3.1. Evaluation by ERA-5 Data
The ECMWF assimilated meteorological data from different sources to obtain the ERA-5 reanalysis dataset, which included the wind speed of 10 m height above the sea surface, so we compared the wind speed of ECMWF reanalysis data with the CYGNSS inversion results. The track of CYGNSS and the location of the typhoon center are shown in
Figure 2. The orange line represents the CYGNSS track, the black dot represents the typhoon center and the green patch represents the land.
In view of the difference in observation time, the CYGNSS and ERA-5 data were matched up within 20 minutes. A 2° × 2° spatial grid was constructed from the longitude and latitude of the typhoon center. Within this space grid, we find the matched data columns of CYGNSS with ECMWF reanalysis data at the same time. The spatial resolution of CYGNSS was 0.2° × 0.2° and the resolution of the ECMWF reanalysis data were 0.25° × 0.25°. In order to get more matching points of CYGNSS and ECMWF for comparison, we had adopted the method of biharmonic spline interpolation for the CYGNSS and ECMWF reanalysis data and the spatial resolution after interpolation was unified to 0.1° × 0.1°.
From September 8th to September 16th, at the same time of day, the collocated CYGNSS and ECMWF reanalysis data at the same latitude and longitude were enabled to be compared. The comparison results are shown in
Figure 3.
The correlation analysis was conducted between the CYGNSS data and the ERA-5 reanalysis data of ECMWF.
Figure 3a shows the correlation coefficient and (R = 0.96). The overall RMSE is 4.12 m/s and the mean error between the CYGNSS and ECMWF observation is 1.36 m/s. The color bar represents the data density in a certain range. The histogram of residual distribution was shown in
Figure 3b, where the highest residual density is generally distributed near zero. The error sources of CYGNSS wind speeds and ECMWF reanalysis data mainly include:
(1) The time resolutions of CYGNSS and ECMWF reanalysis data are not exactly the same. The matching of some sample points is based on extremely approximate time, which may leads to the difference in wind speed.
(2) The ERA-5 is reanalysis data with assimilating a large number of historical data from different sources, which has some errors with large uncertainty.
3.2. Variations of Sea Surface Wind Speed
According to the longitude and latitude data of the Typhoon Mangkhut and the information on the typhoon evolution is obtained from the National Meteorological Center of CMA. With the help of the relevant information of the Hong Kong Observatory (
https://www.hko.gov.hk/en/informtc/mangkhut18/maxWind.htm), CYGNSS and ECMWF reanalysis data at sample points were compared in the variation trend of the sea surface wind speed.
In the wind speed inversion error from −10 m/s to 10 m/s, the Probability density functions (PDF) of all wind speed sample points and wind speed less /greater than 15 m/s are calculated, which are shown in
Figure 4. It can be seen that when the wind speed increases, the positive errors are more than the negative errors, which shows that CYGNSS overestimates the wind speed values when the wind speed is large.
From September 8th to September 16th, the sample points after interpolation are sorted according to the distance from the typhoon center, and the Euclidean metric is used to indicate the distance between the sample points and the typhoon center, and the variation of sea surface wind speed in different life stages of typhoon is analyzed. The time of comparison was 8:00 am every morning from September 8th to September 16th. On September 8th, the average sea surface wind speed of sample points from CYGNSS and ECMWF reanalysis data were 8.5 m/s and 8.4 m/s, respectively, and the peak wind speed of sample points reached 10.3 m/s and 11.8 m/s, respectively.
On September 9th, due to the high sea surface temperature, weak vertical wind shear and excellent upper air divergence, Mangkhut began to increase in intensity. It was upgraded to a severe tropical storm in the early morning of September 9th and was upgraded to “typhoon” at 8:00 am. The average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 12.8 m/s and 13.0 m/s, respectively, and the peak wind speeds of sample points were 15.7 m/s and 16.9 m/s, respectively.
On September 10th, the typhoon moved to the sea with strong wind shear and its development was restrained due to the invasion of high-level dry air. However, it was still upgraded to the level of “strong typhoon” at 8:00 pm. The average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 27.6 m/s and 27.5 m/s, respectively, and the peak wind speeds of sample points reached 33.8 m/s and 33.4 m/s, respectively. The comparison from September 8th to September 10th is shown in
Figure 5a. The
x-axis represents the number of sample points per day, and the
y-axis represents the wind speed.
When it is upgraded to “super typhoon” in the morning on September 11th, the average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 35.4 m/s and 34.5 m/s, respectively, and the peak wind speeds of sample points were 43.9 m/s and 39.5 m/s, respectively.
From September 12th to 14th, the Mangkhut developed steadily and always keep the level of a super typhoon. On the evening of the September 12th, the typhoon entered the “eyewall replacement” cycle, which led to a slight decrease of the wind speed, but the overall intensity was still rising. The average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data on the September 12th were 35.8 m/s and 32.1 m/s, respectively. The peak wind speeds of sample points reached 46.8 m/s and 39.6 m/s, respectively.
On September 13th, the typhoon continued to move towards the northwest. The average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 40.5 m/s and 38.3 m/s, respectively, and the peak wind speeds of sample points reached 55.1 m/s and 51.7 m/s. The comparison from September 11th to September 13th is shown in
Figure 5b.
On September 14th, the typhoon entered the “eyewall replacement” cycle again, and the typhoon intensity has decreased. The average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 38.7 m/s and 32.8 m/s, respectively, and the peak wind speeds of sample points were 41.7 m/s and 34.1 m/s, respectively. On the early morning of September 15th, Mangkhut made landfall in the Philippines. Due to the influence of topography, the typhoon structure was damaged, and the intensity was weakened. At 8:00 am, the average sea surface wind speed of sample points from CYGNSS and ECMWF reanalysis data were 20.8 m/s and 19.8 m/s, respectively, and the peak wind speed of sample points reached 28.5 m/s and 21.8 m/s, respectively. Then the typhoon entered the South China Sea and the circulation structure of Mangkhut in the South China Sea was reorganized, leading to a rebound in typhoon strength and wind speed. On the 8:00 am of the September 16th, the average sea surface wind speeds of sample points from CYGNSS and ECMWF reanalysis data were 36.4 m/s and 35 m/s, respectively, the peak wind speeds of sample points reached 52.2 m/s and 44.1 m/s, respectively. The comparison from September 14th to September 16th is shown in
Figure 5c.
Mangkhut made landfall in Guangdong as a “strong typhoon” on the afternoon of the September 16th, and then the Mangkhut typhoon deeply penetrated inland regions. Due to the impact with topographic factors, the intensity continued to weaken and gradually dissipated.
The average and maximum wind speeds of CYGNSS and ECMWF reanalysis data during the development of typhoon are given in
Table 1, and the latitude and longitude of the maximum wind speed are given in
Table 2. The different stages of Typhoon Mangkhut are shown in
Table 1, clearly showing the average and maximum wind speeds of CYGNSS and ECMWF reanalysis data in the matched sample points. The average value of wind speed was obtained by averaging the sample points at each stage. It can be seen from
Table 2, among all the sample points, the longitude and latitude of the maximum wind speed obtained by CYGNSS and ECMWF reanalysis data were basically similar. From
Figure 5a–c, when the wind speed of these sample points was lower than 15 m/s, the wind speeds given by CYGNSS and ECMWF reanalysis data were generally in good agreement, but there is a difference when the wind speed was above 15 m/s.
The wind speeds observed by CYGNSS below and above 15 m/s were further compared with ECMWF reanalysis data, respectively. The comparison results are shown in
Figure 6 and
Figure 7. From
Figure 6 and
Figure 7 we can see that in general, the CYGNSS wind speeds below 15 m/s had a small mean error from the ERA-5 reanalysis data with 0.05 m/s; the correlation coefficient was 0.91 and the RMSE was 1.02 m/s. For the wind speeds above 15 m/s, the mean error was 1.61 m/s, the correlation coefficient was 0.90 and the RMSE was 4.36 m/s by comparing with the ECMWF reanalysis data. The results show that the accuracy of wind speed inversion of CYGNSS satellite was high at wind speed below 15 m/s and the accuracy of high wind speed above 15 m/s was reduced.
The relative errors of wind speeds below 15 m/s and above 15 m/s were computed, respectively. When wind speed was lower than 15 m/s, the average relative error was 9.8%; when wind speed was higher than 15 m/s, the average relative error was 11.6%.
The statistical characteristics of data errors could be obtained by comparing the wind speed less than 15 m/s and greater than 15 m/s, respectively. In this study, the positive errors indicate that the wind speed of CYGNSS was larger than ECMWF reanalysis wind speed; the negative errors indicate that the wind speed of CYGNSS was less than ECMWF reanalysis wind speed. The color bar represents the data density in a certain range.
As we can see from
Figure 8, when the wind speed was less than 15 m/s, the number of positive errors and negative errors were very close, which indicates that the wind speed of CYGNSS was very consistent with the ECMWF reanalysis data in general. When the wind speed was greater than 15 m/s, the highest density of wind speeds errors was generally distributed near 0 m/s and 5 m/s, but the number of positive errors was much larger than the negative errors, indicating that when the wind speed was large, the overall CYGNSS wind speed was larger than the ECMWF reanalysis data. The statistical characteristics of wind speed errors in different regions are shown in
Table 3.
In order to analyze the performance of CYGNSS under extreme wind speed, the wind speed sample points of Mangkhut during the development from strong typhoon to super typhoon (wind speed greater than 41.5 m/s) were selected. It can be seen from
Figure 9 that the data concentration was not high, and the distribution was relatively scattered. Most of them deviated from the bottom of the diagonal, indicating that when the wind speed was between 40–60 m/s, the overall CYGNSS wind speed was larger than the ECMWF reanalysis data and the errors between them were further increased when compared to the low wind speed.
Figure 8b shows that the wind speeds from strong typhoons to super typhoons have more positive errors, which indicates that in high wind speed ranges greater than 40 m/s, the accuracy of CYGNSS decreased faster than at lower wind speeds.
The statistics resulted show that when a typhoon reaches the level of strong typhoon or above, the RMSE of CYGNSS and ECMWF was 5.07 m/s, the mean error was 3.57 m/s and the correlation coefficient was 0.52. When the wind speed was larger than 41.5 m/s, the average relative error was 11.0%. This average relative error confirmed that the higher the wind speed was, the lower the CYGNSS precision was. However, within the dynamic range of 41.5 m/s and above, the change of average relative error was still relatively low and there was no particularly large deviation value.
3.3. Discussion
The CYGNSS provided wind speed products are compared with ECMWF reanalysis data in this paper. When the wind speed was lower than 15 m/s, the RMSE and mean error of CYGNSS derived wind speed are 1.02 m/s, and 0.05 m/s, respectively, and the correlation coefficient between two wind speed data sets reach 0.91. Whereas, the RMSE of CYGNSS’s wind speed at higher wind speed range is 4.36 m/s, mean error is 1.61 m/s and the correlation coefficient reaches 0.90. This is mainly due to the fact that the ocean scattering cross section lost the sensitivity to accurately respond to the change of heavy wind speed and the random error of obtaining DDM signal was increased. Therefore, the error started to increase slowly when the wind speed was higher than 15 m/s. The higher the wind speed is, the greater the error is. When compared the wind speed of CYGNSS with ECMWF reanalysis data, the sample points are divided into the wind speed below 15 m/s and wind speed above 15 m/s and the comparison results were also consistent with the actual situation.
In order to verify the accuracy of CYGNSS in actual super typhoons, the wind speed during different typhoon development stage are compared. When the wind speed is greater than 41.5 m/s, the RMSE of CYGNSS and ECMWF reanalysis data is 5.07 m/s, the mean error is 3.57 m/s and the correlation coefficient was 0.52, the average relative error account 11.0%. Compared to the uncertainty of low wind speed, the retrieval error of CYGNSS increased at high wind speed further increases, and the range of error is similar to the expected error range of CYGNSS.
In the future, the influence of satellite related parameters on inversion results and the ocean topographic factors on the typhoon will be further analyzed. Moreover, at higher wind speed, the original DDM signal generated a large number of random errors because the ocean scattering cross section could not accurately reflect the wind speed change. This random error was transmitted during the subsequent signal processing and reduced the inversion accuracy of the wind speed. If the error effect of the signal under high wind speed conditions could be removed, the accuracy of high wind speed could be improved. Furthermore, with the development of other GNSS constellations [
23,
24], high resolution will be expected in the future.