The advantages of the increased sampling in the GSSODS dataset and the efficient recovery of observations by the refined thresholds used for bias correction are clearly displayed in Figs. 1, 2; however, their influence on the accuracy of the reconstructed SSTs needs to be more definitively assessed. To quantitatively evaluate the impacts, SSTs are reconstructed with the input dataset of GSSODS but by the unified threshold and refined thresholds separately, and by the refined thresholds but with the input dataset being ICOADS3.0 + GTS or GSSODS. Thereafter, the biases relative to observed buoy SSTs are estimated, and the globally averaged biases from 1999 (the coverage of buoy superobservations becomes sufficient for comparison by no less than 15% from 1999) for each of the two sets of monthly reconstructed SSTs are intercompared (Fig. 3). It can be seen that the biases of the SSTs reconstructed by the refined thresholds are slightly lower than those reconstructed by the unified threshold, and the differences are locally clear from about 2011 (Fig. 3a). Comparatively, the biases of the SSTs reconstructed with the input dataset of GSSODS are obviously lower than those reconstructed with ICOADS3.0 + GTS throughout the analysis period, and the average reduction in bias is about 0.02°C (Fig. 3b). This indicates that the refined thresholds serve as a mediator to retain some efficient observations and have a limited capacity to improve the fidelity of the reconstructed SSTs; whereas the increased in situ sampling in GSSODS has straightforward and more stable influences on the reconstructed SSTs, being able to improve the reconstruction accuracy noticeably. Generally speaking, producing CMA-SST with the input dataset of GSSODS and the refined thresholds for bias correction is sensible.
Figure 3. Globally averaged biases of reconstructed SSTs from 1999. The biases are relative to buoy SSTs and intercompared between (a) SSTs reconstructed with the input dataset of GSSODS but by the unified threshold and refined thresholds, and (b) SSTs reconstructed by the refined thresholds but with the input datasets being ICOADS3.0 + GTS or GSSODS. A 5-month moving average filter is applied in the plots.
The monthly root-mean-square errors (RMSEs) relative to the observed buoy SSTs are estimated for the SSTs reconstructed by the refined thresholds but with the input dataset being ICOADS3.0 + GTS or GSSODS, and the spatial distribution of the average RMSE difference from 1999 is shown in Fig. 4a. It can be seen that the SSTs reconstructed by ICOADS3.0 + GTS have obviously higher RMSE than those by GSSODS in the northern areas of the Pacific and Atlantic, especially in China’s offshore and adjacent sea areas. This spatial pattern is consistent with the simultaneous spatial distribution of the in situ SST sampling increments from ICOADS3.0 + GTS to GSSODS, including the total number of in situ ship SST observations in each 2° × 2° grid box (Fig. 4b) and the total number of monthly 2° × 2° buoy grids containing SST observations (Fig. 4c). This further indicates that the increased sampling in GSSODS is conducive to making the reconstructed SSTs more realistic than those reconstructed with ICOADS3.0 + GTS.
Figure 4. (a) Averaged RMSE (°C) differences between SSTs reconstructed by ICOADS3.0 + GTS and GSSODS from 1999, and the simultaneous number differences between GSSODS and ICOADS3.0 + GTS for (b) in situ ship SST observations in each 2° × 2° grid box and (c) monthly 2° × 2° buoy grid boxes containing observations. The RMSE is relative to the buoy SSTs.
The monthly biases relative to observed buoy SSTs in the period 1999–2019 and CCI reanalysis data in the period 1992–2010 are estimated, and the globally averaged biases are compared among CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 (Fig. 5). Generally, CMA-SST has biases roughly similar to the ERSST.v5, COBE-SST2, HadISST2, and HadSST3 products. All the datasets tend to be cooler than the in situ buoy observations and warmer than the independent CCI reanalysis data. Comparatively, the biases of CMA-SST are closest to those of ERSST.v5, due to the application of similar bias correction and reconstruction techniques. Both are closer to 0°C than the biases of the other products, particularly for the bias estimations relative to the independent CCI reanalysis data. In the main, CMA-SST shows relatively homogeneous biases of less than 0.08°C in relation to buoy SSTs, and positive biases of less than 0.16°C and with an apparent decrease through time in relation to the CCI data. Besides, HadSST3, with the smoothest spatial resolution (5° × 5°), shows substantially higher biases than the other products, particularly for the bias estimations relative to buoy SSTs, possibly because of its non-reconstruction of in situ observations with random noise, which needs further investigation.
Figure 5. Globally averaged biases of monthly CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 data relative to (a) buoy SSTs in the period 1999–2019 and (b) CCI reanalysis data in the period 1992–2010. COBE-SST2, HadISST2, HadSST3, and CCI are resampled to the same 2° grid of CMA-SST, and a 12-month moving average filter is applied in the plots.
The spatial distributions of the averaged biases relative to buoy SSTs over the period 1999–2019 and CCI reanalysis data over the period 1992–2010 are compared among the CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 datasets (Fig. 6). Generally, CMA-SST produces a similar spatial pattern and strength of biases to the other products, especially the larger biases in western boundary current regions and at high latitudes, and its bias distributions are most similar to those of ERSST.v5. Comparatively, CMA-SST and ERSST.v5 are systematically cooler in the Arctic area relative to buoy SSTs, and most of the biases exceed 0.4°C; whereas COBE-SST2, HadISST2, and HadSST3 produce both negative and positive biases to the same degree of magnitude in this region. Besides, CMA-SST, ERSST.v5, and COBE-SST2 have clearly lower biases than HadISST2 and HadSST3 at high latitudes of the Southern Hemisphere. Furthermore, the biases of all datasets are mostly within −0.1 to 0.1°C at low and middle latitudes, except for HadSST3, which has clearly higher biases in the central-eastern equatorial Pacific Ocean. Meanwhile, in relation to CCI reanalysis data, CMA-SST and ERSST.v5 have their lowest biases within −0.1 to 0.1°C in the Arctic area, closely followed by HadISST2 and HadSST3 and then COBE-SST2 with obviously higher biases in this region. Besides, comparable biases among CMA-SST, ERSST.v5, and COBE-SST2 are apparent at high latitudes of the Southern Hemisphere, as well as in low and middle latitudes, while distinctly larger biases in these regions are captured in HadSST3 and slightly higher biases at high latitudes of the Southern Hemisphere are captured in HadISST2. Overall, the accuracy of CMA-SST is no less than other current congeneric products.
Figure 6. Averaged biases (°C) of (a) CMA-SST, (b) ERSST.v5, (c) COBE-SST2, (d) HadISST2, and (e) HadSST3 relative to buoy SSTs over the period 1999–2019. (f–j) As in (a–e), but for the averaged biases relative to CCI reanalysis data over the period 1992–2010. COBE-SST2, HadISST2, HadSST3, and CCI are resampled to the same 2° grid of CMA-SST.
The globally averaged SSTAs of monthly CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 data and twice their standard deviation (2STD) in the period 1900–2019 are estimated and compared (Fig. 7). The SSTAs have been set to be relative to their own climatologies over 1981–2010. Generally, CMA-SST shares similar trends to the other products, with two significant warming trends before the mid-1940s and after the mid-1970s, and a shutdown of the warming trend between them is captured for all datasets. The year-to-year variations of CMA-SST are also similar to those of the other products. For the pre-1965 period, the globally averaged SSTAs of CMA-SST tend to be slightly lower than in the other datasets, but the differences are within the 95% confidence interval of the differences for all the datasets (shown in the lower panel), and these differences do not change the overall character of the SSTA variation through the analysis period. After the mid-1960s, most of the SSTAs of CMA-SST are identical to those of the other products, accompanied by a notable decrease in STD for all the datasets.
Figure 7. Globally averaged SSTAs of monthly CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 data (upper panel) and twice their standard deviation (2STD; lower panel) in the period 1900–2019. The SSTAs have been set to be relative to their own climatologies over 1981–2010, and a 12-month moving average filter is applied in the plots.
The trends of annually and area-weighted averaged SSTAs and their uncertainties at the 95% confidence level are estimated and compared among the CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 datasets (Tables 1–3). The 95% confidence interval was calculated by using the IPCC’s method of accounting for the uncertainty of a trend estimation (Karl et al., 2015). Significant warming rates are detected in all the datasets over the various regions of different latitudes for periods including 1900–2019, 1950–2019, and 2000–2019. In addition, an accelerating warming rate of globally averaged SSTAs is captured in all analyzed datasets by a higher warming rate in the more recent period. Furthermore, the accelerating warming rates are also detected in all latitudinal zones except the region of 60°–20°S in the CMA-SST dataset, which reflects the statistics derived from the other products including HadISST2 and HadSST3. Besides, CMA-SST exhibits an ocean warming rate within the rates of the other four products in all the different regions in the period 2000–2019, the region 60°–90°N in the period 1950–2019, and the regions 60°–90°N and 60°–20°S in the period 1900–2019. Apart from these regions, CMA-SST shows a slightly higher warming rate than the other products; however, the warming rates are within the quantified uncertainties of ERSST.v5.
CMA-SST ERSST.v5 COBE-SST2 HadISST2 HadSST3 90°S–90°N 0.79 ± 0.055 0.74 ± 0.061 0.61 ± 0.046 0.55 ± 0.042 0.71 ± 0.074 60°S–60°N 0.84 ± 0.057 0.78 ± 0.064 0.63 ± 0.049 0.58 ± 0.045 0.73 ± 0.073 60°–90°N 0.42 ± 0.139 0.53 ± 0.164 0.94 ± 0.177 0.55 ± 0.091 0.37 ± 0.158 20°–60°N 0.80 ± 0.149 0.71 ± 0.150 0.68 ± 0.115 0.64 ± 0.114 0.73 ± 0.155 20°S–20°N 0.83 ± 0.079 0.79 ± 0.083 0.61 ± 0.067 0.50 ± 0.069 0.64 ± 0.071 60°–20°S 0.86 ± 0.089 0.80 ± 0.089 0.62 ± 0.055 0.64 ± 0.046 0.88 ± 0.070
Table 1. Trends (°C century−1) over 1900–2019 and their uncertainties at the 95% confidence level of annually and area-weighted averaged SSTAs. The SSTAs have been set to be relative to their own climatologies over 1981–2010, and values in bold indicate trends passing the 0.05 significance level based on the F-test
CMA-SST ERSST.v5 COBE-SST2 HadISST2 HadSST3 90°S–90°N 1.06 ± 0.058 1.01 ± 0.063 0.81 ± 0.058 0.71 ± 0.058 0.88 ± 0.113 60°S–60°N 1.11 ± 0.062 1.05 ± 0.067 0.82 ± 0.062 0.75 ± 0.062 0.89 ± 0.110 60°–90°N 0.93 ± 0.194 1.09 ± 0.254 1.63 ± 0.106 0.92 ± 0.139 0.78 ± 0.280 20°–60°N 0.99 ± 0.277 0.93 ± 0.287 0.74 ± 0.225 0.71 ± 0.257 0.86 ± 0.314 20°S–20°N 1.21 ± 0.096 1.14 ± 0.098 0.95 ± 0.090 0.88 ± 0.094 0.91 ± 0.109 60°–20°S 1.06 ± 0.106 1.01 ± 0.086 0.72 ± 0.066 0.63 ± 0.056 0.93 ± 0.074
Table 2. As in Table 1, but for trends and uncertainties (°C century−1) over 1950–2019
CMA-SST ERSST.v5 COBE-SST2 HadISST2 HadSST3 90°S–90°N 1.54 ± 0.228 1.72 ± 0.208 1.39 ± 0.178 1.12 ± 0.174 1.38 ± 0.195 60°S–60°N 2.13 ± 0.124 2.93 ± 0.129 1.61 ± 0.070 2.16 ± 0.091 2.18 ± 0.117 60°–90°N 2.13 ± 0.124 2.93 ± 0.129 1.61 ± 0.070 2.16 ± 0.091 2.18 ± 0.117 20°–60°N 2.47 ± 0.274 2.57 ± 0.272 1.88 ± 0.246 1.58 ± 0.242 1.95 ± 0.315 20°S–20°N 1.80 ± 0.334 1.91 ± 0.339 1.57 ± 0.313 1.53 ± 0.305 1.30 ± 0.306 60°–20°S 0.84 ± 0.204 1.11 ± 0.127 1.08 ± 0.179 0.49 ± 0.069 0.92 ± 0.123
Table 3. As in Table 1, but for trends and uncertainties (°C century−1) over 2000–2019
The Niño-3.4 index over the period 1950–2019 is estimated and compared among the CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 datasets (Fig. 8a). The Niño-3.4 index acts as one of several ENSO indicators and is based on the average SSTA in the region 5°S–5°N, 170°–120°W (Trenberth, 1997; Yasunaka and Hanawa, 2011). Overall, the SSTA time series of CMA-SST in the Niño-3.4 region shows no significant differences with those of the other products, including the peak times and values associated with ENSO events. In addition, the AMO index and the PDO index of CMA-SST over 1950–2019 are also compared with those of the other products (Figs. 8b, c). The AMO index is based on the average SSTA in the region 0°–70°N, 80°W–0° and has been identified as an indicator of multidecadal SST changes over the North Atlantic Ocean (Yasunaka and Hanawa, 2011). It can be seen that the SSTA time series of CMA-SST shows an identical trend in the AMO region to the other products. A decrease in SSTAs between the early 1960s and the mid-1970s and an increase with fluctuations after the mid-1970s are captured in all datasets in the AMO region. Besides, a cool AMO phase in the 1960s–1980s and a warm phase from the late 1990s are also captured in all datasets. The PDO is often described as a long-lived El Niño-like pattern of Pacific climate variability, and the index is defined as the leading principal component (EOF) of North Pacific (20°–70°N, 110°E–100°W) monthly SSTA variability (Zhang et al., 1997). It shows that the monthly PDO index of CMA-SST is consistent with the other products, which proves the capability of CMA-SST to monitor the climate variation signal from interannual to decadal scales.
Figure 8. Averaged SSTAs in the (a) Niño-3.4 and (b) AMO regions, and (c) the PDO index of CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 in the period 1950–2019. The SSTAs are set to be relative to their own climatologies over 1981–2010, and a 5-month running mean filter is applied in (a).
The correlation coefficients between the monthly SSTAs from CMA-SST and those from ERSST.v5, COBE-SST2, HadISST2, and HadSST3 for the period 1900–2019 are calculated and their spatial distributions are shown in Fig. 9. Generally, CMA-SST shows the best agreement with ERSST.v5, with high correlation coefficients (> 0.8) in most regions of the global ocean, especially at high latitudes of the Southern Hemisphere near sea ice with limited in situ sampling, and this is mainly due to the application of similar analysis techniques. Next is COBE-SST2, with which CMA-SST also has high correlation coefficients (> 0.8) in most of the low–mid-latitudes, especially the central and eastern tropical Pacific, the high-latitude North Pacific and North Indian Ocean, the Atlantic, and the oceans around Australia. Overall, the central and eastern tropical Pacific and North Atlantic regions are detected with similar and relatively high correlation coefficients between CMA-SST and the other four products. Around 60°S, CMA-SST has low correlation coefficients (< 0.2) with COBE-SST2, HadISST2, and HadSST3, which is mainly due to the limited observational data in this area.
Figure 9. Correlations between the monthly SSTAs from CMA-SST and those from (a) ERSST.v5, (b) COBE-SST2, (c) HadISST2, and (d) HadSST3 for the period 1900–2019. The SSTAs are set to be relative to their own climatologies over 1981–2010, and COBE-SST2, HadISST2, and HadSST3 are resampled to the same 2° grid of CMA-SST. The correlation coefficients shown are those passing the 0.05 significance test.
The spatial STDs of the monthly SSTAs from CMA-SST, ERSST.v5, COBE-SST2, HadISST2, and HadSST3 from 1900 to 2019 are calculated and presented in Fig. 10. Although the spatial resolution of HadSST3 is lowest (5° × 5°), its spatial heterogeneity is largest since it is the only dataset without reconstruction and is vulnerable to the random noise of in situ observations (Figs. 5, 6). On the contrary, the spatial heterogeneities of CMA-SST, ERSST.v5, COBE-SST2, and HadISST2 are close and significantly lower than that of HadSST3, because their random noise of in situ observations is smoothed by the reconstruction procedures to some extent. By comparison, HadISST2 shows obviously higher stability of spatial heterogeneity among the four close datasets; its spatial STDs are more stable between the pre- and post-1960s periods, while CMA-SST, ERSST.v5, and COBE-SST2 have higher spatial heterogeneity in the pre-1960s period, when the in situ sampling is relatively sparse. The differences in the strength and stability of the spatial STDs might be due to the reconstruction procedures accounting for the spatial patterns of global and regional SSTAs, which needs further investigation.
|90°S–90°N||0.79 ± 0.055||0.74 ± 0.061||0.61 ± 0.046||0.55 ± 0.042||0.71 ± 0.074|
|60°S–60°N||0.84 ± 0.057||0.78 ± 0.064||0.63 ± 0.049||0.58 ± 0.045||0.73 ± 0.073|
|60°–90°N||0.42 ± 0.139||0.53 ± 0.164||0.94 ± 0.177||0.55 ± 0.091||0.37 ± 0.158|
|20°–60°N||0.80 ± 0.149||0.71 ± 0.150||0.68 ± 0.115||0.64 ± 0.114||0.73 ± 0.155|
|20°S–20°N||0.83 ± 0.079||0.79 ± 0.083||0.61 ± 0.067||0.50 ± 0.069||0.64 ± 0.071|
|60°–20°S||0.86 ± 0.089||0.80 ± 0.089||0.62 ± 0.055||0.64 ± 0.046||0.88 ± 0.070|