Visibility Characteristics over the South China Sea during 1980–2018 Based on Gridded Data Generated by Artificial Neural Network

1.   Introduction

The South China Sea (SCS) is one of the most active area in the world, and has always been a battlefield for many countries, not only for its significant importance for maritime shipping, but also because of its rich natural resources like mineral, fishery, and coral reef resources (Chen and Li, 1994; Jia et al., 2004; Mo, 2004; Liu, 2005; Masrikat, 2012; Zhang et al., 2018). However, the SCS is located in the subtropical and tropical regions, and has a vast territory. Due to the sea area being against the mainland and having a unique geographic location, the weather over the SCS is very complicated (An and Qi, 2014). Therefore, it is necessary to study the natural environment of the SCS to ensure the safety when working in or navigating through the area. Visibility (Vis) is one of the most important factors influencing the safety of navigation or work there (Gultepe et al., 2011; Li et al., 2011), so it is significantly important to study the temporal and spatial characteristics of Vis over the SCS.

Many studies have realized the importance of the SCS and analyzed its climatic characteristics (He and Guan, 2000; Xiang et al., 2020). However, very few studies focus on Vis over the SCS because of limited Vis data. Wen and Zhang (2001) used the data from various stations along the south coast of the SCS, annual marine-survey data, and some ship reports to analyze the variation and distribution of climate factors over the SCS. Because the used data are limited, they only finished the preliminary statistical analysis for the temperature, precipitation, wind speed and direction, sea-level pressure, relative humidity (RH), cloud volume, and other meteorological elements over the sea area. Ren and Wen (2012) analyzed the changes of temperature over the Nansha sea area from 1989 to 2018 with the data of the meteorological observation station in Fiery Cross Reef. Fan et al. (2016) studied the prediction of summer monsoon over the SCS, which also helps to ensure the safety of activities in the SCS. Xu et al. (2019) investigated the characteristics of convective quasi-biweekly oscillation over the SCS and Northwest Pacific in spring, which has significant impact on the weather and climate over the Asian monsoon region. Considering the important impact of tropical cyclones over the SCS on the Southeast Asian Nations, Chen et al. (2019) studied the impacts of decaying La Niña and intraseasonal oscillation on the tropical cyclones over the Northwest Pacific and SCS in summer of 2018. Since that climate change caused the obvious changes of atmospheric characteristics over the SCS in recent years, An and Qi (2014) analyzed the characteristics of conventional atmospheric parameters over the SCS, such as temperature and humidity, as well as temporal and spatial distribution of clouds and Vis. Limited by the data, only the parameters over the coastal area of SCS have been analyzed.

Sea fog is one of the most important phenomenon influencing Vis, which has been widely studied. Yue et al. (2012) studied the chemical composition of sea fog along the South China Sea to find out how high concentrations of these chemical compositions are generated in the sea fog. Lin and Song (1990) analyzed one process of sea fog over the South China Sea in 7–9 January 1989, and they provided four suggestions for the prediction of sea fog in winter over the South China Sea. Sun et al. (2018) studied the mechanism of sea fog influencing inland Vis on the southern China coast. Since Vis data are so limited for us, few studies have studied temporal and spatial characteristics of Vis over the whole South China Sea, which is important for navigating or doing maritime operations in the sea area. In this paper, we first reconstruct the historical gridded Vis dataset based on the methods of Shan et al. (2019a, b). Then, temporal and spatial characteristics of Vis over the whole South China Sea are investigated.

Overall, the objectives of this paper are twofold: (1) to generate reliable historical continuous gridded Vis data over the South China Sea, and (2) to analyze temporal and spatial characteristics of Vis over the whole South China Sea. The paper is organized as follows. Section 2 introduces the methodology and data used in this study, and Section 3 analyzes the error of inferred Vis from artificial neural network (ANN). The spatiotemporal characteristics of Vis over the South China Sea are presented in Section 4. Concluding remarks are given in Section 5.

2.   Methodology and data

2.1   Methodology

The flowchart of technical process for this study is shown in Fig. 1. First, we constructed the relationship between Vis and the meteorological factors related to it based on ANN, and used data from the International Comprehensive Ocean–Atmosphere Dataset (ICOADS) to train the constructed ANN and test its effectiveness (Shan et al., 2019a, b). Then, we used the trained ANN and gridded reanalysis data of meteorological factors related to Vis from the ECMWF to infer gridded Vis data over the South China Sea. Finally, we analyzed spatial and temporal characteristics of Vis over the South China Sea on the basis of generated gridded Vis data. At the same time, the probability distributions of low-Vis and high-Vis events (LVEs and HVEs) over the South China Sea were also analyzed.

Download

Larger image

PowerPoint slide

Fig  1.  Flowchart of the technical process for this study.

The correlation coefficients (R) between Vis and its influence factors are calculated to determine the factor with the least relevance to Vis. The correlation coefficient between the any two arrays is defined as follows:

where X and Y are the two arrays, $\overline {X}$ is the average of array X, and $\overline {Y}$ is the average of array Y. Mean absolute deviation (MAD) and standard deviation (SD; Liu et al., 2019) of the inferred Vis from ANN are calculated to test the effectiveness of the trained ANN. The MAD and SD of the inferred Vis are defined as follows:

where ${\rm{Vis}}\_{\rm{C}}$ is the Vis data from ANN, ${\rm{Vis}}\_{\rm{M}}$ is the Vis data from ICOADS, and N is the amount of data-sequence. The regression coefficient (RC) and standard deviation (SD) are also calculated when investigating spatiotemporal characteristics of Vis over the South China Sea, which are defined as follows:

where ${\overline {{\rm{Vis}}\_{\rm{C}}}}_{i}$ is the average of the inferred Vis data, and ${x}_{i}$ and ${y}_{i}$ are the two data series with the same length. Note that spring is from March to May, summer is from June to August, fall is from September to November, and winter is from December to next year’s February in our study.

2.2   Data

The data used to train the ANN model and test its accuracy were obtained from ICOADS. Data of meteorological elements used to generate daily gridded Vis products over the South China Sea were obtained from ECMWF. According to the result of Shan et al. (2019b), RH, dew-point temperature (DPT), sea-surface temperature (Ts), air temperature (Ta), horizontal wind speed (Uh), sea-surface pressure (Ps), and temperature difference between Ts and Ta (TDsa) have important impact on Vis and should be used as the neural network’s input factors to infer Vis in this paper. However, considering that the aerosol and humidity status over the ocean is closely related to wind direction, Uh was tried to be decomposed into U and V in this paper, representing the component of Uh in east–west direction and north–south direction, respectively. At the same time, we calculated the correlation coefficients between Vis and its influence factors, and performed susceptibility experiments for these influence factors to build better nonlinear relationship between Vis and them. Information of the two datasets was provided in the following subsections.

2.2.1   ICOADS

The ICOADS archives are the largest collection of ocean-surface observational datasets with observations from 1784 to the present, which include global data from ships, buoys, and coastal sites, and these datasets are distributed by the National Climatic Data Center (NCDC) of the United States. Because of the nature of sampling, station number density in a specific area changes with time and location. Since the recorded Vis in the dataset is the Vis level, the final inferred Vis in our study is also the Vis level. The rule classifying Vis level in ICOADS is given in Table 1 (Freeman et al., 2017; Gultepe et al., 2019). As the generated Vis data are the Vis level, and Vis-level data cannot be used to analyze Vis climate characteristics, the middle Vis value in each Vis level was used as the Vis estimation value at that level. The Vis estimation value of each Vis level is also shown in Table 1.

Table  1.  Rule for classifying Vis level and Vis estimation value for each Vis level. The Vis here is divided into 10 levels and each Vis level has its own Vis range and estimation value

Download as CSV

Visibility (Vis) level	Vis range (km)	Estimated Vis value of each Vis level (km)
1	< 0.05	0.025
2	0.05–0.2	0.125
3	0.2–0.5	0.35
4	0.5–1	0.75
5	1–2	1.5
6	2–4	3.5
7	4–10	7
8	10–20	15
9	20–50	35
10	> 50	50

 | Show Table

DownLoad: CSV

2.2.2   ECMWF-Reanalysis Archive (ERA)-Interim data

ERA-Interim data are the gridded data product distributed by the ECMWF that includes both observations and predicted products for the entire world, and they are used for the initialization of numerical weather prediction (NWP) model. ERA-Interim data are the 3rd generation product, and its quality is significantly higher than 2nd generation product ERA-40. The data represent the time period from 1979 to the present, and they are constantly updated. The dataset includes the products with different temporal and spatial resolutions, where the minimum (maximum) temporal resolution is 3 h (1 month), and the minimum (maximum) grid resolutions are 0.125° × 0.125° (3° × 3°).

3.   Analysis of inferred Vis from ANN

To test the performance of taking U and V into consideration, we first calculated the Vis based on ANN in both cases where considering U and V or Uh. The ICOADS data, including Vis, RH, DPT, Ts, Ta, U and V (or Uh), Ps, and TDsa, in each month from 1970 to 2016 over the South China Sea (4°–21°N, 105°–118°E) were used to train the ANN model, and the same data were used to test the validity of ANN model. The amount of ICOADS data in each month from 1970 to 2016 over the South China Sea are shown in Table 2.

Table  2.  The amount of ICOADS data over the South China Sea in each month for 1970–2016. “Number” here means the number of the groups of data with the 10 meteorological elements (Vis, RH, DPT, Ts, Ta, Uh, U, V, Ps, and TDsa)

Download as CSV

Item	Month
	1	2	3	4	5	6	7	8	9	10	11	12
Number	11,632	11,724	12,277	10,606	11,079	10,743	11,231	11,545	10,662	12,325	13,847	11,562

 | Show Table

DownLoad: CSV

According to Shan et al. (2019b), we used the data from one of the 12 months to train the ANN model, and then calculated Vis in that month based on the trained ANN model. The mean MAD and SD of the inferred Vis in each month in case of considering U and V or Uh are shown in Fig. 2. Since the instability always exists in ANN analysis, the average of 10 experiments is used to be the final inferred Vis. The results show that, when considering U and V instead of Uh, the mean SD and MAD of the 12 months are 1.5550 and 0.9221 respectively; and when considering Uh instead of U and V, the mean SD and MAD of the 12 months are 1.3520 and 0.7534, respectively. At the same time, it should be noted that the mean SD and MAD in each month in case of considering Uh are both less than that in case of considering U and V. The results indicate that the ANN model used to infer Vis over the ocean can perform better when not decomposing Uh into U and V, although the wind direction could have influence on Vis. The phenomenon may be because that different wind directions of horizontal wind speed have the same influence on Vis in the marine environment, unless considering the influence of terrain on Vis. Therefore, this paper inputs Uh instead of U and V into ANN model when using it to infer Vis over the South China Sea.

Download

Larger image

PowerPoint slide

Fig  2.  Mean MAD and SD of the inferred Vis in each month in case of considering U and V or Uh.

In addition, to obtain the more accurate Vis from ANN, we first calculated the correlation coefficients between Vis and its influence factors mentioned in Section 2.2 (using Uh instead of U and V), and then determined the factor with the least relevance to Vis. Finally, Vis was calculated based on ANN in both cases where considering the factor with the least relevance to Vis or not. ICOADS data from 1970 to 2016 over the South China Sea were used to calculate the correlation coefficients between Vis and its influence factors. The calculated correlation coefficients are shown in Table 3. The results show that DPT has the least relevance to Vis.

Table  3.  Correlation coefficients between Vis and its influence factors calculated with ICOADS data from 1970 to 2016 over the South China Sea

Download as CSV

Influence factor of Vis	RH	DPT	Ts	Ta	Uh	Ps	TDsa
Correlation coefficient between Vis and the factors	−0.115	0.005	0.037	0.089	0.220	0.097	−0.046

 | Show Table

DownLoad: CSV

The mean MAD and SD of the inferred Vis in each month when considering DPT or not are shown in Fig. 3. The results show that the mean SD of the inferred Vis in each month is less than 1.6, and the mean MAD of the inferred Vis in each month is less than 1 whether considering DPT or not. Since Vis data in this paper are the Vis level, the results indicate that the trained ANN could well fit the relationship between Vis and its influence factors.

Download

Larger image

PowerPoint slide

Fig  3.  MAD and SD of inferred Vis in each month when considering DPT or not, where training and testing datasets are the same. Although the calculated Vis levels in Fig. 3 are not integer, the result could still be used to judge the fitting effect of the trained ANN.

At the same time, the mean SD and MAD of the 12 months are 1.3520 and 0.7534 respectively when DPT is considered. However, when not considering the influence of DPT on Vis, the mean SD and MAD of the 12 months are 1.3529 and 0.7554, respectively, which are a little bigger than that when considering DPT. Therefore, although DPT has little influence on Vis, the inferred Vis could have higher accuracy when considering DPT.

To better test the feasibility of the methods that generate gridded Vis over the South China Sea, the last 1000 sets of the dataset in each month are used as the testing dataset, and the other data in that month are used as the training dataset. Figure 4 shows that the mean SD of the inferred Vis in each month is less than 1.3, and the mean MAD of the inferred Vis in each month is less than 0.8 whether considering DPT or not. At the same time, the mean SD and MAD of the 12 months are 1.0407 and 0.5973, respectively when DPT is considered, which also indicate the high accuracy of the inferred Vis based on the ANN. However, when not considering the influence of DPT on Vis, the mean SD and MAD of the 12 months are 1.0475 and 0.6024, respectively, which are a little bigger than that when considering DPT. The conclusion is the same as that from Fig. 3. Therefore, it can be concluded that ANN could well fit the relationship between Vis and its influence factors, and the more fully the influence factor is considered, the higher the accuracy of inferred Vis is. In this study, RH, DPT, Ts, Ta, Uh, Ps, and TDsa are considered when calculating Vis by ANN.

Download

Larger image

PowerPoint slide

Fig  4.  Mean MAD and SD of inferred Vis in each month when considering DPT or not, where the last 1000 sets of the dataset in each month are used as testing dataset, and other data in that month are used as training dataset. Although the calculated Vis levels in Fig. 4 are not integer, the result could still be used to judge the fitting effect of the trained ANN.

4.   Analysis of Vis characteristics over the South China Sea

To analyze Vis characteristics over the whole South China Sea, this study generated gridded Vis data over the South China Sea at 1200 UTC from 1980 to 2018 based on ANN model. ICOADS data over the South China Sea from 1970 to 2016 were used to train the constructed ANN model, and the trained model and gridded data of the influence factors of Vis from ERA-Interim were used to generate the gridded Vis data. The inferred gridded Vis data had a temporal resolution of one day and a spatial resolution of 0.125° × 0.125°.

4.1   Spatial characteristics of Vis over the South China Sea

The spatial distribution of Vis in each season over the South China Sea is shown in Fig. 5. The result shows that the average annual Vis in most sea areas over the South China Sea is more than 20 km, and only few coastal areas have low Vis (Fig. 5a). Overall, the Vis in the south is higher than that in the north, except for spring, whose best Vis is in the middle of the South China Sea. Since the fog season is from January to May in the northern South China Sea, and the fog center is mostly in the north of Beibu Gulf, Vis is best in the southern South China Sea in summer and worst in the northern South China Sea in winter, which is consistent with the analysis of Zhou and Su (2003). Further analysis reveal that areas with the worst Vis are all along the coast. This could be due to effects of human activity, automobile exhausts, and factory activities that cause changes in atmosphere composition, leading to reduced Vis.

Download

Larger image

PowerPoint slide

Fig  5.  Spatial distributions of Vis (km) over the South China Sea from 1980 to 2018. (a) The average, (b) spring, (c) summer, (d) fall, and (e) winter.

4.2   Temporal characteristics of Vis over the South China Sea

The Vis variation over the South China Sea from 1980 to 2018 is shown in Fig. 6. The results show that Vis variation has obvious differences in different seasons. Overall, Vis is best in summer and worst in winter, which is consistent with Section 4.1, and Vis in spring and winter has slowly improved, while Vis in summer and fall has gradually worsened. Also, the Vis in spring changes little from 1980 to 2018. At the same time, Vis significantly increased since 2010, which needs further study to explain. The seasonal variations of Vis over the South China Sea are similar to the Vis in the coast of the South China Sea (An and Qi, 2014). In addition, it is interesting that the Vis in the fall has obvious interdecadal variations. Firstly, the Vis in fall from 1980 to 1995 is better than that from 1995 to 2010. Then, the Vis improves rapidly after 2010.

Download

Larger image

PowerPoint slide

Fig  6.  Annual Vis (km) variation in different periods. Solid lines show the changes of Vis value, and dashed lines indicate the change trend of Vis in each season.

To better analyze temporal characteristics of Vis over the South China Sea, this study calculated the average and SD of Vis in the four seasons during the 39 years (Table 4). At the same time, the RC between inferred Vis and year in the four seasons was also calculated (Table 4). The results show that Vis varied least in spring, and most in fall. Vis decreases by about 0.84 km in summer and 0.11 km decade⁻¹ in fall, while it increases by about 0.18 km decade⁻¹ in spring and 0.37 km decade⁻¹ in winter every decade. In addition, only the average Vis in summer is higher than 30 km, while the Vis in other seasons is less than 30 km. Also, the Vis in summer has significant variations of intensity due to the obvious interdecadal variation.

Table  4.  The average, regression coefficient (RC), and standard deviation (SD) of Vis over the South China Sea in 39 yr. This paper distinguishes different seasons when calculating the above three values

Download as CSV

	Annual	Spring	Summer	Fall	Winter
Average (km)	25.63	25.15	31.89	24.41	20.96
RC (km decade−1)	−0.1	0.18	−0.84*	−0.11	0.37*
SD (km)	0.92	1.12	1.49	2.12	1.4
Note: * indicates that the regression passes the 90% confidence t-test.

 | Show Table

DownLoad: CSV

The seasonal variation of Vis is shown in Fig. 7. The results show that Vis here has obvious seasonal variation, with low Vis from September to March of the next year, except February and November, and high Vis from April to August. Thus, it indicates that the Vis is best in summer and worst in winter. The result is consistent with the previous analysis (An and Qi, 2014).

Download

Larger image

PowerPoint slide

Fig  7.  Seasonal variation of Vis (km) over the South China Sea. The Vis value represents the average-month Vis from 1980 to 2018.

4.3   Probability distribution of low (high)-Vis events

This study further analyzed the Vis distribution over the South China Sea. First, Vis less than 500 m is defined as the LVE, and Vis greater than 10 km is defined as the HVE. Then, probability of low (high) Vis is defined as the days of LVE (HVE) over the total number of days. Finally, the probability of low (high) Vis at each grid point of the South China Sea is calculated.

Figure 8 shows the probability distribution of LVE. The result shows that the probability of LVE for 39 yr is generally less than 10%, and the probability of LVE is increasingly lower from shore to ocean, which indicates that shore is more prone to LVE than the ocean (Fig. 8a). In spring, LVE probability is significantly higher than that in other seasons, especially in the northwest of Hainan Island and the northwest of Malaysia (Fig. 8b). In summer, LVE probability in the north of the South China Sea appears to be chaotic, while other seasons are region-dependent. LVE distribution in fall and winter is similar, and few LVEs occur in the South China Sea (Figs. 8d, e). It is interesting that although the Vis in winter is the worst, the LVE probability is not the highest.

Download

Larger image

PowerPoint slide

Fig  8.  Probability distributions of low-Vis events (LVEs) from 1980 to 2018. (a) The average, (b) spring, (c) summer, (d) fall, and (e) winter.

Figure 9 shows probability distributions of HVE. During the studied 39 years, HVE probability is more than 60%, which is significantly higher than that of LVE. HVE probability from sea to shore is increasingly lower, and HVE probability in the northern South China Sea is higher than that in the southern South China Sea (Fig. 9a). In spring, HVE probability gradually decreases from the middle of the South China Sea to the coast. In summer, HVE probability gradually increases from the north to the south of the South China Sea. In fall, HVE probability gradually decreases from the north to the south of the South China Sea, which is opposed to that in summer. Note that HVEs are most likely to occur in winter, with the probability of HVE greater than 90% in most areas of the South China Sea. It is also interesting that the Vis in winter is the worst, whereas the HVE probability in winter is the highest. It demonstrates that the Vis in most areas of South China Sea in winter is most more than 10 km.

Download

Larger image

PowerPoint slide

Fig  9.  As in Fig. 8, but for high-Vis events (HVEs).

5.   Conclusions and discussion

This study generates gridded South China Sea Vis data from 1980 to 2018 on the basis of the ANN model, and the accuracy of the generated data is tested. Then, temporal and spatial characteristics of Vis in the area are analyzed by using the generated Vis data. The probability distributions of HVE and LVE are also analyzed. Overall, the following conclusions could be drawn from this study:

(1) The average SD of the inferred Vis in each month is less than 1.6, and the average MAD of the inferred Vis in each month is less than 1, which indicates that the trained ANN could well fit the relationship between Vis and its influence factors.

(2) Vis in the southern South China Sea is better than that in the northern South China Sea. In the past 39 years (1980–2018), Vis in spring and winter has slowly improved, while it in summer and fall has gradually worsened. At the same time, Vis is best in summer and worst in winter on the whole.

(3) Low-Vis probability in spring is significantly higher than that in other seasons, especially in the northwest of Hainan Island and the northwest of Malaysia.

Vis-level data are not easily used to analyze temporal and spatial characteristics. Therefore, in the future, reconstructing historical gridded Vis values is more significant to analyze the temporal and spatial characteristics of Vis over the South China Sea. At the same time, a deeper ANN model could be used to infer Vis, which could have better fitting effect.