OCTOBER 2007 B A R T O N 1773

Comparison of In Situ and -Derived Sea Surface in the Gulf of Carpentaria

IAN J. BARTON CSIRO Marine and Atmospheric Research, Hobart, Tasmania, Australia

(Manuscript received 29 June 2005, in final form 9 November 2006)

ABSTRACT

During 30 days in May and June 2003, the R/V Southern Surveyor was operating in the Gulf of Carpen- taria, northern Australia. Measurements of sea surface (SST) were made with an accurate single-channel infrared radiometer as well as with the ship’s thermosalinograph. These ship-based mea- surements have been used to assess the quality of the SST derived from nine satellite-borne instruments. The satellite dataset compiled during this period also allows the intercomparison of satellite-derived SST in areas not covered by the ship’s track. An assessment of the SST quality from each satellite instrument is presented, and suggestions for blending ground and satellite measurements into a single product are made. These suggestions are directly applicable to the international Global Data As- similation Experiment (GODAE) High Resolution SST Pilot Project (GHRSST-PP) that is currently developing an operational system to provide 6-hourly global fields of SST at a spatial resolution close to 10 km. The paper demonstrates how the Diagnostic Datasets (DDSs) and Matchup Database (MDB) of the GHRSST-PP can be used to monitor the quality of individual and blended SST datasets. Recommendations for future satellite missions that are critical to the long-term generation of accurate blended SST datasets are included.

1. Introduction flying on European has been specifically de- signed to provide accurate estimates of SST suitable for Sea surface temperature (SST) is one of the key pa- applications. Other infrared instruments have rameters used in global numerical modeling of been launched on a series of environmental satellites, and climate processes. Many space agencies have re- specifically the Global Imager (GLI) on the Advanced cently launched satellites that include infrared and mi- Observing Satellite-II (ADEOS-II) and the Mod- crowave instruments designed to provide global or re- erate Resolution Imaging Spectroradiometers (MODIS) gional data on the temperature of the ocean surface. on the Terra and satellites. Advanced Infrared systems that provide operational products in- Scanning Radiometer (AMSR) instruments have been clude the Advanced Very High Resolution Radiometer included on both the ADEOS-II and Aqua satellites. (AVHRR) on the National Oceanic and Atmospheric In this study only data from the Earth Observing Sys- Administration (NOAA) series of polar-orbiting satel- tem (EOS) AMSR (AMSR-E) on Aqua are used. Un- lites and the imaging radiometers on all geostationary fortunately, ADEOS-II has not provided data since No- operational meteorological satellites. In this study we vember 2003, but data from GLI have been made avail- have access to data from the Geostationary Operational able for this study. Further details on the satellite Environmental Satellite-9 (GOES-9) Multispectral Im- instruments used in this study are included in a later ager (MSI) after it was moved from above the eastern section. Pacific Ocean to over the equator north of Australia. The instruments described above all supply estimates The Along-Track Scanning Radiometer (ATSR) series of SST that have a range of spatial resolutions and ac- curacies. However, in most cases, the accuracies of sat- ellite-derived SST estimates are comparable to those Corresponding author address: Dr. I. J. Barton, CSIRO Marine and Atmospheric Research, P.O. Box 1538, Hobart, Tasmania for data collected from ground-based platforms. The 7001, Australia. suite of satellite instruments listed above now provides E-mail: [email protected] a global coverage that is not practical from ships and

DOI: 10.1175/JTECH2084.1

© 2007 American Meteorological Society

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC

JTECH2084 1774 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24

TABLE 1. Satellite radiometers used in the analysis with the number of cloud-free coincidences with the ship location. The overpass times for each satellite over the Gulf of Carpentaria and the channels used for SST determination are also included. UTC ϭ Gulf of Carpentaria overpass times (UTC) day (night); Resolution ϭ spatial resolution (km).

Satellite Radiometer UTC Resolution Channels for SST Total scenes Cloud free NOAA-12 AVHRR 0640–0815 (1830–2000) 1.1 3.7, 10.8, 12 ␮m4936 NOAA-16 AVHRR 0400–0535 (1540–1715) 1.1 3.7, 10.8, 12 ␮m5242 Terra MODIS 0015–0150 (1240–1355) 1.0 3.7, 3.9, 4.0, 11, 12 ␮m4723 Aqua MODIS 0345–0505 (1529–1645) 1.0 3.7, 3.9, 4.0, 11, 12 ␮m4724 ERS-2 ATSR-2 0045–0115 (1300–1330) 1.0 3.7, 10.8, 12 ␮m74 Envisat AATSR 0015–0045 (1230–1300) 1.0 3.7, 10.8, 12 ␮m184 ADEOS-II GLI 0035–0110 (1240–1340) 1.0 3.7, 8.6, 10.8, 12 ␮m 112 22 GOES-9 MSI Geostationary 4.0 10.8, 12 ␮m 640 217 Aqua AMSR-E 0345–0505 (1529–1645) 38, 56 6.9, 10.6, 18.7, 24, 36, 90 GHz 30 8

buoys. With all these data being available to research within the Commonwealth Scientific and Industrial Re- and application communities there is now a need to search Organisation (CSIRO). The radiometer has a develop methods for getting the best SST possible using heritage going back many years and is the culmination all the satellite and ground-based data that are avail- of developments leading to a reliable accurate instru- able. This is the aim of the Global Ocean Data Assimi- ment. Full details of the instrument are provided by lation Experiment (GODAE) High Resolution SST Pi- Bennett (1998). A rotating 45° plane mirror sequen- lot Project (GHRSST-PP) (www.ghrsst-pp.org). The tially views the sea, a hot blackbody calibration target, available data include a range of satellite orbits and the sky, and finally an ambient temperature blackbody overpass times, several different infrared spectral con- calibration target. The incoming radiation is physically figurations, and microwave-based estimates that can be chopped against a second ambient temperature black- used in cloudy conditions. A further complication is body target, and the chopped radiation is focused with that infrared radiometers measure the radiometric skin a45° parabolic front-surfaced mirror onto a pyroelec- temperature of the ocean, microwave radiometers give tric detector. The detector is located behind an inter- subskin measurements, and ship and buoy measure- ference filter that passes radiation with wavelengths be- ments of SST are typically at a depth of 1–3 m. Thus, tween 10.5 and 11.5 ␮m. The temperatures of the two when developing a blended SST from all available data calibration blackbody targets are accurately monitored, all these factors need to be taken into account. Finally, providing excellent absolute radiometric accuracy. any system developed under the GHRSST-PP program During 2001 the DAR011 radiometer was included in needs to be flexible enough to account for loss of sen- the Miami2001 infrared radiometer calibration and in- sors, the introduction of new sensors, and any develop- tercomparison. The radiometer was calibrated against a ment of new blending techniques. National Institute of Standards and Technology In this paper satellite data from the nine instruments (NIST)-designed blackbody target and compared named above (and listed in Table 1) have been as- against other similar instruments used for the validation sembled for comparison with shipborne measurements of satellite-derived surface temperatures, and was of SST. The data analysis shows the performance of found to perform with a high degree of accuracy— each instrument in tropical clear-sky conditions. The better than 0.1 K. Results from the Miami2001 exercise data also allow the intercomparison of the different sat- are reported by Barton et al. (2004) and Rice et al. ellite-derived SST estimates without reference to any (2004). Since Miami2001, regular comparisons with ground-based data. The analysis techniques developed laboratory-based blackbody targets and other ship- here demonstrate how, in GHRSST-PP, in situ and sat- based measurements suggest that the DAR011 radiom- ellite datasets can be used to monitor the accuracy of both individual and blended SST datasets. eter has maintained its accuracy. A further radiometer intercomparison is planned for the near future to en- sure that shipborne infrared radiometers have main- 2. Ship instrumentation tained their capability of providing skin SST measure- ments suitable for validation. a. DAR011 infrared radiometer On the R/V Southern Surveyor the infrared radiom- The DAR011 radiometer is a single-channel, self- eter was mounted above the bridge and viewed the sea calibrating, infrared radiometer developed and built surface outside the ship’s wake with a nadir angle of

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC OCTOBER 2007 B A R T O N 1775

data used in the analysis were obtained within the same minute. For the radiometer the same coincident time applied, except when the radiometer was in a calibra- tion mode, and then the coincidence was always less than 3 min. The satellite instruments are listed in Table 1 along with details of the number of images supplied and the number of cloud-free coincidences with the ship location. It is important to note that the SST algorithms for the Advanced ATSR (AATSR) and ATSR-2 are derived theoretically and give a radiometric temperature of the sea surface, which is often called the skin temperature. In contrast, the SST algorithms for the other satellite instruments are tuned to give a bulk SST, even though infrared radiometers sense the skin temperature and the microwave radiometers sense the subskin tempera- FIG. 1. Voyage track of the R/V Southern Surveyor during ture at a depth close to 1 mm. voyage SS0403.

a. AVHRR on NOAA-12 and NOAA-16 45°. The radiometer was generally operated for two periods of about 4 h each day throughout the voyage. The AVHRR instruments on two NOAA satellites On some occasions, the radiometer was not operated (NOAA-12 and NOAA-16) provided coincident data due to other operational requirements. Typical operat- with the ship measurements during the voyage. ing times were 0800–1200 and 2000–2400 LT (LT is 10 AVHRR instruments have three infrared channels (3, h ahead of UTC). 4, and 5) with a 1-km nadir spatial resolution and cen- ␮ The radiometer measurements of the sea surface tral wavelengths at 3.7, 10.8, and 12.0 m. The data were corrected for nonunity surface emissivity using the were received at the Australian Centre for Remote sky-view measurements and an emissivity of 0.9875 that Sensing (ACRES) station at Alice Springs and relayed is interpolated from the tables given by Masuda et al. to the CSIRO laboratory in Hobart. For NOAA-12 the (1988). The data were then averaged over 1-min inter- SST was derived using the nonlinear SST (NLSST) al- vals for comparison with the satellite and thermosali- gorithm coefficients (available online at http://noaasis. nograph measurements. noaa.gov/NOAASIS/pubs/SST/noaa12sst.asc). These coefficients are based on December 1993 drifting buoy b. Bulk SST and other ship measurements matchups. For NOAA-16 the NLSST coefficients (available online at http://www2.ncdc.noaa.gov/docs/ The bulk SST on the R/V Southern Surveyor is mea- klm/html/g/app-g3.htm#a101901aa) are based on sured continuously by the ship’s thermosalinograph March 2001 global drifting buoy matches. with the intake being at a depth of 3 m. The The algorithms were used to create fields of SST for thermosalinograph temperature is accurate to within 1 comparison with the ship-based measurements. This mK. A full set of meteorological data are recorded, processing includes an initial cloud-clearing based on including wind velocity, air temperature, relative hu- the method of Saunders and Kriebel (1988). The satel- midity, surface air pressure, and insolation. Ship loca- lite pixel covering the ship location was identified, and tion, heading, and speed are provided by GPS, and the a mean value over 5 ϫ 5 pixels about the ship location voyage track is shown in Fig. 1. was derived. The standard deviation of the SST in this small area was calculated, and if this was greater than 3. Comparisons between ship and satellite SST 0.5 K then the data were assumed to be contaminated estimates by cloud and not used in the analysis. Figure 2 shows the differences between the AVHRR and thermosali- Image data from nine different satellite instruments nograph estimates of SST for the entire AVHRR were obtained for comparison with the ship measure- dataset. Data for the two different instruments are ments of radiometric and bulk SST as supplied by the identified in the figure, and the bias and standard de- DAR011 radiometer and the thermosalinograph, re- viation for the comparisons are, respectively, Ϫ0.41 and spectively. Coincident thermosalinograph and satellite 0.79 K for NOAA-12 AVHRR and 0.07 and 0.51 K for

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC 1776 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24

FIG. 2. Comparisons between AVHRR SST and the ship FIG. 3. Comparisons between MODIS SST and the ship TSG thermosalinograph bulk SST. bulk SST. the instrument on NOAA-16. For NOAA-12 the over- c. ATSR-2 on European pass times were close to 0515 and 1730 LT and for Satellite-2 (ERS-2) and AATSR on Envisat NOAA-16 they were 0230 and 1445 LT. The instrument ATSR-2 and AATSR are satellite instruments that on NOAA-12 does not perform as well as that on were designed specifically for the accurate measure- NOAA-16, which may not be surprising as the former ments of SST. Each instrument has three infrared chan- instrument was launched in 1991 and has survived for a nels that view the earth’s surface at two different view period well beyond its design life. Gaps in Fig. 2, for angles by using an offset conical scan. Full details of this example between AVHRR pass numbers 91 and 95, are design are described by Prata et al. (1990). The instru- due to the removal of cloud-contaminated data. ments include two accurate calibration targets and pro- vide data with 12-bit digitization. SST can be estimated b. MODIS on Terra and Aqua by using different combinations of infrared channels MODIS is a 36-channel instrument flying on both the and view angles, and algorithm coefficients have been Terra and Aqua satellites launched by the National derived using a theoretical radiative transfer model (see Aeronautics and Space Administration (NASA) in Závody et al. 1995). The 2- and 3-channel algorithms December 1999 and March 2002, respectively. The use the nadir view data at 11 and 12 ␮m and 3.7, 11, and MODIS channels cover the spectral bands between 12 ␮m, respectively. The 4- and 6-channel algorithms wavelengths of 400 nm and 15 ␮m, with a nadir spatial use the same channels but with both the nadir and for- resolution of near 1 km. Channels at 3.7, 4.0, 10.8, and ward views and are the algorithms used in this analysis. 12.0 ␮m are used to derive the SST. The 3.7-␮m data are only usable at night due to re- SST was derived using the algorithms supplied by the flected solar contamination during the day. Rosenstiel School of Marine and Atmospheric Science The DAR011 radiometer measurements were cor- at the University of Miami for each instrument, and the rected for nonunity sea surface emissivity and then av- values were then averaged over 5 ϫ 5 pixel areas. The eraged over 1-min intervals. Coincidences between the development of the MODIS SST algorithms, similar to AATSR data and the ship measurements were isolated, that used for the AVHRR Pathfinder Project, is de- and the AATSR brightness temperature fields were scribed by Kilpatrick et al. (2001). Again those cases in then scanned to ensure that the sky in the vicinity of the which the standard deviation of the SST in the 5 ϫ 5 vessel was free from clouds. pixel area was more than 0.5 were assumed to be cloud- After this analysis was completed ATSR-2 and contaminated. Finally, the SST average values were AATSR each gave four occasions with coincident sat- compared with the thermosalinograph measurements, ellite and DAR011 data. For these coincidences the and the results are shown in Fig. 3. The Terra satellite standard European Space Agency (ESA) AATSR and views the Gulf of Carpentaria at close to 1100 and 2330 ATSR-2 SST algorithm coefficients were used with the LT, while the Aqua times are nearer 0200 and 1430 LT. brightness temperatures to produce the two SST esti- For the MODIS on Terra the comparisons give a bias mates for the 4- and 6-channel algorithms. These esti- and standard deviation of Ϫ0.15 and 0.43 K, respec- mates were then averaged over areas of 5 ϫ 5 pixel tively, while for Aqua these values are 0.05 and 0.41 K. windows for each ship–satellite coincidence. The differ-

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC OCTOBER 2007 B A R T O N 1777

overpass time of each ADEOS-II satellite pass was ob- tained (typically these were close to 1100 and 2315 LT), and the ship location at that time was then determined. The bulk temperature was obtained from the thermo- salinograph, and the radiometric temperature of the sea surface was derived as described above. SST was calculated using the official algorithms sup- plied by JAXA. These use channels 34, 35, and 36 dur- ing the daytime, and the same three channels plus chan- nel 30 at night. Further details of the GLI SST algo- rithm development and validation are given by Sakaida et al. (2006). The differences between the thermosalinograph FIG. 4. ATSR-2 and AATSR radiometric SST comparisons with measurements and the 5 ϫ 5 pixel averages of the SST the DAR011 measurements. derived using the daytime and nighttime GLI algo- rithms are shown in Fig. 5. For the daytime data the comparisons gave a bias and standard deviation of ences between the DAR011 measurements and the sat- Ϫ0.01 and 0.19 K, while for the night data these values ellite-derived estimates from ATSR-2 and AATSR are were 0.30 and 0.36 K. When combined into a single shown in Fig. 4. Both AATSR and ATSR-2 have over- dataset these statistics were 0.22 and 0.36 K. Radiomet- pass times of approximately 1030 and 2300 LT. ric data from the DAR011 radiometer are also avail- All SST values are obtained with the 6-channel algo- able but have not been compared with the GLI data, as rithm except for the single daytime coincidence when the JAXA algorithms are designed to provide bulk SST the 4-channel algorithm is used. For the dataset used estimates. Suffice it to say that the difference between here DAR011 data were available for all cloud-free the bulk and radiometric SST (the skin effect) is, as ATSR-2 and AATSR passes. The results in Fig. 4 show anticipated, between 0.0° and 0.3°, with the radiometric that all the estimates for both AATSR and ATSR-2 are temperature being cooler. within the 0.3-K target, and the mean difference is Ϫ0.12 K for both instruments. Insufficient data are e. MSI on GOES-9 available for any further statistical analysis. This situa- tion highlights the major limitation of the ATSR instru- The MSI on the geostationary GOES-9 satellite has ments. With a narrow swath of 500 km and a 35-day three thermal infrared channels at wavelengths of 3.8, repeat cycle, tropical areas may be viewed less than 14 10.7, and 12 ␮m, with a spatial resolution of near 4 km. times in the 35-day period and can have periods of up to These channels are capable of providing an estimate of 8 days without coverage. the sea surface temperature in the same manner as the AVHRR instruments on the NOAA polar-orbiting sat- d. GLI on ADEOS-II ellites. Full details and specifications of the infrared The GLI on the ADEOS-II satellite is a 36-channel radiometer that is similar to NASA’s MODIS instru- ment. The radiometer has four infrared channels that can be used to derive SST. These channels (30, 34, 35, and 36) all have a spatial resolution of 1 km and have central wavelengths of 3.7, 8.6, 10.8, and 12.0 ␮m. The latest version (1.6) of the GLI satellite data was kindly supplied by the Japan Aerospace Exploration Agency (JAXA). For the period between 1 May and 10 June, 112 GLI data files were provided, 22 of which provided clear-sky matchups with the ship measure- ments. Cloudless skies were confirmed by manual scan- ning of the satellite radiance fields and confirmed by a steady DAR011 sky-view radiance measurement when these data were available. All local ship times and dates FIG. 5. Comparisons between GLI SST and the ship TSG bulk were converted to UTC to match the satellite data. The SST.

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC 1778 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24 channels are given by May and Osterman (1998). The GOES-9 satellite was moved from its location over the western United States to 140°E early in May 2003, and the Australian Commonwealth Bureau of Meteorology started to receive data on 15 May. The early data were cloud-affected and sporadic, so the GOES-9 analysis in this work was started on 20 May (day 140). GOES-9 images were received approximately hourly, and 640 images were provided for the period between 15 May and 9 June. The GOES-9 data were supplied as separate files for latitude, longitude, and the two thermal infrared bright- ness temperatures. The GOES SST estimates were cal- culated using the two brightness temperatures and the FIG. 6. Comparison between coincident GOES-9 SST estimates operational algorithm presented by May and Osterman and the ship TSG. (1998). When applying the GOES-9 SST algorithm the brightness temperature differences were averaged over f. AMSR-E on Aqua a5ϫ 5 pixel area about the central value. For com- parison with other satellite and ship measurements, the The AMSR-E is an 8-channel radiometer supplied to SST values were finally averaged over a further 5 ϫ 5 NASA by the National Space Development Agency of pixels. This averaging process was required because the Japan (NASDA), now named JAXA, for inclusion in GOES-9 brightness temperatures are provided as 8-bit the payload of the Aqua satellite. (Fields of SST and data, and significant smoothing is required to reduce other AMSR-E geophysical products are supplied by the digitization effects to an acceptable level. As a re- Remote Sensing Systems online at http:// sult of this process, each final individual SST value is www.remss.com/, where information on algorithm deri- then an average over approximately 0.20° in latitude vation and validation is also available.) Orbital SST and 0.12° in longitude (22 km ϫ 12 km). When the fields are provided daily as well as average values over GOES-9 SST values were plotted against time for dif- 3 days, a week, and a month. The daily orbital data, ferent locations it was obvious that further smoothing which are used in this analysis, are mapped onto a 0.25° of the results with time was required. A running mean (approximately 28 km) latitude–longitude grid. Data of five consecutive values (usually 5 h) was found to that are within 100 km of shore are not included due to give consistent estimates, with a steady variation of SST possible land surface contamination in the 38-km foot- with time instead of the unreal sporadic variations from print of the radiometer. Thus the number of coinci- hour to hour. dences with the ship thermosalinograph is limited to A preliminary assessment of the diurnal variations in eight cases, when the ship was in the north of the study the GOES-9 SST estimates suggested that the data had area and well away from the coast. The differences be- not been corrected for the “midnight blackbody cali- tween the thermosalinograph and AMSR-E estimates bration” effect, as described by Weinreb et al. (1997). of SST are plotted in Fig. 7, and a statistical analysis Because this facet of GOES-9 data analysis had not gave a bias of 0.05 K and a standard deviation of 0.67 K. been implemented for the GOES-9 new location during g. Summary of satellite–ship intercomparisons the observation period, only daytime data from GOES-9 are used in this analysis. The previous sections contain an assessment of the Using daytime SST values, a comparison was then performance of each satellite radiometer in the deter- performed between the GOES SST and coincident mination of the SST in the Gulf of Carpentaria during measurements with the ship’s thermosalinograph. Data the north Australian “dry” season. While this is a lim- were not included in this analysis if the standard devia- ited geographic area and a short time period, the results tion of either GOES-9 11- or 12-␮m brightness tem- give a valuable insight into how each radiometer per- peratures over a 5 ϫ 5 pixel area exceeded 0.5 K. This forms and how it may contribute to a blended dataset simple test was used to exclude cloud-contaminated that is compiled from all satellite and ground-based sys- data and provided 217 coincidences for comparison tems. The bias and standard deviation of each compari- with the thermosalinograph temperatures (see Fig. 6). son between the ship and satellite estimates are given in The comparisons gave a bias and standard deviation of Table 2. The polar-orbiting infrared instruments, ex- 0.51 and 0.50 K, respectively. cept for the long-serving NOAA-12 AVHRR, all pro-

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC OCTOBER 2007 B A R T O N 1779

4. Satellite–satellite intercomparisons The Gulf of Carpentaria dataset, with its assembly of images from nine different satellites, provides the op- portunity of comparing satellite-derived SST values without using any surface-based data. This is a valuable technique to develop, as there are many geographical locations, especially in the Southern Ocean, where re- liable surface-based data are nonexistent. Comparisons between the products provided by different instruments will provide useful information for the development of future data-blending techniques. Cloud-free latitude–longitude areas of 1.5° ϫ 1.5° have been selected from the dataset to compare the FIG. 7. Comparison between coincident AMSR-E SST estimates SST fields from different sensors. The areas were cho- and the ship TSG. sen to be clear of islands and to provide cloud-free data from each of the instruments used in the intercompari- vide SST with accuracies near or better than 0.5 K. The son. The SST estimate for each sensor has been benefit of providing data with 12-bit digitization (in- remapped onto a 0.01° latitude–longitude grid giving an stead of the 10 bits of AVHRR) can be seen in the image of 151 ϫ 151 pixels. For AMSR-E no interpola- improved standard deviations of MODIS and GLI es- tion was applied, and the SST value is constant over timates. The MSI provides data with only 8-bit resolu- each (AMSR-E) pixel of 0.25° ϫ 0.25°. For the other tion, and temporal and spatial averages are required to sensors a nearest-neighbor approach was used to select obtain data with acceptable accuracy. With the advent the SST value on each grid point. In each of the first of new geostationary radiometers on the European Me- three comparisons undertaken here the AVHRR teosat Second Generation (MSG) satellites (www.esa. (NOAA-16) estimates were available and were thus int/msg) and Japan’s Multifunctional Transport Satel- used as the base measurement. In the fourth compari- lite (MTSAT-1R) (www.jma.go.jp/en/gms), the provi- son three images from the GOES-9 SMI were used to sion of infrared data with 10-bit digitization will give a identify possible diurnal warming of the ocean surface marked improvement in SST measurements from geo- in low-wind-speed conditions. Figure 8 shows images of stationary orbit. the SST fields used in each comparison, while Fig. 9 The single microwave radiometer in this analysis pro- shows histograms of the differences between the SST vides SST with a low bias and an acceptable standard estimates at each grid point in the remapped images. deviation. The ability of these radiometers to provide Each case is discussed separately below. SST estimates in cloudy conditions ensures that they will be a valuable asset in the development of near-real- a. Night of 25 May time global datasets—especially in those regions that endure long periods of cloud cover. LATITUDE 14.25°–15.75°S, LONGITUDE 139.50°–141.00°E

TABLE 2. Biases and standard deviations for the clear-sky com- The AVHRR data are for approximately 0300 LT, parisons between the ship and satellite estimates of SST. Note that with AATSR and GLI being four hours earlier. All all comparisons are between the satellite-derived values and those three images show cooler water in the southeast of the from the thermosalinograph except for ATSR-2 and AATSR, where the comparisons are with the radiometric temperatures area. The two difference histograms show a sharp peak, measured by the DAR011 infrared radiometer. indicating low variations in the SST differences. In both comparisons the standard deviation of the differences is Radiometer Satellite Scenes Bias (°C) Std dev (°C) 0.13 K. The AVHRR–AATSR difference peak is close AVHRR NOAA-12 30 Ϫ0.41 0.79 to 0.2 K, which is expected as the AATSR algorithms AVHRR NOAA-16 42 0.07 0.51 are derived to give the radiometric (skin) temperature, Ϫ MODIS Terra 23 0.15 0.43 while the AVHRR algorithm gives the bulk tempera- MODIS Aqua 24 0.05 0.41 ATSR-2 ERS-2 4 Ϫ0.12 n/a ture. The GLI shows a warm bias compared to AATSR Envisat 4 Ϫ0.12 n/a AVHRR, in agreement with the results in Fig. 5 and GLI ADEOS-II 7 0.22 0.36 Table 2. The minor striping in GLI that is aligned close MSI GOES-9 217 0.51 0.50 to lines of latitude is due to scan mirror and detector AMSR-E Aqua 8 0.05 0.67 array effects in the instrument. Some secondary diago-

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC 1780 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24

FIG. 8. Common SST fields from different sensors: (a) night of 25 May, (b) night of 22 May, (c) day of 23 May, and (d) three GOES-9 fields for different times on 23 May. The instrument and time of data collection are included below each image. Note: MYDIS refers to the MODIS instrument on the Aqua satellite.

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC

Fig 8 live 4/C OCTOBER 2007 B A R T O N 1781

ment on Aqua.) The difference histograms show MYDIS to be 0.4 K warmer than AVHRR, and AMSR-E cooler by 0.02 K. These nighttime biases are somewhat different to those given for combined day and night in Table 2. The cause of the AVHRR– MYDIS discrepancy is not known at the moment but may be related to the AVHRR data being near the swath edge, while the MYDIS data are close to 750 km off the subsatellite track, but extra datasets would need to be collected to confirm this possibility. The large area pixels for AMSR-E are obvious in the image. The mirror and detector striping of the MYDIS descending node data are also evident in the image. Further details of the image striping are given by Antonelli et al. (2004). As with the GLI striping discussed above, the color scale is rather stretched and the striping anoma- lies are quite small. c. Day of 23 May

LATITUDE 14.25°–15.75°S, LONGITUDE 139.00°–140.50°E The daytime AVHRR, MYDIS, and GOES-9 data displayed here are all for between 1400 and 1430 LT. All three images show a warm feature in the northeast with a weak cooling trend toward the southwest. The difference histograms show peaks with absolute values of less than 0.1 K. MYDIS is slightly cooler than AVHRR and GOES-9 warmer. The warmer tempera- tures in the northern and eastern extremes of the GOES-9 image are only partly evident in the other in- struments. d. Day of 23 May

LATITUDE 14.25°–15.75°S, LONGITUDE

FIG. 9. Histograms of differences in the SST fields shown in Fig. 139.00°–140.50°E 8. Note: MYDIS refers to the MODIS instrument on the Aqua The final three images are all from daytime measure- satellite. ments with the geostationary GOES-9. The first image at noon local time shows temperatures close to 302.5 K nal striping is thought to be due to electronic noise in over the entire area, with a small section in the north- the instrument electronics. Both striping effects are mi- east warmer than 303.0 K. Two hours later, shown in nor anomalies, and the quality of the GLI SST images the central image, temperatures are lower, nearer 302.0 is very high—especially when averaged over a small K, except in the extreme north and east of the image. array of 1-km pixels. Further details of these phenom- By 1800 LT, shown in the third image again, tempera- ena are given by Kurihara et al. (2004). tures are lower over the entire area. One possible cause for these variations in temperature is the presence of b. Night of 22 May some diurnal solar heating of the near-surface layer during a period of low wind speeds in the morning fol- LATITUDE 13.75°–15.25°S, LONGITUDE lowed by an increase of wind speed during the early 139.50°–141.00°E afternoon that progressed from the south of the area to The AVHRR, MYDIS, and AMSR-E data are all for the north. The middle image thus shows a wind-mixed near 0200 LT. (N.B., here the NASA notation is surface layer except in the extreme north. The last im- adopted with MYDIS referring to the MODIS instru- age suggests that the wind-mixing had occurred over

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC 1782 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24 the entire area by the evening. During this day the R/V provide the most accurate satellite-derived SST in Southern Surveyor was operating some 160 km south of cloud-free conditions. The measurements reported in the southern edge of this area, and the wind measure- this paper, and the other reports mentioned above, ments at the vessel confirmed a low wind speed near support this claim. However, the narrow swath and noon but showed no increase in strength during the 35-day repeat orbit limits their coverage in low lati- afternoon. tudes. These data will best be used to validate (and perhaps to adjust) the more frequent SST fields pro- vided by wide-swath instruments such as AVHRR 5. Discussion and MODIS. This may not be as easy as it first ap- The analysis in this paper has shown that all of the pears, as the AATSR algorithms are theoretically de- nine satellite instruments investigated provide SST with rived to give a surface skin SST, while AVHRR and a reasonable accuracy. The digitization of the radiance MODIS algorithms are derived through a regression signal in the satellite data stream partly limits the ac- analysis of satellite infrared brightness temperatures curacy of the derived SST using standard split-window and in situ bulk SST data. Comparisons between algorithms (see, e.g., Dudhia 1989). GLI and MODIS these two different measurements of SST thus re- data have 12-bit digitization, and comparisons with the quire an estimate of the skin–bulk temperature dif- ship measurements show standard deviations of be- ference. Donlon et al. (2002) provide a simple em- pirical formulation that only applies when wind tween 0.35 and 0.45 K. The number of matchups for the Ϫ1 ATSR instruments in this study is small, but other re- speeds are greater than 6 m s . More complicated ports by Mutlow et al. (1994), Corlett et al. (2006), and formulations, such as that of Fairall et al. (1996), re- O’Carroll et al. (2006) confirm that these sensors pro- quire ancillary data that are not always easily ob- vide SST estimates with a standard deviation of less tained. than 0.3 K when compared with in situ data. The • The validated (adjusted) AVHRR and/or MODIS AVHRR provides 10-bit data, and the standard devia- data can next be used to provide an SST in cloud-free tions found in this analysis are close to the single-pixel regions. The comparisons in Figs. 8 and 9 show that theoretical limit of 0.5 K (Pearce et al. 1989). The 8-bit the AVHRR on NOAA-16 gives excellent measure- data of the MSI on GOES-9 severely limits the useful- ments of SST over a wide swath. In contrast, the in ness of the data in many applications because signifi- situ comparisons for the NOAA-12 AVHRR demon- cant temporal and spatial averaging is required to get strate that sensors can degrade with age and that con- reasonable agreement with surface measurements. This tinuous monitoring, validation, and intercomparisons places a severe limit on the usefulness of current geo- are required to ensure that each sensor is in good stationary observations, and future geostationary in- health. For AVHRR and other wide-swath sensors it struments must have at least 10-bit data streams, pref- is important that composite SST fields over several erably 12 bit. The number of matchups between days be developed to provide useful estimates in AMSR-E and the thermosalinograph in this study is cloudy regions. small, but the derived SST accuracy is close to that • The comparisons between AVHRR and the MODIS reported by Wentz et al. (2003). instrument on Aqua in sections 4b and 4c show a The satellite–satellite intercomparisons show evi- difference between the night and day estimates. The dence of detector (and mirror side) striping when the time difference between the measurements in each images are stretched. If future sensors are developed case was less than an hour, and good agreement using this latest detector technology, then improved is obtained during the day. However, at night the techniques for data analysis are required. This may re- MODIS data are colder by 0.4 K, and there is no quire regional algorithms or a dynamic system for re- obvious reason why this difference is so large. These moving this extra source of error. one-off comparisons show that decisions on how to The results given in this paper suggest a way forward best blend the data should not be made on small in developing steps to provide a blended SST product samples; rather, it is important to collect datasets from the abundance of data that is currently available. over different seasons, , and other conditions to fully understand any differences in the individual • As mentioned above, the ATSR series of instruments datasets. have been specifically designed to provide accurate • For areas in which there is persistent cloud cover, and measurements of SST. With a dual-view capability, infrared estimates are not available for some time 12-bit digitization, accurate onboard calibration tar- (say, greater than days), then microwave measure- gets, and detectors cooled to 80 K, these instruments ments should be used to complete a global dataset.

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC OCTOBER 2007 B A R T O N 1783

• Geostationary data should be used in two ways: first, should have a spatial resolution of 5 km or less and to confirm that anomalies in the infrared and micro- provide data with 10- or 12-bit digitization. wave data are due to diurnal solar heating of the • With their two views, 12-bit digitization and accurate upper ocean layers in low-wind situations and, sec- onboard calibration, the ATSR instruments provide ond, to provide input to short-term (Ͻ6 hourly) skin the most accurate SST measurements from space. SST data products if these are required. The com- There is a concern within the wider SST and climate parisons in section 4d demonstrate the importance of community that currently there are no approved hourly estimates in monitoring the diurnal variations plans for a follow-on instrument in the ATSR series. in SST. Such monitoring will be of great assistance in Data from such an instrument are vital in maintaining using daytime satellite data to provide accurate esti- the absolute accuracy of future blended SST prod- mates of bulk SST under light-wind conditions. ucts. • In situ measurements from ships, mooring, and buoys • Improved detector-stripe removal methods are re- should be continually used to validate the blended quired for instruments that use detector arrays. satellite–SST products as well as assessing the perfor- • Future microwave instruments with better spatial mance of each satellite instrument. The GHRSST-PP resolution would provide valuable data close to will provide many Diagnostic Datasets (DDSs) over coastlines. The current restriction of more than 100 selected geographical areas as well as a Matchup Da- km from land limits the usefulness of this valuable tabase (MDB) of combined satellite and surface data data source. for these purposes. Typically, a DDS covers an area • Further research is required to enable better esti- of 2° latitude by 2° longitude that is centered on a mates of skin–bulk temperature differences and their region of interest (e.g., a mooring, a validation site, dependence on short-term surface wind history and an anomalous SST location). Over this area a DDS insolation. Future geostationary satellites with 12-bit data file is generated for each satellite dataset with infrared sensors should provide valuable data in this SST estimates on a grid of 0.01° in latitude and lon- research. In the meantime, comparisons between gitude. The results from the SST comparisons under- AATSR and AVHRR skin and bulk SST measure- taken in this study demonstrate how the GHRSST ments would be simplified if AATSR bulk SST algo- DDS data files may be used in the development and rithms were developed in the same manner as those validation of future blended global SST datasets. for AVHRR. Of course, theoretical AVHRR skin SST algorithms should also be developed at the same Finally, any scheme that is developed for blending time. data must be sufficiently flexible to account for loss of • Following on from the last point, SST algorithms that sensors and the addition of new instruments. are developed from a regression analysis of bright- ness temperatures with in situ bulk SST measure- 6. Conclusions and recommendations ments should be done only with data collected at wind speeds greater than 5 m sϪ1. Inclusion of diurnal The analysis of the in situ and satellite datasets ob- heating events in the derivation of these regression- tained during a voyage of the R/V Southern Surveyor based algorithms currently provides errors in bulk has provided an assessment of each satellite instru- SST estimates that can be avoided. As a corollary, ment’s estimates of SST in the study region. Compari- such algorithms should only be applied to satellite sons of different satellite SST datasets over small geo- data when the local wind speed is above this thresh- graphic areas demonstrate how the GHRSST-PP DDS old. Therefore, estimates of wind speed from satellite can be used to continually monitor the quality of indi- or other sources (including numerical model fore- vidual and blended SST estimates. The analysis also casts) are a vital ingredient in future satellite-derived gives a guide to future development of blending tech- SST data analyses. niques to assemble a single global SST field using all available satellite and in situ measurements. It also Acknowledgments. Pamela Brodie and Stephen Tho- identifies some shortcomings in the existing framework mas operated the DAR011 radiometer during the R/V that should be addressed to assist with the development Southern Surveyor voyage and have provided an excel- of the GHRSST program. These are listed below. lent dataset. The satellite data have come from a variety of sources: AVHRR from the Australia Centre for Re- • Future geostationary meteorological satellites should mote Sensing (ACRES), AATSR from ESA, ATSR-2 include sensors with split-window infrared channels from the Rutherford Appleton Laboratory (United at wavelengths of 11 and 12 ␮m. The instruments Kingdom), MODIS from NASA, AMSR-E from Re-

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC 1784 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 24 mote Sensing Systems (California), GLI from JAXA, pure and sea for the model sea surface in the infrared and the MTI from the Australian Government Bureau window regions. Remote Sens. Environ., 24, 313–329. of Meteorology. All these agencies and institutes are May, D. A., and W. O. Osterman, 1998: Satellite-derived sea sur- face temperature: Evaluation of GOES-8 and GOES-9 mul- acknowledged for their assistance in data provision. tispectral imager retrieval accuracy. J. Atmos. Oceanic Tech- Ken Suber and Chris Rathbone provided the SST nol., 15, 788–797. analyses from the MODIS and AVHRR data. Part of McMillin, L. M., and D. S. Crosby, 1984: Theory and validation of this work was funded by JAXA through ADEOS-II the multiple window sea surface temperature technique. J. Contract A2GCF003. Geophys. Res., 89, 3655–3661. Mutlow, C. T., A. M. Zavody, I. J. Barton, and D. T. Llewellyn- Jones, 1994: The Along Track Scanning Radiometer (ATSR) REFERENCES on ESA’s ERS-1 satellite—Early results and performance analysis. J. Geophys. Res., 99, 22 575–22 588. Antonelli, P., M. di Bisceglie, R. Episcopo, and C. Galdi, 2004: O’Carroll, A. G., J. G. Watts, L. A. Horrocks, R. W. Saunders, Destriping MODIS data using IFOV overlapping. Proc. Geo- and N. A. Rayner, 2006: Validation of the AATSR Meteo science and Remote Sensing Symp. (IGARSS’04), Anchorage, product sea surface temperature. J. Atmos. Oceanic Technol., AK, IEEE, 4568–4571. 23, 711–726 Barton, I. J., P. J. Minnett, K. A. Maillet, C. J. Donlon, S. J. Hook, Pearce, A. F., A. J. Prata, and C. R. Manning, 1989: Comparisons A. T. Jessup, and T. J. Nightingale, 2004: The Miami2001 in- of NOAA/AVHRR-2 sea surface temperatures with surface frared radiometer calibration and intercomparison. Part II: measurements in coastal waters. Int. J. Remote Sens., 10, 37– Shipboard results. J. Atmos. Oceanic Technol., 21, 268–283. 52. Bennett, J. W., 1998: CSIRO single channel infrared radiometer— Prata, A. J., R. P. Cechet, I. J. Barton, and D. T. Llewellyn-Jones, Model DAR011. CSIRO Atmospheric Research Internal Pa- 1990: The Along Track Scanning Radiometer for ERS-1— per 8, 22 pp. Scan geometry and data simulation. IEEE Trans. Geosci. Re- mote Sens., 3–13. Corlett, G. K., and Coauthors, 2006: The accuracy of SST retriev- 28, als from AATSR: An initial assessment through geophysical Rice, J. P., and Coauthors, 2004: The Miami2001 infrared radiom- validation against in situ radiometers, buoys and other SST eter calibration and intercomparison. Part I: Laboratory characterization of blackbody targets. J. Atmos. Oceanic data sets. Adv. Space Res., 37, 764–769. Technol., 21, 258–267. Donlon, C. J., P. J. Minnett, C. Gentemann, T. J. Nightingale, I. J. Sakaida, F., K. Hosoda, M. Moriyama, H. Murakami, A. Mu- Barton, B. Ward, and M. J. Murray, 2002: Toward improved kaida, and H. Kawamura, 2006: Sea surface temperature ob- validation of satellite sea surface skin temperature measure- servation by Global Imager (GLI)/ADEOS-II—Algorithm ments for climate research. J. Climate, 15, 353–369. and accuracy of the product. J. Oceanogr., 62, 311–319. Dudhia, A., 1989: Noise characteristics of the AVHRR infrared Saunders, R. W., and K. T. Kriebel, 1988: An improved method Int. J. Remote Sens., channels. 10, 637–644. for detecting clear sky and cloudy radiances from AVHRR Fairall, C. W., E. F. Bradley, D. P. Rogers, J. B. Edson, and G. S. data. Int. J. Remote Sens., 9, 123–150. Young, 1996: Bulk parameterization of air–sea fluxes for Weinreb, M. P., M. Jamieson, N. Fulton, Y. Chen, J. X. Johnson, Tropical Ocean–Global Coupled Ocean– C. Smith, J. Bremer, and J. Baucom, 1997: Operational cali- Atmosphere Response Experiment. J. Geophys. Res., 101, bration of GOES-8 and -9 imagers and sounders. Appl. Opt., 3747–3764. 36, 6895–6904. Kilpatrick, K. A., G. P. Podesta, and R. H. Evans, 2001: Overview Wentz, F. J., C. Gentemann, and P. Ashcroft, 2003: On-orbit cali- of the NOAA/NASA Pathfinder algorithm for sea surface bration of AMSR-E and the retrieval of ocean products. Pre- temperature and associated matchup database. J. Geophys. prints, 12th Conf. on Satellite Meteorology and Oceanogra- Res., 106, 9179–9198. phy, Long Beach, CA, Amer. Meteor. Soc., CD-ROM, P5.9. Kurihara, S., H. Murakami, K. Tanaka, T. Hashimoto, I. Závody, A. M., C. T. Mutlow, and D. T. Llewellyn-Jones, 1995: A Asanuma, and J. Inoue, 2004: Calibration and instrument radiative transfer model for sea surface temperature retrieval status of ADEOS-II Global Imager. Proc. SPIE, 5234, 11–19. for the along-track scanning radiometer. J. Geophys. Res., Masuda, K., T. Takashima, and Y. Takayama, 1988: Emissivity of 100, 937–952.

Unauthenticated | Downloaded 09/28/21 09:38 AM UTC