatures and other weather conditions. The observed daily maximum surface temperatures are obtained from the NWS Climate Reports which, per the ASOS Users' Guide (http://www.nws.noaa.gov/asos/aum-toc.pdf, accessed on 29 October 2015) have accuracy at the half degree Celsius level. The archived NWS weather reports from 15 June–14 July 2015 are obtained from the Iowa Environmental Mesonet (IEM) site (https://mesonet.agron.iastate.edu/; Todey et al., 2002), which also hosts the NWS-issued Morning Temperature and Precipitation Summary, from which the observed daily maximum surface temperatures for Roseau (48.86∘ N, 95.70∘ W) and Baudette (48.73∘ N, 94.61∘ W), MN, were retrieved, as these were not available from the NWS Climate Reports.

To supply surface observations for comparisons to forecast models over the greater upper Midwest and Upper Mississippi and Ohio River Valley study area, Automated Surface Observing System (ASOS) surface data are obtained from the Iowa Environmental Mesonet (IEM) site (https://mesonet.agron.iastate.edu/) for North Dakota, South Dakota, Nebraska, Minnesota, Iowa, Alabama, Arkansas, Illinois, Indiana, Kansas, Kentucky, Missouri, Mississippi, Oklahoma, and Tennessee (Fig. 3a and e). The ASOS data include surface temperature (2 m standard), dew point (2 m standard), wind speed (10 m standard), and direction (10 m standard) as well as visibility conditions. The surface temperature data used in study have the accuracy on the order of 0.5 ∘C for the normal temperature range of -50 to 50 ∘C (ASOS user's guide, http://www.nws.noaa.gov/asos/aum-toc.pdf, accessed on 29 October 2015).

The next step in this analysis was to compare model midday (12:00–13:00 LST, 18:00 Z) surface temperature forecasts with ASOS observations, and relate differences to the location of the smoke plume. 18:00 UTC was selected because it is near local noon and is only 15 min off the Terra satellite overpass time (17:45 UTC) for North Dakota on 29 June 2015. The primary model set used for comparison is the deterministic forecasts from ECMWF. Two-meter surface temperate forecasts for the 18:00 Z valid times (30 and 54 h forecasts) are examined. The 29 and 30 June 2015 18:00 Z forecasts and ASOS observations are examined in detail. The forecast error statistics are also examined for these ASOS sites from 15 June through 14 July.

Model data from the operational version of the European Centre for Medium-Range Weather Forecasts Integrated Forecast System (ECMWF IFS) were used. Forecast data are available 3-hourly from the 00:00 and 12:00 UTC analysis. Analyses are also available at 06:00 and 18:00 UTC from the four-dimensional variational (4D-Var) system with ensemble-generated flow-dependent background error statistics. The current resolution of the ECMWF IFS is approximately 16 km (T1279 spectral) with 137 vertical levels. More information is available here: https://software.ecmwf.int/wiki/display/IFS/CY41R1+Official+IFS+Documentation.

In addition to ECMWF, two other model data sets were also examined. Forecast surface temperatures at 24 and 48 h forecast intervals from the Global Ensemble forecast System (GEFS) UKMO UM ensemble, at 18:00 UTC, were obtained from the THORPEX Interactive Grand Global Ensemble (TIGGE) data archive (Bougeault et al., 2010). The NCEP GEFS data are available on a global scale, with a regridded 0.5 × 0.5∘ (latitude/longitude) spatial resolution at 00:00, 06:00, 12:00, and 18:00 UTC. Gridded statistical interpolation is included as the data assimilation method for the control analysis (http://tigge.ecmwf.int/models.html). The 2 m air temperatures from the NCEP model runs are used. Note that the NCEP data record is not complete for the selected study period, and missing data are listed in Table 1.

The UKMO data are available at a regridded spatial resolution of 0.5 × 0.5∘ (latitude/longitude) on a global scale. The 4D-Var assimilation scheme is included for the control analysis (http://tigge.ecmwf.int/models.html). The reported 1.5 m air temperature from the UKMO model runs are used in this study. Other details of the UKMO and NCEP models can be found from Bougeault et al. (2010) and the TIGGE website (http://tigge.ecmwf.int/models.html).

To assist the analysis, data from a number of sources are utilized. Descriptions of fire activity were obtained from the Canadian Interagency Forest Fire Center (CIFFC) situation reports (http://www.ciffc.ca/, last accessed 1 December 2015). MODIS fire hotspot data were also used (MOD14/MYD14, Justice et al., 2002). Soundings with temperatures, dew points, and mixing ratios from radiosonde data at Aberdeen, SD, are used (45.45∘ N; 98.4∘ W). To diagnose lower middle troposphere flow patterns, ECMWF reanalysis was utilized (Dee et al., 2011). Finally to assess the transport trajectory of individual smoke parcels, the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model (Draxler and Hess, 1997) is also used. The HYSPLIT model computes trajectories of air parcels, both in forward and backward modes, given the geolocation and altitude of an air parcel, as well as model initiation and spinning times.

The smoke event described here originated in a set of fires in the Northwest Territories and northern Alberta and Saskatchewan that were initiated ∼ 23 June 2015, as discussed by CIFFC and observed in MODIS fire hotspot anomalies. These fires were likely the result of lighting in association with widespread thunderstorm activity in central Canada lasting several days. By 27 June 2015 (Fig. 2a), over 60 individual fires or complexes were visible in the MODIS fire product, with over 30 fires reported greater than 1000 Ha by the CIFFC. 28 June 2015 MODIS imagery (Fig. 2b) showed significantly enhanced fire activity, with thick palls of smoke being visible over central Canada. Comparison of MODIS fire to the CIFFC suggests that a number of major fire complexes were missed in the satellite product, with significant burning being missed in central Saskatchewan and Manitoba. Nevertheless the dense smoke was present. By 29 and 30 June, smoke was clearly being transported across the Midwest, through the Upper Mississippi and Ohio River Valley, and into the Carolinas.

The rapid transport of this smoke event was related to a persistent long-wave high over the western United States, and corresponding trough over the eastern seaboard. The resulting lower free tropospheric winds were west–northwesterly (e.g., see 700 hPa height and wind analysis from the ECMWF reanalysis in Fig. 4). These winds veered to north–northwest at 500 hPa.

Thus, smoke was channeled into the upper Midwest from central Canada. Smoke transport was further enhanced by a fast-moving shortwave and cold front, with 700 hPa winds at ∼ 25 m s-1 (evident from the upper Great Lakes through Iowa and Nebraska in Fig. 4a). This shortwave resulted in the first tongue of smoke entering the United States through central North and South Dakota on 28 June (Fig. 2b). The most dramatic day, 29 June 2015, saw the rapid transport of the major smoke pall from northern Canada into the central Midwest behind the aforementioned shortwave with mid-visible AOTs in the upper Midwest above 4 (Fig. 2c and e). Embedded in this smoke event were a set of smaller disturbances and associated wind enhancements across south central Canada and the upper Midwest (Fig. 4b). At the core 18:00 Z analysis time for this study, peak winds associated with the shortwave ranged from west–northwesterly 10 m s-1 at 950 hPa, veering to northwesterly to 25 m s-1 at 500 hPa.

A major shift in the pattern occurred on 30 June. Smoke from the previous day had now advected into the Upper Mississippi and Ohio River Valley. Indeed, HYSPLIT trajectories suggest smoke over Grand Forks should have advected to South Central Illinois within 24 h. At the same time, a low and occluded front moved into the Dakotas, bringing heavy cloud cover, some rain, and more zonal winds (Figs. 2d, 4c). At the same time, observed fire activity diminished. Over the first week of July, while smoke was still clearly present at moderately high levels in the upper Midwest (Fig. 2e), the plume structure was not as nearly dramatic. Smoke was also frequently embedded in cloud layers. By 6 July, a significant cold front moved through the area, largely putting the smoke event to an end (e.g., Fig. 2e). From 23 June–9 July, CIFFC reported that ∼ 2 000 000 Ha were burned.

Operational radiosonde releases within the 29–30 June main smoke event are rare due to the unfortunate trajectory of the main plume, perfectly between the Bismarck and International Falls stations in the north and the Omaha/Topeka/Springfield corridor and Chahassen/Davenport/Lincoln corridor in the south. Further, the 00:00 Z and 12:00 Z releases are nominally in the plume region in the morning and evening. However, there were two radiosondes related to the event, collected under cloud-free sky conditions, the 29 June 12:00 Z and 30 June 00:00 Z release at Aberdeen (Fig. 5). Even though the site is on the edge of the main plume, the MODIS-inferred τ550 was still high ∼ 2. Clearly, the soundings are dry, with temperature and dew point profiles indicative of relative humidity on the order of 40–50 %. Water vapor mixing ratios dropped to below 2 g kg-1, by 600 hPa, or 4 km.

Unfortunately for ascertaining plume altitudes for this event, no Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) lidar data are available until 30 June due to solar flare activity. Over the remaining days, orbit and clouds prevented clear operations across the axis of the plume. However, we can infer from the early morning and afternoon 1 July overpasses over the east coast that this plume was largely below 5 km in altitude. This is corroborated by the Aberdeen sounding, which showed very low water vapor mixing ratios above 4 km in altitude. In regard to smoke base, despite the very high AOTs, surface PM10 measurements hardly registered the plume passage. Based on all of the above information, we are confident that the plume was confined to the lower to middle free troposphere.

Estimates of particle size and optical properties of the smoke plume were retrieved from the four core AERONET sites used in this analysis (Table 2). These retrievals were collected from 29 June–3 July over the study area. Particle sizes were fairly stable over the United States, with an effective radius of ∼ 0.165 µm, or a volume median diameter of ∼ 0.38 µm. This value is large in comparison to more typical boreal fires (e.g., Reid et al., 2005), but well within values found for mega events from Canada (e.g., 2002 Quebec fire with τ550 > 5; Colarco et al., 2004; O'Neill et al., 2005). Retrieved single scattering albedo was also consistent and within expected values, ∼ 0.94 in the mid-visible. In regard to this analysis of surface temperature, what we are most interested in is forcing efficiencies, which ranged from -48 to -58 W m-2τ550-1 for the top of the atmosphere. For retrieved surface forcing efficiencies, values varied more between sites. Grand Forks, Sioux City, and Bondville all agreed well, ranging from -118 to -124 W m-2τ550-1. Note that top-of-atmosphere (TOA) (surface) aerosol forcing efficiency is defined as the amount of change in upward (downward) shortwave radiation at TOA (surface) for a unit change in AOT. Negative surface aerosol forcing efficiencies indicate a reduction in shortwave radiation reaching the surface and mostly likely linkage to a decrease in surface temperature. However the Ames site had several outlier retrievals, leading to a higher magnitude downward forcing efficiency of -165 W m-2τ550-1, due to noticeably lower near-infrared single scattering albedos and slightly smaller size. This departure was consistent through the event. One explanation of this difference between Ames and other sites is that the averaged AOT (0.5 µm) is around 0.5 for the Ames site, whereas the averaged AOT (0.5 µm) for the other sites ranges from 0.8 to 1.4 (Table 2) . Thus, sampling bias is likely a factor.

Figure 3a, c, and e show the RGB true color images of the smoke event over the upper Midwestern United States on 28 June (17:00 and 17:05 UTC) and 29 June (17:45 UTC), and over the Upper Mississippi and Ohio River Valley on 30 June (16:50 and 16:55 UTC), constructed using the Collection 6, Level 1b Terra MODIS data. Figure 3b, d, and f show the corresponding Terra MODIS C6 DT τ550 for the same study periods as Fig. 3a, c, and e. The observed surface temperatures reported from ASOS stations are overplotted on Fig. 3a, c, and e from North Dakota, South Dakota, Nebraska, Minnesota, and Iowa on 28 and 29 June, and from Alabama, Arkansas, Iowa, Illinois, Indiana, Kansas, Kentucky, Missouri, Mississippi, Nebraska, Oklahoma and Tennessee on 30 June. Each data point in Fig. 3a, c, and e represent the averaged observations within ±10 min from 18:00 UTC of each given day for a given station. The observations from 18:00 UTC are selected as both model analyses and forecasts are available at this time, enabling us to further explore differences between modeled and observed surface temperatures with respect to smoke aerosol properties.

Shown in Fig. 3a, on 28 June, is a stripe of smoke aerosol plume which starts to appear over the upper Midwest region. The overall aerosol loadings are still relatively low (τ550 < 0.8 for the stripe of plume and less than 0.2 for most other regions) across the domain. A mild temperature difference on the order of 1–2 ∘C is observed between eastern and western North Dakota. In comparison, on 29 June, a thick smoke plume is observed over the eastern Dakotas and western Minnesota with significant MODIS DT τ550 values of 2–5. While warmer surface temperatures of 27–32 ∘C are observed over the Western Dakotas where lighter aerosol loadings (less than 0.6) are found, surface temperatures of 22–24.5 ∘C are found over the eastern Dakotas and western Minnesota. The sharp spatial gradient in surface temperature on the order of 5 ∘C between eastern and western North Dakota on 29 June 2015, matching the smoke plume pattern, shows the potential influence of the smoke aerosol particles on the observed surface temperatures.

On 30 June, the smoke plume migrates to the Upper Mississippi and Ohio River Valley, as shown in Fig. 3e and f. Note that surface observations are obtained around 18:00 UTC, and the Terra MODIS overpasses are 16:50–16:55 UTC. Thus, there is ∼ 1 h difference between surface- and satellite-based observations. Still, as shown in Fig. 3e, especially over Missouri (Center of Fig. 3e), lower surface temperatures are visible over regions with heavy aerosol loadings, which again, reinforces the finding from the 29 June case.

To assess the degree to which the smoke event impacted forecast temperatures, we first performed a hand analysis of the difference in forecast and observed surface temperatures between Grand Forks and Bismarck as reported from the National Weather Service for 29 June. These two sites correspond to the middle and just outside of the main plume. Figure 6 shows the forecast maximum surface air temperatures up to 96 h for Grand Forks and Bismarck for 29 June 2015. Filled diamonds represent forecast update time. The final daily maximum temperatures, nominally 25.6 and 33.3 ∘C for Grand Forks and Bismarck respectively, are also shown. For 29 June, an ∼ 8 ∘C difference is seen between sites in and out of the plume even though, typically, the high temperatures between Grand Forks and Bismarck are highly correlated. For the month surrounding the event (15 June–14 July, excluding 29 June), Bismarck was historically warmer than Grand Forks by 1.0 ± 2.0 ∘C, with a correlation of 0.90. Forecasters are well aware of this natural difference and hence account for it in their forecasts. It is also noteworthy that while the daily maximum near-surface air temperature forecasts for 29 June remain unchanged since 27 June for Bismarck, the Grand Forks NWS made a -2.8 ∘C (-5 ∘F) adjustment for their daily maximum near-surface air temperature forecast at around 10:00 local time on 29 June 2015, possibly to compensate for the initial unexpected surface cooling due to the thick smoke aerosol plume. Despite the higher winds in the lower to middle free troposphere, 29 June was a relatively calm day with moderate winds at the surface, (∼ 3–5 m s-1). Taking all of the above factors into consideration, it is hypothesized that the smoke plume with an AERONET-reported daily mean τ550 of ∼ 3.4 introduced a surface temperature cooling for Grand Forks of ∼ 5 ∘C. This is equivalent to a daytime aerosol cooling efficiency of ∼ -1.5 ∘C/τ550, given that the daily averaged τ550 is 3.4 as reported from Grand Forks AERONET station. Meanwhile, the reported MODIS τ550 value over Bismarck was ∼ 0.35.

While observations from Bismarck and Grand Forks represent measurements at the diffuse western edge and the central smoke plume, Roseau and Baudette, MN, which are close to Grand Forks, are selected to represent the eastern diffuse edge of the smoke plume. As listed in Table 3, τ550 values are 0.84 and 1.06 for Roseau and Baudette respectively at 17:45 UTC, 29 June 2015, as approximated from MODIS DT retrievals. Note that using the observed surface temperature differences between Grand Forks and the two selected cities in MN for evaluating aerosol direct cooling effect is not ideal, as surface temperatures from Roseau and Baudette may also be modulated by nearby lakes. Further, lower correlations in daily maximum temperatures, around 0.85, are found between Grand Forks and the other two locations in MN. Still, Grand Forks is around 2 ∘C warmer than Roseau and Baudette on a monthly average (Table 3). However, on 29 June 2015, a much smaller temperature difference of 1.1 ∘C is found between Grand Forks and Baudette, and Roseau is actually 0.6 ∘C warmer than Grand Forks. Both cases may indicate the potential smoke cooling effect. It is noteworthy that the NWS made a -1.7 ∘C (-3∘F) adjustment for the forecasted daily maximum temperatures on 29 June 2015 for both Roseau and Baudette, MN, possibly to compensate for the unexpected smoke-aerosol-induced surface cooling. Lastly, besides the aerosol direct surface cooling effects, surface temperatures could also be impacted by differences in dynamical environments, which adds uncertainties to the study.

The above hand analysis provides a benchmark estimate of the cooling efficiency of the Canadian smoke plume. To test this value through an objective analysis, we compared this finding to surface forecast errors focusing on the ECMWF models, starting with the 29 June case. After this analysis, we extended the study to the NCEP and UKMO models and for the 30 June case as well. A synopsis of findings is provided in Fig. 7, where we show (a) the relationship between recorded 18:00 Z temperature to MODIS τ550; (b) the difference of ASOS observation to ECMWF 30 h forecast against τ550; and (c) and (d), the corresponding overlay of observation minus ECMWF 30 h forecast mapped over the 29 and 30 June investigation domains. The plots are generated using measurements from ground stations as shown in Fig. 3c and e. In addition, over the center of the smoke-aerosol-polluted regions, the smoke plume is so optically thick that the MODIS aerosol retrieval scheme failed to report τ550 values. Thus, the closest MODIS τ550 value within 1∘ latitude/longitude of a given ground station is used to represent the τ550 value of that station where there is no MODIS aerosol retrieval available. Note that this assumption may introduce a bias in the estimated MODIS AOTs.

The question of how important the smoke cooling efficiency is to numerical weather prediction is fundamentally related to the overall skill of the natural model. Models with large root-mean-square errors (RMSEs) will mask the aerosol signal; such models have more important sources of error. Models with high skill, on the other hand, naturally are sensitive to higher order terms. In this section, we examine this phenomenon and by evaluating near-surface air temperature forecasts from ECMWF, UKMO, and NCEP in the upper Midwest region with respect to smoke τ550 for the 29 June case. As the first step, baseline uncertainties in near-surface air temperatures from NCEP, UKMO, and ECMWF model runs are evaluated (Table 6) using surface observations from ground stations, as shown in Fig. 3g. To construct Table 6, 0-, 24(30), and 48(54) h model forecasts at 18:00 UTC from 15 June to 14 July are collocated with ground-based ASOS data (the numbers included in parentheses are for ECMWF). The means of the differences between observed and forecasted temperatures are computed for the 0, 24(30), and 48(54) h model forecasts and are represented by ΔT0h, ΔT24/30h, and ΔT48/54h, respectively, in this study. Indicated in Table 6, similar ΔT48/54h values of around -1 ∘C with similar 1 standard deviation of ∼ 2.5 ∘C are found for the 48 h forecasted near-surface air temperatures from UKMO and NCEP. A smaller ΔT48/54h of -0.4 ∘C, with a smaller 1 standard deviation of 2.0 ∘C, is found for the 54 h forecasted 2 m air temperatures from ECMWF. ΔT24/30h and 1 standard derivation of ΔT24/30h of around -0.8 and 2.3 ∘C are found for the 24 h forecasted 2 m air temperatures for NCEP, and the values are -0.6 and 2.1 ∘C for the 24 h forecasted 1.5 m air temperatures for UKMO. Again, smaller values of ΔT24/30h and 1 standard derivation of -0.2 and 1.9 ∘C are found for the 30 h forecasted 2 m air temperatures for ECMWF. In comparison, the 0 h forecasts of near-surface air temperatures exhibit much smaller standard derivations of the differences to the observed surface temperatures; around 1.5 ∘C from all three models.

The root-mean-square error (RMSE) values for the 0, 24(30), and 48(54) h model-forecasted near-surface air temperatures are 2.3, 2.5, and 2.7 ∘C for NCEP data, 1.3, 2.2, and 2.7 ∘C for UKMO, and 1.6, 1.9, and 2.0 ∘C for ECMWF model runs, respectively. The same analysis has also been conducted for the 30 June 2015 case. Not surprisingly, the reported RMSE values are consistent for both the upper Midwest and the Ohio River Valley regions. For example, the computed RMSE values for the 30 June case are 1.5, 2.0, and 2.2 ∘C for the 0, 30, and 54 h ECMWF forecasts. The RMSE values for the 0, 24, and 48 h NCEP and UKMO model forecasted near-surface air temperatures are 1.9, 2.2, 2.5 ∘C, and 1.3, 2.1, 2.5 ∘C, respectively.

The RMSE values represent the baseline cases for the modeled uncertainty in near-surface air temperatures. Theoretically, the effect of aerosols on weather forecasts can likely be detected if the aerosol-induced surface cooling is larger than the baseline uncertainties in the modeled near-surface air temperatures. Given a rough estimation of ∼ -1.5 ∘C/τ550 for the daytime smoke Cτ, the changes in τ550 need to be above ∼ 1.5–2 for the aerosol-induced cooling effect to be observable from the 48(54) h model forecasts. Similarly, τ550 values of ∼ 1–1.5 and ∼ 1.5 are required for the aerosol-induced cooling effect to be detectable from the 0 and 24(30) h model forecasts.

The deterministic forecasts from ECMWF were obtained from Angela Benedetti. All other data sources were obtained online with data links mentioned in the data set section (accessed on or before 1 December 2015).