| 69 | | Using the model variability does not guarantee that the estimated uncertainty is realistic. While the sampling provides estimates of the covariances that result from the model dynamics, the variances might not have a realistic amplitude. One can check this by comparing the ensemble-estimated variances (i.e. the diagonal of the ensemble sample error covariance matrix, which is computed in the examples like the Lorenz-96 model) with the deviation of the ensemble mean from the observations. Often, one can plot both quantities like model fields (e.g. if ocean surface temperature is observed). Both quantities will be different, but if they are on the same order of magnitude the uncertainty estimate is usually realistic. If the uncertainty is overestimated, one can reduce the ensemble spread (by multiplying the ensemble perturbations by a factor between 0 and 1). Likewise one can increase the ensemble spread, if it underestimates the difference between the ensemble mean and the observations. |
| | 69 | Using the model variability does not guarantee that the estimated uncertainty is realistic. While the sampling provides estimates of the covariances that result from the model dynamics, the variances might not have a realistic amplitude. One can check this by comparing the ensemble-estimated variances (thus, the diagonal of the ensemble sample error covariance matrix, which is computed in the examples like the Lorenz-96 model. It can also be computed with the routine [wiki:DataAssimilationDiagnostics#PDAF_diag_variance PDAF_diag_variance].) with the deviation of the ensemble mean from the observations. Often, one can plot both quantities like model fields (e.g. if ocean surface temperature is observed). Both quantities will be different, but if they are on the same order of magnitude the uncertainty estimate is usually realistic. If the uncertainty is overestimated, one can reduce the ensemble spread (by multiplying the ensemble perturbations by a factor between 0 and 1). Likewise one can increase the ensemble spread, if it underestimates the difference between the ensemble mean and the observations. |
