Case Study of a Morning Convective Rainfall Event over Southwestern Taiwan in the Mei-Yu Season under Weak Synoptic Conditions

There exists a minor, secondary early­morning peak in mei­yu rainfall along the western coast of Taiwan, and this work investigates one such event on June 8, 2012 in southwestern Taiwan under weak synoptic conditions through both observational analysis and numerical modeling, with the main focus on the triggering mechanism of the convection. Observations indicate that the convection developed offshore around midnight near the leading edge of a moderate low­level southwesterly wind surge of 15 – 20 kts and intensified and moved onshore to pro­ duce rainfall. The cold outflow from precipitation also led to new cell development at the backside, and the rain thus lasted for several hours until approximately 07:00 LST. Numerical simulation using a cloud­resolving model at a grid size of 0.5 km successfully reproduced the event development in close agreement with the observations, once a time delay in the arrival of the southwesterly wind surge in initial/boundary conditions (from global analyses) was corrected. Aided by two sensitivity tests, the mod­ el results indicate that the convection breaks out between two advancing boundaries, one from the onshore surge of the prevailing southwesterly wind and the other from the offshore land/mountain breeze, when they move approximately 40 km apart. Additionally, both boundaries are required, as either one alone does not provide suf­ ficient forcing to initiate deep convection in the model. These findings on the initiation of offshore convection in the mei­yu season, notably, are qualitatively similar to some cases in Florida with two approaching sea­breeze fronts (in daytime over land).


Introduction
During the transition from the winter to summer monsoon, repeated occurrence of fronts between the cold continental high and the warm/moist Pacific sub tropical high leads to an earlysummer rainy period in East Asia (Chen 1983;Tao and Chen 1987;Ding 1992).Known as the meiyu season, this unique weather and climate phenomenon typically lasts from midMay to midJune in Taiwan, when the front often extends from the vicinity of Japan through Taiwan and into southern China (Chang and Chen 1995;Chen 2004).Under its influence (e.g., Chen 1993), meso scale convective systems (MCSs) often form along or near the front and bring heavy rainfall to Taiwan (e.g., Kuo and Chen 1990;Chen 1992Chen , 2004;;Yeh and Chen 1998;Chen et al. 2008), particularly when the prefrontal southwesterly monsoon flow strengthens and increases the moisture supply (e.g., Chen and Yu 1988;Chen and Li 1995;Chen et al. 2005a;Jou et al. 2011;Wang et al. 2016).
With steep and complex topography (Fig. 1a), the meiyu rainfall in Taiwan exhibits pronounced diurnal variation with a peak amount in local afternoon and a minimum in late night (Chen and Yang 1988;Yeh and Chen 1998;Johnson and Bresch 1991;Chien et al. 2002;Wang et al. 2017).This diurnal cycle is linked to island circulation and changes in stability (e.g., Akaeda et al. 1995;Georgelin et al. 1996;Chen et al. 1999;Kerns et al. 2010) and can apply to both perturbed and unperturbed periods according to the analysis by Ruppert et al. (2013) on the data from the SouthWest Monsoon Experiment (SoWMEX; Jou et al. 2011).During the perturbed periods when a meiyu front is close by and the island circulation is not as pronounced (compared to the unperturbed periods; cf.Fig. 6 of Ruppert et al. 2013), heavy rainfall events can occur from organized MCSs (or even scattered but vigorous convection) that are often fed by strong, moisturerich prevailing southwesterly flow (e.g., Jou and Deng 1992;Chen and Chen 1995;Chen et al. 2005a;Davis and Lee 2012;Xu et al. 2012;Tu et al. 2014;Wang et al. 2014a) and enhanced by topographic uplift (e.g., Johnson and Bresch 1991;Lin 1993;Ruppert et al. 2013;Wang et al. 2017).Besides the thermal effect that modulates the timing of rainfall (also Wang et al. 2014b), the dynamic effect of terrain blocking of Taiwan may also affect the development of convection and shift rainfall locations (e.g., Li and Chen 1998; Wang et al. 2005aWang et al. , 2016)).All the studies mentioned above, among others, have contributed toward a better understanding of the mechanisms of heavy rainfall in the meiyu season and how Taiwan's terrain affects its development and occurrence.
During unperturbed periods with no meiyu front and generally weaker prevailing flow, the diurnal temperature range in Taiwan becomes larger, and the island circulation is more pronounced (Ruppert et al. 2013).In this regime, localized rainfall events may still develop, and some of them are afternoon showers closely tied to solar heating and valley/sea breezes (e.g., Yeh and Chen 1998;Chen et al. 2001;Lin et al. 2011;Sun and Zhang 2012;Chen et al. 2014).By contrast, one noteworthy phenomenon in rainfall climatology is the presence of a second, minor rainfall peak in early morning along the western coast from central to southern Taiwan (see, e.g., Fig. 9 of Kerns et al. 2010 and Fig. 4 of Chen et al. 1999), at a timing almost 180° out of phase with the primary peak (also Chen and Yang 1988;Chi et al. 1998;Chen et al. 2005b;Huang and Chan 2011;Wang et al. 2017).For this minor and small peak, nighttime land breeze (or mountain breeze, drainage or katabatic flow) is hy pothesized to play an important role in triggering the convection offshore (Chen and Yang 1988;Chi et al. 1998;Kerns et al. 2010).Some recent studies in other tropical/subtropical regions also show results support ive of this reasoning (e.g., Hill et al. 2010;Chen et al. 2016).For precipitation systems that are initiated in southeastern China and then propagate eastward across the Taiwan Strait (e.g., Wang et al. 2005b;Bao et al. 2011;Johnson 2011;Chen et al. 2013), land breeze may also provide additional forcing for their enhancement or development overnight.However, partly due to sparse observations over the upstream ocean, very few case studies on such morning events in Taiwan exist in the literature.Thus, the detailed process of their initiation and early development is unclear, and our understanding of them is much less compared to that of those during the perturbed periods in the meiyu season.
Compared to the land breeze, the generally stronger and thicker seabreeze circulation (Mak and Walsh 1976;Mapes et al. 2003) has been studied more ex tensively, especially over the Florida Peninsula (e.g., Burpee 1979;Cooper et al. 1982;Nicholls et al. 1991;Fankhauser et al. 1995).While the seabreeze front (SBF) exhibits enhanced pockets of convergence favorable for convection (Kingsmill 1995), one noteworthy finding is that deep convective cells are often triggered farther ahead (e.g., Fovell and Dailey 2001), in the region between two approaching SBFs (Nicholls et al. 1991) or one SBF and another existing boundary (such as a gust front) prior to their collision (Fankhauser et al. 1995;Kingsmill 1995).In some cases, the uplifting by the SBF alone is found to be insufficient to trigger deep convection, and additional forcing (e.g., from horizontal rolls and mountain waves) is also needed (Fovell 2005;Joseph et al. 2008).The abovementioned aspects are proved to be relevant in our later discussion.
With the goal to understand the detailed process leading to the deep convection and morning rainfall over the southwestern coastal region of Taiwan in the meiyu season, particularly under weak synoptic forcing during unperturbed periods, a case study was carried out.From a preliminary survey using radar plots, the event on June 7 -8, 2012 (Fig. 1b) is select ed for detailed investigation, mainly through an observational analysis and highresolution model simulation and sensitivity tests.After Section 2, which describes our data and methodology, this event is reviewed through a detailed mesoscale analysis in Section 3. The numerical model and experiments are described in Section 4, and their results are presented and discussed in Section 5. Finally, the conclusions are given in Section 6.

Data and methodology
For the discussion on synoptic conditions and case environment, the observational data used include surface weather maps from the Central Weather Bureau (CWB) of Taiwan, the European Center for MediumRange Weather Forecasts (ECMWF) global gridded reanalyses (ERAInterim, Berrisford et al. 2011;Dee et al. 2011), and the National Oceanic and Atmospheric Administration (NOAA) Ocean2 Scatterometer (OSCAT) ocean surface winds (Jaru watanadilok et al. 2014) during the case period June 7 -8, 2012.The ERAInterim dataset has a resolution of 0.75° × 0.75° latitude/longitude at 37 pressure ( p) levels (1,000 to 1 hPa) and is available every 6 h (at 02:00, 08:00, 14:00, and 20:00 LST).The sounding data prior to the event are also used to evaluate the thermodynamic environment and wind structure.
To document the rainfall event and for a more detailed analysis, reflectivity composites of a vertical maximumecho indicator (VMI) from the operational radars in Taiwan (every 10 min) and infrared (IR) cloud imageries from the Japanese Multifunctional Transport Satellite2 (MTSAT2; every 30 min) are employed.Also used are hourly data from a network of approximately 450 gauges for rainfall distribution in Taiwan (Hsu 1997) and those from selected CWB surface stations for sectiontime and wind/conver gence analyses.In addition to mesoscale discussion, some of the observational data are used to validate model simulations where needed, including reflectiv ity scans every 7.5 min by the Chigu Doppler radar in southern Taiwan.

Synoptic conditions
The synoptic and environmental conditions of the present case are first reviewed in this section.The CWB surface weather maps in Fig. 2 indicate no significant weather systems near Taiwan from 14:00 LST on June 7, 2012 to 08:00 LST on June 8, 2012.However, the pressure gradient near Taiwan gradually increased during this period due to the westward extension of the subtropical high (as indicated by the 1,008 hPa isobar) over the region southeast of Taiwan and the development of a weak low (1,000 hPa) over the East China Sea (Figs. 2a -d).The ERAInterim analyses indicate that the prevailing flow at 02:00 LST on June 8 was from the southwest at 1,000 -850 hPa (Figs.3a, b) and gradually turned into a west south westerly at 700 -500 hPa and then further to northerly at 200 hPa (Figs. 3c,d).This veering of wind with height suggests warm air advection in the lower atmosphere.Coming from an upstream with tighter geopotential height gradients, a surge of lowlevel southwesterly flow of 15 -20 kts from the South China Sea (SCS) toward southwestern Taiwan existed at 02:00 LST (cf.Fig. 3a), also consistent with Fig. 2, in which the horizontal pressure gradient was increasing with time.Except for this moderate strengthening of lowlevel winds, a pressure ridge was either near or directly over Taiwan through deep layers from 850 to 200 hPa, with no trough or upperlevel jet nearby (Figs.3b -d).Roughly 2 h earlier, the stronger southwesterly and southerly flow from the SCS was also observed by OSCAT at 23:48 LST on June 7 (Fig. 4), although data within approximately 50 km offshore were unavail able.Thus, in Figs. 2 -4, it is evident that the present case developed under weak synoptic conditions.
The Pingdong sounding launched at 20:00 LST on June 7, 2012 (Fig. 5, cf.Fig. 4 for location), approxi mately 5 h before the convection moved onshore into its vicinity, exhibits a nearneutral condition to dry ascent below 950 hPa and mostly conditional instabil ity further up from 900 to 200 hPa.For a surface air parcel, the convective available potential energy was 1,844.8J kg −1 , the convective inhibition was only 44.8 J kg −1 , and the lifting condensation level and the level of free convection were 945 and 839 hPa, respectively (Fig. 5).The winds also veered with height from the surface to the midlevel, i.e., they were consistent with Fig. 3, but the wind speed was mostly within 10 -15 kts at this location onshore (cf.Fig. 4).Thus, as ex pected, the thermodynamic conditions around south western Taiwan prior to the event were conducive to deep convection, and the winds indicated weak synoptic conditions.

Mesoscale analysis
In this subsection, results of a mesoscale analysis are presented to provide an overview of the present event on 7 -8 June 2012.The VMI reflectivity from operational radars in Taiwan at 30 min intervals from 21:30 LST on June 7 to 07:00 LST on June 8, 2012, and the corresponding MTSAT2 IR cloud imageries and rainfall over southwestern Taiwan every 1 h, are illustrated in Fig. 6 to depict the development and evolution of the convection.Isolated convection first appeared near 22°N, 120°E, i.e., approximately 60 km upstream from Taiwan at 21:00 LST (not illustrated) then enhanced to over 40 dBZ at 21:30 LST (Fig. 6a1, short arrow).As the convection moved toward Taiwan after formation, it first appeared on the colorenhanced IR cloud imagery at 22:00 LST (Fig. 6a3).At this time, the cloudtop temperature (T B ) was still above 0°C, so the convection had not yet reached the upper levels.As the convection approached southwestern Taiwan, it first weakened prior to 23:30 LST (Figs. 6b, c1), then reintensified into several cells near shore (close to 22.5°N, 120.1°E and approximately 15 km from the coastline) at 00:00 LST on June 8 (Fig. 6c2).During this period, active convection (with T B < −54°C) also developed south of Pingdong and over the ocean upstream from the southern tip of Taiwan (Fig. 6c).Over the next 2 -3 h, the convective cells grew stronger (higher into the upper troposphere) to form a banded structure and gradually moved onshore to produce rainfall along the coast of southwestern Taiwan (Figs. 6d, e).As the convection and rainfall continued to move inland, the banded structure some what weakened, and new cell development was seen near shore, i.e., on the backside (or to the southwest) of the old convection since 03:00 LST (short arrows in Figs.6f, g).After 03:30 LST, the old convection started to weaken, and so did the new cells soon after ward (Figs.6g, h).While the upperlevel cirrus/anvil clouds continued to grow in size after 04:00 LST, both convective cells and rainfall gradually weakened (Figs.6h -j).By 07:00 LST, the rain had largely diminished (Fig. 6j4), and the present event had come to an end.The total rainfall over land was mostly 5 -40 mm (cf.Fig. 1b) and not particularly high for the meiyu season, because the convective cells were migratory without a quasistationary forcing mechanism.None theless, it is clear that oceanic convection developed approximately 60 km upstream from Taiwan near 21:00 LST on June 7 and then moved onshore to cause rainfall in southwestern Taiwan after midnight.Its detailed mechanism for initiation and development is the main focus of the present study and is addressed further.
Although the convection first developed farther off shore, the surface wind and convergence can be exam ined using observations from selected coastal/island stations that enclose an area near shore, as illustrated in Fig. 7a (thick solid lines).Using line integral (e.g., Eq. 2 of Chen et al. 1999), the time series of surface convergence inside this polygon (and inward wind components normal to the segments) are obtained and plotted in Fig. 7b.Before 00:00 LST on June 8, the convergence (dashed black curve) was only approxi mately 2 -3 × 10 −5 s −1 but increased several fold (to 1.3 × 10 −4 s −1 ) right afterward, when the convective cells moved into the polygon and intensified (cf.Fig. 6c2).This enhancement in convergence was mainly due to an intensification of onshore (inward and southwest erly) winds from approximately 1 to 3 m s −1 between Dongji Island (DJI) and Shawliucho (SLC; red), a phenomenon also depicted in Figs. 2 -4.Note also that the convection thus far was not near DJI or SLC (Figs. 6a -c); so, the wind speed increase could not have been convectively driven.The offshore (also inward) wind along the coast between Kaohsiung (KH) and Fanliao (FL; purple and green), meanwhile, also  increased, but with a speed of no more than 1 m s −1 , its contribution to the convergence was only a minor one (Fig. 7b).After 02:00 LST, however, the offshore wind from Tainan (TN) to KH became stronger (blue; to approximately 1.5 -2 m s −1 ), as did that farther south near Dongkang (DK) since 03:00 LST (from 1 to 2.5 m s −1 at 05:00 LST), whereas the onshore wind weakened to some extent.This strengthening in off shore flow along the coast after 02:00 LST coincided well with the initiation of new cells on the backside of the old ones seen in Fig. 6 (during 03:00 -04:30 LST) and likely replaced some of the onshore wind (between DJI and SLC) and thus led to its weakening.After 06:00 LST, all the onshore/offshore winds and their convergence weakened (Fig. 7b), and the event soon ended.
The offshore flow along the coast of southwestern Taiwan during the night had three possible origins: the deflection of prevailing flow by Taiwan topography (or return flow of blocking effect), land breeze (in combination with mountain breeze), and outflow from existing convection to the east.In Figs. 3 -5, at approximately 10 -15 kts, the lowlevel prevailing flow below 800 hPa was from the southsouthwest and nearly parallel to the topography of Taiwan (cf.Fig. 1); thus, the blocking effect was not expected to be signif icant.To investigate further into the other two mech anisms, distance-time plots from stations along two almost straight segments were produced, as illustrated in Fig. 8.The first segment is roughly east-west along approximately 23.1°N and connects DJI, Chigu (CG), Shanhwa (SH), Yujing (YJ), and Giashan (GS), which lies at the foothill of the CMR (cf.Fig. 7a).From 18:00 to 02:00 LST, the surface winds at DJI and CG were nearly parallel to the coastline, while those at SH had a larger offshore (westward) component (mostly at 2 -3 m s −1 ), thus indicating some land breeze (Fig. 8a).During this period, the surface air at the coast (at CG) remained close to 30°C, while that at the foothill (at GS) cooled from 28°C to 24°C.However, there  was little indication of mountain breeze at GS, and the surface air at YJ was also quite weak (≤ 1 m s −1 ; from the northeast) and existed only after 00:00 LST.
During 01:00 -02:00 LST, convective rainfall of at least 20 mm occurred at CG (Fig. 8a; also cf.Fig. 6e), as the temperature dropped suddenly by 3°C and the winds turned into a southeasterly.Easterly or north easterly winds also occurred at SH and YJ after 03:00 LST in association with rainfall and cooling (Fig. 8a).
Obviously, these offshore winds in the early morning of June 8 strengthened in response to cold outflows from nearby convection, in agreement with Fig. 6.Surface manual analysis of streamlines at 00:00 and 03:00 LST, as illustrated in Fig. 9, also support this notion.
The distance-time plot for the second segment that runs approximately north-south (SDM, CZ, DK, and SLC; Fig. 8b) indicates more evident offshore winds (toward the west) at DK (2 -4 m s −1 ; from land breeze) and SDM (approximately 1 m s −1 , likely from the mountain breeze) and a weak and less steady southerly flow upstream at SLC.Both rainfall and some cooling in temperature occurred at CZ and DK after 01:00 LST, whereas surface winds at the latter site also shift ed to northerly since 04:00 LST (Fig. 8b), also clearly due to outflow induced by convection (cf.Figs. 6f,g,9b).
Figures 7 to 9 indicate that the arrival of the strength ening lowlevel prevailing southwesterly winds, likely aided by the offshore land breeze at night, caused the locally generated convection to further develop and intensify near the coastline around midnight (cf.Figs.6c, d).Later on, the cold outflows from the convection that moved onshore, by contrast, played a role in the initiation of new cells at the backside (again, near the coast) around 03:00 -04:00 LST.However, to inves tigate the key question of the present study [how the convection was initiated further upstream in the first place (near 22:00 LST) in this event], highresolution simulation and sensitivity tests using a numerical model are required.The related results and analyses are presented in the next section.

The CReSS model
In the present study, the CloudResolving Storm Simulator (CReSS, version 2.2) was used for the simulation and sensitivity tests.The CReSS model is a nonhydrostatic and compressible cloud model that uses a heightbased terrainfollowing coordi nate (Tsuboki andSakakibara 2002, 2007).In the horizontal and vertical, the ArakawaC and Lorenz staggered grids are adopted, respectively, with no nesting.Prognostic equations are formulated for 3D momentum (u, v, w), p, potential temperature (θ ), and mixing ratios of water vapor (q v ) and other conden sates.A bulk coldrain scheme (Lin et al. 1983;Cotton et al. 1986;Murakami 1990;Ikawa and Saito 1991; Murakami et al. 1994) was used to simulate clouds explicitly without cumulus parameterization (Table 1).This scheme includes a total of six species: water vapor, cloud water, cloud ice, rain, snow, and graupel (Tsuboki and Sakakibara 2002).Subgrid scale turbu lent mixing is parameterized using 1.5order closure with turbulent kinetic energy (TKE) prediction (Dear dorff 1980; Tsuboki and Sakakibara 2007).Radiation and surface momentum/energy fluxes are considered with a substrate model (Kondo 1976, Louis et al. 1982;Segami et al. 1989).Additionally, fourthorder numerical diffusion is applied to prevent spurious waves (e.g., Lindzen and FoxRabinovitz 1989;Persson and Warner 1991).To improve computational efficiency, a timesplitting scheme is used (Klemp and Wilhelmson 1978), with a filtered leapfrog method for large steps (Δt ) (Asselin 1972) and the implicit Crank-Nicolson scheme for small steps (Δτ) for time integration (cf.Table 1).

Model experiments and sensitivity tests
Similarly to the strategy adopted by Wang and Huang (2009) and Wang et al. (2011), a larger domain with a coarser grid was first used to simulate the over all evolution of the current earlymorning coastal rain fall event in southwestern Taiwan.As listed in Table 1, in this coarser run (named CTL25; see Fig. 1a for domain area) driven by the sixhourly 0.75° ERAIn terim analyses from 08:00 LST on June 6 for 60 h, we chose 2.5 km as the grid spacing, which is often fine enough in other events with a stronger forcing (e.g., Wang et al. 2014a, b).Next, using the 2.5 -km outputs as initial condition and boundary condition (IC/BC), including those at the surface level, a fine experiment (CTL05) with a grid size of 0.5 km was performed (Table 1) to reproduce the detailed evolution of the convection during a 12 h period starting from 20:00 LST on June 7 (also cf.Fig. 1a) for investigation, particularly the initiation process.At the lower boundary in both runs, terrain data (every (1/120)°) and analyzed sea surface temperature (Reynolds et al. 2002) were also provided.When the 0.5 km experi ment could capture the event in a realistic fashion as verified against the observations, sensitivity tests were performed to examine the impact of Taiwan's topog raphy and thus the land/mountain breeze in this event.However, these experiments were not so straightfor ward, and they are described and discussed in the next section.

The 2.5 km experiment (CTL25)
Driven by the ERAInterim analyses as described above, the CTL25 experiment produced neither the earlymorning rainfall over southwestern Taiwan nor the convection responsible for it at the right time and was thus unsuccessful.In this run, the arrival of the lowlevel southwesterly winds at DJI, CG, and SH (Fig. 10a) and the increase in surface convergence inside the same polygon as in Fig. 7a, although comparable in magnitude, occurred near 06:00 LST (Fig. 7b) and were thus delayed by approximately 5 h.So, the timing of the produced rainfall is way behind and incorrect (Fig. 10a).This issue is likely related to both the coarse temporal resolution of the global analyses (every 6 h) and inadequate observations over the ocean upstream from Taiwan in the northern SCS to correct the errors through data assimilation.Related to this, the simulated surface flow (Fig. 10a) is also somewhat weaker (and more steady) at DJI and CG, but the opposite is true at YJ and GS, since the observed winds in Fig. 7a are also subject to detailed local effects (at micro scale not resolved in the model) near the mountains.At least partly linked to the wind errors, the temperatures over the coastal plains (CG and SH) before rainfall are also too low, by approximately 2 -3°C.Nonetheless, the differences in temperature and wind (mostly within 2 m s −1 ) are deemed acceptable.

The 0.5 km experiment (CTL05)
As an attempt to remedy the timing error, in the fine experiment of CTL05 driven by the 2.5 km outputs, the IC/BC data were artificially shifted early by 5 h; that is, for example, the output at 05:00 LST on June 8 is treated as that at 00:00 LST instead.Though simple, this practice effectively solved the issue and the convection and subsequent rainfall now occur in the CTL05, as illustrated in Fig. 11 together with the CTL25 results (no shifting) for comparison.Without the time shift, the precipitation in CTL25 did not move onshore until approximately 05:00 LST (Figs. 11a -h), compared to the early morning of 00:00 -06:00 LST in the observation (cf.Figs.6d -j).With IC/BC shifted forward in time, the 0.5 km control run captures the initiation and development of the convection near shore during 22:00 -24:00 LST on June 7 (Figs.11i -l) and some of the later, backside development around 01:00 -04:00 LST on June 8 (short arrows in Figs.11m -p), similar to those in the radar observations (cf.Fig. 6).However, the convective cells in the model  (Lin et al. 1983;Cotton et al. 1986;Murakami 1990;Ikawa and Saito 1991;Murakami et al. 1994) PBL/turbulence 1.5order closure with prediction of turbulent kinetic energy (Deardorff 1980;Tsuboki and Sakakibara 2007) Surface processes Energy/momentum fluxes, shortwave and longwave radiation (Kondo 1976;Louis et al. 1981;Segami et al. 1989) Substrate model 43 levels, every 5 cm to 2.1 m * The vertical grid spacing (Δz) of CReSS is stretched (smallest at the bottom), and the averaged spacing is given in the parentheses.# The ICs/BCs from the 2.5 km outputs are shifted early in time by 5 h.See text for details.move faster than in the observation after their forma tion.In both Figs.6 and 11, convection also developed farther upstream inside the stronger southwesterly winds (cf.Figs. 7b,12), in agreement with Wang et al. (2014a), albeit at a considerably weaker wind speed in our case.
The distance-time section along the east-west segment constructed from CTL05 (Fig. 10b) now also compares better with Fig. 8a, especially for surface winds at DJI and CG.However, the air temperatures from CG to YJ before approximately 03:00 LST on June 8 are still too low, and the offshore winds from SH to GS (at approximately 2 -4 m s −1 ), with a finer grid, become even stronger than those in CTL25 but in general not by much (within 1 m s −1 at most sites before 00:00 UTC, cf.Fig. 10a).Since the ICs/BCs of CTL05 were replaced by outputs at a later time (with colder conditions), an improvement in temperature simulation is unlikely and not expected.The rainfall distribution over land also differs from the observa tion to some extent, but this is reasonable since moist convection, with its highly stochastic and nonlinear nature, is a major source of rapid error growth in cloudresolving models (e.g., Zhang et al. 2003;Walser et al. 2004;Zhang et al. 2006;Hohenegger et al. 2006).Nonetheless, as stated in Section 1, we focus our remaining discussion on the key question of the convective initiation over the upstream ocean, where the simulation results also agree better with the observations (less affected by error growth).
Figure 12 presents the simulated surface winds and convergence (at 10 m above ground) in CTL05 at 20 min intervals from 21:20 LST on June 7 to 00:00 LST on June 8, covering the period of convective initiation.Before the initiation of deep convection (near 22:00 LST on June 7), the southwesterly wind surge (of 15 -20 kts) is approaching Taiwan, with convergence at its leading edge that gradually strengthens (Figs.12a -c), in general agreement with Figs. 2 -4.By con trast, the offshore wind is moving westward (outward from Taiwan) and associated with its own conver gence at the leading edge, i.e., the land breeze front (LBF).As these two arcshaped, bulging convergence zones approached each other, narrow bands of new convergence and divergence developed between them closed to 22.4°N, 119.8°E near 22:00 LST (in a region of confluence, cf.Fig. 11j), in agreement with Figs.
The two zones of convergence collided and merged shortly after 22:40 LST (Fig. 12e).After the merger, the resultant convergence zone became less and less continuous but generally moved toward Taiwan, while divergence existed at its backside (Figs.12f -h).
Note that this divergence is associated with mostly southeasterly winds, which form another convergence zone with the oncoming southwesterly, environmental flow.Besides on the immediate backside of old cells (as also observed in Fig. 6), presumably through the coldpool dynamics (e.g., Weisman and Klemp 1986;Rotunno et al. 1988;Doswell et al. 1996), new cells also developed along this new convergence zone far ther upstream in CTL05 (cf.Figs.11m -p).To confirm the CTL05 simulation to be realistic, including its results on the LBF, the reflectivity observed by the Chigu radar at the lowest elevation angle of 0.5° at three selected times is presented in Fig. 13.Some weak (< 10 dBZ) but persistent echoes existed approximately 35 km off the coastline at 21:45 LST on June 7 (black arrow in Fig. 13a) and gradually extended southward and evolved into a linear shape with time (Fig. 13b), most likely in association with the LBF.At such a range, the radar beam is approx imately 350 m above sea level.Later, at 23:08 LST (Fig. 13c), this cloud band was caught up by another line that was also persistent near the leading edge of the southwesterly wind surge (red arrows).The two lines intersected at the location of the deep convection to form a "v shape", a configuration highly similar to our model's result near 23:00 LST (cf.Fig. 12f).In Figs.12b -d, the LBF in the CTL05 run is also comparable in location to the lowlevel cloud band observed by the radar.Most of these features are also visible in Figs.6a -c, but they are less well depicted.
Vertical cross sections along line AB (from south west to northeast; approximately 180 km in length) are constructed from the CTL05 run to further exam ine the initiation of the deep convection in relation to the two approaching boundaries, as illustrated in Fig. 14.The offshore flow from southwestern Taiwan at 21:00 LST is roughly 500 m thick in depth but becomes shallower to 250 m over the next hour (Figs.14a -d).Its origin from the mountain area indicates that the land and mountain breezes are combined, consistent with Qian et al. (2012) and Chen et al. (2016).The LBF at the leading edge is associated with upward motion (~ 20 -30 cm s −1 ) and slowly advances at approximately 12 km h −1 , with structural characteristics illustrated in Figs. 12 and 14 in general agreement with previous studies on land breezes/LBFs (e.g., Simpson and Britter 1979;Ohara et al. 1989;Sha et al. 1991;Wang and Huang 2009;Chen et al. 2014;Li and Chao 2016).By contrast, the leading edge of the southwesterly wind surge (depicted by the red triangle) is moving about twice the speed, from 119.63°N at 21:00 LST to 119.89°N at 22:40 LST (Fig. 14).As the two boundaries move closer to each other, the upward motion, with a banded (wave) struc ture, near 1 -3 km between them gradually intensifies (Figs.14a -c), exceeds 1 m s −1 and penetrates above 4 km into the upper troposphere close to 22:00 LST (Fig. 14d), clearly enough to trigger deep convection (as the LFC is approximately 1.6 km; cf.Fig. 5).As the two boundaries collide shortly after 22:40 LST, the convection develops well into the upper levels (Figs.14e, f).Thus, Fig. 14 confirms that the offshore deep convection in the model develops between the two boundaries (from a southwesterly wind surge and LBF), rather than along the leading edge of either one of them.Notably, such a finding is similar to the results in some studies on Florida SBFs reviewed in Section 1 (Nicholls et al. 1991;Fankhauser et al. 1995;Kingsmill 1995;Fovell and Dailey 2001), even though the SBFs obviously develop over land during daytime.
In Fig. 15a, the distance-time (Hovmoller) plot of surface winds and convergence and the maximum mixing ratio of precipitating hydrometeors in the air column along line AB depicts the overall evolution of the convection in relation to the two advancing boundaries toward each other.Along the section, there are a southwesterly flow behind the convergence zone at its leading edge, an easterly (offshore) flow behind the LBF, and southeasterly winds in between.
In CTL05, likely to adjust to the initial atmospheric condition provided by CTL25 to a finer grid, a rapid spinup (within ~ 30 min) occurs and the LBF propa gates offshore soon after the initial time (20:00 LST).This timing, nonetheless, is in general agreement with previous studies on island circulation of Taiwan and transition between onshore/offshore flows (e.g., Fig. 9 of Ruppert et al. 2013;Chi et al. 1998;Chen et al. 1999;Kerns et al. 2010).Again, by the time when the two boundaries of convergence move to within approximately 36 km from each other along the section at 22:00 LST, convection has already started to develop.After their collision near 22:40 LST, one convergence zone (the one previously associated with the southwesterly wind surge) continues to advance toward Taiwan steadily at the same speed (approxi mately 24 km h −1 ) as before the collision, while the backside convergence with the oncoming upstream southwesterly flow is also clearly visible (Fig. 15a), as discussed earlier (cf.Figs.12g -i), but its speed is reduced.The two boundaries moving away from each other after their collision are also found by Kingsmill and Crook (2003).

Sensitivity tests
Besides CTL25 and CTL05, two additional sen sitivity tests using the grid size of 0.5 km were also performed, as summarized in Table 2.The first is the "noshiftearly" experiment (NSE05), in which the ICs/BCs from the outputs of CTL25 do not shift early in time, and its purpose is to test the evolution of the convection when the southwesterly wind surge arrives too late (by approximately 5 h) using the fine grid.In the Hovmoller plot, as Fig. 15a that covers the period 20:00 -24:00 LST on June 7, the land breeze and its associated front have nearly identical evolution as in the 0.5 km control run (Fig. 15b).This indifference suggests that the time shift has little effect on the offshore flow from Taiwan, including its propagation speed.Even if the strength of the offshore winds and/ or the phase of the LBF is somewhat altered by the time shift, the evolution in convection is likely to remain similar since the southwesterly surge is ad vancing much faster than the LBF.In Fig. 15b, there exists no southwesterly wind surge (which occurs later) without the time shift; subsequently, the LBF simply continues to move away from Taiwan without the occurrence of deep convection (as no significant mixing ratio values of precipitating hydrometeors is manifested).
In the second sensitivity test, the topography of Taiwan is removed (named NTR05) to test its role on the land breeze in the present case.For consistency, this run requires a 2.5 km run driven by the ERA Interim analyses without the terrain (NTR25) before hand to provide its ICs/BCs (which are also shifted early by 5 h; Table 2).In this NTR05 test, the south westerly wind surge evolves nearly the same way, but both the LBF and the steady easterly surface flow behind it disappears (Fig. 15c).Now from the south or southsoutheast, the surface offshore flow becomes very weak and increases in speed only slowly with time, and this indicates that the steep topography of Taiwan is crucial in producing its island circulation, in agreement with Mahrer and Pielke (1977), Jiang (2012), Qain et al. (2012), and Chen et al. (2016).Again, without a stronger offshore flow, deep con vection does not develop in Fig. 15c.Thus, the two sensitivity tests suggest that for the deep convection near shore and the subsequent morning rainfall over southwestern Taiwan to occur in the present case, both the southwesterly wind surge and the land breeze are required.In the model, either one boundary alone is insufficient to trigger the convection under the weak synoptic conditions.As reviewed in Section 1, several other studies also concluded that multiple features are needed to trigger deep convection in their case (e.g., Fovell 2005;Joseph et al. 2008).For convective lines off the southeastern coast of Taiwan (e.g., Yu and Jou 2005;Yu and Lin 2008;Alpers et al. 2010), Wang and Huang (2009) also found that both terrain blocking and island land/mountain breeze contribute to the initiation of deep convection in their case study.
Of course, the southwesterly wind surge also needs to arrive at the correct time for a successful simulation of convective initiation, as indicated by both the CTL05 and NSE05 experiments.Among the earlier studies, the conceptual model put forward by Fig. 21 of Fankhauser et al. (1995), in particular, with some  Remove Taiwan terrain in the general evolution of the event Test the effect of no Taiwan terrain and the associated local circulation in the event Test the effect of late arrival of the southwesterly wind surge in the environment # The ICs/BCs from CTL25 (or NTR25) are shifted to be earlier by 5 h to remove the timing error of southwesterly wind surge.characteristics of horizontal convective rolls also visi ble in Figs.11 and 12 (close to 22:00 LST) before the breakout of deep convection, appears quite applicable in our case here, despite the fact that the two features that provide the approaching wind surges are obvious ly different.

Conclusions
In the present study, a morning rainfall event in the coastal region over southwestern Taiwan on June 8, 2012 under weak synoptic conditions in the meiyu season was investigated, mainly through a detailed mesoscale analysis of observational data and high resolution numerical experiments.The analysis in dicated that the rainfall (mainly during 00:00 -07:00 LST) resulted from convective cells that developed near the leading edge of a moderate lowlevel south westerly wind surge (of 15 -20 kts) approximately 60 km offshore and further intensified (and organized into a quasilinear shape) around midnight when they moved to within approximately 15 km from the coastline of Taiwan.In this nearshore region, surface and lowlevel confluence and convergence occurred between the strengthening southwesterly flow and offshore winds from the interior of Taiwan.Once the convection matured and moved onshore, the cold outflow reinforced local offshore winds and triggered new convection along the coast, i.e., at the backside (to the southwest) of the old cells.The main focus of the present study was to identify the detailed process that triggered the convection offshore, and this was achieved through numerical experiments using the CReSS model.
In the 2.5 km control run (coarse grid) driven by the ERAInterim analyses and intended to reproduce the overall evolution of the event, the southwesterly wind surge arrives roughly 5 h late.To correct this timing error, its outputs used as the ICs/BCs to drive the 0.5 km control run (fine grid) are shifted to be 5 h earlier, and the latter successfully captures the initiation of convection upstream from Taiwan and the subsequent evolution in close agreement with the observations.From the results of this 0.5 km run and those of two other sensitivity tests (one without forward shift in time in ICs/BCs and the other without the terrain of Taiwan), the main findings are summarized below.1) When the two arcshaped and bulging boundaries with convergence (one at the leading edge of the southwesterly wind surge and the other associated with the LBF) approach each other to a distance of approximately 40 km (Fig. 12c), deep convection breaks out between them at around 22:00 LST, rather than along either boundary.While this result is in agreement with some earlier studies with (qualitatively) similar scenarios, the conceptual schematics of Fankhauser et al. (1995), in partic ular, appear also applicable in the triggering of deep convection in our meiyu event under weak synoptic conditions.2) Both features of the southwesterly wind surge and the LBF are required for a successful simulation of convective initiation in the present event, since neither the test with late arrival of the surge (NSE 05) nor the one without Taiwan terrain (NTR05) can reproduce the convection responsible for the coastal morning rainfall.This finding is also in agreement with some previous studies.
3) The offshore flow behind the LBF (approximately 200 -450 m in depth) has its origin from the interi or of Taiwan and thus forms by the combined land and mountain breezes.The steep topography of Taiwan is found to be crucial in its formation, as it becomes very weak in the NTR05 test with the terrain removed but land mass retained.Since the present event took place under weak syn optic conditions, it is perhaps not surprising that some features are needed, the moderate southwesterly wind surge upstream and the land/mountain breeze in this case, to provide enough lowlevel forcing to trigger deep convection and lead to the morning rainfall.Nev ertheless, the findings of this case study are not seen in the literature and offer new knowledge regarding the cause of morning rainfall in southwestern Taiwan during the meiyu season.For a fuller understanding on the topic, however, more studies of other events are needed in the future.

Fig. 4 .
Fig. 4. NOAA OSCAT observation of oceanic winds (m s −1 ; barbs and color) near Taiwan at 23:48 LST on Jun 7, 2012.Full (half) barbs de note 10 (5) kts.The general areas with strong winds ≥ 15 kts near southwestern Taiwan are circled, and the location of Pingdong sounding is denoted by a solid triangle.

Fig. 8 .
Fig. 8. (a) Distance-time cross section of surface air temperature (°C; color isotherms), hourly rainfall (mm; color shading), and wind (m s −1 ; at 10 m height) from (a) five stations along the dotted orange line (west to east: DJI, CG, SH, YJ, and GS) and (b) four stations along the north to south dotted brown line (N -S: SDM, CZ, DK, and SLC) in Fig. 7a, at 1 h intervals from 18:00 LST on Jun 7 to 06:00 LST on Jun 8, 2012.The isotherms are drawn every 0.5°C, and warmer (colder) colors correspond to higher (lower) values (black = 27.5 °C).For winds, a half barb, a full barb, and a flag represent 0.5, 1, and 5 m s −1 , respectively.Time is vertical axis (top-down) in (a) but horizontal axis in (b).

Fig. 9 .
Fig. 9. Manual analyses of surface streamlines and hourly rainfall (mm; color; scale at the bottom) over and near southern Taiwan, overlaid with topography (km; gray shades) at (a) 00:00 LST and (b) 03:00 LST on Jun 8, 2012.For the stations (dots), the temperature (°C) is plotted to the upper left, and rainfall to the lower right (smaller fonts), whereas for winds, a half barb, a full barb, and a flag represent 0.25, 0.5, and 2.5 m s −1 , respec tively.

Fig. 10 .
Fig. 10.As in Fig. 8a, except for the distance-time cross section from (a) the CTL25 and (b) CTL05 simulation using outputs at the locations of the five stations (west to east: DJI, CG, SH, YJ, and GS).Note that CTL05 starts from 20:00 LST on Jun 7, 2012.

Fig. 13 .
Fig. 13.Reflectivity (dBZ; color) observed by the Chigu Doppler radar (white + sign at the center of the rings), at the lowest elevation angle of 0.5°, at (a) 21:45, (b) 22:15, and (c) 23:08 LST on Jun 7, 2012 (source: SoWMEX data website).The distance between the rings are 50 km, and black and red arrows denote low clouds most likely associated with the LBF and a southwesterly wind surge, respectively.

Table 1 .
The CReSS model configuration (top half) and physical schemes (bottom half, with references) used in the 2.5km and 0.5km control simulations in this study.

Table 2 .
The control and sensitivity tests performed in this study and their purposes.The 2.5km (CTL25) and 0.5km con trol experiments (CTL05) are the same as those listed in Table1.For the sensitivity tests, two experiments (2.5 and 0.5km) of no Taiwan terrain (NTR) and one (0.5km) of no shiftearly (NSE) in ICs/BCs are performed.