Abstract
The cumulus parameterization is widely recognised as a crucial factor in tropical meteorology: this paper intends to shed further light on the effects of convection parameterization on tropical cyclones’ numerical predictions in the “grey zone” (10–1 km grid spacing). Ten experiments are devised by combining five different convection treatments over the innermost, 5 km grid spacing, domain, and two different global circulation model datasets (IFS and ERA-Interim). All ten experiments are finally analysed and compared to observations provided by the National Hurricane Center’s best track record and multisatellite rainfall measurements. Results manifestly point to the superiority of employing no convective parameterization at the scale of 5 km versus the usage of any of those provided by WRF to reproduce the case study of Hurricane Tomas, which hit the Lesser Antilles and Greater Antilles in late October and early November 2010.
1. Introduction
Tropical cyclones’ (TCs) numerical predictive ability has improved in recent decades thanks to a deeper understanding of the physical mechanisms responsible for the tropical cyclogenesis and subsequent spatiotemporal evolution of these deep moist convective structures and also thanks to the improvement in supercomputing capabilities, which make it possible now to simulate TC at very high resolutions [1, 2].
Several studies [3–6] have explored the Weather Research and Forecasting (WRF) model predictive ability sensitivity to its different physical parameterizations, that is, to the effects of varying the computational horizontal and vertical resolutions and the number of domains.
Davis et al. [3] performed, with Advanced Research Hurricane (AHW) component of WRF model, a very comprehensive study of the effects of increased horizontal resolution on hurricane wind forecasts: they compared the accuracy of hurricane track, intensity, and structure in a set of 69 forecasts performed at each of two horizontal grid increments with the AHW-WRF model. These forecasts covered 10 Atlantic tropical cyclones: 6 from the 2005 season and 4 from 2007. The forecasts were integrated from identical initial conditions produced by a cycling ensemble Kalman filter [7]. The high-resolution forecasts used moving, storm-centred nests of 4 km grid spacing and 1.33 km grid spacing, with explicit convection. The coarse-resolution forecasts consisted of a single 12 km domain, with Kain-Fritsch cumulus parameterization. The number of vertical levels was 35, and the adopted PBL scheme was YSU (Yonsei University), while the microphysics one was WSM5 [8]. In a nutshell, storm position errors showed no statistically meaningful differences between the coarse-resolution and high-resolution simulations, while storm intensity, defined as the maximum instantaneous 10 m wind in the model, was slightly better forecast in the nested simulations than in the 12 km forecasts.
In a study of the sensitivity in simulations of TC Jal (2010) to WRF physics parameterizations, Chandrasekar and Balaji [4] used a setup of 3 domains of 90, 30, and 10 km of horizontal resolution, respectively, with the scope of determining the best combination of physics schemes for track and intensity forecasts. They showed that their best set of physics combination worked properly for track prediction but variably for predicted cyclone intensity, with the cumulus parameterization having more impact on intensity prediction than any of the other physics subgrid schemes, whereas for both track and intensity prediction skill, it was the cumulus together with the PBL and microphysics parameterizations that played a larger role than the other physics schemes (land surface and radiation) of the model.
Concerning the vertical grid resolution (VGR) sensitivity, Zhang et al. [6] performed more than 60 five-day numerical simulations of two different idealized hurricane vortices, and they varied the VGR in different portions of the atmosphere with the operational version of the HWRF model. Their results provided some important hints for designing appropriate VGRs to predict reasonably hurricane intensity and inner-core structures under different environmental flows: certain minimum VGRs (around 40) should be configured to produce reliable intensity forecasts with more detailed inner-core structures; relatively higher (lower) VGRs in the lower (middle) levels appear to be more preferable than higher VGRs in the upper levels.
Notably Biswas et al. [5] explored the sensitivity of tropical cyclone track and intensity forecasts to different convective schemes by using the Hurricane Weather Research and Forecasting (HWRF) system. The HWRF configuration similar to the 2012 operational system was used in this study: the parent, middle, and the inner domains have 27, 9, and 3 km horizontal grid spacing, respectively. A control configuration adopting the HWRF Simplified Arakawa Scheme (SAS) was compared with the Kain-Fritsch and Tiedtke schemes, as well as with newer implementation of the SAS. Averaging over almost 250 cases, they showed that the SAS scheme produced the best track forecasts (Figure of [5]) till a lead time of 72–84 hours followed very closely by Kain-Fritsch scheme.
On average, hurricane track forecasts have improved much faster than the prediction of their intensity [10–12]. Such a challenge is strictly connected to the ability of a model to represent the dynamic processes of a TC, which means recognising the role played by the physical parameterizations as a major critical and limiting factor in numerical weather prediction models. Some studies prove that TCs prediction is very sensitive to the choice of the cumulus scheme [13–15]. Other studies say that the explicit treatment of convection provides instead a more realistic representation of the hurricane, eye and eyewall structure, and storm intensity [5, 16].
In this study, the fully compressible and nonhydrostatic Weather and Research Forecasting (WRF) model (version 3.4) (Skamarock et al. 2015) is used to analyse the sensitivity of different convective parameterizations to the prediction of Hurricane Tomas, which occurred in late October and early November of 2010 in the Caribbean Sea.
Hurricane Tomas represents an interesting case study. First of all, it was responsible for a total of 71 deaths in the Lesser Antilles and Greater Antilles, with economic damage in the order of 750 million USD [9].
Secondly, it was the latest hurricane on record (1851–2014) in the calendar year to strike the Windward Islands; thirdly, Tomas was also the fourth-longest lived post-October 28 tropical cyclone in the period of 1945–2015.
Furthermore, the intensity of Tomas was poorly handled in the NHC forecasts; official forecast intensity errors were greater than the mean official errors for the previous 5-year period at all forecast intervals and considerably greater at 48 through 120 h [9].
Finally, Hurricane Tomas experienced also a rapid intensification (http://www.nhc.noaa.gov/aboutgloss.shtml) phenomenon [17] with an increase in the maximum sustained winds of 30 kt in the period from 0600 UTC on 4 November (40 kt) to 0600UTC on 5 November (70 kt). Furthermore, Atlantic basin hurricane activity in 2010 was quite high due to the combination of anomalously warm sea surface temperatures and a rapidly developing La Niña event. Such situation led to favourable dynamic and thermodynamic conditions for tropical cyclogenesis and intensification.
The layout of the rest of the paper is as follows: Section 2 describes the case study and Section 3 provides an overview of the model setup, while Section 4 presents the study’s results. Conclusions are drawn in Section 5.
2. Description of the Case Study: Hurricane Tomas
Tomas was the latest hurricane of the 2010 Atlantic hurricane season, bringing floods and landslides to Haiti [18].
As thoroughly described in Pasch and Kimberlain [9], Hurricane Tomas formed from a tropical wave that moved off the coast of West Africa on October 24. Organised convective activity first showed up on 26 October and gradually evolved until around 06UTC on 29 October, when the system transitioned to a tropical depression with a minimum central sea level pressure (SLP) of 1006 hPa located about 740 kilometers southeast of Barbados. Around 12UTC on 29 October, moving west-northwestwards, the system continued to intensify and to strengthen, becoming a tropical storm due to an enhanced upper-tropospheric outflow. After the disturbance slowly approached the Windward Islands, it headed west-northwest and around 09UTC on 30 October the 997 hPa centre was located over the southern coast of Barbados.
Around 12UTC on 30 October, Tomas became a category 1 hurricane on the Saffir-Simpson scale, and around 20UTC the intense northern eyewall passed over St. Lucia reaching its peak intensity of 155 km/h. Later that day, Tomas became further organised and intensified to category 2 with winds of 160 km/h and minimum central pressure of 983 hPa. After crossing the Windward Islands, the hurricane weakened to tropical storm force by 00UTC on 1 November due to the impact of a moderate southwesterly shear and dry air intrusion at mid-levels that caused convection to wane near the centre. Between 00UTC on 1 November and 00UTC on 3 November, Tomas became increasingly disorganised with an elongated, tilted circulation that degenerated to a tropical depression with a minimum pressure of 1006 hPa while travelling westwards over the central Caribbean Sea, heavily affected on its west by shear and dry air. Tomas then reintensified to tropical storm strength around 18UTC on 3 November, featuring better organised deep convection and low-level circulation, and by 06UTC on 4 November the storm became stronger and attained hurricane status around 06UTC on 5 November through a rapid intensification process, after it turned northward to northeastward ahead of an unusually deep mid-tropospheric trough.
Later on the same day, the category 1 hurricane moved along the Jamaica Channel, about 60 km west-southwest of the western tip of Haiti, and then through the Windward Passage up to the early hours of 6 November when it weakened once again to a tropical storm because of the increasing vertical wind shear. The constant increase in vertical shear and an intrusion of very dry air caused Tomas to downgrade into an extratropical disturbance around the initial hours of 8 November. Best track data are provided in Table 1.
3. Model Setup
The fully compressible and nonhydrostatic Weather Research and Forecasting (WRF) model (version 3.4) [19] is used for this work.
The two adopted domains cover a region including Cuba and Haiti as shown in Figure 1: the outer (the blue area) and inner (the yellow area) domain horizontal resolutions are, respectively, 15 and 5 km. The analyses presented in this paper refer to the results of the nested domain, constituted by a grid of 786 by 504 points. Both domains have 41 vertical layers with a higher concentration (about ten) of them in the PBL, in agreement with the recommendations of Zhang et al. [6] about VGRs.
The innermost domain’s grid spacing, 5 km, was chosen as a good compromise between computational constraints and recent research results [1, 20], which suggest that increasing the horizontal resolution down in the cloud-permitting range (5–1 km) does not necessarily lead to an actual improvement of the predictive ability of a storm’s spatiotemporal evolution, while further exploring the sensitivity to other model features, such as the physical parameterizations, in particular the convective ones, and initial and boundary conditions, can be of great value [1].
On the basis of such considerations, five different approaches to the convection treatment in the innermost domain are considered: varying the cumulus parameterization among four of those available in WRF on one hand and adopting an explicit (no scheme) convection setup on the other hand.
The first convection closure selected has been the Kain-Fritsch (KF, [21]) scheme, which is based on the evaluation of the entrainment and detrainment rate in relation to the local environment profile [22]. The second one has been the Betts-Miller-Janjic (BMJ, [23, 24]) scheme, which is shown to work well in tropical regions [13]. The third scheme chosen has been the Grell-Devenyi (GD, [25]) scheme, which includes an ensemble cumulus scheme in which multiple mass-flux type cumulus schemes with different updraft, downdraft, entrainment, and detrainment parameters and precipitation efficiencies are run within each grid box and then the results are averaged to give feedback to the model [13]. Finally, the last scheme employed is the new Grell ensemble (NGE), similar to the Grell-Devenyi scheme but more suitable for higher-resolution domains, allowing for subsidence in neighboring columns and without the quasi-equilibrium approach among the ensemble members [26].
The fifth approach has been to employ the KF scheme in the outer domain, as one of the reference cumulus schemes for hurricane modelling [3, 5], and then none in the inner domain so as to treat convection explicitly in that scale. Such choice arises from some studies [5, 16] that underline that grid spacing between 8 and 4 km resolves adequately the TC circulation with explicit convection treatment.
Among the several microphysics schemes available in the WRF-ARW 3.4 model, the Thompson scheme [27, 28] has been selected; it is a double-moment scheme, widely used by the hurricane modelling community [29, 30] with 6-class microphysics, with ice and rain number concentrations also predicted and time-split fall terms. In the Thompson scheme, riming growth of snow is required to exceed depositional growth of snow by a factor of 3 before rimed snow transfers into the graupel category, resulting in more realistic values for graupel mixing ratio in the eyewall [1].
For the planetary boundary layer parameterization, the YSU scheme [31] has been adopted, as it is the most chosen PBL scheme in Monthly Weather Review, the Journal of the Atmospheric Sciences, and the Quarterly Journal of the Royal Meteorological Society articles that presented tropical cyclone simulations using MM5 or WRF in the period of 2006–2010 [32].
Each of the five convection treatments has been explored with two different sets of initial and boundary conditions: the operational global meteorological forecasting model, ECMWF’s (European Centre for Medium-Range Weather Forecasts) IFS (Integrated Forecast System, 0.125° and 3-hour forecast boundary conditions, [33]) and ECMWF’s ERA-Interim (0.7° and 6-hour reanalysis boundary conditions, [34, 35]) reanalysis product.
While IFS and ERA have the same parent modelling system, they are very different in terms of grid spacing and forcing data, thus justifying their adoption for this modelling study. The dynamical core of IFS is hydrostatic, two-time-level, semi-implicit, and semi-Lagrangian and applies spectral transforms between grid-point space (where the physical parameterizations and advection are calculated) and spectral space. In the vertical direction, the model is discretised using a finite-element scheme. A reduced Gaussian grid is used in the horizontal direction. The IFS version used for Tomas’ initial and boundary conditions is Cy36r2 ((https://www.ecmwf.int/en/forecasts/documentation-and-support/evolution-ifs/cycles/cycle-36r2-detail)). The data assimilation system used to produce ERA-Interim is based on a 2006 release of the IFS (Cy31r2 (https://www.ecmwf.int/en/forecasts/documentation-and-support/evolution-ifs/cycles/cycle-31r2-summary-changes)). The system includes a 4-dimensional variational analysis (4D-Var) with a 12-hour analysis window. The spatial resolution of the dataset is approximately 80 km (T255 spectral) on 60 vertical levels from the surface up to 0.1 hPa.
All the simulations encompass the time period from 00UTC on 1 November to 00UTC on 6 November 2010, which includes Tomas’ increasingly disorganised phase (from 00UTC on 1 November to 00UTC on 3 November) and the subsequent rapid intensification phase (from 06UTC on 4 November to 06UTC on 5 November). Concerning the choice of the initialization timing, the recommendations of Davis et al. [36] and Xue et al. [37] have been taken into account. Davis et al. [3] found that there was no meaningful difference between storm position errors in the 12 km and nested higher-resolution forecasts for five landfalling Atlantic hurricanes during 2005. However, TC intensity (in terms of the maximum 10 m wind) was somewhat better forecasted on the nested grids than on the single 12 km grid, and the difference was statistically significant at 72 h and beyond. Xue et al. [37] argued that, for tropical cyclones below hurricane strength, more improvement in the track forecasts comes from high resolution (namely, 4 km) at 48 h and beyond. The high-resolution WRF forecasts produced significantly improved intensity forecasts at 24 h and beyond in terms of both maximum 10 m wind speed and minimum sea level pressure. Furthermore, Xue et al. [37] suggested that forecasts of longer ranges than the 48 h forecasts should also be explored: consequently, the 72 h and beyond range adopted here is worth to be explored.
Table 2 summarizes the ten different configurations that are adopted for all experiments presented in this study, which focuses on the temporal period corresponding to Tomas’ transition over the Greater Antilles area, spanning from 00UTC on 4 November 2010 to 00UTC on 6 November 2010, corresponding to the +72 to +96 hours’ time range, when Tomas evolved towards the hurricane conditions.
4. Results
A study on Hurricane Tomas’ sensitivity has been carried out to determine the most appropriate combination of cumulus parameterizations, IC, and BC to reproduce the best track, the minimum SLP evolution, the surface wind, and the precipitation associated with the tropical disturbance, with the highest degree of accuracy achievable. As indicated in Table 2, ten combinations of cumulus parameterizations, IC, and BC have been analysed in order to determine the best model convection treatment for Hurricane Tomas forecast. The results of the 5 km nested domain have been used for the study, spanning the simulations of 4 and 5 November 2010 which were initialized on 1 November 00UTC, and are compared against NHC’s best track data and satellite rainfall measurements.
4.1. Intensity and Trajectory Analysis
Figure 2 displays the tracks of all simulations and NHC’s best track from 00UTC on 4 November to 18UTC on5 November 2010 (black line). This figure shows also NHC’s track forecast cone (TFC). Its represents the probable track of the centre of the tropical cyclone in a way that two-thirds of the 5-year historical official forecast error fall within the cone: the black cone corresponds to the probability up to 3 days and the red one up to 5 days.
At first sight, it becomes clear that virtually all predicted centre locations of the proto-Tomas, except KF-KF with IFS and ERA experiments, fall within the limits of the 5-day TFC that ends at 00UTC on 6 November (note that the tracks’ last time step is 18UTC on 5 November). In addition, virtually all experiments are relatively closer together and either cross through or end their course near the Windward Passage. Differently, KF-KF with IFS and ERA experiments cross through Cuba and are clearly different from the rest.
Another notable characteristic of the tracks in Figure 2 is that, for fixed convective treatment, apart from the KF-E cases, they behave similarly independently by both IC and BC datasets (IFS and ERA-Interim) used. Contrarily, there is less similarity among simulations with the same IC and BC and different convective schemes. The track variance is much larger for different cumulus schemes than it is for ERA versus IFS.
This points clearly to the stronger degree of control exerted by the choice of convection parameterization over the TC spatiotemporal evolution than by the choice of global model data to feed the model. This statement is further supported by the vorticity field (-component) at an altitude of 2500 m predicted by the different experiments at 00UTC on 4 November (Figure 3). Analyzing the results, it is possible to notice that while the large-scale circulation features of the main TC (e.g., elongated convection line around 22.5°N and secondary convective region over the Lesser Antilles area) preserve similarities between the different experiments, even 72 hours after the initialization, the choice of the convective parameterization approach influences strongly the proto-Tomas evolution. Both KF-E cases generate finer scale convection activity around the TC intensifying core; conversely, parameterized convection for 5 km grid spacing domain tends in general to overintensify the predicted TC losing finer scale activity. This is really apparent at 00UTC on 4 November for the KF-KF-IFS, KF-KF-ERA, BMJ-BMJ-IFS, and BMJ-BMJ-ERA cases when even two neighboring rotating cores emerge. Similar considerations hold also for the water vapor field at m at 00UTC on 4 November (Figure 4), where all convection parameterized runs produce a moister TC core, especially KF-KF-IFS, KF-KF-ERA, BMJ-BMJ-IFS, and BMJ-BMJ-ERA cases, characterized also by more intense spiraling bands around the TC core itself as shown by the corresponding vorticity patterns.
In Table 3, the errors of the tracks with respect to the best track are shown: the experiments with the smallest average errors, about 200 km or below, are those with KF-E and GD-GD for both IFS and ERA datasets. In the case of the two experiments with GD-GD, while being very similar to each other, they are also among the closest to the best track for most instants of the period studied, although they tend to recurve towards the east more than the best track, which shows a more northward movement. In the case of the simulations with KF-E, they much better replicate the observed northeast direction of motion, although, for the run with IFS, that causes proto-Tomas to translate farther north than the best track indicates. On the opposite extreme, and as already mentioned before, we find the two experiments with KF-KF performing the worst of the whole study, with average track errors of more than 500 km and travelling outside the TFC for most of their timespan.
It is worth mentioning that the NHC official skill (OFCL, mean errors for the 5-year period of 2005–2009) baseline track forecast errors are about 180 km for 48 hours’ forecast, 270 km for 72 hours’ forecast, 360 km for 96 hours’ forecast, and 470 km for 120 hours’ forecast: this confirms once more the very good performances of experiments KF-E and GD-GD at +96 hours (00UTC on 5 November) and +114 hours (18UTC on 5 November), both below the OFCL target values, also slightly better than OFCL (Tomas) values for NHC operational models [9].
Having identified the experiment subsets of KF-E and GD-GD, both with IFS and ERA datasets, as best-performing from the point of view of the system’s spatial motion, now we turn to the analysis of the minimum SLP and maximum sustained surface wind.
The intensity of Tomas was poorly handled in the NHC forecasts. The official forecast intensity errors were greater than the mean official errors for the previous 5-year period at all forecast intervals and considerably greater at 48 through 120 h [9].
Figure 5 shows their respective evolutions in comparison to the best track for each simulation (black for SLP and red for wind; solid for best track and dashed for prediction) and also presents their average mean absolute errors (MAEs). From a quick visual inspection of Figure 5, it is apparent that the best performers in terms of both minimum SLP and maximum surface wind evolutions simultaneously are the KF-E with IFS and with ERA. Additionally, both KF-E-IFS and KF-E-ERA reproduce accurately the rapid intensification phase of Hurricane Tomas with an increase in the maximum sustained winds of 30 kt in the period from 0600UTC on 4 November (40 kt) to 0600UTC on 5 November (70 kt). They have the two lowest pressure MAEs (1.58 and 4.03 hPa, resp.) and also lowest wind MAEs (4.19 and 3.03 m/s), which is only bettered by the BMJ-BMJ-IFS simulation (2.80 m/s). In fact, BMJ-BMJ-ERA has an acceptable wind MAE of 5.12 m/s, making that experiment the fourth best from this viewpoint; however, both BMJ-BMJ simulations definitely overestimate the pressure fall with respect to the best track and the KF-E experiment group, resulting into exaggerated rapid intensification of the predicted Tomas experiment. Interesting enough, all these 4 experiments have intensity error lower than NHC official skill baseline errors, which is about 9.5 m/s for 96 hours’ forecast.
For the GD-GD experiments, which were among the best for track error, for the IFS run, the pressure and wind errors indicate greater error on 4 November than other experiments shown in Figure 5, while the ERA run significantly underestimates intensity, particularly on 5 November: GD-GD experiments are not able to capture properly the rapid intensification of Hurricane Tomas. Finally, turning to the results for the KF-KF and NGE-NGE groups, they are clearly the two parameterization experiment sets that overintensify the most both pressure and wind persistently during the whole timespan: this is due to the fact that KF and NGE parameterizations promote too strong convection (see Figures 4–6 and 7–9), which results in a greater intensity and makes the storm more resilient to shear and steering from environmental forcing.
(a)
(b)
(c)
Another aspect to note from the overall analysis is the relationship between the translational speed of the protohurricanes and their degree of intensification, an idea pointed out by Mei et al. [38]. Although our simulations do not use the 1D model option for ocean mixed-layer representation and hence do not incorporate wind-induced SST cooling with its negative feedback to the storm, yet, intriguingly enough, our results show that the simulations creating the most intense proto-Tomas perturbations, the groups KF-KF and NGE-NGE, are also those that bring the system furthest from their origin and so create the fastest translational speeds.
4.2. Precipitation
4.2.1. Analysis Methodology
Since this study focuses on the western portion of the Caribbean, where rain gauges and radar data are hardly available, rainfall amounts for Hurricane Tomas are estimated from satellite measurements, and in particular three datasets are considered: TRMM 3B42 (spatial resolution: 0.25 × 0.25 degrees; temporal resolution: 3 hours; [39]), GSMaP (spatial resolution: 0.1 × 0.1 degrees; temporal resolution 1: hour; [40]), and PERSIANN-CDR (spatial resolution: 0.25 × 0.25 degrees; temporal resolution: 3 hours; [41]). Among these, PERSIANN-CDR is the newest dataset, while TRMM 3B42 can probably be considered as the most reliable [42, 43]. The fields shown in Figure 6, for the simulations, spatial grid spacing of 5 km, and for GSMaP dataset, about 10 km, have been aggregated, for ease of comparison purposes, to the same resolution as TRMM 3B42 and PERSIANN-CDR datasets, 25 km.
To support the visual analysis of the 48-hour rainfall fields, a more quantitative (although without the temporal dimension) analysis of the phenomenon has been performed: then the Method for Object-Based Diagnostic Evaluation (MODE, [44–47], developed at Research Application Laboratory NCAR/Boulder (USA) and intended to reproduce a human analysts’ evaluation of the forecast performance, has been adopted.
The MODE analysis is performed using a multistep automated process. A convolution filter radius and a threshold are applied to the raw field to identify the objects. When the objects are identified, some attributes regarding geometrical features of the objects (such as location, size, aspect ratio, and complexity) and precipitation intensity (percentiles, etc.) are computed. These attributes are used to merge objects within the same forecast/observation field, to match forecast and observed objects, and to summarize the performance of the forecast by attributes comparison. Finally, the interest value combines in a Total Interest function all the attributes computed in the object analysis, providing an indicator of the overall performance of matching and merging between observed and simulated objects. In our experiment, we have empirically chosen the convolution disk radius and convolution threshold, so that this choice would recognise precipitation patterns similar to those a forecaster would identify. After a set of experiments, we fixed the value of the convolution radius to one grid point and the threshold of the convoluted field to 100 mm/48 hours.
4.2.2. Analysis Results
The purpose here is to establish which cumulus parameterization and global model combination best represents total rainfall on 4 and 5 November.
Figure 6 shows total rainfall for the two days for all simulations and observation datasets. Looking at the results of all three satellite datasets, it is evident that there is a substantial degree of differences among them. However, there are some important commonalities between them as well. For example, the rainfall maximum in all three satellite datasets is consistently located over the Caribbean Sea between Hispaniola and Jamaica. The region of high precipitation values extends to the south and west, reaching the coasts of Venezuela and Colombia. Also an area of consistent precipitation extends from the Windward Passage to the north in each of the observation fields. In any case, they always coincide to place over the westernmost tips of La Hispaniola the highest rainfall amounts caused by Hurricane Tomas. Now, analyzing our simulations, one of the most important results is that the two experiments with explicit treatment of convection in the inner domain (KF-E-IFS and KF-E-ERA) create well organised patterns. Therefore, rainfall spatial distribution is in good agreement with observations. In addition, KF-E-IFS and KF-E-ERA simulations do not create extensive areas of artificial precipitation in the rest of the domain as do most of the rest of simulations, which all use convective parameterizations in the inner domain. This suggests that getting the convective organisation correct is important for not only intensity but also track, in agreement with results of Li and Pu [13]. Another feature that further favours the no-parameterization simulations with respect to the others is the location of the maximum amounts, especially in the case of KF-E-IFS, and to a lesser degree KF-E-ERA. In this point, also BMJ-BMJ-ERA performs remarkably well, but the same cannot be said of its counterpart BMJ-BMJ-IFS. On the other hand, it is noted that explicit convection runs do not produce enough light-to-moderate rain (yellows).
As described in the previous chapter, the MODE analysis has been applied to support and strengthen the visual analysis of the rainfall fields. The results are summarized in Table 4: BMJ-BMJ and NGE-NGE-ERA outperform all other microphysics schemes in terms of total interest performances, with very good performances also for BMJ-BMJ-ERA, KE-E-IFS, and KF-E-ERA runs, especially when the TRMM dataset, the most reliable one [42, 43], is considered. Negative performances are, instead, shown by GD-GD-IFS, KF-KF-ERA, and GD-GD-ERA.
As final analysis, Figure 7 shows the cumulative distribution function of the rainfall for 4 and 5 November, including both the numerical experiments and observation datasets. It is apparent also in this figure that the three satellite measurements disagree considerably among each other, as already discussed. In any case, the two explicit convection simulations that performed the best in the intensity (wind and pressure) analysis and were among the top ones also from the trajectory, shown here in solid (KF-E-IFS) and dashed (KF-E-ERA) red lines, continue to be among the closest to the observational datasets available. Looking at KF-E-IFS probability line, it is rather close for most rainfall values (except for the highest extreme, not shown here) to the GSMaP curve, especially at the beginning and then from about 0.8 onwards. For KF-E-ERA, its curve is particularly similar to the TRMM 3B42 dataset, and it closely follows it from 0 and up to 0.85, where KF-E-ERA and TRMM 3B42 dataset curves start to diverge somewhat and TRMM 3B42 remains on the lower side. Finally, it is evident that GD-GD-IFS and GD-GD-ERA have too much light precipitation, while the opposite is true for BMJ-BMJ-ERA.
5. Conclusions
Two of the well-known error sources that influence the numerical simulation of severe hydrometeorological phenomena are analysed in this study, that is, the uncertainty associated with initial and boundary conditions and the influence of a parameterization scheme on the subgrid physical processes representation. Ten simulations have been performed combining five ways to represent convective processes and making use of two different global models to feed our regional numerical weather prediction model. For the specific case of Hurricane Tomas and for days 4 and 5 November 2010, when a rapid intensification phase was also observed, the results of this study clearly point to the advantage of not using a convective parameterization at a scale of 5 km while keeping it at the parent domain (15 km). By doing so, critical TC indicators such as maximum surface wind, minimum SLP, trajectory, and rainfall depths are simulated successfully and reasonably approach the values of NHC’s best track and satellite measurements. It is also shown here that the choice of convective parameterization has a higher impact on the intensity and trajectory of the TC, as well as on the overall rainfall pattern, than the choice of global model data used to feed the model. Combining all the results of the analysis in terms of best performance, the configuration with higher positive responses for the Tomas postprocessing evaluation seems to be the one with explicit treatment convection on the smaller domain. On the other hand, it is also evident from this case study that grid resolution of the order of 5 km has to be considered as the higher limit for simulations on which no convective parameterization could be adopted. Since a possible limitation of the present study is due to the fact that the conclusions are based on a single model study for one individual case, a confirmation using additional models and performing more case studies is under evaluation. Then, some preliminary simulations about the recent Matthew Hurricane (October 2016) are hereafter presented, focusing on the best and worst performing settings for Thomas (KF-E-IFS and KF-KF-IFS configurations). The results are reported hereafter in terms of best track and minimum pressure, with the regard to the Matthew simulation initialized at 00UTC on 1 October 2016, when Matthew Hurricane was already a category 2 hurricane: in agreement with Xue et al.’s [37] conclusions, the track and intensity plots are presented for the time period of +24 hours onwards. Figure 8 shows that the KF-E-IFS track is very close to the observed one, making landfall on Haiti around 06UTC on 4 October 2016, as in the observed event. Conversely, the KF-KF-IFS experiment diverges significantly, making landfall on Cuba around 00UTC on 4 October 2016. The KF-KF-IFS experiment’s behaviour for the case of Matthew Hurricane is indeed very consistent with the one that occurred for the case of Tomas Hurricane, when the trajectory diverged significantly towards Cuba once more. KF-E-IFS experiment performs better than KF-KF-IFS one also for the minimum pressure (Figure 9) and maximum sustained 1-minute wind speed (not shown) with regard to the case of Matthew Hurricane. Concerning the minimum pressure, KF-KF-IFS experiment converges more quickly towards the observed values (in 72 hours) than KF-E-IFS (in 96 hours), but because of the wrong trajectory and related landfall on Cuba, the minimum pressure suddenly increases, when the observed one reaches its minimum values at 00-12UTC on 4 October 2016, as correctly captured by KF-E-IFS simulation. Thus, even if these are preliminary results, they seem to validate the Tomas settings chosen as the best ones.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Acknowledgments
Antonio Parodi would like to acknowledge the support of Project FP7 DRIHM (Distributed Research Infrastructure for Hydro-Meteorology, 2011–2015) (Contract no. 283568). Thanks are due to the LRZ Supercomputing Centre, Garching, Germany, where the numerical simulations were performed on the SuperMUC Petascale System (Project ID: pr45de).