Single Sensor Error Statistics (SSES)
The GHRSST-PP is dedicated to Quality control and uncertainty estimation. Single Sensor Error Statistics are assigned to every L2 satellite pixel in the form of a bias error and standard deviation. there are a variety of methods for defining SSES depending on the specific satellite instrument.
GHRSST-PP L2P processing centres take existing SST data products as provided by data producers and enhance them by adding additional information and reformatting so that they can be combined to produce the new generation of products using a strategy that is scientifically sound and technically feasible. The GHRSST-PP has achieved considerable success not just by solving scientific and technical problems, but also by co-operation at an international level to agree on data product definitions and standards acceptable to users and producers. Most of all it has required collaboration between organisations and agencies that produce competing satellite SST data products from different sensor types. Seeking to include all the major players, GHRSST-PP developed a system which exploits the unique contributions of each sensor type and institutional partner.
Uncertainty estimates for each SST observation or analysis grid point is one of the key user requirements for GHRSST-PP SST data products. Uncertainty estimates allow users to select the accuracy level suitable for their application and to make optimum use of the SST observations (e.g in data assimilation). Estimating uncertainties associated with a particular satellite observation is a challenging task especially considering the vast amount of observations, the highly variable characteristics of the atmosphere (affecting the atmospheric correction algorithms applied to satellite observations) and temporal stability of the satellite instruments themselves. Nevertheless, techniques are emerging that consider statistically the uncertainties associated with each observation in each data stream separately as a Single Sensor Error Statistic (SSES). SSES are based on understanding errors associated with a specific satellite instrument and errors associated with the geophysical retrieval of SST for each individual satellite scene. The simplest L2P SST uncertainty estimation is based on matching satellite SST with in situ observations co-located in space and time to within 25 km and 6 hours. A large match-up database of data is required for each satellite instrument which is then periodically analyzed to derive a mean bias and standard deviation for each satellite system.
A more useful spatially and temporally varying uncertainty estimate can be made by analysing the match-up database as a function of sensor specific criteria known to cause errors. In one approach, a proximity confidence value (on a scale of 1-5) is defined based on the most likely source of error for a given satellite instrument. In the generic case of infrared satellite SST retrievals, the most likely source of error is cloud contamination (SST measurements using infrared techniques are not possible when cloud is present), amplification of noise at extreme satellite zenith angles, deviation of SST from climatology and, other specific channel differences. In the case of microwave SST retrievals proximity confidence errors can be defined using knowledge of side lobe contamination in the coastal zone, radio frequency interference and rainfall flagging (rainfall prohibits retrieval of SST) and errors in the estimation of surface emissivity. These criteria are used to classify every pixel within a satellite scene according to the proximity confidence scale. Using the match-up database, a bias and standard deviation (SD) uncertainty estimate is derived for each satellite sensor data set proximity confidence value on a regular basis. Based on the relationship between proximity confidence value, bias and standard deviation, SSES uncertainty estimates can then be assigned to all pixels in a given scene. The proximity confidence map generated for the MSG-SEVIRI SST data set for the period 03/28/04 to 09/02/2004 is shown in the figure. Clearly seen in this example are excellent confidence values in cloud free regions and degraded confidence values in the proximity of clouds. Shown to the right are the SSES bias and SD estimates for each confidence value (1-6 corresponding to cloudy through to excellent in the proximity confidence map legend) for the SEVIRI over a 7 month period showing how higher confidence values are associated with lower bias and reduced SD.
While the SSES process is not able to account for all errors, it provides a method that is functional in a real time environment, caters to the most obvious satellite specific errors and should be better than simply taking the latest published figures from sparsely available in situ validation studies. Furthermore, it is expected that as more experience is gained with the SSES process, better error estimates will be generated based on detailed knowledge of each data set. For example, a new 4-dimensional lookup table of errors dependent on a combination of satellite instrument and geophysical fields (called a SSES hypercube) offers a promising approach for the MODIS sensor and considerable research for new approaches for microwave SST are in progress.
Single Sensor Error Statistics (SSES)
Single Sensor Error Statistics (SSES) are a key component of all GHRSST-PP L2P data files. SSES attempt to capture the time and space varying error characteristics of a specific satellite SST data stream using sensor specific information. SSES are complex and require continuous update based on the mose recent uncertainty estimation. Typically SSES are different for infrared and microwave SST retrievals e.g., for microwave SST retrievals side lobe contamination and rainfall are known as soources of error, but for infrared retrievals clouds and infrared-specific aerosols are important error sources.
The GHRST-PP has agreed a common 6-point scale of SSES Quality values which can be used to filter SST observations. The table below summarises the GHRST-PP SSES Quality Value scale.
QV -- Meaning
0 -- Unprocessed/ (No Data/Land etc)
1 -- Not Useable: Cloudy/Rain/Side lobe contamination
2 -- Bad: Data are considered suitable for qualitative applications only
3 -- Suspect: Follow data provider reccomendations if using these data
4 -- Acceptable: These data are suitable for quantitative applications
5 -- Excellent: These data are the best data availablefor quantitative applications
Documentation
Several documents describing SSES develpments for GDS-v2.0 are in development. contact the SSES-WG coordinator for mor informtion
(Last Updated: 28-01-2010)