Experiments are carried out to assess the potential contributions of two new satellite datasets, derived from the Visible Infrared Imaging Radiometer Suite (VIIRS) on board the Suomi–National Polar-Orbiting Partnership satellite and the Advanced Microwave Scanning Radiometer 2 (AMSR2) on board the Global Change Observation Mission–Water (GCOM-W) satellite, to the quality of global sea surface temperature (SST) analyses at the Canadian Meteorological Centre (CMC). The new datasets are assimilated both separately and together. Verification of the analyses against independent data shows that the VIIRS and AMSR2 datasets yield analyses with similar global average errors, with the VIIRS analysis performing better during some seasons and the AMSR2 analysis performing better in others. Seasonal cloudiness in some regions diminishes the availability of VIIRS retrievals, resulting in better performance by the AMSR2 analysis. Both datasets were assimilated together along with data from the Advanced Very High Resolution Radiometer (AVHRR), ice data, and in situ data in an updated version of the CMC analysis produced on a 0.1° grid. Verification against independent data shows that the new analysis performed very well, with global average standard deviation consistently better than that of the international Group for High Resolution SST (GHRSST) Multiproduct Ensemble (GMPE) real-time system. This analysis is shown to outperform the currently operational CMC SST analysis, with most of the improvement being due to its assimilation of the VIIRS and AMSR2 retrievals and a further small gain being due to changes to the analysis methodology (including higher resolution).
Accurate specification of SST as an initial condition is an important contributor to forecast skill for both numerical weather prediction (NWP) and numerical ocean prediction (NOP; He et al. 2014). Fluxes of heat and water vapor from the sea surface depend on the SST and can significantly impact the formation and development of storms (Booth et al. 2012). In the atmospheric boundary layer, wind stress curl and divergence are linearly related to the local SST gradient (Chelton et al. 2007). The negative covariance between concentration of sea ice and SST has been well documented (Caya et al. 2010), making SST information an important input to assimilation schemes for sea ice (Buehner et al. 2016). At the Canadian Meteorological Centre (CMC), forecast models use the SST analysis from the microscale (Leroyer et al. 2014) to the global scale (Charron et al. 2012).
As part of its operational environmental prediction program, the CMC produces a daily, global, 0.2° analysis of SST using in situ and satellite data. This analysis is an updated version of the 1/3° analysis described in Brasnett (2008). It currently assimilates data from three Advanced Very High Resolution Radiometer (AVHRR) instruments, and from in situ observations and ice data. Recently, excellent-quality SST data have begun to flow from two new instruments, the Global Change Observation Mission–Water (GCOM-W) Advanced Microwave Scanning Radiometer 2 (AMSR2) and the Suomi–National Polar-Orbiting Partnership (S-NPP) Visible Infrared Imaging Radiometer Suite (VIIRS). A new experimental analysis has been developed to assimilate these new data as well and to improve other aspects of the methodology, most notably an improved resolution from 0.2° to 0.1° to better capture thermal fronts and gradients near the ice edge.
This paper discusses the new data sources and the impact they have on analysis quality. Section 2 describes the AMSR2 instrument and the SST retrievals derived from it. Section 3 describes the VIIRS instrument and the VIIRS SST dataset. Section 4 is a description of the CMC SST assimilation method with emphasis on updates associated with the 0.1° version. Section 5 assesses the contribution of AMSR2 and VIIRS retrievals to analysis quality, and a discussion follows in section 6.
2. The AMSR2 SST product
The AMSR-E instrument aboard NASA’s Aqua satellite demonstrated conclusively the capability of microwave remote sensing to produce accurate SST data in clear and cloudy conditions (Chelton and Wentz 2005). The follow on to AMSR-E, AMSR2, was launched on 18 May 2012 aboard the Global Change Observation Mission–Water 1 (GCOM-W1) satellite (Japan Aerospace Exploration Agency 2013). The AMSR2 instrument has a 2-m antenna compared to the 1.6-m antenna of AMSR-E. This gives AMSR2 a footprint of roughly 49 km as compared to 56 km for AMSR-E. The swath width for AMSR2 is 1450 km.
The AMSR2 retrievals used by the experimental 0.1° CMC analysis are produced by Remote Sensing Systems (RSS; see http://www.remss.com/missions/amsre). RSS retrieves SST and averages retrievals onto a 0.25° × 0.25° grid separately for northbound and southbound orbits. An example of the 24-h coverage is shown in Fig. 1, mapped on the same 0.25° grid used by RSS. Although retrievals are generated in both clear and cloudy regions, there are two important limitations to the coverage. Retrievals are not possible either through precipitating cloud or within about 75 km of land.
The CMC analysis described in Brasnett (2008) represents SST at a depth where no diurnal variability is present. This temperature is referred to in the literature as the foundation SST (SSTfnd; Donlon et al. 2007). Unfortunately, the buoy and ship observations used in the analysis and satellite bias correction algorithm are not, in general, measurements of SSTfnd, and steps should be taken to eliminate observations affected by diurnal variability. Usually, this is done by rejecting daytime only, or daytime and nighttime data, if the wind speed is less than 6 m s−1 (Gentemann et al. 2004; Martin et al. 2012). In the case of the AMSR2 dataset, RSS includes retrievals of surface wind speed that can be used for this purpose. Accordingly, daytime retrievals of SST are not used in the analysis if the wind speed is less than 6 m s−1 for retrievals between 25°S and 25°N. Elsewhere, the same criterion is applied but only within 45 days of the summer solstice.
3. The VIIRS SST product
Launched in October 2011, the S-NPP VIIRS instrument is a significant advancement from its heritage instruments, the National Oceanic and Atmospheric Administration (NOAA) AVHRR and the National Aeronautics and Space Administration (NASA) Moderate Resolution Imaging Spectroradiometer (MODIS). When comparing VIIRS and AVHRR, one significant difference is the resolution, 740 m at nadir for VIIRS versus 1.1 km for AVHRR. But the most significant difference in resolution is at the edge of the swath. While the swath width is similar for both instruments (2800–3000 km), technological innovations in the VIIRS sensor result in a resolution of 1.6 km at edge of scan (Miller et al. 2006), a vast improvement from the 6.7-km resolution at the edge of scan for AVHRR (Johnson et al. 1994). Since SST retrievals are not possible through cloud for infrared instruments, a higher-resolution sensor means a higher probability of detecting cloud-free pixels in scenes with patchy cloud cover. Thus, the VIIRS instrument can provide retrievals in some areas where few or none are possible with the lower-resolution AVHRR sensor.
The VIIRS dataset (NOAA OSPO 2014) used here is produced by NOAA using the Advanced Clear-Sky Processor for Oceans (ACSPO). The production of these retrievals includes a regression algorithm (Petrenko et al. 2014), an automatic adaptive destriping algorithm (Bouali and Ignatov 2014), a clear-sky masking process (Petrenko et al. 2010), and monitoring (Liang and Ignatov 2011; Dash et al. 2010; Xu and Ignatov 2014). ACSPO provides observed top-of-atmosphere clear-sky brightness temperature (BT) and SST retrieved from these BTs. It also provides simulated BTs, and their accuracy is important for ACSPO functionalities. Several real-time global analyses of SST have been tested for their potential use as first-guess fields in ACSPO (Saha et al. 2012) and following that study, the CMC SST analysis was implemented as the first guess for ACSPO.
Examples of the VIIRS retrieval coverage compared to that of AVHRR are shown in Fig. 2. The AVHRR data shown in this figure is from the sensor on board NOAA-19. Data from this AVHRR instrument are subsampled on board the satellite. Four out of every five samples along the scan line are used to compute an average value, and the data from every third scan line are processed yielding a resolution of 4 km at nadir and 25 km at edge of scan. These data are referred to as global area coverage (GAC) data. The NOAA-19 AVHRR dataset used in our analysis is produced by the Naval Oceanographic Office (NAVO; May et al. 1998). Binning and averaging of GAC AVHRR data result in a final resolution for NAVO’s product of around 8 km at nadir and around 25 km for a view zenith angle of almost 53°, the cutoff angle for the NAVO product. One of the qualities of VIIRS retrievals produced by NOAA is the preservation of its natural spatial resolution, so the product used here has a resolution of approximately 0.75 km at nadir and 1.5 km at swath edge for a view zenith angle of 68° (NOAA OSPO 2014). In addition to the difference in resolution between the ACSPO VIIRS data and NAVO AVHRR data, there is another difference between the two datasets. NAVO uses a more conservative algorithm to identify cloudy pixels than that used by NOAA (A. Ignatov 2014, personal communication), with the result that the coverage provided by the NOAA product is more optimal. For the boreal summer case shown in Fig. 2, 3.2 times more grid cells contain ACSPO VIIRS data than contain NAVO AVHRR data. North of 50°N, where data are historically quite sparse, this ratio is 5.6. Indeed, on this day, ACSPO VIIRS data also provided better coverage than AMSR2 by a factor of 1.3. The orbital gaps in the AMSR2 data (Fig. 1) resulting from its 1450-km swath width are not present in the ACSPO VIIRS data. Also contributing to the better coverage at high latitudes is the large amount of lake surface temperature data included in the ACSPO VIIRS dataset. Figure 3 shows the coverage for much of Canada and the northern United States for 24 September 2014. Many more lake surface temperature retrievals appear in the ACSPO VIIRS dataset (left panel) than appear in the composite of three NAVO AVHRR datasets: NOAA-18, -19, and MetOp-A (right panel). Note that although only GAC data are available globally for NOAA-18 and -19, the original high-resolution AVHRR data [called full-resolution area coverage (FRAC)] are available for MetOp-A. The MetOp-A dataset used in this study is a GAC product generated by NAVO by subsampling the FRAC product. This yields a dataset with a resolution of almost 8 km, similar to the GAC datasets obtained from the NOAA platforms.
VIIRS data from the ACSPO system have been publicly available from NASA’s Physical Oceanography Distributed Active Archive Center (http://podaac-www.jpl.nasa.gov) since May 2014. Data are provided in a standard format established by the international Group for High Resolution SST (GHRSST; Donlon et al. 2007) and include ancillary data, such as quality flags and surface wind speeds. At CMC, only retrievals flagged as best (quality level 5) are used. In addition, the wind speeds are used to remove those retrievals likely to be affected by diurnal heating. Daytime retrievals are eliminated when the wind speed is less than 6 m s−1 between 25°S and 25°N and elsewhere within 45 days of the summer solstice.
4. Assimilation methodology
The essential components of the method are as described in Brasnett (2008). Briefly, the method of statistical interpolation is applied: to the analysis problem, to the observation quality control, and to estimate satellite biases. The SST assimilation methodology uses anomaly from climatology as the analysis variable. The background is based on simple persistence. In practice, this means taking the most recent (24 h old) analysis and modifying it by a return to climatology for use in the analysis procedure. The return to climatology consists of scaling the anomalies by 0.983, equivalent to an exponential decay with an e-folding time of 58 days. With this field as the background, current observations are assimilated, including retrievals from the two new instruments, AMSR2 on board GCOM-W and VIIRS on board S-NPP. The AMSR2 and daytime VIIRS retrievals are ascribed an observation error of 0.9 K, and nighttime VIIRS retrievals are assimilated with an error of 0.7 K. In this section, updates included with the 0.1° version are described in detail.
a. Adjustment of satellite biases
The method described here seeks to estimate and remove the large-scale biases in retrievals of SST. Following the discussion in Reynolds et al. (2002), it is assumed that these biases are correlated in time as well as in space, that biases for nighttime and daytime retrievals evolve independently, and that biases from each satellite instrument are also independent.
The statistical interpolation method is used to produce estimates of the satellite biases. Using preprocessed observations from moored and drifting buoys and preprocessed ship observations (one observation per platform per day), data from a given satellite are paired with an in situ observation if the spatial separation is 25 km or less and if both observations originate on the same day. If multiple satellite observations satisfy these criteria, the median SST of the retrievals is selected. The apparent bias is then computed as the difference of the satellite retrieval and the in situ observation. These differences are used as the observations in a statistical interpolation analysis of satellite bias performed on a 2.5° grid. This type of analysis is produced for both daytime and nighttime retrievals and for each satellite instrument contributing to the SST analysis. Once the gridded estimates of satellite bias are updated, they are interpolated to the retrieval locations so the bias estimates can be subtracted from the retrievals. These bias-adjusted retrievals are then used in the SST analysis.
Observation errors in these satellite bias analyses are 0.6 K for drifter matchups, 0.8 K for ship matchups, and 0.7 K for moored buoy matchups. The background error standard deviation is taken to be 0.2 K everywhere. The background for each analysis is the analysis from the previous day scaled by 0.8187 which, in the absence of new information, produces an exponential decay of bias estimates with an e-folding time of 5 days. The background error correlation length scale is 750 km over most of the globe but increases to 1000 km south of 53°S and to 3000 km south of 68°S. These length scales were arrived at empirically.
Figure 4 shows an example of a bias field for daytime ACSPO VIIRS retrievals for 1 October 2014. On this day biases were generally less than 0.4 K except for a few areas near the coast of North America and along the coast of Portugal, where biases were slightly greater than 0.4 K. In most cases, the bias correction scheme should remove some or all of the bias in retrievals over areas where reliable in situ data are present. However, when a sensor bias has a spatial scale of less than 750 km and/or a decorrelation time of less than 5 days, retrieval bias will be an issue.
b. Background error spatial scales and retrieval spacing
Along with increasing the resolution of the analysis grid to 0.1°, additional changes were needed to fully benefit from the improved resolution. To resolve smaller spatial scales, the length scales (e-folding distances) of the background error correlations were modified as shown in Fig. 5. The length scales are isotropic and symmetric about the equator, and there is no difference between the low- and high-resolution analyses from the equator to ±37.5°. Different experiments were done varying these length scales, and the values that minimized the analysis error estimates based on independent data were chosen. In this way, it was determined that the analysis could not be improved in the low latitudes by reducing the length scale but modest improvement at high latitudes is possible. Accordingly, at high latitudes where the smallest length scale is used, the value of this length scale is 24 km in the 0.1° analysis compared to 43 km in the 0.2° analysis.
The density of the satellite data assimilated is critically important. The connection between the observation density, observation resolution, and the resolution of the model grid was studied by Liu and Rabier (2002). They showed that the analysis quality decreases if the density of the observational dataset is too large and the error correlations are neglected.
Satellite retrievals are fundamentally different from conventional observations. All retrievals in a satellite swath originate from the same instrument, whereas many instruments are needed to obtain similar coverage from buoys. As discussed in the previous section, even with a satellite bias correction scheme, it cannot be assumed that retrievals from a single sensor are unbiased. Moreover, our implementation of statistical interpolation does not take into account correlated observations errors. As a consequence of these facts, satellite retrievals must be thinned prior to assimilation so that they do not receive undue weight in the analysis. It was found by experimentation that an appropriate spacing for infrared retrievals from a single satellite instrument is 33 km north of 40°N and south of 40°S, increasing to 80 km at the equator. The smallest spacing of satellite data used in the operational 0.2° analysis is 44 km. For the lower-resolution AMSR2 instrument, a 55-km spacing is used everywhere except north of 60°N and south of 60°S, where the spacing is 89 km.
It may appear somewhat contradictory to increase the resolution of the analysis grid from 0.2° to 0.1° while at the same time thinning the satellite retrievals to 33 km (effectively 0.3°). At first glance, it seems this strategy will yield analyses with no features at spatial scales smaller than 0.3°. As shown by Reynolds and Chelton (2010), the grid spacing of an SST analysis is not an indicator of the feature resolution; this actually depends on the coverage from different satellite instruments. The analysis of small-scale features benefits from high-resolution data coverage (Reynolds et al. 2013). However, while high-resolution data provide an increase in feature resolution, there is another important source of small-scale information, namely, the contribution to each analysis from prior observations. As we will show in section 5, in general for the World Ocean, observations that are more than one day old contain substantial information about the current SST.
c. Insertion of ice information
The 0.1° SST analysis adds ice information by inserting proxy SST data at locations where ice is present. As part of its NWP activities, CMC produces a 10-km global ice analysis four times per day. This global analysis uses essentially the same methodology as the regional ice analysis described in Buehner et al. (2016). The global ice analysis was chosen for use with the 0.1° SST product due to its use of new methodology and its higher resolution. For application to SST, the ice analysis valid at 1800 UTC is sampled. Each day, proxy observations are generated at every third grid point along the orthogonal lines of grid points yielding a spacing of 30 km. The proxy observations are produced where the ice concentration is 0.6 or larger. The sampling starts from a reference grid point of the ice analysis grid, which is displaced daily so that a complete sampling of the 10-km grid occurs over a 9-day period. An example of the spatial distribution of ice proxy observations is shown in Fig. 6.
In most cases, the ice proxy value is −1.8°C, the freezing point of seawater with a salinity of 33 psu. When the ice is melting, however, this value is not appropriate. When meltwater is present, the three phases of H2O coexist at the air–water–ice interface. Thus, by definition of the triple-point temperature of H2O, a proxy value of 0°C is used in this situation (Halliday and Resnick 1974). To identify those grid points where the ice is likely to be melting, a time average of the surface air temperature is used. This time average is produced from analyses of air temperature valid at 6-h intervals using a 6-h forecast from the CMC global atmospheric model (Charron et al. 2012) as background, and it incorporates all available air temperature reports from drifters, ships, and land stations. These temperature analyses are exponentially weighted by combining the current temperature analysis with the mean from 6 h earlier, using weights of 0.2 and 0.8, respectively. At locations where this updated running mean is greater than 0°C and the ice concentration is between 0.6 and 0.9, a proxy SST of 0°C is used. Where the ice concentration exceeds 0.9, a proxy value of −1.8°C is used regardless of the average temperature. The proxy SSTs are then assimilated with an ascribed observation error of 1.0°C. This error is greater than the error ascribed to any other observation type.
Several of the changes to the analysis methodology result in better definition of the SST gradient near the ice edge. This is illustrated in Fig. 7, which shows the region of the Beaufort Sea, Banks Island, and western Victoria Island on 1 August 2014. Ice cover on this day was extensive over the northwest quadrant of the region, as indicated by the locations of ice proxy observations assimilated by each analysis (white dots). Besides the area of open water between 70°N and 72°N, there was also an area of open water bounded approximately by 72°N and 74°N and 125°W and 130°W. Each analysis has large gradients to the west and north of this latter region, but the gradients in the experimental analysis (left panel) are 5 K (100 km)−1 larger at several locations. The larger gradients are likely due to a combination of factors, including the smaller correlation length scales at these latitudes (Fig. 5), the denser spacing of observations, the inclusion of VIIRS retrievals, and the finer resolution of the analysis grid. One prominent difference between the two panels in Fig. 7 is the area of large gradient in the experimental analysis near 69.5°N, 135°W. The experimental analysis captured a plume of warm water from the Mackenzie River, which flows into the Beaufort Sea here—a feature missed by the operational analysis (right panel).
d. Capturing sudden SST changes due to tropical storms
High winds associated with the passage of an intense tropical cyclone form a cold wake by turbulent mixing of deep water to the surface, particularly where the wind stress is largest (Price 1981). D’Asaro et al. (2007) state that the reduction of SST values is almost entirely due to the upward mixing of water rather than to air–sea fluxes. This cold wake has been extensively studied for many hurricanes (Price 1981; Price et al. 1986; Cornillon et al. 1987; Price et al. 2008).
One problem noted with the operational analysis is its tendency to reject observations originating from the wake of intense tropical storms. During the quality control of observations, which is performed using the statistical interpolation methodology (Brasnett 1997), an estimate of the SST anomaly is produced at the location of the observation being tested but with the latter withheld from the calculations. The result is an independent estimate against which the observation can be compared. Observations that differ by more than a certain tolerance from the independent estimate are rejected. However, the background error used by the algorithm is essentially a climatological value, unrepresentative of rare events like intense tropical storms, whose passage can cause SST to decrease by roughly 3 K or more. SST changes of 3 K in one day generally do not occur outside of the western boundary currents. As a result of the inappropriate background error values, good-quality observations in the wake of intense tropical storms can be rejected.
To avoid the rejection of these good observations in the 0.1° analysis, wind forecasts from the CMC global atmospheric model are used to identify regions likely to have experienced a sudden decrease in SST. The 6-h forecasts of 10-m wind speed from a 4-day period prior to analysis time are used. The background error used during the quality control phase of the analysis is then amplified in those regions by a factor that varies from 1, where the maximum wind speed is 21 m s−1 or less, to 12, where the maximum wind speed is 24 m s−1 or more. The factor is then filtered to produce a smoothly varying field. This amplified background error is used only to perform the quality control calculations. The unaltered background error is used to compute the analysis at grid points.
The effect of this change to the quality control is illustrated in Fig. 8, which shows the difference between the 0.1° analysis and the operational analysis. Both analyses assimilated data for the period ending at 0000 UTC 9 July 2014. The 0.1° analysis is colder than the operational analysis over a large swath centered at 20°N, 130°E. The operational analysis rejected numerous satellite retrievals in the region. The track of Supertyphoon Neoguri is also plotted in the figure. It is clear that the colder temperatures in the 0.1° analysis are associated with the wind field of the storm. The coldest temperatures are to the right of the track, where the strongest winds typically occur. Differences near the location of the storm on 8 July are small because neither microwave nor infrared satellite retrievals are available in areas with precipitating cloud.
Differences between the two analyses were verified using independent data from Argo floats, which are typically available several days after observation time. Although Argo floats are high-quality measurements of ocean temperature and salinity profiles, they have a 10-day sampling interval, making them less useful for detecting a tropical cold wake region, which can appear and dissipate in less than 10 days. Consequently, only seven observations were available for verification for the period 5–8 July 2014, and most of these were from the periphery of the cold wake region. Nevertheless, based on these seven observations, the modified quality control procedure reduced the bias in the region from 2.1 to 1.1 K and the rms error from 2.4 to 1.3 K.
e. Other changes
The large difference in resolution between infrared and microwave retrievals necessitates an important difference in the processing of the retrievals. With the 0.1° grid length of the background, many grid points fall within the 49-km footprint of the AMSR2 instrument. Accordingly, following the approach of Donlon et al. (2012), the observation operator needed to compute the AMSR2 innovations (observation minus background differences) was modified to average values at grid points within the footprint of AMSR2. In the case of infrared retrievals, the observation operator remains a bilinear interpolation of the background.
Another change to the methodology was in the preprocessing of closely spaced ship observations. Previous versions generated “superobservations,” the average of two or more collocated observations. Occasionally it was found that the averaging of clusters of observations resulted in a contaminated superobservation when one or more outliers contributed to the mean. To avoid this, the strategy was changed to choose one report from the cluster (the report with the median SST) and discard the rest. This approach has a small but positive impact on analysis quality.
5. Assessment of contributions from AMSR2 and VIIRS
The contributions of different observations to the analysis were evaluated by estimating analysis error using the Argo float temperatures discussed in section 4d. These temperature reports, which are not used in the analysis, are used for verification only if they are between 3 and 5 m in depth and within four standard deviations of the climatology interpolated temporally and spatially to the date and location of the Argo float observation. Using these criteria to select the observations, an average of 144 observations per day was available to assess analysis error. Argo floats are quite uniformly distributed over the World Ocean except for marginal seas and areas near the ice edge.
As stated in section 4b, the thinning of high-resolution retrievals from satellite sensors makes it possible to neglect the correlations in the observation errors of these retrievals. Here we show that the analysis may still adequately represent small-scale features due to the contribution to each analysis from prior observations. Figure 9 shows the results of several experiments carried out to assess the ability of the experimental analysis to preserve SST information on the analysis grid when observations are not available. It shows the global average of the analysis standard deviation for the normal analysis, which has used the full complement of data every day during the period (green dotted–dashed line), the climatology (solid black line), an experiment where data were denied on the current day (black dotted–dashed line), an experiment where data were denied for 7 days (red line) and, finally, an experiment where data were denied for 30 days (blue line).1 The climatology is an important reference because the analysis is designed to gradually return to climatology if there are no new observations, as explained in section 4 above. Hence, if data are denied for long enough, the analysis error will approach the error of climatology asymptotically, making the latter the upper bound on the analysis error. It should be noted that the analysis that is denied data on the current day is identical to the background for the normal analysis. It is clear from the figure that the reduction in error achieved from assimilating the current day’s observations (difference between the two dotted–dashed lines) is small compared to the reduction in error achieved from using the most recent analysis as the background instead of climatology (difference between the two black lines). This is an intrinsic property of the analysis described here. The figure illustrates the excellent quality of the background in the normal analysis. This background represents the information retained from all prior observations. Note also that the analysis denied data for 30 days (blue curve) continues to have significant skill over climatology, indicating that even 30-day-old observations contain some information on the current SST and also showing that the analysis has preserved this information over this period. Moreover, the only constraint on the spatial scales of the information stored on the analysis grid is the grid length and therefore, one should expect all spatial scales that can be resolved on the grid to be present. These results suggest that if information with spatial scales smaller than those of the thinned observations accumulates in the analysis through random sampling, the system will preserve this information on the grid for weeks.
It should also be noted that the results from Fig. 9 confirm the validity of some choices made in the design of the analysis methodology. The excellent quality of the analysis that was denied data on the current day justifies its use as the background for the normal analysis. Also, the fact that an analysis denied data for 30 days has significantly more skill than climatology indicates that the global average decorrelation time for this field is likely greater than 30 days, bolstering confidence in the value of 58 days reported by Reynolds (1978) and employed in the return to climatology algorithm.
a. Evaluation of new datasets assimilated independently
To quantify the contribution from the different sources of satellite data, two experimental analyses were produced with the same methodology on a 0.2° grid. In addition to in situ and ice data used in both experiments, the first experiment assimilated AMSR2 data and the second used VIIRS data. Verification results for these analyses appear in Fig. 10. VIIRS data were not available from 1 April 2014 to 19 May 2014 and so this period is omitted from the results for the analysis with VIIRS. To facilitate comparison of the results, the month of May was defined as 20–31 May for both experiments.
As Fig. 10 shows, the analysis errors of the two experiments were generally similar. However, the VIIRS experiment was significantly better than the AMSR2 experiment at the 95% level of confidence during January, February, October, and November, while AMSR2 was significantly better than VIIRS during June–August. The variation in the errors between the January–March (JFM) period and the June–August (JJA) period is explored further in Fig. 11. The figure shows the percentage of days during both periods where retrievals were available from VIIRS (left panels) and AMSR2 (right panels). For JFM (top panels), the VIIRS dataset is characterized by excellent coverage everywhere except over the Gulf Stream, the intertropical convergence zone, and the marginal ice zone. However, during the JJA period (bottom panels), persistent cloudiness affects VIIRS coverage quite dramatically. The regions most affected include the North Pacific; the Atlantic north of about 45°N; the Arabian, Greenland, Norwegian, and Barents Seas; the waters west of Peru and Angola; and the Bay of Bengal. In all of these regions, VIIRS retrievals were available 30% of the time at most, while the AMSR2 retrievals were generally available at least 60% of the time. This advantage for AMSR2 is manifested in lower JJA analysis errors in Fig. 10.
b. Evaluation of new datasets assimilated together
In Fig. 12, a 12-month time series of analysis standard deviations and biases is shown for the 0.1° analysis, the CMC operational analysis, and the GHRSST Multi-Product Ensemble (GMPE). The 0.1° analysis used NAVO AVHRR data from NOAA-18 and -19 and MetOp-A, as well as data from RSS AMSR2, ACSPO VIIRS, the new 10-km CMC ice analysis, ships, and buoys. The operational analysis during the same period used data from in situ sources, the operational CMC ice analysis (resolution: 37 km), and NAVO AVHRR data from NOAA-18 and -19 and MetOp-A. The GMPE product, described in Martin et al. (2012), is the median of several (typically 10 or 11) real-time analyses and was found to be more accurate than any of the contributing analyses. The 0.1° product is consistently more accurate than the operational CMC analysis and the GMPE product. Only in April, a month for which no ACSPO VIIRS data were available for this study, was the error difference between the experimental analysis and GMPE not statistically significant.
Figure 13 shows average analysis errors for the global ocean and several regions. As in Fig. 12, the results are for the period 1 January 2014 to 31 December 2014, and show the same three analyses. The global statistics confirm what was seen in Fig. 12, with the 0.1° product showing the smallest standard deviation and bias. Only for the Indian Ocean was the error difference between the experimental analysis and GMPE not statistically significant.
The results of Figs. 12 and 13 raise the question of whether the gain in skill of the experimental analysis is primarily due to improvements in the analysis methodology (including analysis resolution) or to the addition of the RSS AMSR2 and ACSPO VIIRS datasets. This question is addressed by the results of Fig. 14. The figure shows the monthly average standard deviations and biases for three analyses. The first (blue curves) is the operational 0.2° analysis as described above. The second (red curves) is the same 0.2° analysis but with RSS AMSR2 and ACSPO VIIRS data assimilated, in addition to the NAVO AVHRR, in situ, and ice data assimilated in the operational analysis. The same data are assimilated in the third analysis (green curves), which is the 0.1° experimental analysis and includes the modifications to the analysis methodology described in section 4. The results clearly show that most of the reduction in analysis standard deviation results from the addition of AMSR2 and VIIRS data. However, the changes to the analysis methodology do consistently provide a small gain over the 0.2° analysis using the same data, although this improvement is only statistically significant for February, April, and July.
c. Wavenumber spectra
As mentioned in section 4, neither the increase in resolution from 0.2° to 0.1° nor the assimilation of new satellite retrievals guarantees an increase in feature resolution. To determine if more small-scale information is resolved, Reynolds and Chelton (2010) examined zonal wavenumber spectra. Their study compared the zonal spectral density of six SST analyses averaged over two 1-month periods for several regions. The authors were able to make statements about the information content of different analyses over a range of spatial scales, regions and seasons. While such an extensive study is beyond the scope of the present work, it is of interest to examine the spectra of the two analyses with the different resolutions discussed previously.
The spectral decomposition used here was obtained with the discrete cosine transform discussed in Denis et al. (2002). This transform is well suited to the calculation of two-dimensional analysis spectra for a regional domain without contamination by aliasing of the large-scale modes. Unlike the zonal spectra calculated by Reynolds and Chelton (2010), the discrete cosine transform is a two-dimensional transform and therefore produces a two-dimensional wavenumber spectrum.
Spectral variance was computed for the domain bounded by 70°N, 72°N, 132°W, and 142°W, a subregion of the region shown in Fig. 6, for each day in July and November 2014. These variances were then averaged over the month. Following Denis et al. (2002), the monthly average variances were binned by a two-dimensional wavenumber band and summed within each band to compute the total variances shown in Fig. 15. The solid lines represent the spectral variance of the 0.1° analysis described in section 4, and the dashed lines are the spectral variance of the 0.2° analysis with AMSR2 and VIIRS data assimilated, as described in the previous section. In July, the peak variance at large wavelengths gradually declines for both analyses starting at about 110 km. For wavelengths smaller than about 75 km, the 0.1° analysis has more variance than the 0.2° analysis. At a wavelength of about 45 km, roughly the Nyquist wavelength for the 0.2° grid, the spectral variance for this analysis drops off as expected. For both analyses the spectral variance is lower during the month of November because the region is mostly covered with ice over this period, and consequently the small-scale features of the analyses are present only over a smaller area. Even in November, however, the 0.1° analysis has more variance than the 0.2° analysis for wavelengths less than 75 km and the variance drops again for scales smaller than 45 km. These results confirm that the 0.1° analysis indeed contains more small-scale information than the 0.2° analysis.
d. Impact of VIIRS data on the analysis of lake temperature
As stated in section 3, the addition of ACSPO VIIRS retrievals dramatically increases the number of retrievals of lake surface temperature available to the analysis compared to the NAVO AVHRR products. The effect of the additional data can be substantial. Figure 16 shows the 0.1° analysis and the operational analysis for the world’s tenth largest lake by surface area, Great Slave Lake, on 15 August 2014. Large differences are seen in the northeastern arm of the lake. The operational analysis, relying on the NAVO AVHRR datasets, assimilated no observations in this region, while the experimental analysis assimilated many ACSPO VIIRS retrievals here. As a result, climatological values prevail in the operational analysis, resulting in large differences between the two analyses. Another significant difference is on the southern shore of the lake near 113.5°W, where a plume of warm water measuring approximately 50 km across can be seen in the experimental analysis but not in the operational analysis. This plume is likely the warm water from the Slave River flowing into Great Slave Lake at this location.
SST retrievals from two new satellite instruments, the S-NPP VIIRS and the GCOM-W1 AMSR2, were assimilated by an updated, 0.1° resolution, CMC analysis system. The results were encouraging, with the experimental product showing more skill than the existing CMC analysis and indeed more skill than the GMPE product. It was also demonstrated (Fig. 14) that most of the improvement of the new analysis relative to the operational analysis was due to the addition of the RSS AMSR2 and ACSPO VIIRS datasets. It should be pointed out, therefore, that due to the recent release of the new datasets, it is unlikely that very many of the member analyses of the GHRSST multiproduct ensemble were assimilating the two new datasets during the period considered. Thus, perhaps it is not surprising that the experimental analysis showed more skill than the GMPE product. A more meaningful comparison will be possible when a majority of the members of the ensemble are assimilating the new datasets. Nevertheless, since the GMPE product is recognized as the most accurate global SST product available in real time, it remains an important benchmark for assessing analysis accuracy.
Figure 10 revealed a seasonal dependence in the relative skill of the VIIRS analysis and the AMSR2 analysis. Further investigation showed that the ACSPO VIIRS dataset was diminished by persistent cloudiness in several regions during the JJA period. This effect reduced sampling frequency to less than 30% of days in these regions. However, the RSS AMSR2 dataset continued to have high sampling frequency in these regions, illustrating how it complements the infrared observing systems nicely. During the JFM period, when data loss due to cloud cover is generally less of an issue, the VIIRS analysis tended to outperform the AMSR2 analysis, suggesting that in clear skies the higher-resolution ACSPO VIIRS retrievals are of more value than those of RSS AMSR2. However, both the ACSPO VIIRS and RSS AMSR2 datasets are clearly needed to produce a high-quality analysis.
Differences in spatial distribution between various satellite datasets were illustrated in Figs. 1–3, and the preceding discussion on the seasonal variation of analysis skill illustrates the impact that poor sampling can have. Therefore, the spatial coverage provided by competing datasets must be one criterion to consider when choosing datasets for assimilation. In fact, significant differences in coverage exist between the two VIIRS datasets currently available, produced by NOAA and NAVO. Ideally, retrieval producers should generate datasets with the most complete coverage possible. The GHRSST format for distribution of these data includes ancillary information, such as the satellite zenith angle, which could be used to preselect the best-quality retrievals.
The experimental analysis described here will be implemented operationally in the near future. Future work with SST data at CMC will focus on the assimilation of SST data directly into an ocean model. The aim of the work is to improve forecasts of ice, ocean, and atmosphere by capturing variability caused by ocean dynamics and the diurnal cycle. Nevertheless, the 0.1° analysis may well be needed for some time since it provides lake surface temperatures for the NWP models. Another application of this methodology is for the reanalysis of SST to produce high-quality, daily, global analyses for past decades, particularly for the satellite era (starting in 1981). For climate studies, it is important to process the data in a consistent way throughout the time series. Our experience with the ACSPO VIIRS product suggests that generating similar ACSPO products from all historical AVHRR and MODIS sensors back to 1981 and 2000, respectively, will be beneficial to the CMC reanalysis effort, which currently goes back to 1991, the start of the Along-Track Scanning Radiometer (ATSR) SST record.
We thank Hersh Mitchell for his careful review of the manuscript, encouragement, and support. We are grateful to Alexander Ignatov and his colleagues at NOAA for providing the VIIRS data for January–March 2014. We appreciate the help and encouragement of Wei Yu and three anonymous reviewers for their constructive comments. The ACSPO VIIRS SST product is produced by NOAA under the Joint Polar Satellite System (JPSS) program. Data are available from the Physical Oceanography DAAC (podaac.jpl.nasa.gov) and the NOAA National Oceanographic Data Center (www.nodc.noaa.gov). AMSR2 data are produced by Remote Sensing Systems and is sponsored by the NASA Stand-Alone Missions of Opportunity (SALMON) program and the NASA Earth Science MEaSUREs DISCOVER project. Data are available online (at www.remss.com).
To produce a data-denial analysis, one can take the analysis from 1, 7, or 30 days before the analysis date and apply the return to climatology for this period. The result will be the analysis for this date.