1. Introduction
Sea surface temperature (SST) is an important measure of the earth’s climate and is essential for understanding climate variability and change (Stocker et al. 2014). The SST analysis based on observations, therefore, is a critical component in climate monitoring and projection. Several SST analyses are currently available but may differ in their choice of input data and analysis methods. For example, in situ data are used in the Extended Reconstructed (ER) SST (ERSST) (Smith and Reynolds 2004; Huang et al. 2015), the Hadley Centre SST (HadSST; Kennedy et al. 2011), and the Centennial In Situ Observation-Based Estimates of SST (COBE-SST; Hirahara et al. 2014). In situ and satellite data are used in the daily optimum interpolation (OI) SST (OISST), version 2 (DOISST; Reynolds et al. 2007), and the weekly OISST (Reynolds et al. 2002). This work is part of an effort to better understand the design characteristics that cause differences between two of these datasets.
In situ SST observations are typically made below the surface (2–7 m; Kent et al. 2007). Before the 1980s these observations predominately came from ships. Buoy observations became increasingly common since that time and are the primary in situ source after 2000 (Woodruff et al. 2011). The in situ observations, however, do not cover the entire global oceans. In particular, the data are sparse in the remote regions of the Arctic and Southern Oceans. Therefore, it is very difficult to reconstruct a time-consistent and spatially gridded product in those regions. One approach to overcome the sparse distribution of in situ observations is to include satellite-based observations from the Advanced Very High Resolution Radiometer (AVHRR) after 1981. Moreover, AVHRR SSTs has been reprocessed using the Pathfinder algorithm to produce a more temporally consistent record (Kilpatrick et al. 2001). The AVHRR provides global coverage, including in high-latitude oceans during the wintertime when in situ observations are sparse. However, AVHRR SSTs may contain systematic biases. The AVHRR SSTs retrieved from the radiances measured in various wavelengths may be different from the in situ SSTs measured by thermometers. The infrared instrument of AVHRR can only penetrate the top millimeters of the sea surface. The AVHRR measurements therefore tend to be more variable and subject to diurnal heating, although the measurements are calibrated by a regression with buoy data within a narrow time window. Moreover, the AVHRR SST observations can be contaminated by clouds and continental aerosols and therefore exhibit systematic biases of −0.2° to −0.5°C relative to in situ SSTs (Fig. 1), as also seen in Zhang et al. (2004). In particular, volcanic aerosols can cause biases exceeding 1°C (Reynolds 1993). Therefore, it is critically important that these biases are adjusted before the AVHRR SSTs are merged with in situ SSTs to produce a consistent SST dataset. For example, biases of AVHRR SSTs are adjusted based on in situ SSTs to produce DOISST (Reynolds et al. 2007).
Averaged (1982–2012) bias of AVHRR SSTs from DOISST. Contours are −0.1°, −0.2°, −0.3, −0.4°, and −0.5°C, and contours lower than −0.3°C are shaded.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
Near-real-time climate monitoring and analyses indicated that the SSTs from two popular reconstructed operational datasets—namely, DOISST and the operational monthly ERSST, version 3b (ERSST.v3b)—can differ substantially in the tropical Pacific when in situ observations are lacking, such as when some TAO buoys stopped providing data for an extended time period between May 2012 and March 2013 (Huang et al. 2013). One of the differences between the two datasets is that DOISST uses AVHRR SSTs, while ERSST.v3b does not. However, the analysis of Huang at al. (2013) showed that an experimental ERSST that included AVHRR SSTs (ERSSTsat) still differed significantly from DOISST in the tropical Pacific. It was therefore suggested that the SST difference between DOISST and ERSST.v3b is associated with the bias adjustments of AVHRR SSTs used in DOISST rather than whether AVHRR SSTs are included in the analyses. The biases of AVHRR SSTs are adjusted once per day using daily SSTs within a data window of 3–15 days in DOISST and adjusted once per month using monthly SSTs in ERSSTsat. Since more in situ data are used in the monthly bias adjustment in ERSSTsat, Huang et al. (2013) proposed that the difference in AVHRR bias adjustments is associated with the length of the in situ and AVHRR SST data window used in the calculation of AVHRR bias adjustments. Huang et al. (2013) also assumed that the SST difference between DOISST and ERSSTsat is not associated with the analysis methods of ER or OI. However, the question regarding the relative contributions from the two analysis methods and the AVHRR bias adjustments had not been quantified, which is the subject of this paper.
In addition to the length of data window, other internal parameters (see details in section 3) of the AVHRR bias adjustment may also contribute to the SST difference between DOISST and ERSSTsat. These include the selection of base-function empirical orthogonal teleconnections (EOTs; Van den Dool et al. 2000), the ship SST adjustment relative to buoy SSTs (Reynolds et al. 2002), and zonal SST adjustment (Table 1). In this paper, we present detailed analyses on these aspects, upon which recommendations are made for future improvements in both DOISST and ERSSTsat. Twelve progressive experiments (EXPs) are designed to examine the role of each of the 12 parameters in the AVHRR SST bias adjustment. The impact of analysis methods is compared with two experimental SST analyses of ERSSTsat and DOISST. Each experimental analysis is further examined by applying two extreme bias adjustments from those 12 experiments.
Progressive EXPs in quantifying differences of SST biases. Terms w7, w15, w31, and wMon represent 7, 15, 31 days, and monthly data window, respectively. Data window is defined as the centralized data length of both in situ and AVHRR SSTs used for bias calculation.
In situ and AVHRR SSTs datasets are described in section 2. AVHRR SST bias adjustment schemes are described in section 3. The 12 progressive tests on AVHRR biases are described in section 4. The roles of these 12 experiments in the AVHRR bias adjustments are presented in section 5. The impacts of the bias adjustments and analysis methods are assessed in section 6. Finally, a summary and discussion are described in section 7.
2. Datasets
In this study, which focuses on the AVHRR bias adjustment and its impact on SST analysis, the same in situ and AVHRR datasets are used to produce SST datasets with two different analysis methods, namely, the DOISST and ERSSTsat. The analysis methods of DOISST and ERSSTsat are briefly described in appendixes A and B, with details described in Reynolds et al. (2007) and Smith and Reynolds (2004). The in situ SSTs are from the International Comprehensive Ocean–Atmosphere Data Set, release 2.5 (ICOADS R2.5; Woodruff et al. 2011), from 1981 to 2007 and from the data collected via the Global Telecommunications System (GTS) by NOAA’s National Centers for Environmental Prediction (NCEP) from 2008 to 2013. The AVHRR SSTs are from Pathfinder, version 5.2 (Casey et al. 2010), between September 1981 and December 2011, and operational U.S. Navy SSTs (May et al. 1998) from the NOAA-19 and MetOp-A between January 2012 and May 2013. The analyzed SSTs from operational DOISST are used as the first guess in the quality control procedure. The in situ and AVHRR SSTs are filtered out when they exceed the first guess by 4.5 times the SST standard deviation (STD) derived from monthly SSTs reprocessed from weekly OISST (Reynolds et al. 2002) between 1982 and 2011.
3. Bias adjustment schemes
a. EOT decompositions
b. DOISST application
As described in Eq. (5), sufficient superobservations are needed to support an EOT mode used in the bias calculation in Eq. (6). However, in DOISST, the number of in situ superobservations is too low to support a valid bias adjustment on daily data. Therefore, the superobservations within a centralized data window in time are used to calculate the AVHRR biases. The width of the data window is 15 days (7 days pre- and postcurrent analysis date) (w15; Table 1). In addition to the 15-day data window, the fitting coefficient 
c. ERSSTsat application
In contrast to DOISST, the AVHRR bias adjustment in ERSSTsat is much simpler. Monthly SSTAs from in situ and AVHRR superobservations are applied to Eqs. (3)–(6) without considering the running data window, 5-day fitting coefficient smoothing, zonal SST bias adjustment, or ship–buoy SST adjustment (0.14°C; Reynolds et al. 2002). However, the SSTAs are calculated according to Eq. (2) on the regular even-grid in ERSSTsat, while according to Eq. (1) on the irregular SST locations in DOISST. The EOTs are trained from weekly OISST between 1982 and 2005 in the even-grid system in ERSSTsat, while they are trained from weekly OISST between 1982 and 2000 in the odd-grid system in DOISST. The AVHRR SSTAs are adjusted by subtracting the biases in Eq. (6) and merged with in situ SSTAs (Appendix B) to produce ERSSTsat.
4. Experiment design
Analyses show that the AVHRR biases in DOISST and ERSSTsat are different due to their unique parameter settings, which result in the SST differences between DOISST and ERSSTsat (section 6). To demonstrate how the AVHRR biases evolve from the DOISST to ERSSTsat parameter settings (section 3), 12 experiments (EXP A–L; Table 1) are designed and analyzed progressively. The experiment duration is from January 1982 to May 2013. Starting from the DOISST parameter settings, one parameter is changed at a time in each additional experiment until all the parameter settings match the monthly ERSSTsat parameter settings. The bias difference between two consecutive experiments represents the contribution of that parameter to the total bias difference between DOISST and ERSSTsat. In DOISST, a 15-day data window (w15 in EXP-A) is used in the AVHRR bias calculation. To assess the role of the data window width, additional 7- and 31-day data windows (w7 and w31, respectively) are tested in EXP-B and EXP-C, respectively. The daily biases in EXP-A, EXP-B, and EXP-C are averaged to monthly biases that will be compared with the monthly outputs from experiments EXP-D–K.
In w31, up to 15 days of observations from the previous or following month are used when the analysis is at the beginning or ending of a month. Therefore, the bias for the current month also may be impacted by pre- and postmonth observations. To reduce the impact of pre- and postmonth observations, the daily bias at the middle of the month (either the 15th or 16th day; midmonth) in w31, which essentially uses the observations of the current month, is saved to represent the monthly bias in EXP-D. To further reduce the impact of pre- and postmonth observations, the 5-day running mean filter on the fitting coefficients is removed in EXP-E. In EXP-F, the zonal SST bias adjustment is excluded, since it is not included in the monthly ERSSTsat.
The monthly averaged bias in EXP-F, however, remains different from the one in ERSSTsat, as will be shown in section 5. To further assess the factors causing the difference, an exact monthly window (wMon) is used in EXP-G. In EXP-H, the monthly SSTC is used instead of the daily SSTC to diagnose their effects on SSTAs, since the monthly SSTC is used in ERSSTsat. In EXP-I, the EOTs trained by SSTs between 1982 and 2000 (used in DOISST) are replaced by those trained between 1982 and 2005 as in the ERSSTsat. Further, the constant ship–buoy SST adjustment of 0.14°C (Reynolds 2009) used in DOISST is removed in EXP-J, as it is not used in ERSSTsat and the current operational ERSST.v3b (but will be implemented in the next operational version). In EXP-K, the calculation of SSTAs at in situ locations [Eq. (1)] is revised so that it is calculated at regular grid points [Eq. (2)]. Finally, the even grids rather than the odd grids are used in EXP-L. The EXP-L is the same as the standard monthly ERSSTsat and its bias parameters are equivalent to those in ERSSTsat (Table 1).
5. Biases in AVHRR SSTs
a. Biases in November 2012
Our analyses indicate that the differences in AVHRR SST biases [Eq. (6)] among the experiments vary in space and time, and sometimes they can be as large as 0.5°C in some regions. To demonstrate how these biases differ in DOISST and ERSSTsat, we first show a typical example of the biases in the tropical Pacific in November 2012, when the bias difference is large, and will further analyze the biases in the global oceans in the next subsection.
In November 2012, the series of experiments show that the AVHRR bias is approximately −0.6°C in the central-eastern equatorial Pacific in DOISST (EXP-A; Fig. 2a), while it reaches approximately −1.0°C in the ERSSTsat (EXP-L; Fig. 2l). This suggests that the SST difference between the DOISST and ERSSTsat analyses can be as large as 0.4°C, which may critically impact El Niño–Southern Oscillation (ENSO) monitoring. In the western tropical Pacific, the bias is 0.4°–0.6°C in both DOISST and ERSSTsat, but it stretches toward the southwest more in DOISST than in ERSSTsat.
(a)–(l) Satellite-based SST biases from EXP-A to EXP-L (Table 1) in November 2012. Contour intervals are 0.2°C.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
Comparisons of progressive experiments show that almost all 12 parameters play a role in the AVHRR SST bias to a different degree. As suggested by Huang et al. (2013), the width of the data window plays an important role in the AVHRR biases. When a width of 7 days is applied (Fig. 2b), the bias weakens from −0.6° to −0.4°C in the central equatorial Pacific when compared with the 15-day window (Fig. 2a). When the data window is widened to 31 days, the bias gets larger and the −0.4°C contour stretches westward (Figs. 2b,c). It is important to note that these monthly biases may have been smoothed due to the use of the observations in the pre- and postmonths, as described in section 3. Therefore, the daily bias at the midmonth is used to represent the monthly bias that is as low as −0.8°C in the central equatorial Pacific (Fig. 2d), which is very close to that when an exact monthly data window is applied (Fig. 2g). This shows that the estimated AVHRR biases in DOISST and monthly ERSSTsat are very different and can lead to large differences in SST analyses as suggested by Huang et al. (2013). In addition, by removing the 5-day running mean in the fitting coefficient, the bias weakens from −0.8° to −0.6°C (Figs. 2d,e), which may result in a smaller bias estimate in DOISST.
However, using an exact monthly window (Fig. 2g), the AVHRR bias in the central equatorial Pacific (−0.8°C) remains higher than that of ERSSTsat (−1.0°C; Fig. 2l). This difference may be attributed to the combined effects of the selection of EOTs, ship–buoy SST adjustment, and replacement of odd grids by even grids. When the EOTs trained by the latest data are applied to the bias adjustment, the negative bias in the central-eastern tropical Pacific weakens from −0.8°C to −0.6°C (Figs. 2h,i), while the positive bias in the western equatorial Pacific becomes narrower and is located near 170°E. When the ship–buoy SST adjustment is removed, the bias strengthens from −0.6° to −0.8°C (Figs. 2i,j) in the central equatorial tropical Pacific. When the even-grid system is applied, the bias strengthens further from −0.8° to −1.0°C (Figs. 2k,l). In contrast, parameters such as zonal SST bias (Figs. 2e,f), using monthly instead of daily climatology (Figs. 2g,h), and 2° × 2° climatology (Figs. 2j,k) have little effect on the bias.
b. Biases from 1982 to 2013
The relative importance of the 12 parameters (Table 1) can now be assessed in the global oceans. Comparisons suggest that the following five parameters have the most influence on the AVHRR bias assessment: using a wider data window of 31 days, removing the zonal SST adjustment, selecting the updated EOTs, removing ship–buoy SST adjustment, and switching to a different grid system. This subsection describes their effects in terms of zonal averages from 1982 to 2013 and their impacts on historic SST analyses.
The zonally (0°–360°E) averaged bias increases by 0.05°–0.1°C between 50°S and 60°N when a wider data window of 31 days is used (Fig. 3b) and decreases by −0.05°C when a narrower data window of 7 days is selected (Fig. 3a). The increase of the bias in the 31-day data window means the weakening of the negative bias shown in Fig. 1. The weakened bias may result from the fact that the original AVHRR SSTs were calibrated using “monthly coefficients” (Casey et al. 2010), and therefore the bias becomes smaller when in situ observations within a wider data window of near monthly is used to assess the bias.
Zonally averaged satellite SST biases between two progressive EXPs (Table 1). (a),(b) Difference of EXP-B and -C relative to EXP-A. (c)–(k) Difference of EXP-D, EXP-E–L relative to EXP-C, EXP-D–K. (l) Difference of EXP-L relative to EXP-A. Contour intervals are 0.1°C.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
The bias increases by approximately 0.4°C in the high latitudes south of 60°S and north of 60°N, when prior zonal SST adjustment is not included (Fig. 3e). The reason for the increased bias is that the biases in the AVHRR SSTs, which are −0.3° to −0.4°C (Fig. 1), cannot be resolved in the high latitudes south of 60°S and north of 60°N using the bias calculation [Eqs. (3)–(6)] with the damped EOTs in the high latitudes (Smith and Reynolds 2004). The impact of the damped EOTs is also confirmed by the fact that the change in the biases becomes very small in the high latitudes when the updated EOTs (also damped in high-latitude oceans) are used (Fig. 3h). In contrast, in the lower latitudes, the bias can mostly be resolved by EOTs and therefore is sensitive to the selection of EOTs. The bias increases when updated EOTs are used (Fig. 3h), which means that the negative bias of AVHRR SSTs weakens. As shown in Huang et al. (2015), the updated EOTs can better represent and resolve the spatial structures of in situ observations and therefore reduce the bias between decomposed in situ and AVHRR SSTs at a monthly time scale, since the AVHRR SSTs were originally calibrated by monthly coefficients (Casey et al. 2010).
When ship–buoy adjustment is not considered as in the operational ERSST.v3b, the ship SSTs are 0.14°C higher (Reynolds 2009). The warmer ship SSTs increase the merged SSTs from ships and buoys by 0.1°C or less depending on the ratio of the numbers of observations from ships and buoys. This results in a smaller bias according to Eq. (7). In the early 1980s, in situ observations are mostly from ships, and therefore the bias decreases by approximately 0.1°C (Fig. 3i). After the later 1990s, the number of in situ observations is mostly from buoys, and therefore the effect of ship–buoy adjustment has a smaller effect.
When the even-grid system is selected (Fig. 3k), the changes in the biases vary from −0.2° to 0.4°C between 60° and 70°N and from −0.1° to −0.4°C between 60° and 70°S. The impact of the grid system change on the biases may result from the following factors: (i) the change in grid system may modify the ocean–land masks, which may further affect the acceptance/rejection of the observations near the coasts; (ii) the SSTA superobservation may be modified due to the change of SST climatology at different grid boxes; and (iii) to calculate the SSTAs in the even grid, the monthly SST climatology and base-function EOTs need to be interpolated from the odd grid to the even grid, which may have smoothed the minimum/maximum values of the climatology and EOT modes. These factors may result in differences in the AVHRR bias and final SST analysis in the data-sparse regions of the equatorial and high-latitude oceans, where the SST analysis is very sensitive to these factors (Huang et al. 2015). However, the individual roles of these factors are hard to assess, since they cannot be isolated easily. Future studies are needed to clarify their individual contributions to the bias changes.
In contrast, the bias does not change much if the daily midmonth bias is used to represent the monthly bias (Fig. 3c), or if the 5-day running filter to the fitting coefficient is applied (Fig. 3d). If the exact monthly data window is applied (Fig. 3f), then the bias does not change much either, because the difference is relative to the data window (31 days) that is close to monthly. The changes in the biases are small if the monthly SST climatology is used instead of daily climatology (Fig. 3g), or if the SST climatology on regular 2° × 2° grids is applied instead of in situ locations (Fig. 3j). Therefore, the difference of the biases between DOISST and ERSSTsat (Fig. 3l) is largely associated with the zonal bias (Fig. 3e) and the grid shift (Fig. 3k) in the high-latitude oceans, and with the width of the data window (Fig. 3b), the EOT selection (Fig. 3h), and the ship–buoy SST adjustment in the lower-latitude oceans.
It is interesting to note that the bias difference due to those parameter changes is clearly dependent on the seasons in the high-latitude oceans (e.g., Fig. 3l). The reason for the seasonal dependence is that the total number and areal coverage of in situ observations are low in winter and higher in summer, which is particularly true in the high-latitude oceans.
c. RMSDs of biases
The root-mean-square differences (RMSDs) between the AVHRR biases of two progressive experiments are calculated to assess the individual role of each of the 12 parameters of the biases in the global oceans (Fig. 4). For example, Fig. 4h shows the RMSD between EXP-I and EXP-H, which represents the impact of updated EOTs to the bias, since EXP-I uses new EOTs of ERSSTsat and EXP-H uses old EOTs of DOISST. On global average, as listed in Table 2 (the global column), the four dominant parameters that affect the biases are (i) the use of the updated EOTs (EXP I–H); (ii) the wider data window of 31 days (EXP C–A); (iii) the even-grid system (EXP L–K); and (v) the a priori zonal SST adjustment (EXP F–E). The importance of these four parameters is mostly consistent with the analysis in zonally averaged biases in section 5b.
RMSD of satellite-based SST biases (1981–2013) between two progressive EXPs (Table 1). (a),(b) RMSD of EXP-B and EXP-C relative to EXP-A. (c)–(k) RMSD of EXP-D, EXP-E–L relative to EXP-C, EXP-D–K. (l) RMSD of EXP-L relative to EXP-A. Contours are 0.1°, 0.2°, 0.3°, and 0.5°C.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
Averaged RMSD (°C) of biases between a pair of two progressive EXPs in global and Niño-3.4 regions. Values larger than 0.1 are in boldface.
As indicated in Huang et al. (2015), the updated EOTs have a significant impact on the analysis of in situ SSTs. Therefore, it is easy to understand that the selection of the latest EOTs can first impact the in situ SSTA decomposition in Eq. (4) and then impact the bias in Eq. (7). The RMSD due to updated EOTs is 0.1°–0.2°C in almost all the global oceans (Fig. 4h) except the Southern Ocean south of 65°S and the Arctic Ocean where EOT modes are damped. The EOT damping was designed to filter out potential artificial modes due to sparse observations in the high-latitude oceans (Huang et al. 2015). This is also the reason why the zonal SST adjustment [Eq. (7)] is introduced in DOISST to overcome the limitation of the bias calculation due to the damped EOT decompositions in the high-latitude oceans, which results in an RMSD of the bias of 0.5°C in the Arctic Ocean and 0.3°C in the Southern Ocean (Fig. 4e).
In addition to the impacts from the updated EOTs and zonal SST adjustment, the RMSD due to the wider data window of 31 days is approximately 0.2°C in the Indo-Pacific region, Southern Ocean, northwestern North Pacific, and northwestern North Atlantic (Fig. 4b). In these regions, either in situ observations are sparse or the SST daily variability is large, and therefore a wider data window may have smoothed the variability or noise and can directly impact the bias. Similarly, when the even-grid system is applied instead of the odd-grid system, the RMSDs of 0.2°–0.3°C are mostly located in the Southern Ocean, the northwestern Atlantic, and the northern North Pacific (Fig. 4k) where in situ observations are sparse, so that the superobservations become different by shifting the grid system by 1°. In contrast, the impacts from other parameters are relatively small, that is, using the narrower 7-day data window (Fig. 4a), using the midmonth daily bias to represent the monthly bias (Fig. 4c), removing the 5-day running filter (Fig. 4d), using the monthly climatology instead of the daily climatology (Fig. 4g), removing the ship–buoy SST adjustment (Fig. 4i), and applying the climatology at 2° × 2° instead of in situ locations (Fig. 4j). The RMSD between biases of ERSSTsat and DOISST is 0.2°–0.5°C in most of the global oceans (Fig. 4l).
The impacts of these 12 parameters to the AVHRR bias in the Niño-3.4 region (5°S–5°N, 120°–170°W; Bamston et al. 1997) are similar to those in the global oceans (Table 2; Niño-3.4 column). Three parameters [i.e., using the wider window of 31 days (EXP C–A), selecting the updated EOTs (EXP I–H), and applying the even-grids system (EXP L–K)] remain the dominant contributors to the AVHRR bias. The difference is that the impact from the zonal SST adjustment (EXP F–E) becomes less important, while the impact from using the exact monthly window (EXP G–F) becomes slightly more important.
It needs to be noted, however, that the impacts of these 12 parameters to the AVHRR biases are nonlinear, as seen in Table 2. For example, the total change of global averaged bias (0.29°C; third column of Table 2) is much smaller than the sum of 12 individual changes. This means that the impact of a parameter switching from the DOISST setting to the ERSSTsat setting may vary when the sequence of parameter switching is changed.
6. Impacts of satellite AVHRR bias adjustment on SST analyses
a. Niño-3.4 region
As shown in section 5, the estimated AVHRR biases are different in DOISST and monthly ERSSTsat. For example, in the tropical Pacific Niño-3.4 region, the difference of the estimated biases between EXP-L and EXP-A may be as large as 0.5°C at the monthly time scale (Fig. 5a), and the averaged (1982–2012) difference is approximately 0.16°C. To understand the relative impact of bias adjustments within the context of OI and ER analysis methods, four SST analysis experiments are conducted: both OI and ER analyses using two sets of the biases of EXP-A (ERSST15 and OISST15; Table 3) and EXP-L (ERSSTmon and OISSTmon; Table 3) between January 1982 and May 2013. The reason for using the biases of EXP-A and EXP-L is that they have included all possible differences of the biases between DOISST and ERSSTsat. The operational codes of ERSST.v3b are used as the starting point for the ERSSTsat analysis, and the operational codes of DOISST are used as the starting point for the DOISST analysis.
Satellite-based SST biases in EXP-A (black line), EXP-L (red line), and their difference (green line; sign is flipped to match with SSTA) in (a) Niño-3.4 region and (b) global oceans.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
SST analysis EXPs (boldface) using different analysis methods of ER and OI and using different AVHRR biases in EXP-A and EXP-L. ER is based on ERSST.v3b, and OI is based on DOISST.
Figure 6a shows the differences of Niño-3.4 SST among the four SST analyses. It is found that the SSTA differences between two SST analyses are large when different biases of EXP-A and L are applied. The SSTA difference between ERSSTmon and ERSST15 (black line) is 0.1°–0.2°C before 2000, and is 0°–0.2°C after 2000, which is very close to the SSTA difference between OISSTmon and OISST15 (green line). Their averaged differences between 1982 and 2012 are 0.10° and 0.12°C, respectively, which results mostly from the difference in the bias of 0.16°C (Fig. 5a). A smaller SSTA difference versus a bias difference is expected, since the biases are only applied to the AVHRR SSTs, while SST analyses include both in situ and AVHRR SSTs. Therefore, the SSTA differences are expected to be zero in the area covered by in situ observations only and to be close to the difference of the biases in the area covered by AVHRR observations only. It should be noted that the SST differences due to bias adjustment (Fig. 6a, black and green lines) decrease slightly after 2000, which is associated with the decrease in the difference of AVHRR SST bias (Fig. 5a, green line). The decrease in the bias difference may result from the decrease of in situ SST coverage in the tropical Pacific (Huang et al. 2013). The lower coverage of in situ SST may result in less difference in the accepted EOT modes and therefore less difference in estimated AVHRR SST biases.
SSTAs differences in (a) Niño-3.4 region and (b) global oceans between ERSSTmon and ERSST15 (black line), OISST15 and ERSST15 (red line), OISSTmon and OISST15 (green line), and OISSTmon and ERSSTmon (blue line). A 5-month running mean has been applied.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
In contrast, the SSTA differences between the other two SST analyses are small when the same bias adjustment is applied. The SSTA differences are −0.1° to 0.1°C between OISST15 and ERSST15 (Fig. 6a; red line) and between OISSTmon and ERSSTmon (blue line), and their averaged (1982–2012) differences are −0.01° and 0.01°C, respectively. This suggests that the SSTA differences resulting from the different analysis methods (OI and ER) are much smaller than those resulting from the different bias adjustments (EXP-A and L). This also implies the importance of a correct AVHRR bias adjustment when AVHRR SSTs are used in SST analyses such as DOISST and ERSSTsat.
These conclusions can be seen more clearly in the example of SSTA differences in November 2012. An SSTA difference of 0.4°C occurs between ERSSTmon and ERSST15 in the central tropical Pacific Niño-3.4 region when the biases of EXP-L and EXP-A are applied to ERSSTmon and ERSST15, respectively (Fig. 7a). A similar SSTA difference is also found between OISSTmon and OISST15 (Fig. 7c). It is clear, therefore, that the positive SSTA differences directly result from the difference of the biases shown in Figs. 2a,l. In contrast, when the same bias of EXP-A is used, the SSTA difference between OISST15 and ERSST15 exhibit more spatial variability in the central tropical Pacific (Fig. 7b); and the averaged SSTA difference is approximately 0.1°C in the Niño-3.4 region. Similarly, the SSTA difference has greater spatial variability between OISSTmon and ERSSTmon (Fig. 7d) when the EXP-L bias adjustments are used; and the averaged SSTA is small (0.1°C) in the Niño-3.4 region.
SSTA differences between (a) ERSSTmon and ERSST15, (b) OISST15 and ERSST15, (c) OISSTmon and OISST15, and (d) OISSTmon and ERSSTmon in November 2012. Green rectangle boxes indicate the Niño-3.4 region. Contour intervals are 0.2°C.
Citation: Journal of Atmospheric and Oceanic Technology 32, 2; 10.1175/JTECH-D-14-00121.1
It should be noted, however, that the magnitude of SSTA differences (0.4°–0.6°C; Figs. 7b,d) that are due to differing analysis methods is as large as that result from bias adjustments (Figs. 7a,c). The SSTA differences associated with differing bias adjustments are largely determined by the difference of the biases whose spatial scale is large, since the biases are calculated at 2° × 2° resolution in all experiments. In contrast, for those in which the bias adjustments are the same, the SSTA differences are largely determined by the analysis methods, whose spatial resolutions are 0.25° × 0.25° in OISST15 and OISSTmon and 2° × 2° in ERSST15 and ERSSTmon. Therefore, the spatial scale of the SSTA difference with greater spatial variability is mainly controlled by the small spatial scales of OISST15 or OISSTmon.
b. Global oceans
In the global oceans (Fig. 5b), the variability and magnitude of the averaged AVHRR SST biases in EXP-A and -L are reduced in comparison with those in the Niño-3.4 region (Fig. 5a). The weaker variability and magnitude of the bias are mostly associated with averaging over the large global ocean domain. The bias is weaker in EXP-L (red line) than in EXP-A (black line) in the global oceans, while it is stronger in EXP-L than in EXP-A in the Niño-3.4 region. The averaged difference is approximately −0.11°C (green line; its sign is flipped to match with SSTA) over the global oceans. The weaker bias in EXP-L than in EXP-A in the global oceans is largely associated with the exclusion/inclusion of zonal SST bias adjustment in high latitudes in EXP-L/EPX-A, which is indicated by the positive bias difference in Figs. 3e,l.
Calculations also indicate that the role of satellite AVHRR bias adjustments is dominant over that of analysis methods in the global oceans (Fig. 6b), as indicated in the Niño-3.4 region. The SSTA differences resulting from different analysis methods (red line for OISST15 − ERSST15; blue line for OISSTmon − ERSSTmon) are near zero due to SSTA cancellations in the global oceans. In contrast, the SSTA differences resulting from different bias adjustments (black line for ERSSTmon − ERSST15; green line for OISSTmon − OISST15) have a magnitude of −0.05°C at monthly time scale. On average (1982–2012), the SSTA differences due to analysis methods are −0.004° and 0.006°C for OISST15 − ERSST15 and OISSTmon − ERSSTmon, respectively. In contrast, the SSTA differences due to bias adjustments are approximately −0.020° and −0.011°C in ERSSTmon − ERSST15 and OISSTmon − OISST15, respectively.
7. Summary and discussion
The bias adjustments of AVHRR SSTs are studied using 12 experiments, starting with DOISST parameters (EXP-A) and ending with ERSSTsat parameters (EXP-L). It is found that the bias adjustment is not only sensitive to the width of the data window, as indicated in an earlier study of Huang et al. (2013), but is also sensitive to the zonal SST adjustment, the EOT training period, the ship–buoy SST adjustment, and the grid system of superobservations. The RMSD of the bias adjustment can reach 0.3°C in the tropical and subtropical oceans. The RMSD may represent a measure of uncertainty in the bias adjustment of AVHRR SSTs.
The differences in bias adjustments have a direct impact on the SST analyses using both ER and OI methods, which is confirmed by two SST analyses of the ER and OI using two bias adjustments from EXP-A and EXP-L. In the tropical Pacific Niño-3.4 region, the mean difference of the biases between EXP-A and EXP-L is approximately 0.16°C, which results in an SST difference of 0.12°C between the ER (ERSST15 and ERSSTmon) and OI (OISST15 and OISSTmon) analyses. In contrast to the impact of the bias adjustments on the SST analyses, the impact of analysis methods (ER and OI) on SSTs is as low as 0.01°C. This suggests the important role of bias adjustments of AVHRR SSTs in SST analyses, while the impact of the different analysis methods is small in the tropical Pacific. Similar conclusions are valid for the global oceans. Therefore, one has to assess the uncertainty of the AVHRR bias adjustment when AVHRR observations are ingested into an SST analysis.
Our study suggests that the SST analyses of DOISST and ERSSTsat may be improved by selecting appropriate parameter values in the calculation of the AVHRR bias adjustment. For the AVHRR bias adjustment in DOISST, the EOTs trained with data from the period 1982–2000 should be replaced with EOTs trained with the longest period possible (e.g., 1982–2014) data, which may increase the capability of EOTs in explaining SST variability. The zonal SST adjustment in the lower-latitude oceans should not be used in DOISST when an EOT-based adjustment can be computed. The other reason for removing the zonal SST adjustment in the lower-latitude oceans is that the zonal SST adjustment may distort the spatial patterns in the Pacific, Atlantic, and Indian Oceans. The odd-grid system could be replaced by an even-grid system, so that the in situ observations on the equator can be processed with less error. For the AVHRR bias adjustment in ERSSTsat, it is important to introduce the zonal SST adjustment that can greatly improve the SST analysis in the high-latitude oceans. Alternatively, new EOTs without damping in high-latitude oceans are needed to avoid the zonal SST adjustment. The inclusion of ship–buoy adjustment helps improve the SST analysis in the lower-latitude oceans. The use of daily SSTC may also improve the SSTA and its final analysis.
Given the existence of different SST analysis products, it is frequently asked which SST analysis is the best measure of ocean surface temperatures. In the absence of an SST dataset representing the “truth,” this is a difficult question to answer. However, our comparisons indicate that, relative to the pointwise tropical atmosphere ocean (TAO; McPhaden et al. 1998) observations, the differences of analyzed SST using the OI method are much lower than those using the ER method regardless of which bias adjustment scheme is applied (not shown in figure; refer to Fig. 6 in Huang et al. 2013). The smaller difference in the OI analysis is mostly associated with a much higher spatial (0.25° × 0.25°) and temporal (daily) resolution than in the ER analysis (2° × 2° and monthly). Therefore, the temporal and spatial variabilities are close to the pointwise TAO observations.
Comparisons show that the averaged SST differences relative to the Along-Track Scanning Radiometer’s (ATSR) (Merchant et al. 2012) observations in the Niño-3.4 region are generally close between the SSTs analyzed with the ER and OI methods no matter which bias adjustments are applied (not shown in figure; refer to Fig. 7 in Huang et al. 2013). However, during ENSO events, the SST differences are slightly larger using the bias of EXP-L than using the bias of EXP-A, no matter which analysis methods are applied. The large differences may be associated with a wider data window in EXP-L in comparison with EXP-A. The wider data window may act as a temporal filter smoothing the rapid changes in SSTA during ENSO events. In the high-latitude oceans, the SST differences relative to the ATSR observations are generally smaller in the OI analyses than in the ER analyses.
Therefore, the best selection among different SST products depends on the applications of unique spatial and time scales, as well as unique geographic locations. It should be noted that the ER analysis was designed for large-scale long-term variations of SSTs, while the OI analysis was designed for higher-resolution SSTs during the modern satellite era. It is therefore suggested that users carefully choose a specific SST product according to their unique applications.
Acknowledgments
The authors are thankful for the comments from four anonymous journal reviewers and two NOAA internal reviewers, which greatly improved the manuscript. We appreciate the discussion with John Bates at the initial stage of this study.
Appendix A
DOISST Analysis System
The DOISST methodology is documented in Reynolds et al. (2007), with minor modifications introduced in the current version 2 (Reynolds 2009). The spatial resolution is 0.25° in longitude and latitude over the global oceans, and the temporal resolution is daily. The in situ SSTs and original AVHRR SSTs on the irregular locations are bin averaged and processed to the SST superobservations on 0.25° × 0.25° grid boxes. The SSTAs of the in situ and AVHRR SSTs are derived by subtracting the SST climatology. The biases of AVHRR SSTs relative to in situ (ship and buoy) SSTs within a specific data window (15 days) are assessed on the 2° × 2° odd grid (section 2) using EOT decompositions (section 3), and bilinearly interpolated to the 0.25° × 0.25° grid boxes of AVHRR SSTs. The AVHRR SSTs are adjusted by subtracting the biases.
The superobservations of ship, buoy, and AVHRR are processed by the OI procedure. The OI procedure forms a weighted sum of SSTs using linear weights. The linear weights are determined by solving regression equations of the correlation error and the noise-to-signal standard deviation ratio. The correlation error between two adjacent observations (ship, buoy, or AVHRR) is assumed to be Gaussian with damping scales of 151 and 155 km in the zonal and meridional directions, respectively. The noise-to-signal ratio is 1.94 for ship observations, and 0.5 for buoy and AVHRR observations.
Appendix B
ERSSTsat Analysis System
The ERSSTsat analysis is based on Smith and Reynolds (2003) and Smith et al. (2008). The spatial resolution is 2° in longitude and latitude over the global oceans, and the temporal resolution is monthly. The SSTs from ship, buoy, and AVHRR observations are bin averaged to the 2° × 2° even grid. The biases of monthly AVHRR SSTs relative to monthly in situ (ship and buoy) SSTs are assessed on the 2° × 2° even grid (section 2) using EOT decompositions (section 3). The AVHRR SSTs are adjusted by subtracting the biases. The monthly SSTAs of the in situ and AVHRR SSTs at regular even grid are derived by subtracting the monthly SST climatology. The ship-, buoy-, and bias-adjusted AVHRR SSTAs are merged by the weightings of 1, 7, and 12, respectively, according to their averaged error variances. The merged SSTAs are then decomposed into low- and high-frequency SSTAs. The low-frequency SSTA is constructed by a running mean filter of 26° × 26° in space and 15 yr in time. The high-frequency SSTA, defined as the difference between the original and low-frequency SSTAs, is reconstructed by fitting SSTAs on the global domain to the 130 leading EOTs. The EOTs are similar to the empirical orthogonal functions, except that the EOTs are restricted in domain to a spatial scale of 5000 and 3000 km in longitude and latitude, respectively. The high-frequency SSTA is then merged with the low-frequency SSTA to produce the final SSTAs of ERSSTsat. SSTs are retrieved by adding the monthly climatology to SSTA.
REFERENCES
Bamston, A. G., Chelliah M. , and Goldenberg S. B. , 1997: Documentation of a highly ENSO-related SST region in the equatorial Pacific: Research note. Atmos.–Ocean, 35, 367–383, doi:10.1080/07055900.1997.9649597.
Casey, K. S., Brandon T. B. , Cornillon P. , and Evans R. , 2010: The past, present, and future of the AVHRR Pathfinder SST program. Oceanography from Space: Revisited, V. Barale, J. F. R. Gower, and L. Alberotanza, Eds., Springer, 273–287, doi:10.1007/978-90-481-8681-5_16.
Hirahara, S., Ishii M. , and Fukuda Y. , 2014: Centennial-scale sea surface temperature analysis and its uncertainty. J. Climate, 27, 57–75, doi:10.1175/JCLI-D-12-00837.1.
Huang, B., L’Heureux M. , Lawrimore J. , Liu C. , Zhang H.-M. , Banzon V. , Hu Z.-Z. , and Kumar A. , 2013: Why did large differences arise in the sea-surface temperature datasets across the tropical Pacific during 2012? J. Atmos. Oceanic Technol., 30, 2944–2953, doi:10.1175/JTECH-D-13-00034.1.
Huang, B., and Coauthors, 2015: Extended Reconstructed Sea Surface Temperature version 4 (ERSST.v4). Part I: Upgrades and intercomparisons. J. Climate, doi:10.1175/JCLI-D-14-00006.1, in press.
Kennedy, J. J., Rayner N. A. , Smith R. O. , Parker D. E. , and Saunby M. , 2011: Reassessing biases and other uncertainties in sea surface temperature observations measured in situ since 1850: 2. Biases and homogenization. J. Geophys. Res., 116, D14104, doi:10.1029/2010JD015220.
Kent, E. C., and Challenor P. G. , 2006: Towards estimating climatic trends in SST. Part II: Random errors. J. Atmos. Oceanic Technol., 23, 476–486, doi:10.1175/JTECH1844.1.
Kent, E. C., and Kaplan A. , 2006: Towards estimating climatic trends in SST. Part II: Systematic biases. J. Atmos. Oceanic Technol., 23, 487–500, doi:10.1175/JTECH1845.1.
Kent, E. C., Woodruff S. D. , and Berry D. I. , 2007: Metadata from WMO Publication No. 47 and an assessment of Voluntary Observing Ship observation heights in ICOADS. J. Atmos. Oceanic Technol., 24, 214–234, doi:10.1175/JTECH1949.1.
Kilpatrick, K. A., Podestá G. P. , and Evans R. , 2001: Overview of the NOAA/NASA advanced very high resolution radiometer Pathfinder algorithm for sea surface temperature and associated matchup database. J. Geophys. Res., 106, 9179–9197, doi:10.1029/1999JC000065.
May, D. A., Parmeter M. M. , Olszewski D. S. , and McKenzie B. D. , 1998: Operational processing of satellite sea surface temperature retrievals at the Naval Oceanographic Office. Bull. Amer. Meteor. Soc., 79, 397–407, doi:10.1175/1520-0477(1998)079<0397:OPOSSS>2.0.CO;2.
McPhaden, M. J., and Coauthors, 1998: The Tropical Ocean-Global Atmosphere observing system: A decade of progress. J. Geophys. Res., 103, 14 169–1414 169240, doi:10.1029/97JC02906.
Merchant, C. J., and Coauthors, 2012: A 20 year independent record of sea surface temperature for climate from Along-Track Scanning Radiometers. J. Geophys. Res., 117, C12013, doi:10.1029/2012JC008400.
Reynolds, R. W., 1993: Impact of Mount Pinatubo aerosols on satellite-derived sea surface temperature. J. Climate, 6, 768–774, doi:10.1175/1520-0442(1993)006<0768:IOMPAO>2.0.CO;2.
Reynolds, R. W., 2009: What’s new in version 2. NOAA, 10 pp. [Available online at http://www.ncdc.noaa.gov/sst/papers/oisst_daily_v02r00_version2-features.pdf.]
Reynolds, R. W., and Smith T. M. , 1994: Improved global sea surface temperature analyses using optimum interpolation. J. Climate, 7, 929–948, doi:10.1175/1520-0442(1994)007<0929:IGSSTA>2.0.CO;2.
Reynolds, R. W., Rayner N. A. , Smith T. M. , Stokes D. C. , and Wang W. , 2002: An improved in situ and satellite SST analysis for climate. J. Climate, 15, 1609–1625, doi:10.1175/1520-0442(2002)015<1609:AIISAS>2.0.CO;2.
Reynolds, R. W., Smith T. M. , Liu C. , Chelton D. B. , Casey K. S. , and Schlax M. G. , 2007: Daily high-resolution blended analyses for sea surface temperature. J. Climate, 20, 5473–5496, doi:10.1175/2007JCLI1824.1.
Smith, T. M., and Reynolds R. W. , 2003: Extended reconstruction of global sea surface temperature based on COADS data (1854–1997). J. Climate, 16, 1495–1510, doi:10.1175/1520-0442-16.10.1495.
Smith, T. M., and Reynolds R. W. , 2004: Improved extended reconstruction of SST (1854–1997). J. Climate, 17, 2466–2477, doi:10.1175/1520-0442(2004)017<2466:IEROS>2.0.CO;2.
Smith, T. M., Reynolds R. W. , Peterson T. C. , and Lawrimore J. , 2008: Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J. Climate, 21, 2283–2296, doi:10.1175/2007JCLI2100.1.
Stocker, T. F., and Coauthors, 2014: Climate Change 2013: The Physical Science Basis. Cambridge University Press, 1535 pp. [Available online at www.climatechange2013.org/images/report/WG1AR5_ALL_FINAL.pdf.]
Van den Dool, H. M., Saha S. , and Johansson A. , 2000: Empirical orthogonal teleconnections. J. Climate, 13, 1421–1435, doi:10.1175/1520-0442(2000)013<1421:EOT>2.0.CO;2.
Woodruff, S. D., and Coauthors, 2011: ICOADS Release 2.5: Extensions and enhancements to the surface marine meteorological archive. Int. J. Climatol., 31, 951–967, doi:10.1002/joc.2103.
Xue, Y., Smith T. M. , and Reynolds R. W. , 2003: Interdecadal changes of 30-yr SST normals during 1871–2000. J. Climate, 16, 1601–1612, doi:10.1175/1520-0442-16.10.1601.
Zhang, H.-M., Reynolds R. W. , and Smith T. M. , 2004: Bias characteristics in the AVHRR sea surface temperature. Geophys. Res. Lett., 31, L01307, doi:10.1029/2003GL018804.
