High-frequency turbulent data (u, v, w, t, ρv) were collected by the ECF system installed at a height of 12 m from 1 February 2016 to 29 March 2016. Direct measurements of turbulent data were further used to calculate the fluxes using the eddy covariance (EC) method in a specified time period (30 or 60 min). Meanwhile, direct measurements of turbulent fluxes using the ECF system were used only to verify the applicability of version 3.0 of the Coupled Ocean–Atmosphere Response Experiment (COARE3.0) over the SCS.

The selected 30 min averages of the bulk variables (U measured at a height of 10 m and Ta, Qa and Ts measured at a height of 5 m) used for the bulk flux calculations range from 1 February 2016 to 31 January 2017. Note that Ts was measured using an SI-112 infrared radiation thermometer manufactured by Campbell Scientific Company, USA, installed at a height of 5 m, and therefore, we consider Ts representative of the sea surface temperature at a depth of 0.05 m. The value of Qa was derived using Eq. (1) as described in COARE3.0 using Ta, the relative humidity (Rh) and the air pressure (Pa). Furthermore, this paper also adopts SHF and LHF averages within 30 min intervals derived via COARE3.0 using the input observed bulk variables. The heights of Ta and Qa in the OAFlux dataset are both 2 m, while the measurement heights for these two parameters on the YXASFT are both 5 m. Thus, prior to conducting a comparison, we corrected the corresponding heights of the in situ data to correspond to the heights in the OAFlux dataset using COARE3.0. In addition, downward longwave radiation (DLR) data measured using an NR01 net radiometer manufactured by Hukseflux were used in this paper as an indirect variable to infer the cloud cover in the sky. es=6.112e17.502TaTa+240.97(1.0007+3.46Pa⋅10-6)Qa=621.97es(Pa-0.378)Rh/100

In this paper, the OAFlux reanalysis data were selected for two reasons. First, a previous study showed that the OAFlux dataset is the most preferable among five different products (i.e., ERA-1, NCEPS, JRA55, TropFlux and OAFlux) with regard to LHF data over the SCS (Wang et al., 2017; Zhang et al., 2018). Second, OAFlux represents the most recently updated data product (as of July 2017) accessible for the study period. OAFlux is an ongoing global flux product compiled by WHOI with a spatial resolution of 1∘ × 1∘. OAFlux utilizes an integrated analysis method to combine satellite data with modeling and reanalysis data, and it employs COARE3.0 to calculate heat fluxes (Yu et al., 2008). In this study, the daily mean OAFlux datasets include U, Qa, Ts, Ta, LHF and SHF, and YXASFT observations during the same time period were used for a comparison.

The bulk algorithm utilized in this study is based on the Monin–Obukhov similarity theory, which is widely considered to be an advanced bulk algorithm (Fairall et al., 1996). To keep consistent with the bulk method of calculating flux in OAFlux, the COARE3.0 was used to calculate the heat fluxes using the in situ observation bulk parameters in this paper. Compared to COARE2.5, the updated COARE3.0 has some noted improvements as follows. First, the range of wind speed validity is now extended to 0–20 m s-1 after modifying roughness representation. Second, the COARE 3.0 is shown to be accurate within 5 % for wind speeds of 0–10 m s-1 and 10 % for wind speeds between 10 and 20 m s-1 (Fairall et al., 2003). In this method, the calculation equations for the SHF and LHF can be written as follows: SHF=ρaCpChU(Ts-Ta),LHF=ρaLeCeU(Qs-Qa), where ρa represents the air density, Cp represents the latent heat of evaporation, Cp represents the constant-pressure specific heat, U represents the sea surface wind speed (measured at a height of 10 m in this study), Ce and Ch correspond to the turbulence exchange coefficients for the latent heat and sensible heat, respectively, Qs and Qa correspond to the air saturation specific humidity at the sea surface and the air specific humidity near the sea surface, respectively, and Ts and Ta correspond to the sea surface skin temperature and the air temperature near the sea surface, respectively. In Eqs. (2) and (3), only U, Ts, Ta and Qa are independent measurement variables, while the remainder of the variables must be calculated based on the four independent variables.

The EC method is one of the most direct ways to measure and calculate turbulent fluxes (Crawford et al., 1993). The Reynolds decomposition is utilized to break raw data down into their means and deviations. Furthermore, the values of SHF and LHF can be calculated as the covariance between w and scalar values (t, ρv) using the following formulas, respectively: SHF=ρCpw′t′‾,LHF=λw′ρv′‾, where ρ is the dry air density, Cp is the specific heat of dry air at a constant pressure (where 1004.67 Jkg-1 K-1 is used in the calculation) and λ is the latent heat ratio of water vapor evaporation. The overbar represents the Reynolds ensemble average, and the prime symbol denotes the instantaneous deviation from the ensemble average.

To match the timescale of the OAFlux daily data, we derived the daily means of the YXASFT-observed bulk variables and heat fluxes by averaging all of the 30 min datasets from each day. In addition, we used bilinearly interpolated OAFlux values (inversely weighted by the distance) from the surrounding four grid points (16.5∘ N, 111.5∘ E; 16.5∘ N, 112.5∘ E; 15.5∘ N, 112.5∘ E; 15.5∘ N, 111.5∘ E) to represent the corresponding OAFlux value at the YXASFT observation site.

The comparison between the YXASFT and OAFlux datasets (described in Sects. 3 and 4) was quantitatively analyzed by using the mean bias (Bias, defined in Eq. 6), root mean square error (RMSE, defined in Eq. 7), coefficient of determination (R2) and linear regressions. Bias=1N∑i=1N(xi-yi),RMSE=1N∑i=1N(xi-yi)2, where x and y denote the OAFlux values and YXASFT observations, respectively.

Time series of the bulk variables and heat fluxes are given in Figs. 5 and 6, respectively. As shown in Fig. 6, there is an obvious overestimation in both SHF and LHF in OAFlux compared with the YXASFT observations, and this overestimation demonstrates an evident seasonal variation. The time series of LHF from the YXASFT observations and OAFlux data show essentially consistent variation trends and agree with one another better during the spring (February to March) and winter (December to January) than during the summer and autumn (April to November) (Fig. 6b). The SHF variation trend appears to be opposite to that of LHF, since the deviations during the winter and spring are clearly larger than those during the summer and autumn (Fig. 6a). For the bulk variables in Fig. 5, the OAFlux data maintained a higher consistency with the YXASFT observations with regard to the overall variation trend. Furthermore, U and Qa seemed to match better during the winter and spring periods, while an overestimation (underestimation) in U (Qa) is more evident during the summer and autumn periods (Fig. 5a and b). Some abrupt drops (i.e., variations of 3 to 5 days) in the YXASFT Ts observations were obviously not captured by OAFlux (Fig. 5d). In the next section, we divide the annual study period into three periods, namely, spring (1 February 2016–31 March 2016), summer–autumn (1 April 2016–31 November 2016) and winter (1 December 2016–31 January 2017), to conduct a detailed comparison of their seasonal variations.

The heat fluxes from both OAFlux and YXASFT were derived using COARE3.0. Thus, we can further analyze the origin of the seasonal deviations in the heat fluxes by conducting seasonal comparisons of the bulk variables. The scatter plots of U, Qa, Ts and Ta constructed using the YXASFT and OAFlux data for the three separate periods are shown in Fig. 7, and a quantitative statistical summary for each variable is listed in Table 3.

U. The spring, summer–autumn, and winter periods on Yongxing Island represent the monsoon transition, southwest monsoon and northeast monsoon periods, respectively. Previous studies indicated that the northeast monsoon in the northern SCS is much stronger than the southwest monsoon (Yan et al., 2005). In this study, the observed mean wind speeds during the three periods were 6.40, 4.97 and 9.40 m s-1, respectively. It can be seen from Fig. 7 (first row) that the R2 values of U between the OAFlux and YXASFT data during the three periods are 0.90, 0.79 and 0.92, respectively. OAFlux overestimates the values of U in the spring, summer–autumn, and winter periods with mean biases of 0.96 (15 % of the YXASFT-observed mean value), 1.19 (24 %) and 0.67 m s-1 (7 %), respectively.

Qa. The southwest monsoon is often accompanied by a high amount of water vapor and cloudy skies (Chen et al., 2012). Therefore, the Qa value during the summer–autumn period was the highest throughout the year, with an observed mean of 21.08 g kg-1. The R2 values of Qa between the OAFlux and YXASFT data during the three periods are 0.81, 0.68 and 0.80, respectively (Fig. 7, second row). In contrast to U, OAFlux exhibits an overall underestimation of Qa in the spring, summer–autumn, and winter periods with dry biases of 0.33 (2 %), 0.75 (4 %) and 0.11 g kg-1 (1 %), respectively.

Ta. The OAFlux Ta values are highly consistent with the YXASFT observations, with R2 values of 0.92, 0.84 and 0.89 in the spring, summer–autumn, and winter periods, respectively (Fig. 7, fourth row). As shown in Fig. 5c, both the seasonal trends and day-to-day variations are effectively captured in the OAFlux data. The OAFlux reanalyzed Ta data have a warmer bias of 0.52 ∘C (2 %) in the spring and colder biases of 0.10 (0.3 %) and 0.57 ∘C (2 %) in the summer–autumn and winter periods, respectively. Consequently, the OAFlux-estimated Ta can be considered as the most reliable variable in this study.

Ts. The OAFlux-estimated Ts only captures the seasonal trend, and the estimates exclude some special synoptic signals, such as abrupt drops during cold air temperatures and typhoons or gradual temperature increases induced by the passage of a warm eddy. The R2 values of Ts between the OAFlux and YXASFT data are relatively small when compared with those of U, Qa and Ta, suggesting that the reliability of the OAFlux-analyzed Ts is generally low. In contrast to U and Qa, the OAFlux Ts performance was better in the summer–autumn period (R2=0.70) than in the spring (R2=0.47) and winter (R2=0.54) periods, as shown in Fig. 7 (third row).

In summary, the seasonal performances of the OAFlux-estimated U and Qa seem to be highly correlated with the monsoon system in the SCS. This manifests as a better performance of the OAFlux-estimated U (Qa) during the spring and winter periods, characterized by a stronger (drier) northeast monsoon than during the summer–autumn period, characterized by a relatively weaker (wetter) southwest monsoon. The significant difference between the Ts estimates may stem largely from the fact that the OAFlux Ts estimates are retrieved using a advanced very-high-resolution radiometer (AVHRR), which is easily affected by the presence of clouds. Therefore, the available OAFlux Ts estimates were dramatically reduced during the abovementioned special synoptic processes. With the onset of the southwest monsoon, the average total cloud cover, low cloud cover and precipitation all increase throughout the SCS (Yan et al., 2003), and the Ts retrieved via the AVHRR should correspondingly exhibit a lower quality. However, this trend is not observed in the result of this paper. We further utilized in situ observations of the DLR to infer the sky cloud cover. There is an evidently greater fluctuation in the DLR during the winter and spring periods than in the summer–autumn period, indicating that the winter and spring seasons possess greater probabilities of cloudy days (Fig. 8). This interesting phenomenon may be caused by the fact that the intensity of the summer monsoon in 2016 was weaker than those in preceding years; this hypothesis will be further explored hereafter.

The scatter plots of the LHF and SHF estimates obtained from the YXASFT and from OAFlux during the three periods are shown in Fig. 9, and a quantitative statistical summary of each variable is also listed in Table 3. Note that an upward (downward) heat flux is positive (negative) in this paper, and a positive (negative) value represents the loss (gain) of ocean heat to (from) the atmosphere.

LHF. Compared with the YXASFT observations, the OAFlux-estimated LHF is overestimated by a mean bias of 50.95 (70 %) in the spring, 42.43 (76 %) in the summer–autumn and 63.29 w m-2 (74 %) in the winter. The R2 values are 0.80 in the spring, 0.66 in the winter and 0.40 in the summer–autumn (Fig. 9, first row). This is also consistent with the R2 values for U and Qa, which are the two key input factors in the LHF calculations.

SHF. Large SHF variations during the spring and winter are not evident in the YXASFT-derived SHF time series (Fig. 5e). Compared to LHF, the OAFlux-estimated SHF has the smallest R2 values for all three individual periods, as shown in Table 3 for the spring (0.01), summer–autumn (0.31) and winter (0.14). In comparison, the OAFlux-estimated SHF is more reliable during the summer–autumn, with a mean bias of 1.07 w m-2 than in the spring (16.83 w m-2), or winter (23.56 w m-2).

Overall, we can infer that the OAFlux-estimated LHF product is more reliable during the spring and winter periods than during the summer–autumn period, which is consistent with the key input variables U and Qa, and that the product is further affected by the monsoon system in the SCS. Meanwhile, the SHF estimates exhibit opposite characteristics relative to those of LHF, as the OAFlux SHF product is more credible during the summer–autumn than during the spring and winter periods, which is consistent with the seasonal OAFlux Ts performance and is highly correlated with the cloud cover.