13 Sea Surface Temperature Measurements from Thermal Infrared Satellite
215
Statistics (SSES) designed to take into account uncertainties for specific instrument/platforms (Donlon et al., 2007). Bias and uncertainty estimates are generally
derived from near contemporaneous match ups between satellite and in-situ SST
measurements which are periodically analysed to provide SSES. The EUMETSAT
OSI-SAF has developed a statistical method to derive SSES bias and standard deviation estimates by associating a confidence level assigned to the retrieved SST
estimate. The confidence levels are based on tests to the reliability of the cloud mask
and the SST algorithm conditions. Regional (and seasonal) characteristics need to
be accounted for in this scheme although it is successfully used in operations. 1
An alternative approach called the Hypercube has also been developed based
on a match-up data base for the Aqua and Terra MODIS sensors. In this case, the
MDB includes near-contemporaneous, co-located satellite brightness temperatures,
in-situ buoy and radiometer SST, auxiliary data from model or satellite observed
fields, and the satellite viewing geometry. A series of quality tests is applied during processing of the MODIS data to identify cloud and dust aerosol contaminated
retrievals and assign pixels to one of four different quality levels with quality 0 being
the best quality possible. The relative immunity of the MODIS 3.95 and 4.05 μm
bands to both water vapour and aerosols as compared to the increased sensitivity
to both in the MODIS 11 and 12 μm bands is used to identify aerosol data. After
eliminating records with quality levels greater than 1, each match-up database is
partitioned into a multi-dimensional array with the following 7 dimensions: time by
season (4 values), latitude bands (5 steps in 20 ◦ increments from 60 ◦ S to 60 ◦ N),
surface temperature (8 increments in 5 ◦ steps), satellite zenith angle (4 increments),
brightness temperature difference as a proxy for water vapour (4 intervals for 4 μm
and 3 intervals for 11–12 μm SST), retrieved satellite SST quality level (2 intervals) and day/night selection (2 intervals). The bias (satellite-in-situ) and standard
deviation are then computed for each element to define a hypercube look up table
(LUT). The LUT is then used during satellite data processing to predict the SSES
bias and standard deviation of the SST retrieval. The hypercube approach provides
more control over the specification of uncertainty estimates and is being actively
developed within the framework of GHRSST.
Finally, it is important to recognize that more work is required to ensure that
uncertainty values, where possible, are traceable to accepted international reference
standards and SI units. Satellite TIR instruments and ground truth instrumentation
should also be traceable to the same reference standards. More effort is required in
this area.
13.3.5 New SST Retrieval Techniques Using TIR Data
A single-view TIR imager with channels at roughly 3.7, 11 and 12 μm can demonstrate global 1 km accuracy approaching 0.3 K at night-time (i.e., when all three
1 See http://www.osi-saf.org
215
Statistics (SSES) designed to take into account uncertainties for specific instrument/platforms (Donlon et al., 2007). Bias and uncertainty estimates are generally
derived from near contemporaneous match ups between satellite and in-situ SST
measurements which are periodically analysed to provide SSES. The EUMETSAT
OSI-SAF has developed a statistical method to derive SSES bias and standard deviation estimates by associating a confidence level assigned to the retrieved SST
estimate. The confidence levels are based on tests to the reliability of the cloud mask
and the SST algorithm conditions. Regional (and seasonal) characteristics need to
be accounted for in this scheme although it is successfully used in operations. 1
An alternative approach called the Hypercube has also been developed based
on a match-up data base for the Aqua and Terra MODIS sensors. In this case, the
MDB includes near-contemporaneous, co-located satellite brightness temperatures,
in-situ buoy and radiometer SST, auxiliary data from model or satellite observed
fields, and the satellite viewing geometry. A series of quality tests is applied during processing of the MODIS data to identify cloud and dust aerosol contaminated
retrievals and assign pixels to one of four different quality levels with quality 0 being
the best quality possible. The relative immunity of the MODIS 3.95 and 4.05 μm
bands to both water vapour and aerosols as compared to the increased sensitivity
to both in the MODIS 11 and 12 μm bands is used to identify aerosol data. After
eliminating records with quality levels greater than 1, each match-up database is
partitioned into a multi-dimensional array with the following 7 dimensions: time by
season (4 values), latitude bands (5 steps in 20 ◦ increments from 60 ◦ S to 60 ◦ N),
surface temperature (8 increments in 5 ◦ steps), satellite zenith angle (4 increments),
brightness temperature difference as a proxy for water vapour (4 intervals for 4 μm
and 3 intervals for 11–12 μm SST), retrieved satellite SST quality level (2 intervals) and day/night selection (2 intervals). The bias (satellite-in-situ) and standard
deviation are then computed for each element to define a hypercube look up table
(LUT). The LUT is then used during satellite data processing to predict the SSES
bias and standard deviation of the SST retrieval. The hypercube approach provides
more control over the specification of uncertainty estimates and is being actively
developed within the framework of GHRSST.
Finally, it is important to recognize that more work is required to ensure that
uncertainty values, where possible, are traceable to accepted international reference
standards and SI units. Satellite TIR instruments and ground truth instrumentation
should also be traceable to the same reference standards. More effort is required in
this area.
13.3.5 New SST Retrieval Techniques Using TIR Data
A single-view TIR imager with channels at roughly 3.7, 11 and 12 μm can demonstrate global 1 km accuracy approaching 0.3 K at night-time (i.e., when all three
1 See http://www.osi-saf.org