start:hype_tutorials:assimilation_intro
Differences
This shows you the differences between two versions of the page.
| Both sides previous revision Previous revision Next revision | Previous revision | ||
|
start:hype_tutorials:assimilation_intro [2023/09/04 15:45] jmusuuza |
start:hype_tutorials:assimilation_intro [2024/01/25 11:37] (current) |
||
|---|---|---|---|
| Line 22: | Line 22: | ||
| The classic ensemble Kalman filter method assumes a linear model, an infinite ensemble size and Gaussian error distribution. These assumptions are usually violated since the use of very many ensemble members quickly becomes a computational bottleneck in large domains or with long simulation runs. The normally-distributed error assumption has also been found to be invalid for physically bounded variables like soil moisture. The lognormal and logit-normal transformations have been successfully applied to semi-bounded e.g., precipitation and bounded e.g., soil moisture, respectively. | The classic ensemble Kalman filter method assumes a linear model, an infinite ensemble size and Gaussian error distribution. These assumptions are usually violated since the use of very many ensemble members quickly becomes a computational bottleneck in large domains or with long simulation runs. The normally-distributed error assumption has also been found to be invalid for physically bounded variables like soil moisture. The lognormal and logit-normal transformations have been successfully applied to semi-bounded e.g., precipitation and bounded e.g., soil moisture, respectively. | ||
| - | At each time step in the assimilation run, an ensemble of random but finite perturbations is introduced in the forcing (precipitation and temperature) and observations (state variables) that results into randomly generated model trajectories. The perturbations in the forcing are spatially correlated fields that help to propagate information to locations where it may be missing, with the spatial correlations themselves controlled by the so-called spatial localization, which suppresses superfluous covariances that can arise e.g., in regions with complex terrain. | + | At each time step in the assimilation run, an ensemble of random but finite perturbations is introduced in the forcing (precipitation and temperature) and observations (state variables) that results into randomly generated model trajectories. The perturbations in the forcing are spatially correlated fields that propagate information to locations where it may be missing, with the spatial correlations themselves controlled by the so-called spatial localization, which suppresses superfluous covariances that can arise e.g., in regions with complex terrain. |
| In the next step, the ensemble of model states and fluxes is propagated through the dynamic model and transformed to the observation space, where the predicted observation and observation error terms mentioned above are obtained. The localization is included by modifying the error covariance matrix terms | In the next step, the ensemble of model states and fluxes is propagated through the dynamic model and transformed to the observation space, where the predicted observation and observation error terms mentioned above are obtained. The localization is included by modifying the error covariance matrix terms | ||
start/hype_tutorials/assimilation_intro.txt ยท Last modified: 2024/01/25 11:37 (external edit)
Except where otherwise noted, content on this wiki is licensed under the following license: CC Attribution-Share Alike 4.0 International
