Evolution of precipitable clouds

In this part, this study will analyze the local thermodynamic conditions that led to the D18 event. Figure 3.10 shows that at 1200 UTC 8 Dec, the equivalent potential temperature at 925 hPa over central Vietnam and the SCS is commonly between 325 K and 340 K, the relative humidity is over 90% exists between 106E and 110E and reached the 800 hPa level (Fig. 3.10b). The strongest moisture air vertical upward occurs around 108E (Fig. 3.10c). The satellite image also shows a series of convective cells (red, white, black colors) first formed over the study area's northern and central parts. These convective clouds are composed of several big and small isolated cells. Their distribution separates into two main directions. One direction is northeast-southwest over the northern part, and the other direction is east-west over the central part of the study area. Over the next 6h, most small isolated cells were weakened significantly and dissipated after 1800 UTC while moving slowly offshore. The life span of these small convective clouds cells was from 1 to 3 hours. On the contrary, the large convection cloud in the north of the study area appears to be intensified and reach better structured at 2000 UTC, then weakened significantly and dissipated after 0100 UTC 9. These convective cells' life span was about 12 hours—however, this large cell distributed over the ocean. After 18 UTC 8 to 11 UTC 9 Dec, new small cells continued to appear and scattered develop along the coast and move slowly offshore. Besides, there is a remarkable convection cloud, forming over the coastline of the southern part of the study area at 2300 UTC 8 Dec. This convective cell's development is very vigorous and matured at 0400 UTC 9 Dec, then gradually weakened while moving slowly eastward over the coastal sea (Fig. 3.11).

Besides developing these convective cells, this area is also covered by the Stratus clouds (green colors), Including the nimbostratus clouds, the Altostratus cloud, etc. These clouds first formed over northern parts of the study area at 1200 UTC 8, and then grows and expands southward. Especially along the coast. This cloud also covered the entire study area from 2300 UTC 8 to 1100 UTC 9.

Furthermore, the Column Maximum radar reflectivity (Cmax) over the Central Vietnamese area with 1-h intervals shows that the first rainfall occurred along the coast from 1200 UTC to 1700 UTC 8. After that, the rain zone has extended to both the coastal sea and inland. In particular, during 2000 UTC 8 to 0200 UTC 9, Cmax reached the highest value (40 dBZ), which shows that the rainfall intensity was the greatest during

14

this period. After 0200 UTC, the intensity of rain tends to decrease gradually. In general, the Cmax value of December 8 fluctuated quite steadily between 15 dBZ and 35 dBZ, indicating that rainfall was not too intense. However, steady rain falls over many hours, leading to the cumulative rainfall of this day is very large (Fig. 3.12). At 1200 UTC 9, these thermodynamic conditions of the previous day tend to strengthen. Particularly, the high equivalent potential temperature (>340 k) expands from the southern part to the northern part of the SCS (Fig. 3.13a). The easterly wind expands to 750 hPa (Fig. 3.13b).

The high relative humidity (>90%) cover from 104E to 112E and reached 700 hPa level converges with equivalent potential temperature decreases with altitude, indicating the area that occurs instability in saturated air is larger than it in the previous day. Besides, the strong moisture air vertical upward occurs not only around 108E as a previous day but also expand to 111E. The strong vertical downward occurs around 106E (Laos) (Fig.

3.13c). These thermodynamic conditions have led to the formation of precipitable clouds along the coastline and coastal sea, observed by satellite imagery. Concretely, figure 3.14 shows the infrared enhanced color images of the Himawari satellite over the Central Vietnamese area from 1200 UTC 9 December 2018 to 1100 UTC 10 December 2018. It clear to see that the most prominent is the formation and strong development of the Stratus clouds (green colors). This cloud has covered the study area during 1200 UTC 9 to 2300 UTC 9, then gradually disintegrated. Besides, it clear to see that a small convective cell formed and developed over the area of Da Nang (Da Nang is one of the main extreme rainfall centers of the D18 event) at 1200 UTC 9. The life span is about 5 hours (1200 UTC 9 to 1600 UTC 9), and a few cells have formed and developed over the southern part of the study area. In which some cells formed from the previous day and existed until this day. Some of them appear on this day. All these convection cells' common features are formed and developed in the coastal area, then gradually weakening and disintegrating while moving to the sea. (Fig. 3.14). From another point of view, the Cmax data show that the rain occurred 24 hours of this day. However, Intense rain occurs mainly from 1200 UTC 9 to 2100 UTC 9. after that, the rain gradually decreased in both intensity and area. Besides, the Cmax data also indicating that rainfall is mainly distributed over the coastal plain and coastal sea (Fig. 3.15).

At 1200 UTC 10 Dec, the vertical motion tends to decrease in intensity compared to the previous day. High humidity (>90%) areas also narrow significantly in the same

15

comparison. The easterly wind at 750 hPa level is replaced by southeasterly wind (Fig.

3.16). These changes reduce the development of convection clouds, which have been observed by satellite imagery (Fig. 3.18). At 1200 UTC 11 Dec, the vertical upward motion still exists over central Vietnam. However, relative humidity has decreased considerably more than in the previous day, especially at upper levels (> 500 hPa). The southeasterly wind at 750 hPa is replaced by southeasterly wind. As a result, environmental conditions are unfavorable for the development of convection clouds in the area, resulting in reduced rainfall (Fig. 3.17). Particularly, figure 3.18 shows that the precipitable cloud has been decreased significantly or only exists at sea. Furthermore, the number of convective cells has been greatly reduced compared to the previous two days.

Only two small convection cells were formed and developed in the study area's coastline.

One cell appeared over the central coastline of the study area at 1400 UTC 10, then gradually weaken while moving eastward to the SCS and dissipated after 1800 UTC 10.

Another cell formed at 2000 UTC 10 over the southern coastline. This cell grew very vigorously and matured at 2200 UTC 10, then weakened significantly and completely disintegrate at 0200 UTC 11. However, the lifetime of these convective cells on land is only 1-3 hours. The size of these cells is very small. The decrease of the precipitable cloud leads to a significant decrease in rainfall in this day. This can be seen in Figure 3.19.

Particularly, Cmax data show that rainfall occurs mainly from 1 hour to 3 hours, then decreased significantly with time.

In general, during the D18 event, due to the interaction of the low-level cold surge, originating in China, with the low-level easterlies wind over the South China Sea (SCS), led to the formation of a strong low-level convergence and then local deep convections.

Besides, the strong easterly and strong southeasterly anomaly winds also played an important role in transporting moisture from the tropics across the SCS toward central Vietnam. As analyzed above, these conditions, combined with the local thermodynamic conditions, led to the formation of precipitable clouds, including the nimbostratus clouds and convective cells. However, these convective cells are not evenly distributed. They are concentrated mainly in the northern sea or the southern coastline of the study area.

The size of these cells is also very different, composed of several big and small isolated cells. Small cells' lifespan is very short (from 1 to 3 hours), while several big cells can exist up to over 10 hours. Most of these convection cells tend to move slowly offshore.

16

This movement is due to strong southwest or westerly winds at higher pressure levels (refer Fig. 3.6 and Fig. 3.8c). Besides, it is noteworthy that these convection cells mainly form and develop during the nighttime. Meanwhile, the formation and strong development of the Stratus clouds during the D18 event led to steady rains over many hours.

17

CHAPTER 4.

RESULTS

In this chapter, all results of the daily time-lagged forecasts executed by the CReSS for the D18 event are present, and these results will be compared with observations, validated, to investigate the predictability of CReSS model for the D18 event.

4.1. 24-h accumulated rainfall for individual days and three days 4.1.1. 24-h accumulated rainfall for 9 Dec 2018

With four 8-day forecasts per day. In which, the first forecast performed on 1200 UTC 3 Dec 2018, and the last forecasts are on 1200 UTC 8 Dec 2018. Fig. 4.1 showed 21 scenarios of possible future states of the atmosphere (24 hours of rainfall and wind surface fields) for 9 Dec 2018 by CReSS. In general, most scenarios predicted 24-h of rainfall closer to observation. Particularly, these scenarios indicate 24-h rainfall up to over 500 mm (shading by green color), which is the same in observation. However, the location and spatial distribution of the 24-h rainfall are very different from the observation data and scenarios. Along the central Vietnamese coast, where the observed rainfall exceeds 500 mm, most scenarios predicted rainfall below 100 mm. These differences may be due to the model having mispredicted other atmospheric variables, such as wind surface field.

In Fig. 4.1, some scenarios mispredicted the wind direction over the SCS and mid-central Vietnam. For example, scenarios on 7 Dec 2018. Some scenarios have incorrectly predicted the magnitude of the wind field, such as scenarios on 3 Dec 2018. The accuracy of forecasts decreases with lead times. At the shortest-range prediction (1200 UTC 08), the direction and magnitude of the surface wind field are more similar to the reanalyzed data than other predictions at other times. At the longest-range prediction (1200 UTC 03), The direction and magnitude of the surface wind field are not the same as the reanalysis data. The skill scores in Fig. 4.2a also shows that the member that was run at the shortest-range prediction (1200 UTC 8) has the highest score (0.66). The member that was run at the longest-range prediction (1200 UTC 3) has the lowest score (0.04). These scores are consistent with the previous analysis.

18

4.1.2. 24-h accumulated rainfall for 10 Dec 2018

Figure 4.3 presents 25 possible scenarios of 24-h rainfall and average wind surface fields for 10 Dec 2018. The first forecast runs on 1200 UTC 3 Dec 2018, and the last forecast ran on 1200 UTC 9 Dec 2018. It can be clearly seen that several members have not only a very good 24-h rainfall forecast but also a good forecast for the location and spatial distribution of rainfall. Namely, members executed on the 8th and the 9th of December. Moreover, there is an impressive member that ran at 1800 UTC on 4 December. This member has a good forecast of rainfall even though the forecast made at a long forecast range before the target date (approximately five days before the target date). A common feature among these members is that they both well predict the direction and magnitude of the surface wind field. Besides, many members cannot forecast rainfall.

Most of them executed at forecast periods longer than two days before the target date. In general, most of them cannot predict the direction or magnitude or both of these characteristics of a surface wind field. As well as the analysis for the previous day, this may be the reason why they cannot predict the rain accumulation field of 24 hours. The fraction skill scores in Fig. 4.2b also indicated members ran on the 8th and the 9th of December, and at 1800 UTC 4 December reach scored higher than the rest of 25 members.

4.1.3. 24-h accumulated rainfall for 11 Dec 2018

Figure 4.4 show 29 possible scenarios of 24 hours rainfall and wind surface fields for 11 December 2018, executed between 1200 UTC 3 December and 1200 UTC 10 December. It is clear to see that most scenarios forecast quite good the quantitative rainfall in 24-h, although the 24-h rainfall observed of December 11 has decreased significantly compared to the previous two days. However, most of the scenarios predicted the spatial distribution of inland rainfall is smaller than the observed data, which may be related to the fact that members inadequately predicted the direction and magnitude of the wind field, resulting in the incorrect prediction of moisture transport and moisture convergence.

For example, members executed on December 3. These members did not predict the surface wind field; as a result, they could not predict the accumulative rain field of December 11. Besides, the fraction skill sore in Fig. 4.5a also shows that the score of each member in 25 members is low. The member has the highest score (approximately 0.4), was executed at 1800 UTC 7 Dec.

19

4.1.4. 72-h accumulated rainfall between 1200 UTC 8 and 1200 UTC 11 Dec 2018

Due to the long duration of the D18 event, the 72-h accumulated rainfall between 9 and 11 December is also analyzed. A total of 21 members performed from 1200 UTC 3 December to 1200 UTC 8 December. Compared to the observation data, the results show that three members running at 0000 UTC 8, 0600 UTC 8, 1200 UTC 8 well-forecasted rainfall field, both in quantitative rainfall and spatial distribution of rainfall.

Besides, there is a remarkable member that ran at 1800 UTC 4 December. This member has a good forecast of accumulated rainfall in 72 hours, although this member is run approximately four days before the target period. The rest of 21 members forecast accumulative 72-h rainfall inland below 200 mm, which is much lower than the observed data. It is clear that these members did not correctly predict the direction of the wind or the magnitude of the wind field, or they did not correctly predict both the direction and magnitude of the wind field (Fig. 4.6). These assessments are supported by the skill scores in Fig. 4.5b. Particularly, three members ran on Dec 8 has scored higher than 0.7. the member executed at 1800 UTC 4 Dec has score 0.63. The rest of the 21 members has a score lower than 0.5 (Fig. 4.5b).

4.2. Ensemble mean

As we know, an ensemble weather forecast is a set of forecasts that present the range of future weather possibilities obtained from multiple separate members. Hence, the simplest way to use the ensemble forecasts is by computing the ensemble mean.

Besides, some studies showed that the ensemble mean will have a smaller error than the individual ensemble members. This error reduction occurs because high predictability features that the members agree on are emphasized by the mean, while low-predictability features that the members do not agree on are filtered out or heavily dampened (e.g., Leith 1974; Murphy 1988, Surcel et al. 2014). Therefore, this section will show the ensemble mean of rainfall scenarios and its spread for every single day and three days of the D18 event.

20

4.2.1. Results from daily ensemble

In order to see how good Cress can predict the D18 event with the time-lagged strategy (Units are days). From possible scenarios of 24 hours of rainfall for single days and 72 hours for three days during the D18 event that produced by CReSS, this study has grouped scenarios and compute into different lead times, from day 1 (0 h), day 2 (6-24h), and until day 8 (150-168h) depend on the numbers of the member that made for individually days of the D18 event.

Figure 4.7 shown scenarios of 24-h mean rainfall by CReSS and observed rainfall for 9 December 2018. It shows that CReSS generated rainfall at the shorter lead time (day 1, day 2) much more than it at extending ranges (day 3, till day 6). Concretely, at the lead time day 01 and day 02, the rainfall amount produced by CReSS is approximately the observed rainfall with observed peak rainfall up to 639 mm. However, the area covered by rainfall higher 200 mm is much smaller than it by observed data. The rest ranges have accumulated rainfall common between 50 and 100 mm, which is much lower than the observed data. Besides, Fig. 4.7 show also that most rainfall occurred over the ocean.

Besides, examination more in-depth by statistic scores showed in detail the model’s predictability at various ranges within days from 1-6 and at different rainfall thresholds (Fig. 4.8). It shows that at rainfall thresholds lower than 50 mm, model has well-predicted at all lead times (except day 3) with minimum grades at 50 mm: ̴ 0.2 (day 6) ≤ TS ≤ 0.7 (day 1), ̴ 0.2 (day 6) ≤ POD ≤ ̴ 0.8 (day 1), ̴ 0.2 (day 6) ≤ FB ≤ ̴ 0.9 (day 1), 0 (day 6) ≤ FAR ≤ ̴ 0.3 (day 5). At 50 mm < thresholds ≤ 100 mm, model prediction skills decrease to 0 at threshold 100 mm at prediction ranges longer than two days. At thresholds ≥ 100 mm, only day 1 and day 2 show the forecasting skills. In which, day 1 have TS > 0 at 400 mm, while the commonly observed rainfall is between 50 and 450, and the observed peak rainfall is 639 mm (Fig. 4.8d).

Overall, it shows that the model has high-quality QPFs at shorter prediction ranges (day 1 and day 2) with FSS scores is 0.7 for day 1 and 0.4 for day 2. Day 3 range has the worst quality QPFs in forecast ranges (Fig. 4.8e). This worst quality may be related to the model incorrectly predicted the surface wind field in every single run on that day.

Figure 4.9 shown scenarios of 24-h rainfall mean by CReSS and observed rainfall for 10 December. It can be clearly seen that the rainfall produced by model on day 1, days 2, and day 3 is quite similar to observed data not only rainfall amount but also the

21

locations of significant rainbands and regions that concentrate mainly rainfall. For lead times after day 3, the model made the rainfall scenarios much lower than observed rainfall data. In which, day 4 and day 5 are the lowest. It can be relevant to the model did not predict well the wind fields in every single run at extend ranges (after day 3) as analyzed previous. The rainfall of day 6 is the highest these days due to the model have single good-predict at 1800 UTC 04 [Fig. 4.3. (1800 UTC 04)].

Skill scores also show their relevance to the above analysis. In particular, day 1, day 2, day 3 has skill scores higher than the rest at all thresholds. Besides, day 1 has skill scores highest at threshold 350 mm with TS = 0.1, POD = 0.1, FB = 0.8, FAR = 0.9.

Furthermore, day 1 still shows skill scores at 400 mm < threshold < 500 mm (TS >0, POD

>0, FB >0.7, 0.9< FAR <1, due to observed daily rainfall amount is extreme with the observed peak rainfall is 644 mm, so skill scores >0 is possible. At extend ranges after day 3, the only day 6 have skill scores at threshold >100 mm (TS = 0.1, POD = 0.1, FB

= 0.2, FAR = 0.5. Day 4, day 5, and day 7 show skill scores at thresholds lower than 50 mm, and these scores equal to zero at threshold 50 mm (Fig. 4.10).

Overall, the model has a high skill in predict QPFs at shorter ranges (< 72 hours) with FSS = 0.4 for day 1, FSS = 0.6 for day 2, FSS = 0.7 for day 3 (Fig. 4.10e), and this skill decreases with extending ranges due to the errors that tend to grow with time.

Figure 4.11 indicates scenarios of 24-h rainfall mean by CReSS for 11 Dec 2018. With an ensemble 08 day predict, the model produced rainfall scenarios has rainfall amount is quite closer to observed rainfall data commonly at ranges days 1– 6. However, the spatial distribution of rainfall in these scenarios is smaller than in reality. Special, Scenario day 3 did not predict rainfall distribution in the eastern half of the study area. The remaining scenarios did not predict the distribution of precipitation in the southern half part and a small region in the northern part of the study area. Besides, Scenarios day 2 and day 4 considered to have the best quality in these scenarios. Scenario day 7 predicted rainfall is much lower than observed rainfall data. Day 8 did not predict the rainfall field and was considered the scenario to have the worst quality.

Besides, skill scores show detailed information about the model’s predictability at different forecast ranges and rainfall thresholds. The model can predict QPFs at all forecast ranges (except day 8) at thresholds <50 mm, this is rainfall commonly in reality.

However, these skill scores decrease (increase) rapidly with the increase of the rainfall

22

thresholds. In which, rainfall scenarios at day 2 and day 4 have high-quality QPFs than

thresholds. In which, rainfall scenarios at day 2 and day 4 have high-quality QPFs than