Surface winds (2012–2014) from 5 moored buoys were first obtained. However, due to anthropic or natural factors, buoys may drift unpredictably, and their locations need to be checked first. Data from moored buoys that have drifted more than 1 degree of latitude or longitude are excluded from the dataset during quality control and the subsequent evaluation. Considering the continuity and amount of buoy data (Table 1), the surface winds of Buoys 301, 302, and 303 from 2012 to 2013 are obtained for further quality control.
Buoy ID 2012 2013 2014 301 22.17°N, 115.36°E 22.17°N, 115.36°E – 302 21.30°N, 113.60°E 21.30°N, 113.60°E – 303 21.07°N, 112.38°E 21.07°N, 112.38°E 22.17°N, 115.36°E Note: – means no observation was conducted by this buoy.
Table 1. Variation of buoy location from 2012 to 2014
The buoy data undergo extensive quality control analysis to ensure that they meet stringent accuracy standards. Table 2 summarizes the preliminary procedures for real-time buoy data.
Measurement Range error checking Extreme-anomaly checking Consecutive anomalies checking Wind velocity speed values outside 2–
speed changes more than 3.0 STD
from the previous hour
difference between the original and 7-hour-running
mean wind speed outside –0.8 STD to 0.8 STD
Wind direction following the check of the
wind speed; the wind
direction outside 0°–360°
following the check of the wind
speed; direction changes more than
10 STD from the previous hour
following the check of the wind speed; difference
between the original and 7-hour-running mean
wind direction outside –5.0 STD to 5.0 STD
Table 2. Major data quality control procedures
The missing values of hourly wind speed or direction are first complemented with default values for time continuity. The range check ensures that all measurements fall within reasonable upper and lower limits. The lower limit of wind velocity is set at 2 m/s to exclude unreliable wind directions at a low wind speed. The original time series of three buoy wind observations are shown in Figs 2 and 3, indicating that most buoy wind speeds are less than 25 m/s. The second check is the extreme-anomaly check with a threshold of maximum allowable difference (M). M is defined by 3 times the standard deviation (STD) of the wind speed. If the value difference with the last observation is greater than M, this observation is regarded as an erroneous value that should be replaced with default value. It is acknowledged that rapid changes could occur in the wind speed due to tropical cyclones. However, the value of M assigned here is large enough that the hourly changes during tropical cyclones are well expressed after quality control. Finally, owing to the functional errors of sensors, erroneous values last for limited periods and require consecutive anomaly checks. The original observation is smoothed to obtain 7-hour running mean wind speed data. It is checked whether the difference between two time series is limited within
$ \pm $0.8 STD. If the difference is outside of this range, the corresponding observation is recorded as the default value.
Figure 2. Time series of the original buoy wind speed during 2012 (a), the buoy wind speed during 2012 after quality control procedures (b), the original buoy wind direction during 2012 (c), and buoy wind direction during 2012 after quality control procedures (d). The black, red and blue lines correspond to Buoy 301, Buoy 302 and Buoy 303, respectively. BQF means the buoys in shallow waters.
Figure 3. Time series of the original buoy wind speed during 2013 (a), the buoy wind speed during 2013 after quality control procedures (b), the original buoy wind direction during 2013 (c), and buoy wind direction during 2013 after quality control procedures (d). The black, red and blue lines correspond to Buoy 301, Buoy 302 and Buoy 303, respectively.
Quality control of wind direction is first dominated by checking the wind speed. It turns out that many erroneous values of wind direction are related to missing (e.g., wind measurements from Buoy 303 from February to May in Fig. 2 or erroneous (e.g., wind measurements from Buoy 301 from February to April in Fig. 3 wind speed observations that should be filtered. However, extreme values remain in the wind direction after the check above, which are close to 360°. These records would amplify the errors in the evaluation by affecting the data stabilization. As a result, these extreme values are removed by setting a higher value of M considering the frequent and significant changes of wind direction. Similarly, consecutive anomalies are also examined with higher threshold.
It is acknowledged that surface wind vectors in the SCS are characterized by significant seasonal cycles controlled by monsoon activity. As shown in Figs 2 and 3, quality-assured wind speeds are relatively high in boreal wintertime and the corresponding wind directions are fixed as prevailing northerlies. In contrast, wind speeds in summer are rather small except during the specific weather processes. These low wind speeds are accompanied by frequent shifts in the wind direction. All the characteristics mentioned above are in accordance with seasonal patterns but are partly covered by abnormal values in the original wind data. Besides, it is obvious that many error signals have been deleted in the time series of wind direction. However, due to the limited threshold values, it seems impossible to delete all the noise of wind direction especially during boreal wintertime. But in boreal summer and autumn, the disordered wind directions may be reasonable because the real winds vary rapidly associated with the lower wind speed. After preliminary validation techniques, the quality of winds observed by buoys is highly elevated, although many degraded measurements are deleted (Figs 2 and 3).
Due to the close location of these three buoys, data comparison between different buoy measurements can be implemented as an additional check. This procedure is conducive to selecting the best quality buoy data. First, the amount of buoy data was compared. Long data gaps in the time series of wind measurements from Buoy 303 in 2012 and Buoy 302 in 2013 may lead to a small sample size for evaluation, suggesting a higher probability of failure in sensor functioning. Furthermore, two typhoons took place close to the three buoys during the observation time period: Typhoon Vicente (July 22 to 25, 2012) and Typhoon Usagi (September 17 to 23, 2013). It is clear that the rapid variation in the wind speed and direction during the typhoons are well expressed by the buoy measurements except Buoy 302 winds in 2013 since there was no record. However, there are also some other rapid variations, which may be erroneous measurements as there were no occurring of special weather processes. For example, compared to the wind measurements from the other two buoys, the data from Buoy 303 show anomalously high wind speeds during August in both 2012 and 2013, when there was no typhoon or other specific weather processes. These potentially erroneous winds are replaced with the default value as in the quality control.
The SCS is controlled by monsoon circulation, which exhibits prevailing northerlies in boreal winter. The data from Buoy 302 in 2012 and Buoy 301 in 2013 show the most stable winds among all the buoy measurements. According to all these quality control procedures, the best quality buoy winds are from Buoy 302 in 2012 and Buoy 301 in 2013.
The scatter diagrams of comparisons of wind speed and direction between selected buoy winds and CCMP surface winds are shown in Fig. 4. In general, buoy winds are quite consistent with CCMP surface wind vectors. Apparently, there are no significant systematic deviations in either wind speed or wind direction. The wind directions of two datasets are agreed for almost all different directions. A further comparison was conducted using statistical parameters of errors, such as the mean bias, RMSE, and the correlation coefficient (Table 3). The average biases of wind speed (0.20 m/s) and wind direction (3.95°) show small deviations which are close to the estimated accuracies of buoy winds. In addition, the average RMSEs are 1.59 m/s and 32.70° for wind speed and wind direction, respectively. The correlation coefficients of both wind speed and wind direction are larger than 0.8 in each year. It is confirmed that the statistical errors above are of the similar magnitude to that of surface winds along the west coast of North America (Tang et al., 2004). The overall results indicate that buoy winds after quality control could be used in the evaluation of reanalysis wind products as true surface winds.
Figure 4. Scatterplots for the wind speed and direction of the comparisons between the CCMP and wind data from Buoy 302 during 2012 and Buoy 301 during 2013. a and b show the comparison of wind vectors during 2012, while c and d show wind comparison during 2013.
Scatterometer dataset Number of collocations Wind speed Wind direction Bias/(m·s–1) RMSE/(m·s–1) R Bias/(°) RMSE/(°) R CCMP (2012) 1 210 0.47 1.61 0.80 2.31 29.78 0.89 CCMP (2013) 903 –0.08 1.57 0.82 5.58 35.61 0.86
Table 3. The error statistics of the comparison between CCMP and buoy wind data
The reanalysis surface wind vectors during 2012 and 2013 are compared with the quality-assured buoy wind data. Data pairs totalling over 1 200 (over 900) from the selected buoys in 2012 (2013) are collocated by the method introduced in Section 2. The scatter diagrams of the comparisons of the wind speeds and directions are illustrated in Figs 5 and 6. Generally, the buoy wind speeds were largely underestimated by the ERA-I winds. Moreover, the dispersive scatterplots of the NCEP-2 wind speeds and directions indicate less consistence with the buoy winds than the other two reanalysis products.
Figure 5. Scatterplots for the wind speed and direction of the comparisons between the reanalysis and wind data from Buoy 302 during 2012. a. CFSv2 wind speed; b. ERA-I wind speed; c. NCEP-2 wind speed; d. CFSv2 wind direction; e. ERA-I wind direction; and f. NCEP-2 wind direction.
Figure 6. Scatterplots for the wind speed and direction of the comparisons between the reanalysis and wind data from buoy 301 during 2013. a. CFSv2 wind speed; b. ERA-I wind speed; c. NCEP-2 wind speed; d. CFSv2 wind direction; e. ERA-I wind direction; and f. NCEP-2 wind direction.
The statistical parameters of errors between buoy and reanalysis wind products are given in Tables 4 and 5. The biases (RMSEs) of different reanalysis wind speeds compared with buoy data are between –1.57 m/s and –0.01 m/s (1.61 m/s and 2.45 m/s), of which NCEP-2 wind speeds present the largest RMSEs in both years. Due to the consistent scatters of CFSv2 wind speeds, statistical errors between CFSv2 and buoy data reveal the smallest RMSEs. However, the mean biases of wind speeds between ERA-I and buoy data reveal notable negative values, implying a systematic shift in ERA-I wind products. Chelton and Freilich (2005) found that 10-m wind speeds from ECMWF are systematically lower than satellite observations. Such bias could lead to wind stress bias of more than 10%, which potentially induces erroneous ocean currents in models driven by these wind products. In terms of wind directions, there are pronounced differences in the biases between reanalysis and buoy data, ranging from –6.80° to –1.76°. As suggested in the scatterplots, NCEP-2 wind directions show the largest bias of approximately –7°, which is far beyond the error estimation of buoy data. The RMSEs of wind directions show great deviation from observations, the overall average of which is 38.89°. The correlation coefficients demonstrate that reanalysis winds are consistent with buoy winds, but the values of coefficients vary widely. For example, ERA-I wind direction are more consistent with buoy wind direction than other two reanalysis wind products. However, the correlation coefficients between CFSv2 and buoy winds show accordance in both wind speed and wind direction, the values of which are slightly smaller compared with the evaluation results in open oceans (Peng et al., 2013; Schmidt et al., 2017). The worst correlation is between NCEP-2 and buoy winds, including both wind speeds and wind directions. The average correlation coefficient between NCEP-2 and buoy wind speeds is only 0.59, suggesting its limited ability of representing surface winds.
Reanalysis dataset Number of collocations Wind speed Wind direction Bias/(m·s–1) RMSE/(m·s–1) R Bias/(°) RMSE/(°) R CFSv2 1 224 –0.40 1.54 0.82 –1.36 34.48 0.87 ERA-I 1 216 –1.70 1.67 0.76 –2.10 29.46 0.90 NCEP-2 1 200 –0.92 2.29 0.63 –7.26 41.10 0.82
Table 4. The error statistics of the comparison between the reanalysis and buoy wind data in 2012
Reanalysis dataset Number of collocations Wind speed Wind direction Bias/(m·s–1) RMSE/(m·s–1) R Bias/(°) RMSE/(°) R CFSv2 925 0.38 1.67 0.83 –2.16 40.42 0.82 ERA-I 911 –1.44 1.69 0.78 –4.88 36.36 0.86 NCEP-2 913 0.40 2.61 0.55 –6.33 51.51 0.74
Table 5. The error statistics for the comparison between the reanalysis and buoy wind data in 2013
Table 4 and Table 5 offer a general view of the deviation between different reanalysis and buoy wind vectors, while the distribution of the errors is not entirely clear. To analyse the errors of reanalysis winds at different wind speeds, the biases and RMSEs of wind speeds and directions are calculated in bins of buoy wind speeds of 1 m/s. The results are illustrated in Figs 7 and 8, in which the error bars indicate the biases and RMSEs.
Figure 7. Dependence of the collocation numbers and wind speed and wind direction residuals (reanalysis-buoy) on the buoy wind speeds during 2012. Buoy wind speeds over 15 m/s (less than 1% of the whole sample) are ignored due to their small statistical value.
Figures 7 and 8 show that the wind speed residuals (reanalysis-buoy) decrease with increasing buoy wind speed for ERA-I and NCEP-2 wind products. Besides, the biases when buoy wind speeds over 8 m/s are all lower than 0, which means that these two wind datasets underestimate buoy winds at middle to high wind speeds. Buoy winds are most underestimated by the ERA-I product, suggesting a systematic deviation at almost all the wind speed intervals and a requirement of techniques for the improvement of the accuracy. On the other hand, wind speeds from CFSv2 are the closest to those from the buoys with a maximum bias less than 1 m/s in both years. The deviations at different wind speeds are limited at a rather small scale. However, the RMSEs of the reanalysis wind speed show no linear variation with increasing buoy wind speed. This result indicates that there is no dependence on the buoy wind speeds regarding the RMSEs of the wind speed. The RMSEs between the NCEP-2 and buoy wind speeds have the largest amplitude for most of the wind speed intervals.
Compared with the biases in the wind speed, the biases between ERA-I (NCEP-2) and buoy wind directions are insignificant. But there are remarkable RMSEs in wind directions (up to 90°). CFSv2 also demonstrates large RMSEs in wind direction at lower wind speed. The RMSEs of wind direction generally decrease with increasing buoy wind speed, which is the same as the results of Yang and Zhang (2018). This distribution suggests that there is no effective pattern for describing the buoy wind directions at low wind speed.
The errors in the reanalysis wind vectors at different months in 2012 and 2013 are also analysed (Figs 9 and 10, respectively). The comparisons of the RMSEs of the reanalysis wind speeds at different months show consistent accuracy levels during the entire period. But there is a significant difference in the biases of the wind speeds according to the time and reanalysis products. Thereinto, wind speeds from the ERA-I products have negative biases for all 12 months in both years. However, the biases of NCEP-2 wind speeds are not consistent in different years. It maintains negative biases in 2012, while there are positive biases in 2013, especially during the second half of that year. CFSv2 wind speeds, on the other hand, keep close to buoy winds at any time of the year. Therefore, the errors in all three reanalysis wind speed datasets show weak dependence on the time. With respect to the wind direction errors, it is illustrated that significant biases in the wind direction correspond to low wind speeds but with exception of NCEP-2 winds. Significant biases are shown between the NCEP-2 and buoy wind directions during autumn in 2013, when strong northerlies began to prevail. The possible reason is the manifest wind speed errors in the second half of 2013, during which the NCEP-2 winds overestimated the buoy observations. From the above, it is deduced that the accuracy of the wind direction primarily depends on the magnitude of the wind speed as well as the wind speed biases.
Figure 9. Errors of the three reanalysis wind vectors (reanalysis-buoy) and the average wind speeds in different months of 2012.
Evaluation of reanalysis surface wind products with quality-assured buoy wind measurements along the north coast of the South China Sea
- Received Date: 2020-10-22
- Accepted Date: 2020-11-02
- Available Online: 2021-04-07
Abstract: Three archived reanalysis wind vectors at 10 m height in the wind speed range of 2–15 m/s, namely, the second version of the National Centres for Environmental Prediction (NCEP) Climate Forecast System Reanalysis (CFSv2), European Centre for Medium-Range Weather Forecasting Interim Reanalysis (ERA-I) and NCEP-Department of Energy (DOE) Reanalysis 2 (NCEP-2) products, are evaluated by a comparison with the winds measured by moored buoys in coastal regions of the South China Sea (SCS). The buoy data are first quality controlled by extensive techniques that help eliminate degraded measurements. The evaluation results reveal that the CFSv2 wind vectors are most consistent with the buoy winds (with average biases of 0.01 m/s and 1.76°). The ERA-I winds significantly underestimate the buoy wind speed (with an average bias of –1.57 m/s), while the statistical errors in the NCEP-2 wind direction have the largest magnitude. The diagnosis of the reanalysis wind errors shows the residuals of all three reanalysis wind speeds (reanalysis-buoy) decrease with increasing buoy wind speed, suggesting a narrower wind speed range than that of the observations. Moreover, wind direction errors are examined to depend on the magnitude of the wind speed and the wind speed biases. In general, the evaluation of three reanalysis wind products demonstrates that CFSv2 wind vectors are the closest to the winds along the north coast of the SCS and are sufficiently accurate to be used in numerical models.
|Citation:||Jing Cha, Xinyu Lin, Xiaogang Guo, Xiaofang Wan, Dawei You. Evaluation of reanalysis surface wind products with quality-assured buoy wind measurements along the north coast of the South China Sea[J]. Acta Oceanologica Sinica. doi: 10.1007/s13131-021-1746-x|