The ensemble dataset was generated by first perturbing all of the input values (recorded counts, temperatures etc.) as well as model values (calibration coefficients and ICT temperature corrections) using their individual uncertainties as the standard deviation used in the random generation process. For single error sources which are uncorrelated a random generation scheme was used which samples the Gaussian distribution in 10 equal area segments and enables a small number of samples (in this case 10) to give a mean and variance that is closer to the expected values of zero and one than if a purely Gaussian random number generator was used. In the case of those parameters where a covariance matrix exists (the calibration parameters themselves) this was also sampled so as to capture the correct covariance with a small number of samples. The actual values of the Harmonisation/calibration parameters used for the infrared channels were generated in a two stage process. First, estimates of the non-linearity and instrument temperature terms were determined using SNO matchups using the AATSR as a reference. Then in a separate process small corrections to the bias and emissivity correction terms were found using sea surface temperature (SST) retrievals compared to the drifting buoy network to ensure that the Ensemble SST would be accurate. This was needed because the SNO matchups were concentrated at cloud like temperatures (240K-250K) and so had not sampled many SST like temperatures (>270K) leading to a small error in the Harmonised calibration for SST retrievals. When the new FCDR values are combined with the perturbed input values from the Ensemble files it is possible to generate a new set of perturbed Radiances/Brightness Temperatures which capture the underlying sources of uncertainty. These were then used as input into the AVHRR ESA CCI SST/Lake Water Surface Temperature (LWST) algorithms which also included a perturbation to the cloud detection threshold to derive a new baseline SST/LWST data together with perturbed SST/LWST values to create the SST/LWST Ensemble dataset.

Relationship between the SST CDR and the ensemble
The FCDR ensemble is used to generate an SST ensemble product where the complete SST retrieval process is applied to each of the ensemble members directly generating 10 SST perturbations to the baseline SST. Also included in the SST ensemble is a variation in the cloud detection threshold which adds extra uncertainty to the ensemble in an attempt to capture uncertainties in the cloud detection process.