HTML
--> --> -->Since 2010, accuracy in humidity and LWC modeling has improved. For example, a moist boundary-layer scheme based on eddy diffusion combined with mass-flux transport (K?hler et al., 2011) was introduced in the European Centre for Medium-Range Weather Forecasts (ECMWF) model. The new scheme helps to substantially increase the modeled boundary-layer cloud cover, improving accuracy. It also improves the accuracy in total cloud cover as well as the temperature and humidity at the 2-m level in parts of the Northern Hemisphere and the tropics (Dee et al., 2011; Haiden et al., 2015). The increase in accuracy of humidity at the 2-m level means that the initial humidity profile is more accurate, which improves accuracy in the initial conditions. The greater accuracy in the initial humidity profile over sea should allow the development of a useful NWP model with higher forecast scores. Producing such a model is the goal of this study.
Sea fog in China mainly occurs in three regions. In the South China Sea, such fog is common from January to May, mainly occurring on the southern China coast. In the East China Sea, it is common from March to July. And in the Yellow Sea, such fog is common from March through August, trending northward as the year progresses. Among these areas, sea fog has greater extent and longer duration in the East China Sea and Yellow Sea than in the South China Sea (Wang, 1985). As a result, most studies of sea fog in these areas focus on the East China Sea and Yellow Sea (Hu and Zhou, 1997; Fu et al., 2002; Gao et al., 2007; Zhang et al., 2009; Heo and Ha, 2010; Kim and Yum, 2012; Li et al., 2012), which are also the areas with an established sea-fog forecast method (Table 1; Yu et al., 2007; Huang et al., 2009a; Gao et al., 2010; Wang et al., 2014). However, severe sea-fog-related disasters also occur in the South China Sea, a region that still does not have an accurate NWP model forecast method (Zhang, 2016).
The southern China coast is affected by the East Asian monsoon (Tao and Chen, 1987). The winter part of this monsoon has wind from the northeast, cooling the sea surface off the southern China coast to more than 3°C below the annual average (Locarnini et al., 2006). This cooling results in a large gradient in SST over the northern part of the South China Sea (Chen, 1983). Then, when warm, moist air moves north from the warmer SST area in winter and spring, advection sea fog usually forms (Wang, 1985; Huang et al., 2011a; 2015).
In 2007, the Institute of Tropical and Marine Meteorology (ITMM) established the Marine Meteorological Science Experiment Base (MMSEB) at Bohe, Maoming, China. As part of this base, the Integrated Observation Platform for Marine Meteorology (IOPMM) was set up in 2008 and has since acquired many observation datasets of sea fog (Huang and Chan, 2011; Huang et al., 2011a, 2015). Recently, the ITMM developed a regional sea-fog forecast method based on both the Global and Regional Assimilation and Prediction System (GRAPES), an NWP model, and the model output statistics (MOS) method (Huang et al., 2011b, 2016a). Although the operational GRAPES-MOS forecast method by itself gives a better forecast score than previous statistical methods, it focuses only on the Guangdong coastal area and does not show how the sea fog changes, a feature needed by forecasters (Table 1). Hence, a forecast method based on a regional model like GRAPES is urgently needed.
Towards this goal, we spent 2014 and 2015 running tests with GRAPES to determine how it can be incorporated into a method to predict sea fog. This GRAPES model, hereafter referred to as GRAPES-3km, because it has a horizontal resolution of about 3 km, models the LWC of the air near the sea surface, and then uses this measure to determine if fog is present. We describe and evaluate GRAPES-3km for sea-fog prediction here, also comparing it with previous results from the Weather Research and Forecasting (WRF) model (Yuan and Huang, 2011; Huang et al., 2016b). We show that this new GRAPES model, i.e., GRAPES-3km, produces an average ETS score of 0.20 and a Heidke skill score (HSS) of 0.335 for sea-fog events at 24 h in 2016-17. These scores are improvements over those from the previous GRAPES-MOS forecast method by about 5.6% (ETS) and 6.4% (HSS) for the west and east Guangdong areas (Table 1; Huang et al., 2017).
The present paper is organized as follows: Section 2 describes the model system and the configuration of GRAPES-3km. Section 3 explains the evaluation data and method for GRAPES-3km. Section 4 is a detailed model evaluation of two fog cases. Section 5 analyzes the prediction performance of GRAPES-3km at individual stations. A summary of our findings is provided in section 6.
2.1. Model system
The GRAPES model is a new-generation NWP model developed by the China Meteorological Administration. The model is a fully compressible, non-hydrostatic, global-regional unified model. It presently has three-dimensional variational assimilation with a traditional semi-Lagrangian advection scheme (Chen et al., 2008; Xue et al., 2008). The latest generation of the GRAPES model includes several improvements, such as a new model algorithm, a cloud-base mass flux in the cumulus parameterization, and a new orographic drag parameterization scheme (Xu et al., 2014; Zhang et al., 2015; Zhong and Chen, 2015; Chen et al., 2016). As such, GRAPES has been incorporated into three systems in the Guangdong Provincial Key Laboratory of Regional Numerical Weather Prediction. One such system is called GRAPES-MARS, which is a mesoscale atmospheric regional model system. We use GRAPES-MARS here in our operational sea-fog forecast model (GRAPES-3km). GRAPES-MARS is a stand-alone model with a 0.03° horizontal resolution and 56 vertical levels (Table 2).2
2.2. Initial conditions
The initial conditions are extracted from the analysis field from the ECMWF (https://www.ecmwf.int/). The data have a horizontal resolution of 0.125°×0.125° and 17 vertical levels.2
2.3. Establishing GRAPES-3km
Establishing GRAPES-3km involves the steps shown in Fig. 1. As a background study, we ran tests to simulate the sea fog on the southern China coast by using the WRF model [version 3.5.1, with a 9-km horizontal resolution (Skamarock et al., 2008; Huang et al., 2016b)]. These tests examined the effects from increasing the horizontal resolution of initial SST, increasing the number of vertical levels in the boundary layer, and simulation of the evolution of the main meteorological parameters. Results were compared with an observation dataset that included observatory data, cloud-top temperature from MODIS, GPS sonde data, and data from the IOPMM. We also used the WRF-derived model-level LWC as an indicator of sea fog, and found good agreement with observations (Huang et al., 2016b).
Figure1. Method to establish the GRAPES-3km model for predicting sea fog. (Region within dashed line corresponds to the present study).We then applied our experience with the WRF model to the GRAPES model. With GRAPES-3km, we ran initial tests to ensure that the use of 56 vertical levels was sufficiently accurate. As an indicator of sea fog, we used the GRAPES-derived model-level LWC. We also ran both GRAPES-3km and WRF to compare to several typical sea-fog cases on the southern China coast. Both models gave very similar patterns for low-level LWC (Fig. 2). The resulting parameterization is shown in Table 2. We found the Yonsei University (YSU) planetary boundary layer scheme and the WSM6 microphysics scheme to be the most accurate schemes for fog simulation with the GRAPES model. The role of GRAPES-3km is summarized in Fig. 3.
Figure2. Predicted values of LWC (units: g kg-1) at the 20-m level over sea at 2000 LST 21 February 2015: (a) values from the WRF model (two-nested grids, inner grid is a 9-km horizontal resolution) output, wherein colored dots depict the observed visibility in km at 2000 LST 21 February 2015 at the coastal meteorological stations; (b) as in (a) but from the GRAPES model (this study) output.
Figure3. Flowchart of the GRAPES-3km model.We also examined 15 case studies from 2011-15 by simulating them with GRAPES-3km, and then testing the simulations against retrieved fog and cloud patterns from Himawari-8 satellite data, buoy data, IOPMM data, and coastal-station observations. The mostly similar patterns from the tests gave us more confidence in the model.
2
2.4. Output data
For fog detection, we examine the output LWC (g m-3) at the first model level (20 m), and calculate the visibility Rm (km) using the conversion formula determined during a field experiment in 2007 (Huang et al., 2009b): \begin{equation} R_{\rm m}=0.085{\rm LWC}^{-0.61} . \ \ (1)\end{equation}The field experiment ran from 16 March to 29 April during the fog season at the MMSEB in Bohe. We found that, after forming, the sea fog remains as long as the 10-m wind speed remains below 12 m s-1. The above relationship is based on 51 sea-fog droplet distributions (including light fog and dense fog) and manual visibility observations. Following the World Meteorological Organization (WMO, 2008), fog is defined when the visibility goes below 1.0 km. According to Eq. (1), this visibility occurs when LWC=0.017 g m-3. However, at the observation stations, the relative humidity (RH) is measured, not the LWC. So, for these stations, we use a previous finding from MMSEB in which the visibility goes below 1.0 km as RH goes above 98% according to the GPS sonde data (Huang et al., 2011a). To help forecasters see the entire sea-fog process, GRAPES-3km can now make predictions up to 72 h ahead and produce hourly outputs over a sea-fog event.
3.1. Station observations
We focus on the area in southern China from 17°N to 26°N that has shoreline on the South China Sea. This area extends from the Beibu Gulf in the west to the Taiwan Strait in the east (Fig. 4). In this area, sea fog tends to appear about 100-200 km offshore (Wang, 1985; Huang et al., 2015). To test the ability of GRAPES-3km in this area, we use 13 representative meteorological stations, two buoy datasets (Maoming buoy #59765 and Dongguan buoy #G5942), and the IOPMM dataset. These stations are distributed almost uniformly along the northern shoreline of the South China Sea, divided into four sea areas: (I) Beibu Gulf (stations 1-3); (II) west Guangdong (stations 4-9); (III) east Guangdong (stations 10-12); and (IV) Taiwan Strait (stations 13-16). The locations are given in Fig. 4.
Figure4. Range (17°-26°N, 105°-122.5°E) and observation stations of the operational GRAPES-3km model. Sea areas are marked as: (I) Beibu Gulf (1-3); (II) west Guangdong (4-9); (III) east Guangdong (10-12); and (IV) Taiwan Strait (13-16). Stations are marked as follows: (1) Fangchenggang; (2) Beihai; (3) Weizhou Island; (4) Haikou; (5) Zhanjiang; (6) Maoming Buoy (59765); (7) IOPMM; (8) Yangjiang; (9) Zhuhai; (10) Dongguan Buoy (G5942); (11) Shanwei; (12) Shantou; (13) Dongshan; (14) Xiamen; (15) Chongwu; (16) Pingtan. Location A (red) is MMSEB.In this area, fog occurs most frequently around the Leizhou Peninsula, followed in frequency by the Beibu Gulf and Taiwan Strait. The main fog season is from January to April for stations 1-12, and this season extends to May for stations 13-16 (Wang, 1985). We assume that fog occurs when the observed visibility goes below 1.0 km at the representative meteorological stations.
2
3.2. Buoy and IOPMM data
We use RH observations from the buoys and IOPMM. Each buoy lies about 100 km south of the shore and has a 10-m diameter circular iron base. Their RH sensor is installed about 10 m above sea level. The IOPMM station lies over water about 15 m deep and about 6.5 km from the nearest coast (Fig. 5). At this station, the nearest humidity sensor to the 10-m level of the buoys is the sensor at 13.4 m. So, we use the RH data at the 13.4-m height above sea level for fog detection. For comparison to model forecasts, we instead use the RH sensor at the 16.4-m height. [see (Huang et al., 2015) for more information about IOPMM]. For the GPS sonde data and model forecast results, we assume fog occurs when RH ≥ 98% (Huang et al., 2011a; 2015). But due to aging of the RH detectors on the buoys and IOPMM, we assume that fog occurs when the RH ≥ 95%. (The value of 95% is not exact, but after ageing at the buoys, we found that the maximum reading always decreased from 100% to about 97%-98%).
Figure5. The IOPMM station (left) and a 10-m diameter buoy (right). (Both photos courtesy of Weikang MAO. Left-hand photo taken on 27 June 2014; right-hand photo on 24 September 2015.)2
3.3. Satellite data
We use Himawari-8 satellite data to determine fog and cloud extent near the southern China coast. Himawari-8 is the 8th Himawari geostationary weather satellite operated by the Japan Meteorological Agency after MTSAT-2, becoming operational on 7 July 2015. For the Asia-Pacific region, Himawari-8 provides 16-channel cloud and clear-sky imagery in both the visible and infrared bands (Bessho et al., 2016). We use hourly imagery with a 0.05° horizontal resolution, which is freely available from Kochi University (http:// weather.is.kochi-u.ac.jp/wiki/archive).The daytime sea-fog retrieval includes: (1) cloud and clear detection based on the visible channel; (2) sea-fog and other cloud-detection variables based on the temperature difference between cloud top and sea surface (Zhang and Yi, 2013) as well as the standard deviation in the thermal infrared channel (Wu and Li, 2014); and (3) nighttime sea-fog detection that mainly relies on the dual-channel difference test (Gao et al., 2009; Wu and Li, 2014).
For sea-fog observations, stations at sea, on land, and from orbit, each have advantages and disadvantages. (1) Observations over sea are essential to verify a sea-fog forecast; however, maintaining accurate remote sensors at sea is difficult and expensive, making their use rare. (2) Observations from coastal stations can help us understand how a given sea-fog event changes; however, as some coastal stations are not very close to the coastline, they sometimes mistake radiation fog as sea fog (Lamb, 1943). In addition, coastal stations cannot observe the entire extent of the fog, especially the parts further out to sea. (3) Fog retrieval from satellite data can, in principle, detect the fog over its entire extent, but the method remains prone to misidentification. In particular, it is impossible for satellite detection methods to distinguish between low-level stratus and sea fog due to their similar top characteristics (Wang et al., 2014). Thus, uncertainty in observed sea-fog data remains.
Finally, evaluating GRAPES-3km relies on both the RH observations over sea, which can provide definitive evidence of sea fog, and retrieved fog and cloud patterns from Himawari-8 satellite data. The satellite data can show a continuous evolution of the fog and cloud patterns, which can effectively determine the forecast range of GRAPES-3km. After establishing GRAPES-3km in this way, we apply the model to operational forecasting at the Southern China Regional Meteorological Center in the sea-fog season from 2016 to 2017.
2
3.4. Reanalysis data
For the analysis of SSTs of the sea-fog cases, we use the real-time global (RTG High Res 0.083, http://polar.ncep.noaa.gov/sst/) data from the National Centers for Environmental Prediction (NCEP) (Thiébaux et al., 2003). We also depict the wind vector and mean sea-level pressure data by using the NCEP global analysis software on 1°× 1° grid spacing datasets. The research data archive is at the National Center for Atmospheric Research, Computational and Information Systems Laboratory (http://rda.ucar.edu/datasets/ds083.2, accessed 12 March 2012).2
3.5. Evaluation method
To evaluate the forecasts, we use the following criteria: (1) The forecast and observation periods are on the same day, from 2000 LST to 1959 LST of the next day. (2) All fog cases within a one-day period are considered as one case. (3) To be used in the analyses, the station must have at least eight cases of fog during that year's fog season (from January to May). (4) The first day period of the evaluation period is 0-24 h, the second day period is 24-48 h, and the third day period is 48-72 h. The spin-up time of GRAPES-MARS is about 3 h (Zhang et al., 2016).We evaluate the prediction of sea fog as a simple yes or no. The description of the 2× 2 problem with all possible outcomes (contingencies) for a fog event is put in the contingency table of Table 3. Thus, a prediction can be correct by being either a "hit" (a) or a "correct rejection" (d), and incorrect by being either a "false alarm" (b) or a "miss" (c). For example, with values from Zhanjiang station (Table 3), we calculate the following scores (Murphy, 1996; Jolliffe and Stephenson, 2012):
The percentage of correct forecasts (PCF): \begin{equation} \label{eq1} {\rm PCF}=\frac{(a+d)}{n}=\frac{(22+72)}{121}=78\% . \ \ (2)\end{equation} The probability of detection or hit rate (H): \begin{equation} \label{eq2} H=\frac{a}{(a+c)}=\frac{22}{(22+20)}=0.52 . \ \ (3)\end{equation} The false alarm rate (F): \begin{equation} \label{eq3} F=\frac{b}{(a+d)}=\frac{7}{(22+72)}=0.07 . \ \ (4)\end{equation} The false alarm ratio (FAR): \begin{equation} \label{eq4} {\rm FAR}=\frac{b}{(a+b)}=\frac{7}{(22+7)}=0.24 . \ \ (5)\end{equation} The critical success index or threat score (TS), which is commonly used as a measure of rare events: \begin{equation} \label{eq5} {\rm TS}=\frac{a}{(a+b+c)}=\frac{22}{(22+7+20)}=0.45 . \ \ (6)\end{equation} The Gilbert skill score, also called the ETS: \begin{equation} \label{eq6} {\rm ETS}=\frac{a-a_{\rm r}}{(a+b+c-a_{\rm r})} . \ \ (7)\end{equation} Here, ar is the expected value for a random forecast, calculated as \begin{equation} \label{eq7} a_{\rm r}=\frac{(a+b)(a+c)}{n}=\frac{(22+7)(22+20)}{121}=10.07 , \ \ (8)\end{equation} so that the ETS score is \begin{equation} \label{eq8} {\rm ETS}=\frac{a-a_{\rm r}}{(a+b+c-a_{\rm r})}=\frac{22-10.07}{(22+7+20-10.07)}=0.31 . \ \ (9)\end{equation}
The ETS has been used in previous studies for evaluating fog forecasts (Zhou and Du, 2010; Wang et al., 2014; Román-Cascón et al., 2016) and is a standard measure for binary forecasts (Jolliffe and Stephenson, 2012), as is the case here. (Hogan et al., 2010) pointed out that the expected ETS of a random forecasting system falls below 0.01 only when the number of samples exceeds about 30. As our values of n exceed 30, this potential issue is not a concern. However, as the ETS can have biases, we also use two scores that are equitable for all values of n and, in case other measures may become more widely accepted in the future, we also give our values of a, b, c, and d in Tables 4-7. The first one is the HSS: \begin{equation} \label{eq9} {\rm HSS}=\frac{{\rm PCF}-E}{1-E} , \ \ (10)\end{equation} where E is the proportion of forecasts that would have been correct if forecasts and observations were independent and if the model predicts the same proportion of forecasts of occurrence to non-occurrence (Jolliffe and Stephenson, 2012). For this 2× 2 contingency table, \begin{eqnarray} \label{eq10} E&=&\left(\frac{a+c}{n}\right)\left(\frac{a+b}{n}\right)+\left(\frac{b+d}{n}\right)\left(\frac{c+d}{n}\right)\nonumber\\ &=&\frac{42}{121}\frac{29}{121}+\frac{79}{121}\frac{92}{121}=0.58 , \ \ (11)\end{eqnarray} so that the HSS is \begin{equation} \label{eq11} {\rm HSS}=\frac{{\rm PCF}-E}{1-E}=\frac{0.78-0.58}{1-0.58}=0.47 . \ \ (12)\end{equation} Finally, the last one is the Peirce skill score (PSS): \begin{equation} \label{eq12} {\rm PSS}=H-F=0.45 . \ \ (3)\end{equation}
2
4.1. Case 1: fog over most of the southern China coast
According to the RH data from buoy #59765, case 1 of sea fog occurs from 1740 LST 10 February to 1820 LST 14 February 2016. RH data from the IOPMM give a slightly different sea-fog period, starting 4 h and 20 min later, and ending 11 h and 50 min earlier. The time plots of RH are shown in Fig. 6a. As buoy #59765 is further out to sea, the results suggest that the fog forms earlier further off the coast and dissipates later than that near the coast. According to the wind vectors at 10 m and mean sea level pressure, there is a relaxed pressure-gradient distribution and relatively low wind speed along the coast at 0200 LST 12 February 2016 (Fig. 7a). At this time, a cold SST belt has formed around the southern China coast. Then, warm, moist air moves north from the warmer SST area, which helps to sustain the fog and to increase its extent (Fig. 7b). The forecasting visibility pattern of GRAPES-3km at 0500 LST 12 February 2016 shows that sea fog occurs over nearly all of the southern China coast. This is confirmed by the stations around the southern China coast, with many stations simultaneously measuring visibilities below 1.0 km (Fig. 7c). The satellite data also confirm this large-scale sea-fog pattern, with mostly fog and low clouds near the stations. Furthermore, the patterns of retrieved fog and low clouds are very similar to the forecast fog patterns (Fig. 7d).
Figure6. Observed RH from buoy #59765 and IOPMM: (a) values for case 1, from 0000 LST 10 February to 2300 LST 14 February 2016; (b) values for case 2, from 0800 LST 9 March to 2300 LST 14 March 2017.
Figure7. Test case 1. (a) Wind vectors (units: m s-1) at 10 m and mean sea level pressure (units: hPa) at 0200 LST 12 February 2016. Location A is buoy #59765 and B is IOPMM. (b) Wind vectors (units: m s-1) at 1000 hPa at 0200 LST 12 February and SST (units: °C) on 11 February 2016. Panels (a) and (b) are both from reanalysis data. (c) Forecast visibility at 0500 LST 12 February 2016 (model initial time is 0800 LST 11 February 2016). Colors on the graphic show visibility (km). Colored dots depict observed visibility in km (only where values ≤ 1.0 km) at 0500 LST 12 February 2016 at the coastal meteorological stations. (d) Observed fog and cloud from satellite data at 0500 LST 12 February 2016.To better understand the sea-fog process in this case, we compare the 72-h temperature in the surface layer and the RH to the values in our model forecast. The model initial time is 0800 LST 11 February 2016. First, we compare the values between the model forecast's 20-m level and the buoy's 10-m level. The observed RH indicates continuous fog at buoy #59765, which agrees with the forecast RH (Fig. 8a). However, the forecast RH is a little higher than the observed RH, with a root-mean-square error (RMSE) of 2.6% and mean absolute deviation (MAD) of 2.0%. Meanwhile, the forecast temperature is generally lower than observations, with an RMSE of 1.2°C and a MAD of -0.9°C (Fig. 8a).
Figure8. Case 1 values of 72-h temperature and RH at the surface layer over the sea: (a) from model predictions at the 20-m level and buoy #59765 at the 10-m level; (b) from model predictions at the 20-m level and IOPMM at the 16.4-m level. Model initial time is 0800 LST 11 February 2016.Next, we compare the values between the forecast 20-m level values to those of the 16.4-m level of IOPMM. The observed RH indicates fog, which agrees with the forecast RH (Fig. 8b). In this case, the forecast RH and the observed RH have an RMSE of 3.0% and a MAD of 2.2%. Compared to the RH, the forecast temperature agrees better with observations, having an RMSE of 1.2°C and a MAD of 0.5°C (Fig. 8b).
To investigate how well the model predicts the atmospheric boundary-layer structure, we examine the GPS sonde data. We compare the observed and forecast evolution of the main meteorological variables in the atmospheric boundary layer at the MMSEB in Bohe from 1800 LST 11 February to 0900 LST 12 February 2016. First, consider the wind barb below 2000 m. The forecast agrees well with the wind direction and wind speed above 200 m, but below 200 m the wind direction is not as well simulated (Figs. 9a and b).
Figure9. Main meteorological variables in the atmospheric boundary layer at MMSEB of case 1 from 1800 LST 11 February to 0900 LST 12 February 2016. Observed values on the left-hand side; forecast values on the right. (a, b) Wind direction and wind speed (a full bar of wind barb is 4 m s-1). (c, d) Mixing ratio (color; units: g kg-1) and temperature (contours; units: °C). (e, f) RH (%).Second, we consider the temperature and the mixing ratio, which can be used to indicate warm, moist advection. The GPS sonde data depict a warm, moist advection center between about the 150- and 400-m levels, which agrees with previous analysis (Huang et al., 2011a; 2015; Fig. 9c). However, the forecast fails to reconstruct this center region. Although the model's high mixing-ratio level agrees well with observations, the forecast vertical temperature pattern drops quicker than observed, which indicates that the environmental lapse rate in the GRAPES model is larger than in the observation. Especially after 0500 LST 12 February 2016, it appears that extra cold air advects from somewhere in the model, though such cold air does not appear in the observations (Fig. 9d). Finally, the RH observations show a low cloud over the fog layer, which is a common feature here (Huang et al., 2015). However, GRAPES-3km fails to simulate the evolution of the low cloud (Figs. 9e and f). The forecast does well in simulating the fog layer, but not the low cloud. In addition, the model seems to dissipate the fog around 0600 LST 12 February, in contrast to the persistence in the observations. This dissipation might be due to an influence of the erroneous "extra" cold advection in the simulation.
Figure10. Case 1 profiles of virtual potential temperature, mixing ratio, and equivalent potential temperature. The top row are observed profiles from the GPS sonde data at MMSEB at 1800 LST 11 February, as well as 0000 and 0900 LST 12 February 2016. The bottom row are model results at the same three times. (a, d) Virtual potential temperature (units: K). (b, e) Mixing ratio (units: g kg-1). (c, f) Equivalent potential temperature (units: K).The structure of fog in the atmospheric boundary layer can be revealed in profiles of virtual potential temperature, mixing ratio, and equivalent potential temperature. Consider the profiles at MMSEB during the onset, the developing period, and the ending of sea-fog case 1. The measured virtual potential temperature profiles in Fig. 10a show the case to have a strong stable boundary-layer structure, which is similar to the stable structure over the Yellow Sea (Yang et al., 2019). However, the model-predicted profiles in Fig. 10d do not accurately simulate the observed boundary layer structure. In particular, the predicted profile misses the sudden change at around 1.0 km. This might be due to the predicted mixing ratio profile missing the abrupt decrease at about 0.8-1.0 km at the fog top (Figs. 10b and e). In addition, the predicted profile poorly reproduces the increasing mixing ratio close to the sea surface. Also, below about 0.8 km, the mixing-ratio profile at 0900 LST is significantly less than in the observations (Figs. 10b and e). Similar discrepancies between model prediction and observation also occur for the equivalent potential temperature (Figs. 10c and f).
Figure11. As in Fig. 7, except for test case 2. (a) Wind vectors (units: m s-1) at 10 m and mean sea-level pressure (units: hPa) at 0200 LST 11 March 2017. (b) Wind vectors (units: m s-1) at 1000 hPa at 0200 LST 11 March and SST (units: °C) on 10 March 2017. Both (a) and (b) are from reanalysis data. (c) Model visibility at 0500 LST 11 March 2017 (model initial time is 2000 LST 09 March 2017). Colors on the graphic show visibility (km). Colored dots depict observed visibility in km at 0500 LST 12 February 2016 at the coastal meteorological stations. (d) Observed fog and cloud from satellite data at 0500 LST 11 March 2017.2
4.2. Case 2: fog over the western part of the southern China coast
Case 2 differs from case 1 in several ways. At buoy #59765, the fog forms at 0100 LST 10 March and dissipates two days later at 0600 LST. Thus, in contrast to the previous case, case 2 starts at night, dissipates in the morning, and lasts only about half as long. But at IOPMM, the fog forms 15 h later, and dissipates 57 h later, than that at the buoy. Thus, the fog lasts much longer closer to shore (Fig. 6b).Also unlike case 1, case 2 is a relatively small-scale sea-fog case, with fog occurring mainly from the Beibu Gulf to the west Guangdong area. To help understand the formation of the fog, consider the wind vectors at 10 m and the mean sea level pressure at 0200 LST 11 March 2017 (Fig. 11a). At this time, a cold, dry northeast monsoon overwhelms the area over east Guangdong and the Taiwan Strait, resulting in no sea fog. However, the warm, moist south-southeast wind moves towards west Guangdong and the Beibu Gulf, producing sea fog there (Fig. 11a). Wind vectors at 1000 hPa (Fig. 11b) agree with those at 10 m. Thus, the data suggest that both the cold SST pattern near the coastal area and a continuous wind from the warmer sea (Fig. 11b) are needed for the sea-fog formation.
The predicted visibility pattern of GRAPES-3km at 0500 LST 11 March 2017 shows that sea fog occurs mainly over the west Guangdong and Beibu Gulf sea areas. This model result is supported by stations around the southern China coast, with several stations simultaneously measuring visibilities below 1.0 km (Fig. 11c). The satellite data over Beibu Gulf also shows fog in the same areas as GRAPES-3km (Fig. 11d). The satellite also shows east Guangdong and the Taiwan Strait having fog, but here the satellite retrieval method has confused low cloud for fog because stations in these areas do not show visibilities below 1.0 km.
The close relation between fog and low cloud makes distinguishing them difficult. Fog can form by a lowering of the cloud base, and then later while the cold, dry northeast monsoon comes, the lower parts of the fog dissipate, leaving only low cloud. Fog can also transform into stratus when a large surface wind often occurs over the China Sea (Huang et al., 2011a; Zhang et al., 2014). As the lowest fog transforms into low cloud, the fog top does not significantly change, so the retrieved satellite data do not distinguish the fog and low cloud (Fig. 11d).
Consider the performance of the 72-h model prediction. The model initial time is 2000 LST 9 March 2017. For the buoy, the observed RH shows three short fog periods, while the modeled RH misses the first fog period (Fig. 12a), showing only two short fog periods. At IOPMM, the model's RH values indicate fog about 8 h ahead (Fig. 12b). For the buoy, the RMSE of RH is 3.5% and the MAD is 1.1%, whereas the corresponding values for IOPMM are slightly higher at 4.0% and 3.1%.
For the temperature, the model values are lower than measurements at the buoy, with an RMSE of 1.4°C and a MAD of -1.1°C (Fig. 12a). In contrast, the model values at the 20-m level are higher than those at the 16.4-m level of IOPMM (Fig. 12b), with RMSE and MAD values of 1.6°C and 1.2°C respectively——slightly larger than those for the buoy.
Figure12. As in Fig. 6, except for Case 2 values of 72-h temperature and RH at the surface layer over the sea: (a) from model predictions at the 20-m level and buoy #59765 at the 10-m level; (b) from model predictions at the 20-m level and IOPMM at the 16.4-m level. Model initial time is 2000 LST 9 March 2017.Next we examine how well the model prediction does in terms of the main meteorological variables of the atmospheric boundary layer at MMSEB in case 2. The examined period is 1700 LST 10 March to 0500 LST 11 Marcu 2017, as shown in Fig. 13. We make three points:
Figure13. Main meteorological variables in the atmospheric boundary layer at MMSEB of case 2 from 1700 LST 10 March to 0500 LST 11 March 2017. Observed values on the left-hand side; model values on the right. (a, b) Wind direction and wind speed. (c, d) Mixing ratio (color; units: g kg-1) and temperature (contours; units: °C). (e, f) RH (%).
Figure14. As in Fig. 10, but for case 2 at 1700 and 2300 LST 10 March, and 0500 LST 11 March 2017.
Figure15. Two-year-average (a) ETSs and (b) HSSs of four sea areas in 2016-17 using the GRAPES-3km and the GRAPES-MOS methods. Locations are: (I) Beibu Gulf; (II) west Guangdong; (III) east Guangdong; and (IV) Taiwan Strait, which appear on the abscissa axis in order of southwest to northeast along the coast (see Fig. 2 for their exact locations).First, the wind direction and windspeed are well predicted, especially as the wind changes direction from southeasterly at the surface to westerly at the upper levels. However, the upper westerly wind is too weak (Figs. 13a and b).
Second, concerning temperature and the mixing ratio, the model does not perform as well. For example, the model does not capture the warm, moist advection center at 100-300 m (Figs. 13c and d). Moreover, the predicted high mixing-ratio level is about 400-500 m lower than observed, and the predicted temperature drops too quickly with elevation, as we found for case 1. Thus, GRAPES-3km does not accurately predict the mixing ratio and temperature profiles.
Third, the model does not accurately predict some key processes in the cloud-fog structure. For example, from 1700 LST to 1940 LST 10 March, the base of a stratus cloud lowers to the surface, transitioning into fog (Fig. 13e), but the model does not show this transition (Fig. 13f). This failure might be due to the YSU planetary boundary layer scheme tending to form fog easily, but not being able to lift the fog (Wang et al., 2014; Wilson and Fovell, 2018; Yang et al., 2019). The forecast fog layer is also systematically lower than the observed layer. The thickest fog layer occurs around 0400 LST 11 March, with a value of nearly 1730 m [such thick fog is a special characteristic of the southern China coast (Huang et al., 2011a; 2015)]. Although GRAPES-3km does not predict the thickest fog layer, it successfully models a nearly 1400-m-thick fog layer at 0500 LST 11 March 2017 (Figs. 13e and f).
Consider the profiles. As with case 1, the predicted virtual potential temperature profile in Fig. 14a also depicts a strong stable boundary-layer structure, but has better vapor conditions than in case 1. Overall, these modeled profiles match observations better than those in case 1, but the discrepancies are similar. For example, the predicted virtual potential temperature profiles also miss the sudden change between 1 and 2 km (Figs. 14a and d). Also, the predicted mixing-ratio profiles are less than in the observations (Figs. 14b and e). In addition, the model does a poor job of capturing both the increasing mixing ratio close to the sea surface and the quick drop at the fog top (Figs. 14b and e). The predicted profiles of equivalent potential temperature have similar discrepancies as the mixing-ratio profiles (Figs. 14c and f). That is, they are less than in the observations. Because the low-level wind comes from the south, the discrepancies might arise from the initial conditions over the sea. As a test, we also examined how the profiles from observed sounding data at Xisha station (16.83°N, 112.33°E) compared with the ECWMF initial data. We found that the initial profiles of equivalent potential temperature over sea were less than in the observations at the boundary layer in both case 1 and case 2 (figure not shown).
Overall, the forecast scores are high in 2016. There are 16 stations with the required eight fog events to be used as verification stations. For the 0-24 h predictions (shown in Table 4), the average PCF score is 76%, which is as high as that for the operational sea-fog forecast system in both the Yellow Sea and East China Sea (Yu et al., 2007; Huang et al., 2009b). Also, for 0-24 h predictions, the average ETS is 0.21. This ETS is lower than that in eastern China for land-fog forecasting with a multi-model-based mesoscale ensemble prediction system (Zhou and Du, 2010), and is also lower than that using an extended three-dimensional variational data assimilation (3DVAR) method based on the WRF model (Wang et al., 2014; Table 1). (The lower score may be due to GRAPES-3km being based on an LWC-only approach). For the 24-48-h predictions, the average ETS is also 0.21, which indicates a convincing long-term forecast from GRAPES-3km (Table 5). However, for 48-72-h predictions, the average ETS is only 0.15 (not shown), which means the ability of GRAPES-3km declines after 48 h.
Next we consider the scores of some individual stations. The high-incidence area of sea fog around the Leizhou Peninsula has good scores for 0-24 h (Table 4). For example, Weizhou Island, Zhanjiang, and buoy #59765, have relatively high ETSs of 0.27, 0.31 and 0.24, respectively. However, coastal stations in the Beibu Gulf area, such as Fangchenggang and Beihai stations, have relatively low ETSs of 0.21 and 0.13, respectively (Table 4). The performances of these stations are similar for 24-48 h (Table 5).
The year 2017 had a typical number of fog days on the southern China coast, with 11 stations having the required eight fog days. We now examine their scores, as we did for 2016.
Overall, the model predictions are worse in 2017, having an average ETS of only 0.19 for 0-24 h (Table 6). However, stations around the Leizhou Peninsula still have good scores, with Weizhou Island, Zhanjiang, buoy #59765, and IOPMM being 0.22, 0.30, 0.22 and 0.31, respectively. The worse ETSs come from stations near the Taiwan Strait; for example, Dongshan, Xiamen, and Chongwu have low ETSs of 0.10, 0.04 and 0.13, respectively, which are much lower than their scores in 2016. This shortcoming indicates that GRAPES-3km still needs improvement. The average ETS for 24-48 h is 0.18 (Table 7), and for 48-72 h it is 0.17 (not shown), both of which are lower than that for 0-24 h.
But how does this method, i.e., GRAPES-3km, compare to the previous method, i.e., GRAPES-MOS? To find out, and to determine which sea area may yield better forecasts, we calculate the two-year-average ETS and HSS forecast scores of four sea areas in 2016-17. For GRAPES-3km, we examine four sea areas, whereas for GRAPES-MOS we examine just two. The results (Fig. 15) show that the GRAPES-3km model has higher ETSs and HSSs for the west Guangdong and east Guangdong sea areas than those for the Beibu Gulf and Taiwan Strait. For 0-24 h, the two-year-average ETSs from GRAPES-3km equal those from GRAPES-MOS for the west Guangdong area, but are slightly higher (10%) than in GRAPES-MOS for the east Guangdong area. Similarly, the 0-24-h two-year-average HSS of GRAPES-3km is higher that of GRAPES-MOS by 3% for the west Guangdong area, and higher by 12% for the east Guangdong area. In contrast, for 24-48 h, GRAPES-MOS has the higher ETS for the west Guangdong area and the higher HSS for the west and east Guangdong areas (Fig. 15).
By examining two sea-fog cases in detail, we found that GRAPES-3km can roughly predict the formation, evolution, and dissipation of typical sea-fog events in this area. The modeled scope of sea fog coincides well with observations of meteorological stations and satellite-retrieved products of sea fog at the same time. However, the model poorly predicts the onset and dissipation times in some cases. Also, sometimes, the predicted environmental lapse rate exceeds that of the observation, and it does not accurately predict the changing relations between low cloud and fog. Finally, it does not accurately predict the depth of the fog layer.
Stations near the particularly foggy region around the Leizhou Peninsula had good performance scores for forecasting up to 48 h in advance. For the entire area, GRAPES-3km gave an average ETS of 0.21 and an HSS of 0.35 for 0-24 h forecasting in 2016. For 2017, the scores were 0.19 (ETS) and 0.32 (HSS). These average ETS scores were nearly the same for the 24-48 h forecasts. The ETS scores of the west Guangdong and east Guangdong sea areas were higher than those for the Beibu Gulf and Taiwan Strait. In addition, the two-year-average ETS and HSS at 24 h for GRAPES-3km were slightly higher (5.6% for ETS and 6.4% for HSS) than for the previous method, GRAPES-MOS.
To improve this method such that it may be useful for sea-fog forecasting, some of the following suggestions may help. First, a multi-model-based mesoscale ensemble prediction system may help reduce the forecast uncertainty (Zhou and Du, 2010), but would require more computer resources, especially to forecast several days with a higher horizontal resolution over a larger domain. Second, as some errors result from insufficiently accurate initial conditions (particularly profiles of equivalent potential temperature over sea), using a finer-scale resolution with greater accuracy, or an advanced data assimilation method, for the model's initial conditions, particularly over sea, should improve the forecasting result (Ballard et al., 1991; Gao et al., 2007; Tang, 2012; Wang et al., 2014). Third, as GRAPES-3km does not accurately depict the physics processes of fog, the physics parameterization scheme of the model could include more microphysics (e.g., the droplet number concentration, the effective radius, and the application of a sedimentation scheme). Similarly, a modified planetary boundary layer parameterization scheme may also improve the forecast score (Bergot et al., 2005; Gultepe et al., 2007; Zhou, 2011; Ghonima et al., 2017; Wilson and Fovell, 2018).
However, despite the need to improve GRAPES-3km, forecasters should still find the present model useful, because it allows them to foresee the formation, evolution, and dissipation of sea fog on the southern China coast. Meanwhile, we will continue to work on improving this operational model.
