13 Sea Surface Temperature Measurements from Thermal Infrared Satellite
213
and has led to increased scrutiny, research, development and operational uptake of
TIR data. Over 26 Gb of data are provided in NRT every day by GHRSST Services,
and over 25,500 international users have accessed GHRSST products. This framework needs to be maintained and evolve as new satellite TIR instruments come on
line in the coming decade.
13.3.2 Cloud Flagging of SST Derived from TIR Data
The SST fields obtained TIR sensors are corrupted by clouds, with the temperature of cloud contaminated pixels generally colder than the actual SST. Inclusion
of contaminated pixels in final products renders data inaccurate and difficult to use.
For these reasons, flagging of cloud contaminated pixels in SST fields has received a
great deal of attention over the past 30 years. Despite the effort devoted to such algorithms, significant problems and challenges remain. For applications in which the
absolute accuracy of the retrieved SST values is central to their use, it is important
to exclude any pixel that is even slightly cloud contaminated.
In contrast, applications in which the location of oceanographic features is
important make use of the relative accuracy of adjacent SST values and have
some tolerance to cloud contamination. Most cloud screening algorithms are sensitive to large gradients in the retrieved fields and pixels in a high gradient region
are generally flagged as cloud contaminated. Approaches make use of the structural characteristics of fronts to either reset the quality mask for those pixels that
are believed to be frontal pixels that were falsely flagged as clouds (Cayula and
Cornillon, 1996) or add a new flag. One advantage of this test is that it can be
applied after the SST retrieval and quality fields have been obtained.
Development of cloud screening algorithms has focused on applications for
which the absolute accuracy of the SST value is paramount and typically makes
full use of both visible and TIR data available from the sensor in the day time.
Only the IR channels are available at night further complicating cloud detection.
Algorithms rely on differences in emissivity, reflectivity, temperature and spatial
structure between the ocean surface and clouds. Some work well in identifying
cloud-contaminated pixels under most open ocean conditions. However, because
screening is based on thresholds associated with specific parameters and the underlying distributions are in most cases continuous, there will be ambiguity when one or
more of the parameter values is close to a threshold value. The problem is therefore
intrinsically probabilistic, with a trade-off between false alarms and hits, a balance
that depends critically on the user’s application.
Many SST fields are now provided with a separate “quality” field, which is often
derived from the cloud screening portion of the retrieval algorithm. This field allows
users to mask SST values based on the quality threshold that meets their specific
needs. Quality fields are derived differently by different data providers with different meanings that are not always described in sufficient detail making it difficult
for the user to apply them consistently. This challenge requires careful attention in
the future. In addition to providing quality fields with the SST data, there is a trend
213
and has led to increased scrutiny, research, development and operational uptake of
TIR data. Over 26 Gb of data are provided in NRT every day by GHRSST Services,
and over 25,500 international users have accessed GHRSST products. This framework needs to be maintained and evolve as new satellite TIR instruments come on
line in the coming decade.
13.3.2 Cloud Flagging of SST Derived from TIR Data
The SST fields obtained TIR sensors are corrupted by clouds, with the temperature of cloud contaminated pixels generally colder than the actual SST. Inclusion
of contaminated pixels in final products renders data inaccurate and difficult to use.
For these reasons, flagging of cloud contaminated pixels in SST fields has received a
great deal of attention over the past 30 years. Despite the effort devoted to such algorithms, significant problems and challenges remain. For applications in which the
absolute accuracy of the retrieved SST values is central to their use, it is important
to exclude any pixel that is even slightly cloud contaminated.
In contrast, applications in which the location of oceanographic features is
important make use of the relative accuracy of adjacent SST values and have
some tolerance to cloud contamination. Most cloud screening algorithms are sensitive to large gradients in the retrieved fields and pixels in a high gradient region
are generally flagged as cloud contaminated. Approaches make use of the structural characteristics of fronts to either reset the quality mask for those pixels that
are believed to be frontal pixels that were falsely flagged as clouds (Cayula and
Cornillon, 1996) or add a new flag. One advantage of this test is that it can be
applied after the SST retrieval and quality fields have been obtained.
Development of cloud screening algorithms has focused on applications for
which the absolute accuracy of the SST value is paramount and typically makes
full use of both visible and TIR data available from the sensor in the day time.
Only the IR channels are available at night further complicating cloud detection.
Algorithms rely on differences in emissivity, reflectivity, temperature and spatial
structure between the ocean surface and clouds. Some work well in identifying
cloud-contaminated pixels under most open ocean conditions. However, because
screening is based on thresholds associated with specific parameters and the underlying distributions are in most cases continuous, there will be ambiguity when one or
more of the parameter values is close to a threshold value. The problem is therefore
intrinsically probabilistic, with a trade-off between false alarms and hits, a balance
that depends critically on the user’s application.
Many SST fields are now provided with a separate “quality” field, which is often
derived from the cloud screening portion of the retrieval algorithm. This field allows
users to mask SST values based on the quality threshold that meets their specific
needs. Quality fields are derived differently by different data providers with different meanings that are not always described in sufficient detail making it difficult
for the user to apply them consistently. This challenge requires careful attention in
the future. In addition to providing quality fields with the SST data, there is a trend