38 Ensemble-Based Data Assimilation and Forecasting of Volcanic Ash
247
majority of cells being free of volcanic ash. The probability of detection increases
from 55 to 84% when the average value is replaced by the 98th percentile, at the cost
of the forecast area growing in size by a factor of three. However, even in such a case,
only 1.25% of the total number of cells within the simulation domain are wrongly
predicted to contain an amount of ash exceeding the EPZ threshold.
As the initial estimate of the eruption rate is assumed to have a large uncertainty
and as the transport model used in the assimilation run is the same as in the simulation, the inability of the ensemble to cover all areas is presumably partly due
to the perturbation of the meteorological data being too limited. In principle better
results could be obtained using an actual meteorological forecast ensemble instead
of a perturbed single forecast, but in a real time application such an approach could
be limited by the data transfer capacity.
The simulation results help to verify the functionality of the EnKF despite that
all of its mathematical requirements are not met. For example, the forward model
is generally not linear and the relationship between the eruption rate and the in-air
ash concentration can in principle be very complex. It is worth noting that our test
does not take into account all important sources of uncertainty, such as inaccuracies
and nonlinearities of the observation operator. Additionally, real measurements are
in general not fully capable of distinguishing different types of aerosols from each
other and can also be affected by water clouds.
QUESTIONER: Pieter De Meutter
QUESTION: Which criteria were used to choose the perturbations in the meteorological ensemble, and did you validate the spread in the meteorological ensemble?
ANSWER: The meteorological ensemble was created from a single ECMWF forecast by perturbing only the forecast time. Thus, the ensemble members are valid
meteorological fields from a physical point of view, barring issues that may rise
from the diurnal cycle being misaligned. The spread of the meteorological ensemble
was tested in the actual data assimilation run, which showed that perturbing the forecast time of the ECMWF forecast with a standard deviation of 3 h was not sufficient
to fully cover the actual dispersion of ash acquired with the ERA Interim dataset
(given an ensemble size of 80).
QUESTIONER: Richard Menard
QUESTION: Taking into account that the correlation structures you get with this
problem are non-isotropic and elongated and that you need to used localization
which makes the correlation structures more isotropic, have you considered using a
much large number of ensemble members?
ANSWER: Yes, larger ensembles were considered, but the available storage size
prevented testing those. Since the presentation, more space efficient methods for
generating larger ensembles have been implemented in our transport model (SILAM),
and the effect increasing the size of the ensemble will be studied.
247
majority of cells being free of volcanic ash. The probability of detection increases
from 55 to 84% when the average value is replaced by the 98th percentile, at the cost
of the forecast area growing in size by a factor of three. However, even in such a case,
only 1.25% of the total number of cells within the simulation domain are wrongly
predicted to contain an amount of ash exceeding the EPZ threshold.
As the initial estimate of the eruption rate is assumed to have a large uncertainty
and as the transport model used in the assimilation run is the same as in the simulation, the inability of the ensemble to cover all areas is presumably partly due
to the perturbation of the meteorological data being too limited. In principle better
results could be obtained using an actual meteorological forecast ensemble instead
of a perturbed single forecast, but in a real time application such an approach could
be limited by the data transfer capacity.
The simulation results help to verify the functionality of the EnKF despite that
all of its mathematical requirements are not met. For example, the forward model
is generally not linear and the relationship between the eruption rate and the in-air
ash concentration can in principle be very complex. It is worth noting that our test
does not take into account all important sources of uncertainty, such as inaccuracies
and nonlinearities of the observation operator. Additionally, real measurements are
in general not fully capable of distinguishing different types of aerosols from each
other and can also be affected by water clouds.
QUESTIONER: Pieter De Meutter
QUESTION: Which criteria were used to choose the perturbations in the meteorological ensemble, and did you validate the spread in the meteorological ensemble?
ANSWER: The meteorological ensemble was created from a single ECMWF forecast by perturbing only the forecast time. Thus, the ensemble members are valid
meteorological fields from a physical point of view, barring issues that may rise
from the diurnal cycle being misaligned. The spread of the meteorological ensemble
was tested in the actual data assimilation run, which showed that perturbing the forecast time of the ECMWF forecast with a standard deviation of 3 h was not sufficient
to fully cover the actual dispersion of ash acquired with the ERA Interim dataset
(given an ensemble size of 80).
QUESTIONER: Richard Menard
QUESTION: Taking into account that the correlation structures you get with this
problem are non-isotropic and elongated and that you need to used localization
which makes the correlation structures more isotropic, have you considered using a
much large number of ensemble members?
ANSWER: Yes, larger ensembles were considered, but the available storage size
prevented testing those. Since the presentation, more space efficient methods for
generating larger ensembles have been implemented in our transport model (SILAM),
and the effect increasing the size of the ensemble will be studied.