Note that we perform the satellite retrievals (themselves being a type of inversion) and the source inversion separately, rather than choosing the emissions that, using forward modelling of the atmospheric dispersion and the radiative transfer, fit the satellite measured radiances. As far as we are aware, this two-step method is in line with all other approaches to volcanic ash inversion. The method has logistical and computational-speed advantages, with the two steps being separately maintained and developed, and with the ability to take advantage of the linearity of the dispersion in the source inversion. The approach is also assisted by the nature of the satellite retrieval scheme used. This seeks to give results only where these are robust and does not seek to extract low-grade information from the satellite observations. It seems possible that a one-step approach might be more effective at making use of low-grade information at locations where currently the satellite retrievals give no information, but we do not explore that here.
2.1. The Satellite Retrievals
We use data from the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) on the Meteosat Second Generation geostationary satellite. At the latitude of Iceland (∼ N) the pixel size for the infrared channels used for ash detection and retrieval is of the order of 10 km. Scans are completed every 15 min.
Ash properties are estimated from SEVIRI data using the approach described in [
32]. The approach first applies a detection algorithm to estimate which SEVIRI pixels contain detectable ash. The retrieval then applies an optimal estimation technique (a one-dimensional variational analysis) to each of these “ash pixels” which estimates the ash column load (
L), the effective ash particle radius (
) and the atmospheric pressure at the ash layer (
), as well as the size of the errors in these quantities. The approach assumes an infinitesimally thin ash layer, which is why only a single pressure value is retrieved. If the layer has significant depth then the retrieved pressure is likely to be between the layer centre and the layer top, depending on the optical thickness of the ash. Note our inversion method does not currently use the information on
and
.
Clear sky pixels are also diagnosed. A satellite pixel is classified as clear sky and assigned a column load of zero if the detection algorithm fails to detect ash and no cloud (which might obscure any ash) is identified based on the scheme described by Saunders and Kriebel [
37]. The minimum detectable column load depends on the height of the ash and the atmospheric temperature profile. A conservative estimate of the minimum detectable level for the approach [
32] is 0.5 g m
. As a result, it is possible that pixels with less than 0.5 g m
of ash could be labelled as clear sky pixels. Hence, we assign an error estimate of 0.5 g m
to the column loads in the clear sky pixels.
Finally, those pixels which are not classified as either ash or clear sky pixels remain unclassified. These pixels might contain ash hidden by cloud or might not.
The ash column loads and associated errors are then regridded to a resolution of 0.5625 longitude and 0.3750 latitude (approximately 40 km) and hourly averages are computed. The purpose of regridding is both to reduce the data volume and to match the output resolution chosen for the NAME dispersion simulations. It also helps with the treatment of correlations in the satellite data. For Icelandic volcanic eruptions the domain of the regridded satellite data extends from 50 W to 34 E and from 35 N to 80 N. The column load for a 40 km grid cell is calculated by averaging the column loads for the ash and clear sky pixels in the cell (assuming zero for the clear sky pixels). The error estimate for the cell column load is evaluated by averaging the analysis errors for the ash and clear sky pixels. This method of combining the error estimates (interpreted as standard deviations) amounts in effect to assuming a high correlation between the errors in the pixels within the cell. It is important to assume some correlation between the pixels to avoid greatly underestimating the error for the cell as a whole due to the tendency of uncorrelated errors to cancel.
There remains the problem of representativity errors, which arise when the ash and clear sky pixels are not representative of all pixels. We treat this by only using the 40 km cells for which at least 50% of the pixels are ash pixels or for which 90% of the pixels are either ash or clear sky pixels. This also has the advantage of reducing the total number of 40 km cells used in the inversion and thus reduces the computational cost.
We then classify these 40 km grid cells as ash or clear sky cells according to whether the estimated column load is greater than zero or equal to zero. The purpose of this is to ensure that we can compare the effect of using only the ash cells (i.e., excluding the zero values) or using both the ash cells and the clear sky cells (i.e., including the zero values). For the ash cells detected during the 2010 Eyjafjallajökull and 2011 Grímsvötn eruptions, the average column load from the satellite data is 1.9 g m and the average error estimate is 6.1 g m. Error estimates tend to be smaller in the high column load parts of the ash cloud.
For the inversion used to estimate the emissions, we will need to assume something about the correlation between the errors in the different cells. Following Stohl et al. [
14], we take these as zero. This has the advantage of simplicity and reflects the idea that correlation decays with separation. While the sharp cut-off in the correlation (i.e., the high correlation assumed within the cells as discussed above and the zero correlation between cells) is not ideal, this is a pragmatic choice. This is similar conceptually to the approach that is sometimes used in data assimilation for numerical weather prediction, where satellite retrievals are ‘thinned’ (on a scale typically in the range 10–100 km) and zero correlation is assumed for the errors in the remaining retrievals [
38]. If in the future we wish to exploit higher-resolution satellite information but avoid the satellite data, by virtue of the large quantity of data, having too much weight in the inversion relative to the a priori distribution, then a more sophisticated treatment of correlations would be needed.
2.2. The NAME Model
InTEM requires information on how erupted material is transported and dispersed from the volcano. This is provided by NAME, a Lagrangian stochastic atmospheric dispersion model. In NAME, model particles trajectories are calculated using winds from a numerical weather prediction (NWP) model. A stochastic component is added to each trajectory to account for atmospheric turbulence. Model particles carry a mass of the dispersing material (in our case volcanic ash) calculated from the mass emission rate. Each model particle also has a sedimentation velocity which is calculated from the diameter and density of the particles it represents. A density of 2300 kg m
is assumed for volcanic ash [
10,
39]. NAME also models wet deposition and non-sedimentation-driven dry deposition, such as that caused by diffusion to the surface [
10].
We do not attempt to model the fraction of ash that, as a result of large grain sizes or the aggregation of smaller grains, falls out close to the volcano. Instead, NAME uses an effective source strength corresponding to the emission rate of the fine ash particles that undergo long-range transport [
10]. NAME does not model aggregation.
In modelling particulates, each NAME model particle represents numerous real particles of a particular diameter. The particle diameters modelled are described by a particle size distribution. Here we use a particle size distribution based on data provided by Hobbs et al. [
40]. The range of diameters in this size distribution is 0.1
m to 100
m and is not intended to represent all the ash particles from the volcano but just the fine ash particles that undergo long-range transport.
NAME can use NWP meteorological data from a range of NWP models. Here we use meteorological data from the Met Office’s Unified Model (MetUM) in its global configuration. This configuration has a horizontal resolution of about 25 km in mid latitudes and a vertical grid with 70 unevenly spaced levels with a spacing of 300–400 m in the mid-troposphere. We use data obtained from the model at three hourly intervals.
Because NAME’s ash concentration is linear in the emissions, the concentrations for a linear combination of sources can be obtained from those for the individual sources by combining the predicted concentrations. Here we use sources at different times and heights above the volcano with each source having a release rate of 1 g s, and we separately store the results due to each source. InTEM then attempts to select the best linear combination (defined using the Bayesian formalism described below) of these sources (the ‘source profile’). We note that, although the satellite data used in the inversion do not provide information on the vertical distribution of the ash, we expect the combination of the dispersion model and the satellite data to be able to constrain the height variation of the emissions because of the changes with height of the wind, and of the wind direction especially. The accuracy of the height distribution of the emissions (or indeed other aspects of the emissions) will, however, be affected by errors in the wind data. The individual sources are defined on a time–height grid with particle release locations for a given grid cell evenly distributed across the cell.
To compare with the one-hour averaged column loads estimated from the satellite data, the ash concentrations from NAME are averaged over 1 h and vertically integrated. The horizontal grid used for the NAME output is chosen to coincide with that of the regridded satellite retrievals.
The particle size distribution used in NAME covers diameters between 0.1
m and 100
m but the diameters detectable by the SEVIRI satellite are limited. The satellite retrieval assumes a log-normal size distribution which can extend beyond the size range to which the satellite is sensitive and provides an estimate of the ash mass in a diameter range of approximately 1
m–30
m [
14,
32,
41].
Figure 1 shows the distribution used in NAME and where this overlaps with the range that the satellite ‘sees’. About 95% of the mass in the NAME emissions is detectable by the satellite retrievals. To allow for this, we output from NAME only the part of the column load due to the particles within the detectable range of 1
m–30
m.
For computational efficiency, particles are not followed beyond six days after release. It is expected that, by this time, most of the ash will have been removed from the atmosphere or widely dispersed [
10]. Hence, we expect column loads due to particles older than six days to be small, at least for eruptions comparable in size to the 2010 Eyjafjallajökull and 2011 Grímsvötn eruptions. The six-day limit could of course be changed if required.
2.3. The Prior Source Distribution
As we use a Bayesian approach, we need to make assumptions about the probability distribution of the emissions, prior to using the satellite data to refine our assumptions. For this, we make use of the observed plume rise height. The prior assumptions should help regularise the problem, avoid overfitting to the satellite data, and ensure that the observations of plume rise are taken into account. If we just look for a best fit to the satellite data, then the resulting source profiles may have emissions at heights which are too high to be consistent with the observed plume rise height and its uncertainty. While we expect the variation of the wind (and especially the wind direction) with height to enable emission heights to be inferred from satellite data, this may not always be possible, and so it is important to try to make optimal use of all available information. Additionally, sometimes there may be little satellite data available, making the use of the a priori source description even more crucial.
In principle, the observations of plume rise could be regarded as more observations to be used in the inversion. The a priori description would then need to reflect the range of possible volcano behaviours prior to any observations of plume rise. With a wide range of observation types, this might have advantages. However, with the main sources of data being just the plume rise height and satellite data, it is easiest to treat these separately and include the former in the a priori description.
Our starting point in defining the a priori description is the default source specification that is currently used by the London VAAC [
10,
42]. This expresses the mass eruption rate in terms of the plume rise height using
where
M is the mass eruption rate in kg s
and
H is the plume rise height above the volcano vent in km. This expression is based on an empirical power law best fit to many volcanic eruptions presented by Mastin et al. [
7]. The emission profile is taken as uniform between the volcano summit and the plume rise height. While the true effective source profile is more likely to be weighted towards the top of the eruption column, this choice is made by the London VAAC to avoid the risk of predicting no ash where there is a significant risk [
10]. Since we are only interested in the fine ash that undergoes long-range transport, the Mastin emission formulae is multiplied by a constant
representing the distal fine ash fraction, i.e., the mass fraction of the ash that survives the near source fall out of particles larger than 100
m and the near source fall out of smaller particles due to aggregation or due to wet deposition associated with the eruption. The effective mass emission rate
m of fine ash which reaches the distal ash cloud is then given by
. The factor
is taken to be
[
10,
42].
For the inversion calculation, we need a probabilistic description which describes the uncertainties in the emissions. It is not sufficient to simply take the London VAAC emission profile as the mean emission and add uncertainty information to that because, for example, this gives zero mean emissions above the observed rise height and, if there is a risk that the rise height observation is an underestimate, the mean emission should be greater than zero at these heights. Moreover, it is desirable to include estimates of the correlations between the emissions in the various time–height grid cells. This is because, for example, if the rise height is higher than estimated or the mass eruption rate is greater than that given by the Mastin relationship, this is expected to have an impact on all the emissions. If the correlations are ignored in the probabilistic description, then the implied variance of the overall emission will be too small for a given variance in each grid cell due to the tendency of the fluctuations from the mean in the individual cells to cancel. Either the a priori description will overconstrain the overall emission or it will allow too much uncertainty in individual grid cells leading to very noisy emission profiles. Neglecting correlations would also mean that the uncertainty in the shape of the emission profile will be intrinsically linked to the uncertainty in the total emission which is undesirable. Including correlations should also result in an approach that gives results independent of the resolution of the height–time grid as the resolution increases.
In order to estimate the means, variances and covariances of the emissions in the various height–time grid cells, we propose a stochastic model of the source emission profile. In principle, this could be used to generate many realisations of the profile from which the mean and variance of the emission in each cell and the covariance between the emissions in each pair of cells can be calculated. However, it turns out that the model is susceptible to analytic study and (provided and , defined below, are piecewise constant) we can calculate the required quantities with the aid of some numerical integrals, with the benefit of reduced computational cost and a more accurate calculation.
We assume that the plume rise height
has mean
, our best estimate of the rise height as assessed from the local observations, and varies in the range
on a time scale
, where
is an estimate of the error in the rise height. We choose hard limits on
H to avoid any potential difficulties with negative
H such as might occur with Gaussian processes and, to make this easy to model, we treat the time dependence as involving random jumps to new values which are uniformly distributed in
, with exponentially distributed waiting times between jumps and mean waiting time
. Between jumps,
remains constant. For a given rise height
H we assume an effective mean emission rate per unit height (in kg s
km
for
H in km) given by
where we have based this on (
1) but have linearised about
to keep the mean total emission per unit height in the bulk of the column (up to
) equal to the default value used by the London VAAC. The emission rate per unit height in a single realisation is then modelled by multiplying this by two random factors. The first factor,
, which is independent of height, accounts for the variation in the mass emission rate in time which is not accounted for by (
2). We take
r to be exponentially correlated in time with mean zero, standard deviation
, and time scale
. The second factor,
, accounts for variations in the shape of the emission profile but does not alter the total mass emitted.
s is taken to be
where
q has standard deviation
and is exponentially correlated in
t and
with correlation time scale
and (non-dimensional) correlation height scale
. The relationship between
s and
q is defined so that
s has zero vertical integral and hence does not alter the total mass emitted. We also assume
q decorrelates completely in time whenever there is a jump in
as this is interpreted as reflecting a significant change in volcano behaviour. Unlike the time series of
H, there is no need to consider whether
r and
q are Gaussian processes or not because this does not affect the first and second moments of the emissions.
Details of the method used to obtain values for the means, variances and covariances of the emission rates integrated over each grid cell using only numerical integrals and without actually simulating an ensemble of realisations of the stochastic model are given in [
27]. The results obtained using this analytic method were compared with results using an ensemble of realisations of the emission profile and were found to agree within the sampling error expected for a finite number of realisations. The means, variances and covariances are expressed using units of g s
for the emission rates for consistency with the nominal release rate of 1 g s
used in the NAME runs.
Of course there are many aspects of this stochastic model that are greatly simplified compared to reality and there is a lot of uncertainty in the values to assign to the model parameters (discussed below). However we believe that including correlations in this way has the potential to give significant benefits in achieving a better balance between the a priori constraint on the total emission and on the individual cell emissions, as well as in decoupling the uncertainty in the emission profile shape from the uncertainty in the total emission. Parametrizing the various components of the uncertainty separately also makes clear how the assumptions relate to the physical situation and makes explicit where things could be improved. This should enable the improvement of the approach through further studies (e.g., detailed simulations or observations of the eruption column for a range of conditions or experience in applying the approach to a wide range of eruptions). The inclusion of the correlations between the cell emissions should also help to achieve greater benefit when there are few satellite observations, by allowing the observations to inform the emissions from cells that are not directly linked (via the NAME model) to the observations as well as informing the emissions from the cells that are so linked. In contrast, when there are substantial quantities of satellite data, we expect results to be insensitive to the details of the a priori emission description. There might be benefit in assuming an a priori distribution for the uncertain parameters and attempting to refine this as part of the inversion. However, we have not attempted that here.
Our approach can be regarded as lying somewhere between an approach that neglects the correlations in the a priori emissions (e.g., [
14]) and an approach that uses high correlations or (similarly but not quite equivalently) that uses an emission model with a small number of degrees of freedom (e.g., [
17,
19,
20,
21]). In designing the a priori model we have adopted a conservative choice in staying close to the default London VAAC emission specification, but there may be value in moving away from this in future (e.g., using plume rise models or choosing a non-uniform vertical distribution for the mean emissions).
We now discuss the choice of parameters for the stochastic model. Many volcanoes, including those in Iceland, are observed by various methods such as radar and webcams, and these observations can be used to estimate the plume rise [
3,
43]. These estimates are subject to an uncertainty (e.g., resulting from the radar’s discrete scanning elevation angles) which can itself be estimated. Here we assume values of
and
can be estimated as a function of time. We take the time scale
to be 12 h, similar to the observed rate of change of plume height in the Eyjafjallajökull 2010 eruption, using the idea that the plume height and the errors in the plume height are likely to evolve on comparable time scales. For very short eruptions this value will be too large, but the value is then less important than for sustained eruptions like Eyjafjallajökull.
Mastin et al. [
7] indicate that in 50% of cases the mass emission rate falls within a factor of four either side of (
1). However, it is likely that this error estimate is at least partly due to errors in the measured rise height which we are treating separately via
, and so this error estimate may be too large for use in estimating
in our model. On the other hand, the error in the effective mass emission rate is likely to be increased by the uncertainty in the distal fine ash fraction. On balance, we judge that
is a reasonable estimate. This allows emissions of several times the value calculated from Equation (
2) at
and also allows very small values, reflecting the large uncertainties.
is also taken to be 12 h, based on the idea that the relationship between mass emission rate and the plume height is likely to evolve on a similar time scale to the plume height due to changes in volcano behaviour, or on a time scale related to changes in meteorology which influence the relationship between the emission rate and rise height [
12,
13].
To allow for substantial departures from the uniform height profile of the emission we take and we take hours and to allow substantial variations in the vertical profile shape on time scales that are a little faster than the total emission time scale and on height scales significantly smaller than the plume rise height.
2.4. The Inversion Technique
The inversion calculation requires a way to predict the column loads (at the locations of the satellite retrievals) in terms of the emissions. For a given set of emissions, we will call the model predictions corresponding to the satellite observations the ‘model observations’. We wish to compare these with the satellite retrievals. Suppose there are
k satellite retrievals of column load at space-time locations
,
. If we consider the
ith satellite retrieval, the corresponding model observation
for a given set of emissions can be expressed in terms of the emissions and the NAME column loads obtained from individual emission cells as
where
is the NAME prediction of column load at
for a 1 g s
emission rate in the
jth source cell and
is the effective emission rate from the
jth source cell in units of g s
(for
where
n is the total number of model sources). This can be expressed as the matrix equation
where
is the NAME source dilution matrix, i.e., the
matrix of the model column loads (
),
is the vector of source strengths and
is the vector of model observations. The analogous vector of satellite retrievals will be indicated by
and the vector of a priori mean emissions by
.
The goal of the inversion technique is to find emissions which result in a fit of both the model observations to the satellite retrievals, i.e., , as well as of the emissions to the a priori mean emissions, i.e., . The optimal choice uses information on the uncertainty in the satellite retrievals and in the a priori values to decide on the relative priority for fitting and . We do this by maximising the probability density of the emissions given the satellite retrievals, , which Bayes theorem relates to the probability density of the satellite retrievals given the emissions, , and the (a priori) probability densities for the emissions and the satellite retrievals, and .
With our Gaussian assumption, the probability density
of
given
is proportional to
and the a priori probability density
of
is proportional to
Here
is the error covariance matrix for the satellite retrievals and
is the error covariance matrix for the a priori source profile. In principle,
could be regarded as including the errors in the transport model. However, such errors are likely to have a linear dependence on
and a complicated correlation structure (e.g., a typical error might involve the bodily displacement of the ash cloud) which is difficult to treat well except through using ensembles of meteorology. Hence, we do not attempt to account for transport model errors here.
is a
matrix with the satellite retrieval variances on the diagonal and zeros off the diagonal in line with the assumptions described in
Section 2.1. Similarly, the matrix
is filled with the variances and covariances of the a priori emissions derived from the model in
Section 2.3 to give an
(non-diagonal) matrix. Combining
and
by Bayes theorem gives an expression proportional to
. Taking the log of this expression allows the problem to be re-expressed in terms of minimising the cost function
As discussed above, our best estimate of the emissions is taken to be the which minimises J under the constraint that the emissions are non-negative.
To find the cost function minimum we use the ‘non-negative least squares’ (NNLS) algorithm of Lawson and Hanson [
44]. This is similar to an approach presented in [
45] and is an ‘active set method’ in that it iterates over choices of which components of
have the non-negative constraint active, i.e., which components have value zero. To apply this approach the cost function needs to be expressed in the form
for some matrix
and vector
. This can be achieved (up to the addition of a term independent of
) by expanding the cost function in the form
equating the quadratic and linear terms, and solving for
using Cholesky decomposition. We note that after some preliminary calculations to compute the coefficients in the quadratic function (
9), the size of the matrices and vectors, and hence the computational cost, depends only on the number of source cells and is independent of the, usually much larger, number of satellite retrievals.
An advantage of assuming joint Gaussian distributions is that, for a linear forward model
as assumed here, the a posteriori distribution is also Gaussian. The a posteriori error covariance matrix for
,
, is then given by the inverse of the coefficient of the quadratic term in (
9),
Of course the Gaussian assumption is not exact because it implies a possibility of negative emissions, and here it has been supplemented by a truncation of the negative values. Hence, there is some uncertainty in the precise interpretation of
. As a result, it is more appropriate to regard (
10) as giving a general indication of uncertainty.
2.5. Solving the Inverse Problem during an Eruption
Our approach is designed for use during an eruption with results being updated as more data become available. By default, we produce source profile estimates every six hours (using satellite retrievals made for times up to 00:00, 06:00, 12:00 and 18:00 UTC each day). We denote these data cut-off times by , etc., and denote the eruption start time by .
The procedure used to estimate the source profile up to time is as follows:
- 1.
Run NAME with the latest available analysis meteorology valid until time , using all the source components that can emit up to time and tracking the particles from any given source up to the earlier of and six days after release. For the runs use stored results from in order to reduce cost.
- 2.
Estimate the means and covariances of the a priori emissions from these source components using the observed plume rise heights from to .
- 3.
Run InTEM to estimate the a posteriori source profile given the satellite retrievals from to . Note in particular that, for , we do not use the InTEM a posteriori source profile from , nor do we consider only satellite data for times between and . Instead, the inversion starts afresh from the a priori source profile and accounts for all the satellite data up to .
This is repeated every 6 h.
Generally, the NAME runs dominate the cost of the calculation. However, at each update the NAME runs only need to be advanced by 6 h for 6 days’ worth of sources, whereas the inversion calculation (i.e., the calculation of the a priori emission statistics, the cost function coefficients and the optimal emissions) needs (at least with our current setup) to cover the whole eruption up to the current time . Hence, for a long eruption such as the 2010 Eyjafjallajökull eruption, the cost of the inversion can become comparable to the cost of the NAME runs. Towards the end of the Eyjafjallajökull eruption, the time taken using 4 cores on a linux server is about 19 min for the NAME simulations and 13 min for the inversion calculation. This does not include the cost of performing the satellite retrievals and of producing the met data to drive the dispersion model. These timings are for the configuration used in producing the results below for the case with ash and clear sky satellite data.
The time resolution of the source profile must exactly divide the source profile update interval (6 h). For computational speed, we choose, as default, the source discretisation to be 3 h in time and 4 km in height. Some tests of the sensitivity of inversion results to the source resolution for the 2010 Eyjafjallajökull and 2011 Grímsvötn eruptions are presented in [
28].