M. Zhu, School of Earth, Atmospheric and Environmental Science, University of Manchester, Manchester M13 9PL, UK. E-mail: Min.Zhu@res-ltd.com
During 2005–2006, ACTIVE and two other joint field campaigns were conducted near the Tiwi Islands, north of Darwin, Australia where a local thunderstorm, known as Hector, was very well observed. The dataset collected during the campaign are used to validate a meso-scale model using different set-ups.
Four cases of the Hector storm were simulated by running the WRF model in two modes: with or without incorporating local enhanced observations. In the ‘operational forecast mode’ (OP) the model was run in triply nested domains, with model initial and lateral boundary conditions provided solely by operational global analyses, and the finest domain having a horizontal resolution of 1 km. In the ‘observed data incorporation mode’ (DI) the model used a single domain at 1 km horizontal resolution, with either radiosonde or aircraft data incorporated during the run. All runs were validated against available observational data. The results showed that only one in four Hector cases was well simulated by the OP mode run, while three other Hector cases were simulated well in the DI mode. For all Hector-produced runs the general evolution and microphysical structure of the storm were reasonably simulated with some notable differences.
A large deficiency in all the simulated storms was their smaller size and weaker intensity in comparison with the observed storms.
Island thunderstorms in the tropics play a key role in the release of latent heat in the maritime continent and act to redistribute water vapour, trace gases and aerosols in the Upper Troposphere Lower Stratosphere (UTLS) region (Keenan et al., 2000), therefore modulating the Earth's climate. As climate and weather forecasting models are run at higher and higher resolution there is a need to evaluate how well they represent tropical island convection. In this study, a state-of-the-art weather forecasting model is compared with observations of such convection. The focus is on a much studied and regularly occurring thunderstorm, called Hector.
Hector is an isolated maritime continental thunderstorm that occurs over the Tiwi Islands, north of Darwin, Australia. During the transition season (October to December) and monsoon breaks, Hector occurs on 65–90% of days, reaching a height greater than 15 km, and has a horizontal extent of radar echoes > 30 dBZ of at least 200 km2 (Keenan et al., 1990; Wilson et al., 2001). The Tiwi Islands are 150 km in the east–west direction and 60 km in the north-south direction, with Bathurst Island in the west and Melville Island in the east (see Figure 1). They are located to the south of the Indonesian archipelago which is one of the main regions of intense tropical cloudiness and precipitation. The Tiwi Islands have been used as a natural laboratory to study deep convection, e.g., the Island Thunderstorm Experiment (ITEX) in 1988 (Keenan et al., 1989), the Maritime Continent Thunderstorm Experiment (MCTEX) in 1995 (Keenan et al., 2000), the Darwin Waves Experiment (DAWEX) in 2001 (Hamilton et al., 2004) and the Egrett Microphysics Experiment with Radiation Lidar and Dynamics (EMERALD) in 2002 (Whiteway et al., 2004). In summary, two mechanisms for Hector genesis and evolution have been identified (Carbone and Wilson, 2000; Crook, 2001; Wilson et al., 2001), which are (1) a strong diurnal forcing and island-scale sea breeze circulation system, and, (2) complicated interactions among the ambient wind, sea breeze front and gust fronts from adjacent cells.
More recently, during 2005–2006, the ACTIVE (Aerosol and Chemical Transport In tropical conVEction) campaign (Vaughan et al., 2008) and two other international experiments, SCOUT-O3 (Stratospheric-Climate Links with Emphasis on the Upper Troposphere and Lower Stratosphere) and TWP-ICE (Tropical Pacific Warm Pool International Cloud Experiment) (May et al., 2008, 2008) were conducted from Darwin, 50 km south of the Tiwi Islands. One of the scientific objectives of ACTIVE was to identify whether the rapid upward transport in deep convection such as Hector determines the composition of the TTL (Tropical Tropopause Layer). During the experiments, a wealth of meteorological, aerosol and chemical measurements were sampled in the inflow and outflow of Hector from both ground-based and aircraft platforms. This valuable dataset provides a unique opportunity to investigate the dynamics, microphysics and chemistry of the storm and in particular, the interaction between the storm and its environment.
From a subjective perspective, Hector should be very predictable in that it occurs almost every afternoon during the wet season, except during monsoon bursts. Yet there have not been many studies that model real Hector cases. Most numerical studies are restricted by either their simplified model physics or the poorly defined lateral boundary conditions (Golding, 1993; Keenan et al., 2000). Saito et al. (2001) used a non-hydrostatic mesoscale model of the Meteorological Research Institute (MRI), Japan with 1 km horizontal resolution to simulate the evolution of a Hector case during MCTEX successfully. They modelled the diurnal evolution of the convection over the island and classified it into five stages: (1) the dry stage with the appearance of a sea breeze front along the coast and Rayleigh–Bernard convection over the interior of the islands, not exceeding the condensation level; (2) the condensation stage, characterized by the appearance of shallow clouds; (3) the precipitation stage, with the intrusion of sea breezes farther inland and increasing upward motion; (4) the merging stage with cloud mergers (Tao and Simpson, 1984, 1989) and explosive convection, and, (5) the decay stage following the decrease of solar heating. Nevertheless, the model still had difficulty in reproducing the timing and location of deep convection. Tao and Simpson suggested that using data assimilation in future work would improve the quality of the simulations. More recently, Chemel et al. (2009) used the ARW (Advanced Research core of Weather Research and Forecasting) and the U.K. Meteorological Office Unified Model with 1 km horizontal resolution to simulate a Hector case during ACTIVE and study the water content exchange between the Upper Troposphere (UT) and Lower Stratosphere (LS). Their results are discussed in Section '4. Discussion' of the present paper.
The present study focused on: (1) assessing the use of operational analyses as initial conditions for a high resolution mesoscale model in simulating maritime continental deep convection; (2) assessing the impact of observed data incorporation on model simulation, and, (3) assessing model performance in simulating maritime continental deep convection. In Section '2. Data and methods', data and the methods used in the study are described; Section '3. Results' presents the simulation results; Section '4. Discussion' briefly discusses the Hector simulation issues and Section '5. Summary and conclusions' summarizes.
2. Data and methods
The ACTIVE field experiment comprised two phases, from 7 November to 9 December 2005 and from 16 January to 17 February 2006. Six distinct periods were identified by Allen et al. (2008), based on meteorological conditions and how polluted the boundary layer was. These were: biomass burning periods (13–19 November 2005 and 25–30 November 2005), a mini-monsoon (20–24 November 2005), pre-monsoon (30 November 2005 to 5 December 2005), active monsoon (16–22 January 2006), convectively inactive monsoon (25–27 January 2006) and break (4–14 February 2006). During the two phases, a total of 36 single-cellular and multi-cellular Hectors over the Tiwi Islands were captured by the observation network. Four cases were chosen in the study: 16 and 30 November 2005 within the biomass burning and pre-monsoon periods, respectively, and 6 and 10 February 2006 within the monsoon break period, so that cases across the different regimes could be compared. A further reason why these four cases were chosen is because the low-level ambient wind direction was different for each case. Keenan et al. (2000) found that changes in the low-level flow were particularly important in determining which sea breeze circulation was dominant and, as a consequence, where on the islands Hector first developed. This process coupled with the steering flow thus determined the areal distribution of rainfall.
The field measurements available for this study are summarized below and in Figure 1:
the TWP-ICE six-site balloon-borne sounding network data with 3-h resolution (only 6 hourly Darwin sounding data were available during the first phase);
the Natural Environment Research Council (NERC) Dornier Do-228 (0–5 km above sea level (a.s.l.)) aircraft measured temperature, relative humidity and wind with 1 s resolution. During the second phase, due to instrument failure, the relative humidity could not be used;
the Airborne Research Australia (ARA) Egrett (0–14 km a.s.l.) aircraft measured temperature and wind with 1 s resolution;
C-POL (C-band polarimetric radar) reflectivity data (Z) at 2.5 km horizontal, 0.5 km vertical and 10 min temporal resolution;
hourly C-POL derived rainfall data with 1 km horizontal resolution (May et al., 1999), and,
half-hourly 1.25 km resolution MTSAT (Multifunctional Transport Satellites) visual channel images from JMA (Japan Meteorological Agency).
The model used was the ARW version 2.2 developed by National Centre for Atmospheric Research (NCAR). The ARW is a fully compressible, Eulerian and nonhydrostatic model with Arakawa-C grid in the horizontal plane and terrain-following, hydrostatic-pressure coordinates in the vertical plane. The model supports full physics, one-way, two-way and two-way moving nests, analysis and observation nudging and 3-D variational data assimilation (Skamarock et al., 2005). The model physics applied in this study includes: the Rapid Radiative Transfer Model (RRTM) for longwave radiation; the NASA/Goddard shortwave radiation scheme; the Mellor-Yamada-Janjic planetary boundary layer (PBL) scheme; a Monin–Obukhov surface layer scheme; a five-level soil thermal diffusion scheme; the Thompson microphysics scheme (Thompson et al., 2004) and the Grell–Devenyi cumulus scheme. Grell–Devenyi's scheme was only used for nests which had a horizontal grid-length greater than 3 km.
In order to test how well the WRF model simulated Hector, two model run modes were designed, one using observations and one without. The first mode was termed the operational forecast (OP) mode. For this mode, the model was run in a two-way triply-nested domain at 9, 3 and 1 km horizontal resolution and with 115 vertical levels. The domain sizes are 1935 km × 1935 km, 540 km × 540 km and 201 km × 201 km, respectively, with the domains' centre at 11.6°S, 130.8°E. The lowest model level was at 10 m above surface, with the top at approximately 28 km. The initial and lateral boundary conditions for the coarsest domain were taken from 6 hourly 0.5° longitude × 0.5° latitude ECMWF (European Centre for Medium Range Weather Forecasts) operational analyses (also known as real-time analyses). The WRF model was run for a 42 h period (from 1800 UTC day-2 to 1200 UTC day + 0) with half-hourly output. During this run, no observed data were incorporated into the simulations. For the OP mode run, the aim was to assess whether WRF would produce a Hector storm with some degree of realism using only operational analyses to drive the model initial and lateral boundary conditions.
For the four cases studied in this paper, no significant differences were found in the quality of the OP simulation for initialization times of either ∼34 or 10 h before the storm outbreak.
The second mode has been termed the observed data incorporation (DI) mode. Unlike the OP mode, this mode was run with a single 1 km horizontal resolution domain only. The domain size is 256 km × 256 km, which was larger than the 1 km domain used in the OP mode. The initial and lateral boundary conditions for the 1 km domain were again from 6 hourly 0.5° longitude × 0.5° latitude ECMWF operational analyses, except for the 6 February 2006 case where a radiosonde profile only was used. The model was run for an 18 h period (from 1800 UTC day-1 to 1200 UTC day + 0). During the run, either radiosonde or aircraft data were incorporated in the following manner. The data incorporation methods include observation nudging and direct replacement of model fields by observed profiles. The model initialization time for the DI mode runs were approximately 10 h before the Hector storm outbreak, which allowed constraint of the model to the radiosonde and aircraft measurements as the model spun up (using observational nudging). The domain centre, model vertical levels and model output frequency for DI mode are the same as that used in the OP mode. As insufficient observed data were available for running DI mode for 30 November 2005, no DI mode run was carried out for that case. Therefore, only three cases (16 November 2005, 6 February 2006 and 10 February 2006) were run in both OP and DI modes. The terrain height and land-use for the 1 km domain were derived from 30 arcsecond resolution Shuttle Radar Topography Mission (SRTM) and U.S. Geological Survey (USGS) data, respectively. The Sea Surface Temperature (SST) used in the modelling was based on the daily analysis of ship, buoy and satellite observations from NCEP (the National Centre for Environment Prediction) in a 0.5° longitude × 0.5° latitude grid. Cloud Condensation Nuclei (CCN) were assigned to 300 per c.c. for 16 November 2005 and 30 November 2005 cases and 100 per c.c. for 6 February 2006 and 10 February 2006, which were based on aerosol measurements made by the Dornier (Allen et al., 2008).
For the 10 February 2006 case, the simulation was nudged to the 3 hourly TWP-ICE six-site balloon-borne sounding network data in the first 9 h of model integration, i.e. observational nudging was switched off 1 h before the outbreak of deep convection. For the 6 February 2006 case, a strong dry bias was found in the lower troposphere over the Tiwi Islands in 6 hourly ECMWF analyses (compared with the radiosonde data) which could not be mitigated by observational nudging. Considering the weak synoptic winds and low level ambient wind direction (from the southwest) on this day, it was deemed valid to substitute the model initial and lateral boundary conditions with the Pirlangimpi radiosonde data at 1800 UTC 5 February 2006 and 6 hourly Southern Surveyor radiosonde data respectively (refer to Figure 1). For the 16 November 2005 case the TWP-ICE six-site balloon-borne sounding network data were not available. However, during 1500–1600 LST (LST = UTC + 9.5; LST will be omitted hereafter.), the Dornier aircraft flew around Bathurst Island at approximately 700 m a.s.l., where Hector initiated and developed. Therefore, observational nudging to the Dornier data was attempted in the simulation to see whether the 1 h single-level dataset had any impact on the modelled Hector. Table 1 summarizes the implementation of DI mode run in the study.
Table 1. Implementation of DI mode run in the study
DI mode on/off
30 November 2005
Insufficient observed data to run DI mode
10 February 2006
Nudged to 3 hourly sounding data in the first nine simulation hours
6 February 2006
Sounding data as initial and lateral boundary conditions
16 November 2005
Nudged to Dornier data during 11.5–12.5 simulation hours
Convective systems are often highly stochastic in their detailed structure so that detailed inter-comparison of the modelled and observed field on a grid-resolution basis is of limited use. May and Lane (2008) proposed a method for using C-POL radar data to test model simulation of deep convection. This method was adopted in the present study. The method involves the calculation of a ‘Statistical Coverage Product’ (SCP) where the fraction of a specified grid of both observed and simulated radar data covered by various reflectivity thresholds are calculated along with the profiles of maximum reflectivity. More specifically, at each altitude level of the SCP domain (40 vertical levels with 0.5 km interval, see Figure 1 for SCP horizontal area) the following properties were calculated: (1) the fraction of the area covered by Z > [10, 40] dBZ and (2) the maximum reflectivity. (1) effectively represents the area fraction of precipitation particles e.g. rain, snow and graupel. (2) is a proxy for the maximum intensity of the cloud system. The WRF simulated equivalent reflectivity factor was derived from Stoelinga's algorithm (2005).
3.1. 30 November 2005 case
3.1.1. Case overview
This was a pre-monsoon case with moderate pollution from biomass burning (Allen et al., 2008). At 1100, the MTSAT visual channel image showed that Rayleigh–Bernard (RB) convection was well developed over the whole Tiwi Islands (Figure 2), driven by the strong surface heating. After 1130, with the aid of a low-level ambient wind from the northwest, the north coast sea breeze front (NCBF) began to intrude inland while the south coast sea breeze front (SCBF) stagnated near the shore. The area covered by the shallow RB cumulus became clear of cloud as the NCBF passed by (refer to Figure 2(a)). Around 1400, deep convection initiated over the southeast of Melville Island, when the NCBF and SCBF were about 20 km apart (see Figure 2(b)). Over the next 2–3 h, the Hector storm developed into a squall line propagating quickly west-southwestward (Figure 2(c)) steered by a 4–6 m s−1 easterly flow at 2–4 km (as measured by the Dornier, refer to Figure 4(a)). At 1500, the storm reached the maximum area of radar echoes > 30 dBZ of 1113 km2, with the cloud top reaching above 19 km (refer to Figure 3(b) and (c)). Light upper-level winds (according to the Egrett measurements) meant that the anvil flowed radially out from the storm complex, with a slight north-eastward drift (not shown). The storm decayed over the southeast coast of Bathurst Island around 1800.
3.1.2. OP mode simulation
From the temporal variation of simulated maximum updraft and downdraft over the Tiwi Islands (not shown), four stages of Hector evolution were discernible which are consistent with the previous study of Saito et al. (2001).
The condensation stage (0930–1200). At this stage, the maximum updraft and downdraft speeds were about 2–4 and 2 m s−1 respectively. This corresponds to the occurrence of RB cells over the land and the sea breeze circulation over the coastal area. At 1200, the land–sea surface temperature difference was over 7 °C and the NCBF had already penetrated inland about 20 km. The average convective available potential energy (CAPE) and convective inhibition (CIN) below 950 hPa over the land were 2580 and 7 J kg−1 respectively, which favours the occurrence of deep convection.
The precipitation stage (1200–1400). At this stage, the maximum updraft and downdraft speeds increased up to 15 and 6–8 m s−1 respectively and the precipitation rate reached 1–2 mm h−1. At 1230, the NCBF was about 25 km inland. In contrast to the observed stagnation of the SCBF over the south Melville island, the simulated west part of the SCBF also penetrated inland (Figure 2(d)).
The merging stage (1400–1600). At this stage, the maximum updraft and downdraft speeds reached 42 and 16 m s−1 respectively, which are associated with the strong mixing of the updraft with the environment and evaporative cooling effects that intensified the downdrafts in the Hector storm. The maximum precipitation rate was over 7 mm h−1. The timing of the initiation of deep convection over southeast Melville Island was well simulated. However, the model also produced deep convection over southwest Melville Island (see Figure 2(e)). At 1500, the simulated storm reached its maximum echo area > 30 dBZ, with a size of 541 km2, less than half of the observed maximum size. Correspondingly, the simulated Z > 10 and Z > 40 dBZ areas were also much smaller than the observed areas (see Figure 3(d) and (e)). Around 1530, two isolated Hector storms, one over the southwest and the other over the southeast of Melville Island, can be seen on the simulated column-integrated cloud hydrometeors and precipitation mass map (Figure 2(f)). Unlike the observed west-southwestward propagating squall line, the simulated storms were almost stationary. Clouds overshooting into the TTL (∼17 km) were simulated around 1530 (see Figure 3(f)). The merging stage in the simulation ended about 1 h earlier than the observation.
Decay stage (∼1600). At this stage, the maximum updraft and downdraft speeds both decreased dramatically to O(1) m s−1. A large anvil persisted over Melville Island. This simulation produced only 40% of the observed storm rainfall.
The simulation for this case in the OP mode was good in that the general structure of precipitation from the storm (see Figure 3) was captured, e.g. a maximum in deep convection followed by anvil production clearly evident in both observed and simulated Z > 40 and Z > 10 dBZ areas respectively. There were two main deficiencies with the simulation. First, the simulated storm was almost motionless while the real Hector propagated west-southwestward. This could possibly be explained by the discrepancies in lower tropospheric ambient wind between the simulation and the observation in the pre-storm environment, leading to different precipitation and merging processes. Specifically, the simulated low-level northwest ambient wind over the west part of Melville Island was too weak to restrain the intrusion of the west part of the SCBF (refer to Figure 2(d)). As the west part of the SCBF moved inland, the upward motion at the head of the front could trigger the deep convection as shown in Figure 2(e) and release convective energy. Once the two isolated Hectors developed simultaneously in the simulation, the gust fronts associated with the two storms' cold pools may have opposed each other, which could explain the storms' stagnation (refer to Figure 2(f)). The other possible reason for the motionless Hector in the simulation is that the simulated steering level (2–4 km) flow was too weak: only about 1–2 m s−1 in comparison with the observed 6–7 m s−1 easterly wind (Figure 4(a) and (b)). Second, compared with the observation, the simulated storm was generally smaller in size and shorter in lifetime, the simulated cloud overshooting into TTL was also weaker (see Figure 3), e.g. around 1600, the observed area fraction for the Z > 10 dBZ echo was over 15% between 3 and 6 km while the simulation only shows 2%. One possible cause is the shortage of water vapour in lower troposphere in the simulation which may result in the evaporation of the cloud and precipitation particles, e.g. the simulated relative humidity at 3.3 km height is about 50–60%, drier than the aircraft data of around 80–100% (see Figure 4(a) and (b)). Dornier data were not suitable for data incorporation for this case due to the aircraft's short period near the storm and its relatively high measurement height (>3 km).
3.2. 10 February 2006 case
3.2.1. Case overview
This was a monsoon break case with much cleaner boundary layer air (Allen et al., 2008). On this day the low level ambient wind was generally from the west but without a well-defined north-south component over the Tiwi Islands area. In the late morning, RB convection commenced. At 1330, two different regimes over Melville Island could be identified: one was the north and south coastal cloud-free regions due to the intrusion of cold moist marine air, while the other was the widespread RB convection area over the interior of the island (Figure 5(a)). By 1530, the central part of the NCBF accelerated southward while the SCBF invaded 15–20 km inland. Soon after, a latitudinal bow-shaped deep convection belt spanning Melville Island initiated ahead of the SCBF when the NCBF and SCBF were about 15 km apart (Figure 5(b)). Thereafter, the deep convection organized itself into a single, large, Hector storm (Figure 5(c)) with the cloud top above 19 km (Figure 6(c)), and driven west-northwestward by the 8–10 m s−1 steering level east-southeasterly flow. The storm had a maximum echo area > 30 dBZ of 1244 km2 at 1620 (refer to Figure 6(b)). Around 1800 the storm moved over the Timor Sea and then decayed.
3.2.2. OP mode simulation
For this case, the simulation produced precipitation and deep convection over the land too quickly, at least 3 h earlier than the observation (see Figure 6(d)–(f)). At 0930, the average CAPE and CIN below 950 hPa over the Tiwi Islands reached 1479 and 3 J kg−1 respectively, in contrast with the observed 609 J kg−1 (CAPE) and 73 J kg−1 (CIN) (from the Pirlangimpi radiosonde data). Such high simulated CAPE and low CIN arose from an excessively moist boundary layer. Below 950 hPa, the simulated average specific humidity was 18.3 g kg−1, about 2 g kg−1 higher than the observation. Under such circumstances, as solar heating increases in daytime, boundary layer turbulent mixing and thermal structure were able to trigger deep convection. This was evident from the numerous small-scale centres in the accumulated rainfall map from the simulation (not shown). These small scale rainfall centres were distributed at corners of polygons, which was linked to the RB convection. Without the merging process (Tao and Simpson, 1984, 1989), the simulation failed to reproduce a Hector storm.
3.2.3. DI mode simulation
Synthesis analysis of the model output shows that there are four stages of Hector development.
The condensation stage (0930–1230). Before 1230, the modelled fields were nudged to the five 3 h radiosonde sites' data (refer to Figure 1). This greatly improved the simulated boundary-layer thermodynamic structure, which was too moist in the OP run as described above. At 0930, the simulated CAPE and CIN were 622 and 54 J kg−1 respectively, matching the observed data well (see Section '3.2. 10 February 2006 case'). The timing and location of simulated shallow convection also agreed well with the MTSAT visual channel images during the period (not shown).
The precipitation stage (1230–1430). Precipitation started from the southeast coast of the Melville Island as the SCBF moved in (Figure 5(d)), then it moved northward. By 1430 the SCBF passed the central latitude of Melville Island, with the NCBF about 10 km inland of the north coastline.
The merging stage (1430–1730 h). At about 1500, deep convection initiated over north of Melville Island when the NCBF and SCBF were about 20 km apart. At 1530 a single, large Hector formed with maximum updraft and downdraft speeds reaching up to 32 and 10 m s−1 respectively (Figure 5(e)). As the storm moved west-northwestward (refer to Figure 5(f)), its expanding area of echoes > 30 dBZ reached a maximum of 691 km2 at 1600, about 56% of the observed area (refer to Figure 6(h)). The radar echo top also reached its highest altitude of over 16.5 km at this time (Figure 6(i)).
After 1800 the storm entered the decay stage with much reduced rainfall and vertical motion. During the simulation, the model only produced 40% of the observed precipitation.
With radiosonde data nudging, the simulation of Hector for this case was much improved not only in the timing of precipitation but also the general profile of hydrometeors (Figure 6(a), (b), (g) and (h)). The large problem, however, is that the simulated Hector was still smaller and weaker than its real counterpart. For instance, the simulated radar echo area of precipitation hydrometeors (Z > 10 and Z > 40 dBZ areas) was smaller by a factor of 2 (see Figure 6(a), (b), (g) and (h)). The observed cloud overshooting into the TTL from 1600 to 1700 was also missing in the simulation (see Figure 6(c) and (i)). The precipitation and merging stages between model and observation are also different. The observed southward movement of the central part of the NCBF between 1330 and 1430 and the subsequent bow-shaped deep convection belt, in particular the eastern part of the belt, were not well simulated. One possible reason is that it is the model lateral boundary conditions that introduced the northward component to the low-level ambient wind which forced the SCBF to the north of Melville Island, starting the deep convection there. Figure 7 shows that the simulated low level southerly wind component over south Melville Island was stronger than that from the Dornier data. This discrepancy may affect the cumulus merging process (Tao and Simpson, 1984) in the simulation and storm characteristics.
3.3. 6 February 2006 case
3.3.1. Case overview
Synoptic conditions for the monsoon break were established by 6 February with easterly flow at 700 hPa overlying westerlies at the surface. The boundary layer air was clean after the previous active monsoon period in January (Allen et al., 2008). At 1130 the ubiquitous RB convection had already covered the Tiwi Islands (Figure 8(a)). Within the next 2 h, the low-level west-southwest flow opposed the NCBF at the north coast of Melville Island, allowing the SCBF to move inland. Around 1430, an isolated single-cell storm broke out over the northeast of Melville Island with radar echoes reaching 17 km (Figures 8(b) and 9(c)) and the maximum echo area > 30 dBZ of 952 km2 appeared around 1600. As the storm dissipated over the north coast, at 1630, successive deep convection over central Melville Island soon intensified into a Hector storm and propagated to the northwest (Figure 8(c)). The steering level flow was about 6–8 m s−1 from the southeast. The storm decayed over the Timor Sea after 2000.
3.3.2. OP mode simulation
Only about 7% of the observed rainfall was simulated by the OP mode, which had very weak updrafts (∼O(1) m s−1) producing no penetrating cloud and echo top height generally < 5 km (Figure 9(d)–(f)). At 1230, the simulated boundary layer was approx. 1 g kg−1 drier than the radiosonde data, generating much lower CAPE (∼150 J kg−1) and higher CIN (∼40 J kg−1) in contrast with 500 J kg−1 CAPE and 10 J kg−1 CIN at Cape Don. Above the boundary layer, a dry layer with 400–500 m depth and 20–30% relative humidity was simulated, which was not observed in the radiosonde profiles. This further inhibited convection from extending any higher than 5 km. Furthermore, the simulated low-level ambient wind was from the west-northwest, causing most of the rain to fall in the wrong place, i.e., the south of Melville Island. Hence, the simulation was not successful.
3.3.3. DI mode simulation
With much improved initial and lateral boundary conditions, this simulation produced nearly 70% of the observed precipitation from Hector. Four stages of the storm development can be identified.
The condensation stage (0930–1130). Over land, the convection was arranged in rows parallel to the low-level southwest ambient wind, more akin to horizontal convective rolls (HCR) than RB convection (see Figure 8(d)).
The precipitation stage (1130–1330). At 1230, the simulated thermodynamic profile over the land compared well with the observed data (not shown). At 1330, the NCBF lingered over the north coast of the Tiwi Islands, while the SCBF had moved inland about 20 km. During this period, the maximum updraft speed increased steadily from 4 to 10 m s−1.
The merging stage (1330–1830). At 1330, a convective cell started at the northeast of Melville Island, almost the exact same time and location as the observation (Figure 8(e)). Nevertheless, the convection did not deepen into a Hector storm until 1530, which was about 1 h later than observed. At 1630, the simulated storm reached its maximum intensity (Figure 8(f)) with the strongest vertical motion, highest cloud top and largest echo area > 30 dBZ (Figure 9(h) and (i)). The maximum updraft and downdraft speeds in the storm were 20 and 10 m s−1 respectively. The cloud top extended above 15.5 km (about 1.5 km lower than the observation) and the echo area > 30 dBZ was about 936 km2 which is comparable with the observed maximum 952 km2. Within the next 2 h, the storm approached the northwest coast of the Tiwi Islands following the easterly steering level flow.
The decay stage (∼1830). The simulated storm decayed over the northwest of Melville Island at least 1 h earlier than that observed.
The largest discrepancy between the simulation and the observation is that two successive single-cellular Hectors were identified from C-POL radar images on this day, one initiating over the northeast of Melville Island with a 2 h lifetime, the other over the centre of the island with at least a 4 h active period. However, the simulation only generated one Hector storm. The discrepancy may be related to the simulated low-level ambient wind having a more northerly component than the observed west-southwest wind (see Figure 10(a) and (b)). Such a difference may lead to different dynamics of the sea breeze front and convergence and therefore influence the merging stage of the storm. Alternatively, the discrepancy may have arisen by applying a single radiosonde profile as model lateral boundary conditions for a 256 km × 256 km domain, which did not capture the spatial variation in observed boundary conditions.
Comparison of the simulation against radar statistics shows that the general profile of precipitation hydrometeors was reasonably well simulated (Figure 9(a), (b), (g) and (h)) though below 5 km, the simulated echo area Z > 10 dBZ was only half as large as the observation. In spite of the discrepancy, this simulation shows the feasibility of using radiosonde data as initial and lateral boundary conditions for deep convection modelling when operational analysis are problematic and synoptic conditions are relatively homogeneous over the area of interest.
3.4. 16 November 2005 case
3.4.1. Case overview
This was a pre-monsoon case within the biomass burning period (Allen et al., 2008), and having a low-level east-southeast wind. At 1100, RB shallow convection was scattered over the Tiwi Islands except for the southeast of Melville Island due to intrusion of the sea breeze (refer to Figure 11(a)). At 1430, the SCBF lay across the Tiwi Islands from west-southwest to east-northeast, about 20 km away from the stationary NCBF over the north coastline. Around 1510, a single-cell Hector initiated over central Bathurst Island and was then advected west by the strong easterly steering-level winds (10–15 m s−1) (see Figure 11(b)). At 1600, the overshooting cloud reached 19 km altitude (Figure 12(c)), which by 1640, had a maximum echo area > 30 dBZ of about 912 km2 (refer to Figure 12(b)). The prevailing northerly winds at 150–200 hPa detached the storm anvil and advected it southward as Hector moved into the Timor Sea (Figure 11(c)).
3.4.2. OP mode simulation
This simulation produced only 18% of the observed Hector rainfall. At 1200, the temperature (relative humidity) below 700 m a.s.l. over the land was underestimated (overestimated) by the model by 2–3 °C (20%), compared with Dornier data. Although high CAPE (1810 J kg−1) and low CIN (7 J kg−1) were in favour of convection, the simulated high wind speed within the boundary layer (overestimated by 2–3 m s−1, compared to the Dornier data) and low relative humidity above 1 km (underestimated by 20%, compared to the Dornier data) are both unfavourable for the development of deep convection and merging processes (Tao and Simpson, 1984). As a result, only some short-lived convective clouds occurred around 1400–1600 over Bathurst Island, but they failed to merge into a Hector storm (Figure 12(d)–(f)).
3.4.3. DI mode simulation
Prior to the data incorporation period (1500–1600), the model simulated the condensation stage reasonably well (before 1100) with RB convection (Figure 11(d), and the precipitation stage (1100–1400) with the SCBF moving inland. During 1500–1600, the model was nudged to the observed temperature, humidity and wind data sampled by the Dornier aircraft at 700 m a.s.l. (refer to Figure 13). Around 1530, a deep convective cell occurred over central Bathurst Island, agreeing quite well with the observations (Figure 11(e)). It soon developed into a Hector storm (Figure 11(f)) with the largest echo area > 30 dBZ of 370 km2 (40% of the observed size) and highest cloud top below 15 km (4 km lower than the observation) (refer to Figure 12(g)–(i)), with the maximum updraft speed exceeding 20 m s−1. After 1800, the simulated storm decayed as it moved westwards into the sea. In contrast with the OP run, this simulation produced 35% amount of the observed rainfall.
Inspection of the observed data shows that at 700 m a.s.l. there is an obvious convergence area over Bathurst Island around 1500–1600 (see Figure 13). With such information and associated temperature and humidity incorporated into the model, it may provide a mechanism for the simulated convection to develop into a Hector. Despite this, the simulated storm was still smaller in size and weaker in intensity than that observed.
4.1. Use of operational analysis as initial and lateral boundary conditions for Hector simulation
From a numerical prediction point of view, the WRF modelling is an initial-boundary value problem given perfect model dynamics and physics. For nested runs such as the OP mode in this study, the WRF modelling for innermost domain approximately becomes an initial value problem if the coarsest domain is large enough. For runs in OP mode, which used ECMWF operational analyses as initial conditions, the WRF model had very limited skills in simulating the Hector storm, i.e. for only one in four cases was Hector simulated even if the diurnal forcing and detailed surface characteristics had already been included in the model. For the single domain DI mode runs, however, the lateral boundary conditions become more critical. For instance, for the 10 February 2006 case, after the 9 h observation nudging period was over, an unrealistic southerly component in the low-level ambient wind was introduced, which forced the SCBF to the north of Melville Island and started the deep convection there whilst the observed deep convection occurred over the middle of Melville Island. This was attributed to the errors in the ECMWF analysis propagating into the domain from the lateral boundary. Likewise, using idealized lateral boundary condition can also deteriorate Hector simulation. Golding (1993) initialized the U.K. Met Office non-hydrostatic mesoscale model with a single radiosonde profile to simulate the Hector storm. Golding used idealized zero-gradient lateral boundary conditions for the single domain run and found that only the early stage of the storm could be roughly simulated.
All these highlight two points: tropical island thunderstorms are highly sensitive to the fields of temperature, humidity and wind, and it is difficult for an operational analysis system such as ECMWF to capture the detailed atmospheric structure over a remote island without assimilating additional observations such as those collected during the ACTIVE campaign. The other limitation for operational analysis is the cut-off time for the observations to be incorporated into the operational analysis system due to the data not being ready for assimilation.
4.2. The impact of data incorporation on model simulation
The fact that all three DI mode runs successfully produced Hector storms demonstrates that using observed data in the simulation improves it dramatically.
In some cases the modelled storm started too early, so increasing the low level temperature and reducing the low level humidity (in accordance with the observations) improved the simulation. In other cases the storm failed to initialize and so increasing the low level humidity (in accordance with observations) enabled for a better simulation. Modifying the low level ambient wind, improved the interaction between sea breeze fronts, which resulted in the location of the storms being more in accord with observations. Modifying the steering level flow made the propagation phase of the storm more consistent with the observations.
All these again demonstrate that temperature, humidity and ambient winds in the planetary boundary layer (PBL) over land were very important in simulating deep convection and subsequent development of the storm. Therefore, one probable implication from this study is that accurate forecasts on the scale of Hector will only be possible if suitable observing systems are in place. Nevertheless, a further implication is that even when local observations are available and incorporated at the analysis time, multi-day forecasts will be deficient if the erroneous lateral boundary conditions are used.
4.3. Model performance in simulating maritime continental deep convection
For Hector-produced simulations, the simulated diurnal evolution of convection over the island generally follows several stages (condensation, precipitation, merging and decay) which have been confirmed by satellite and radar data. Regardless of which running mode (OP or DI) is used notable differences were found between the simulation and the observations, with common differences being that all of the simulated storms are smaller in size, weaker in intensity and weaker in overshooting of convective cloud into TTL in comparison with the real storms. The striking difference in the model capability in producing a Hector storm between the OP and DI mode runs may be attributed to a problem with model initialization fields. In this study, deficiencies in the ECMWF operational analysis were found when compared with field measurements so the evidence points to the fact that to simulate convective storms accurately, the model initial and lateral boundary conditions should be improved. As Wilson et al. (2001) pointed out, assimilating the high resolution data, e.g. boundary layer winds and temperature from Doppler radar (Sun and Crook, 2001), may be required: radiosonde and aircraft data are restricted by their representativeness and spatial sparseness. Using reanalysis data as model initial conditions would confirm this, since the quality is generally much better; nevertheless, it usually takes a long time for the reanalysis dataset to be made available.
The coupling of the model PBL, microphysics and radiation schemes should also be important. Chemel et al. (2009) used the same model and a similar model configuration to simulate the Hector storm on 30 November 2005, with the main differences being PBL and radiation schemes. Their simulation results are similar to ours but the storm positions are different. Nevertheless, neither simulations reproduce the squall line behaviour of the real Hector storm on that day. Lang et al.'s recent work (2007) highlighted the importance of model microphysics in simulating tropical convective systems. They found that the observed broader areas of moderate intensity echo (i.e., between 20 and 40 dBZ) can only be simulated if the model microphysics is improved. In the Thompson microphysics scheme, which was used in the simulations, some parameters in the assumed snow-size distribution are suitable for simulating midlatitude weather systems, but are not valid for tropical systems. On the basis of aircraft data analysis, Field et al. (2007) proposed a moment estimation parameterization, which is applicable to both midlatitude and tropical ice clouds: hence, assessing how this scheme improves the performance of the spatial coverage of the anvil region is an avenue for future work.
The modelled weaker intensity may also be related to the horizontal resolution used. Bryan et al. (2003) pointed out that 1 km horizontal resolution might be too coarse to represent faithfully physical processes of cloud turbulence that occur in deep convection and thus recommended O(100 m) grid spacing. Balaji and Clark (1988) and Dailey and Fovell (1999) also discussed that resolution of 500 m or less is necessary to produce the horizontal convective rolls, the sea breeze front and the interaction between them. In fact, Large Eddy Simulation (LES) may be required, which removes the need for the PBL parameterization scheme currently used in the 1 km resolution simulation. Chemel et al. (2009) found that a resolution of 250 m in WRF resolves the convection better than the 1 km resolution and produces more storm rainfall, which is closer to the radar observations. Although they conceded that the characteristics of the Hector storm are generally similar in space and time to those obtained in the 1 km horizontal resolution simulations, it is likely that without data assimilation, the benefit from the horizontal resolution increase will be limited. A related topic for future study is the way in which improvements in model physics impact the Hector simulation.
5. Summary and conclusions
In this study, four Hector cases were chosen during the ACTIVE campaign in three different boundary layer aerosol regimes, i.e. 16 November 2005 within the biomass burning pollution period, 30 November 2005 with moderate aerosol pollution, 6 February 2006 and 10 February 2006 with clean marine air, to carry out the high resolution numerical simulation experiments using the WRF (ARW) model. Two modes were designed to test how well the model could simulate the storm with or without incorporating observed data. One mode termed ‘operational forecast mode’ involved running the model in triply nested domains with a long integration time (42 h). The finest domain had a horizontal resolution of 1 km. During the run, no observed data were incorporated. The other mode termed ‘observed data incorporation mode’ was to run the model in single domain at 1 km horizontal resolution with short integration time period (18 h). During the run, either radiosonde or aircraft data were incorporated. All runs were validated against available observed data, including aircraft data, satellite images, radar reflectivity and accumulated rainfall. The results show that only one in four Hector cases (30 November 2005) can be simulated by running the model in OP mode, while the three other Hector cases (16 November 2005, 6 February 2006 and 10 February 2006) can be simulated if the model was run in DI mode. For all the cases that Hectors were simulated, the general evolution and microphysics structure of the storm agree with the observations but with some notable differences.
A major deficiency in the simulated storms is their smaller size and weaker intensity in comparison with the observations. Therefore, for transport studies, the simulated weak storm may lead to underestimate of water content, aerosol and chemical exchanges between troposphere and stratosphere.
This work was supported by the UK Natural Environment Research Council (Grant NE/C512688/1) and the NERC Airborne Remote Sensing Facility. May was supported by the US Department of Energy Atmospheric Radiation Measurement (ARM) Program.