Validation for the years 2017–2019
Figure 1 shows scatter plots between SST datasets, and temporally averaged indices are summarized in Table 1. First, the GMPE and the Sumatra buoy for the period 6 Dec. 2017–30 Nov. 2018 are compared in Fig. 1a. GMPE is very consistent with the in situ observations (a temporally averaged RMSD of 0.17 K, time correlation (TC) of 0.95, regression coefficient (RC) of 1.0, and variance of 0.15 K). Near the shallow coastal sea (approximately 1000 m depth), short-term (within a week) fluctuations in the ocean current direction due to coastal tides and eddies were observed14. As shown in Table 1, the TC and RC values against the Mirai CTD were smaller than those in the open ocean at the Sumatra buoy. These values show that the use of GMPE as a reference is reasonable in the area off the western coast of Sumatra. The indices of OIv2 (Fig. 1b) off the western coast of Sumatra are the worst among the selected L4 datasets. The large RMSD is due to a large negative bias. Although OIv2.1 (Fig. 1c) shows a negative bias, it is significantly improved from OIv2. However, the TC value is low, and the variance is still large in the shallow coastal sea. The RMSD of ECCO2 (Fig. 1d) is due to a positive bias, and ECCO2 shows an overestimation trend for low SSTs less than 29°C. Notably, the variance is quite small, and the TC value is higher than that of OIv2.1.
Table 1
The root mean squared difference (RMSD), time correlation (TC), regression coefficient (RC), and variance of GMPE, OIv2, OIv2.1, and ECCO2 with respect to the Sumatra buoy (5°S, 100°E) averaged for 6 Dec. 2017–30 Nov. 2018 [R/V Mirai CTD (4°S, 101°E) averaged for 6–31 Dec. 2017].
L4 SST dataset | RMSD (K) | TC | RC | Variance (K) |
GMPE | 0.16 [0.07] | 0.94 [0.60] | 1.10 [0.46] | 0.16 [0.05] |
OIv2 | 0.93 [1.43] | 0.51 [0.46] | 0.51 [2.16] | 0.56 [0.40] |
OIv2.1 | 0.33 [0.42] | 0.83 [0.22] | 0.83 [0.83] | 0.29 [0.39] |
ECCO2 | 0.37 [0.49] | 0.87 [0.35] | 0.75 [0.17] | 0.17 [0.04] |
Figure 2a shows the RMSD distribution of OIv2 with respect to GMPE averaged for the period of 16 Sep. 2017–31 Dec. 2019. The large RMSD of OIv2 is not localized in the area off the western coast of Sumatra, and the RMSD of 0.5-1 K is widely distributed in the eastern Indian Ocean, especially in the Southern Hemisphere from 0–30°S. Focusing on the tropics from 15°S-15°N, areas with large RMSD are seen in the coastal regions off western Africa, South America, and northern Australia, which could correspond to coastal upwelling regions. Over the Indian Ocean, multiple areas with large RMSD are distributed in the western, central, and eastern parts. These areas with large RMSD would be a significant issue when summarizing various results that use different L4 SST datasets. For example, different SST datasets could lead to different processes, results, and theories, even for the same phenomenon.
Figure 2b shows the time series for the RMSDs of OIv2, OIv2.1, and ECCO2 with respect to GMPE averaged for the area from 3–8°S, 95–105°E, where diurnal convection originating from Sumatra propagates15; this region is the main target area for air-sea interaction research related to YMC. The verification target area is selected because the suitability of the L4 SST datasets has a large impact as a premise for various YMC studies. Although OIv2 repeatedly shows peaks that exceed 1 K and last for 1 week − 1 month, the RMSD of OIv2.1 is generally less than 0.5 K throughout the entire period. The RMSD of ECCO2 has a peak around Oct. each year associated with the seasonal variation in the large meridional gradient of SST from 3–8°S. The RMSD variation in ECCO2 occurs because the position of the large SST gradient in ECCO2 is shifted from that in GMPE.
Validation for Dec. 2017 during the period of the MR17-08 cruise
Figure 3 shows the time series of SST (a-b) at the Sumatra buoy and in the fixed-point R/V Mirai CTD profiles during the MR17-08 cruise from 6 Dec. − 31 Dec. 2017. The in situ observed SST variation is smaller than approximately 1 K because the mixed layer was well developed (80–120 m in depth). The mixed layer deepening could have been due to the strong westerly winds (5–10 m/s) that were dominant after the MJO passage. OIv2 shows a large negative bias for the entire period of the MR17-08 cruise, and its variation is quite different from that of the in situ SST. OIv2.1 is much improved in terms of temporal variations, although it still has a negative bias with respect to the Sumatra buoy. In addition, OIv2.1 still has some negative bias after 17 Dec., although it is very consistent with the R/V Mirai CTD data before 17 Dec. ECCO2 shows a positive bias, and its variation amplitude is smaller than that of the in situ observations.
Figure 4 shows the average SST distributions of GMPE, OIv2, OIv2.1, and ECCO2 for Dec. 2017. During this period, which includes the MR17-08 cruise, there are remarkable differences in the SST distributions in the area off the western coast of Sumatra in the Southern Hemisphere. In the OIv2 distribution (Fig. 4b), low SSTs less than 28°C, which are clearly different from those of GMPE, are distributed off the western coast of Sumatra and extend to the equator from the south. The large RMSDs of OIv2 exceeding 1 K are distributed in the area off the western coast of Sumatra, the southern coast of Java, and the inland sea between the islands of Borneo and Java. As mentioned on the website providing OIv2 and OIv2.116, the quality of the satellite data used in OIv2 may be degraded in such areas, for example, due to the continuous existence of deep convection. In the distribution of OIv2.1 (Fig. 4c), the unrealistically low SST in OIv2 is corrected in many areas, although the RMSD of approximately 0.4 K due to the unrealistically low SST off the western coast of Sumatra partly remains. Over the open ocean of the Indian Ocean, the RMSD of OIv2.1 is generally less than 0.2 K, and OIv2.1 basically presents a distribution that is very similar to that of GMPE. Focusing on the SST front, as indicated by the contours of 28.4–28.8°C, the distribution of ECCO2 (Fig. 4d) is more similar to that of GMPE than to that of OIv2, although a positive bias with respect to GMPE is seen in the vicinity of the MC. Although Moteki et al.4 showed that ECCO2 data has a positive bias off the coast of western Sumatra, we found here that the explanation for several large RMSDs of 0.4–1.2 K with positive bias is the difference between the magnitude of the meridional SST gradient from 6–10°S and that of GMPE.
Figure 5 depicts SST, SSH, the sea surface geostrophic current calculated from SSH, and outgoing longwave radiation (OLR) on 13 Dec. 2017. The area with large RMSDs exceeding 1 K corresponds well to OLRs less than 240 W/m2. This fact suggests that a large-scale convective system developing continuously over the MC is one of the factors contributing to the quality degradation of OIv2, as indicated by its large RMSD with respect to GMPE. Note that the RMSD of OIv2 does not have a simple proportional relationship with OLRs; thus, their variations do not correlate with each other. The ensemble spread of the 16 datasets used in GMPE shows that the estimated errors of 0.3–0.5 K, which are larger than those of globally averaged values shown in previous reports (e.g.,12,13), are distributed in the area to the north of 6°S off the western coast of Sumatra, although large estimated errors exceeding 0.5 K are found in the area to the south of 6°S, 95–115°E. The large RMSD of OIv2 in the area off the western coast of Sumatra is considered to be caused by inherent problems in the OIv2 estimation process.
OIv2.1 is significantly improved from OIv2 over the whole tropical eastern Indian Ocean. There are several large RMSDs in OIv2.1 exceeding 1 K in the areas off the southern edge of Sumatra (5–8°S, 102–105°E), along the eastern edge of the eddy off the southern coast of Java (10–15°S, 105–110°E), and in the inland sea (0–4°S, 107–109°E).
Because the SST of ECCO2 is estimated after dynamically considering the ECCO2-estimated ocean current, the estimated error of the ECCO2 ocean current is one of the main factors causing the ECCO2 SST error. There are some shifts in position and size between the eddies obtained with ECCO2 and observed SSH to the south of 10°S, and these shifts are considered to be a factor causing the large RMSD of ECCO2 SST in the vicinity of the eddies.
In summary, YMC, a multiyear (from 2015 to present) international program with participants from more than 15 countries, is being conducted. Understanding the characteristics of and differences in SST datasets in the vicinity of the MC is very important as a premise for discussion. This study validated the L4 SST datasets of GMPE, OIv2, OIv2.1, and ECCO2 in the area off the western coast of Sumatra against in situ observations.
GMPE is the most reasonable L4 SST dataset according to validation with Argo floats worldwide. GMPE shows very small RMSD < 0.2 K according to validation with the Sumatra buoy and R/V Mirai CTD data obtained during the MR17-08 cruise, even in the area near the western coast of Sumatra. Although the RMSD of OIv2 off the western coast of Sumatra is very large (1-1.5 K) and associated with a significant negative bias, that of OIv2.1 is improved (RMSD < 0.4 K) in many areas. The RMSD of ECCO2 is less than 0.4 K and associated with a positive bias. Notably, ECCO2 shows a slightly higher TC than OIv2.1.
Among the RMSD results of the OIv2 distribution with respect to the GMPE reference, large RMSDs were found to be widely distributed across the eastern Indian Ocean rather than representing a local feature off the western coast of Sumatra. Multiple areas with large OIv2 RMSD (0.5-1 K) were distributed in the western, central and eastern regions of the Indian Ocean. These large RMSDs indicate a serious problem that can affect interpretations when conducting various analyses and numerical experiments related to YMC. Atmospheric simulation results can vary significantly depending on the choice of SST dataset.
In areas of the Northern Hemisphere and to the east of 115°E, the monthly averaged SST distributions of the 4 SST datasets were generally similar to each other. However, there were significant differences between the datasets in the areas off the western coast of Sumatra, the southern coast of Java, and the Indonesian inland sea. This feature was consistent with the ensemble spread distribution from GMPE. Although low SSTs less than 28°C, which were clearly different from those of GMPE, were located off the western coast of Sumatra and extended to the equator from the south in OIv2, this distribution was improved in OIv2.1. The area with large RMSDs corresponded well to the continuous occurrence of deep convection. Changes in the satellite data used for estimation were suggested to have contributed to the improvement in OIv2.1.