1. Introduction
Volcanic eruptions can inject clouds of solid ash microscale particles into the atmosphere. These particles remain suspended in the atmosphere for periods from hours to days (Rose and Durant 2009) and have the potential to cause significant damage to aircraft engines (Heinold et al. 2012), endangering lives (Sears et al. 2013) and requiring expensive repairs (Mccarthy et al. 2008). Knowledge of the location of a volcanic ash cloud is therefore valuable because it allows flight paths to be adjusted or, in some cases, flight operations to be suspended to avoid the potential of damage to aircraft while airborne (Langmann et al. 2012). The location of an ash cloud may be observed (Tupper et al. 2007) and predictions made to provide information about its possible future location (Webley et al. 2012). Although the ash cloud is potentially very hazardous, overprediction of the hazard may disrupt flight operations unnecessarily, which is inconvenient to passengers (Langmann et al. 2012) and impacts the aviation industry financially (Turnbull et al. 2012). Therefore, accurate predictions of the movement of the ash cloud are required.
The Australian Volcanic Ash Advisory Centre (VAAC), operated by the Australian Bureau of Meteorology, is responsible for monitoring, forecasting, and issuing advisories with regard to the presence of volcanic ash clouds in the atmosphere over a region that includes Australia and much of the Maritime Continent (Mastin et al. 2009). One of the tools used by the VAAC to predict the movement of a volcanic ash cloud is the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) dispersion model (Draxler and Hess 1997, 1998). As with other dispersion models, it depends on meteorological fields generated by a numerical weather prediction (NWP) model. Uncertainties exist within these fields as a result of uncertainties in the NWP initial conditions and model errors (Weigel et al. 2008). Consequently, the dispersion model will produce a forecast that is based on meteorological fields that contain uncertainties. In addition to the uncertainties inherited from the NWP model, the definition of the pollutant source within the dispersion model may introduce further uncertainties (Webster et al. 2012).
Of the three points mentioned above (uncertainties in the NWP initial conditions, NWP model errors, and pollutant source definition uncertainties), the first two are addressed in the work presented here. The first two points benefit from many years of development of ensemble methods applied to NWP (Buizza and Palmer 1995; Bowler et al. 2008). Uncertainties associated with the definition of the volcanic ash pollutant source will be addressed in future work.
Uncertainties in model predictions can be addressed through the use of a meteorological ensemble prediction system (EPS). Multiple NWP forecasts may be generated from different initial conditions that are varied in an attempt to capture the range of uncertainties (Molteni et al. 1996). Alternatively, multiple forecasts may be generated using a range of model formulations, such as one model with varied representations of physics (Bowler et al. 2008), or several models that have each been constructed with a unique combination of dynamics and physics schemes (Weigel et al. 2008). In this work, the ensemble system that is used represents uncertainties in both the NWP initial conditions and the model formulations (Bowler et al. 2008). A set, or ensemble, of NWP model forecasts is generated. Each of the forecasts in the ensemble provides meteorological conditions to a single dispersion model simulation. An approach such as this, with a meteorological EPS driving dispersion model simulations, follows the works of Straume et al. (1998), Straume (2001), Warner et al. (2002), Stunder and Heffter (2004), and Stunder et al. (2007). The work by Scheele and Siegmund (2001) was somewhat similar, with a meteorological EPS driving a trajectory model rather than a dispersion model. The works of Merrill et al. (1985) and Baumann and Stohl (1997) also involved an ensemble approach to trajectory modeling, but did not involve a meteorological EPS. Draxler (2003) used the same approach as Merrill et al. (1985) and Baumann and Stohl (1997), but modeled both the trajectories and dispersion of a pollutant. Galmarini et al. (2010) considered a multimodel ensemble approach to forecasting pollutant dispersion.
To demonstrate the performance of a meteorological EPS driving a dispersion model for the prediction of volcanic ash movement, the atmospheric dispersion of the volcanic ash cloud that was produced by the eruption of Kelut in Indonesia on 13 February 2014 is simulated. The production of probabilistic predictions from an EPS may be useful for decision-makers as they attempt to mitigate risk. In the following section, details of the meteorological EPS will be described, with particular reference to its application to the current task and coupling of the EPS to the HYSPLIT dispersion model. In section 3, observations of the ash cloud produced by the Kelut eruption will be presented. These are used to qualitatively evaluate the performance of the coupled meteorological–dispersion ensemble prediction system (DEPS). Results from the DEPS are presented in section 4, with comparisons against observations and a deterministic forecast. The level of probability that should be used in interpreting the forecasts is considered, and the sensitivity of results to selected thresholds is demonstrated. With the aim of presenting forecast information concisely and clearly, the inclusion of additional information from individual ensemble member forecasts and the ensemble mean is considered in section 5. Conclusions are presented in section 6.
2. Ensemble prediction system
a. Meteorological ensemble prediction system
The meteorological EPS used at the Australian Bureau of Meteorology (BoM) is based closely on the United Kingdom’s Met Office Global and Regional EPS (MOGREPS), described by Bowler et al. (2008) and O’Kane et al. (2008). The set of NWP systems operated by the BoM (BoM 2012) is named the Australian Community Climate and Earth System Simulator (ACCESS). The global ensemble component of this set is referred to as ACCESS-GE. This system represents uncertainties in both the initial conditions and the model formulations. Perturbations to the initial conditions are generated using an ensemble transform Kalman filter (Bishop et al. 2001). This approach produces initial perturbations that are relatively larger than those produced using the singular vector approach (Buizza and Palmer 1995). An advantage of using larger initial perturbations is that less simulated time is required for an adequate level of spread to develop between the respective member forecasts of the EPS. This is particularly valuable for forecasting out to periods as short as 1–2 days. VAAC forecasts of the movements of volcanic ash clouds cover periods of up to 18–24 h, so it is particularly important when employing an EPS to ensure that initial perturbations are large enough to generate useful results over this short period.
The region for which the VAAC is responsible covers areas of the southern Maritime Continent, including Indonesia, where there are many active volcanoes. It is therefore important that useful forecast information be provided for this tropical region. However, Bowler et al. (2008) show that the initial spread of the ensemble is relatively small in the tropics. Although there is an advantage in using an ensemble transform Kalman filter rather than the singular vector approach, thereby enhancing initial perturbations, the relatively smaller spread in the tropics compared with higher latitudes is a point of concern.
In addition to the perturbations to initial conditions, the version of ACCESS-GE used here includes two stochastic physics schemes to represent uncertainties due to model errors. The first scheme accounts for uncertainties in empirical parameters used in physical parameterizations of large-scale precipitation, convection, the boundary layer, and gravity wave drag. The second adds a potential vorticity dipole to model the effects of unresolved convection on the resolved flow. The two stochastic-physics schemes are described in more detail by Bowler et al. (2008).
b. The coupled meteorological–dispersion ensemble prediction system
ACCESS-GE is configured with 24 members, each producing a unique meteorological forecast. Each of these provides meteorological conditions for a single dispersion model simulation, resulting in the generation of a total of 24 unique predictions of volcanic ash cloud dispersion from a single eruption. This approach is similar to the works of Straume et al. (1998), Straume (2001), Warner et al. (2002), Stunder and Heffter (2004), and Stunder et al. (2007).
The horizontal grid spacings used by ACCESS-GE are 0.833° longitude by 0.553° latitude. From the 70 model levels used in ACCESS-GE, 37 pressure levels up to 1 hPa are derived for use in the dispersion model. Meteorological fields are provided to the dispersion model every 3 h. The dispersion model simulations rely on the variations provided by ACCESS-GE. Note that although the HYSPLIT dispersion model is used here as part of an EPS, the ensemble method discussed by Draxler (2003), in which the location of the pollutant source is varied, thereby subjecting the pollutant to a range of meteorological conditions, is not activated in this current work.
In previous versions of the HYSPLIT model, the Stokes formula has been used to predict the sedimentation of pollutant particles. In the version used here, a single modification is made to the HYSPLIT model, with the sedimentation of volcanic ash particles predicted using the scheme of Ganser (1993), as discussed by Dare (2015). The density of volcanic ash particles is 2500 kg m−3 and the particle sphericity (Wadell 1932) is defined as 0.8. The particle size distribution used here is based on the observations of Hobbs et al. (1991) and the modeling experiments of Heffter and Stunder (1993), Dacre et al. (2011), and Devenish et al. (2012). Particle sizes range from 0.21 to 93.1 μm across 14 bins.
3. Observations of the volcanic ash cloud from the Kelut eruption
The Kelut volcano (7.930°S, 112.308°E), also known as Kelud, located in Java in Indonesia, erupted for approximately 4 h, from 1600 to 2000 UTC 13 February 2014. The plume reached a height of approximately 19 km (Kristiansen et al. 2015).
Three sets of information are used to qualitatively evaluate forecasts produced by the coupled EPS. The first set consists of 11-μm brightness temperatures (BTs) observed by the Japanese Multifunctional Transport Satellites (MTSAT). Figures 1a–d show the areal expansion of an ash plume (hourly from 1632 to 1932 UTC) with relatively low BTs, over the volcano, the location of which is identified by the black dot. The dashed-line polygons shown here have been added manually to identify the volcanic plume and will be used later to evaluate the performance of the model forecasts. Although the plume becomes particularly elongated in a westward direction, as a result of the influence of easterly winds, the plume is also observed to expand to the north and south (cf. Figs. 1a and 1b). This is consistent with the observed growth of other so-called umbrella clouds produced from volcanic eruptions (e.g., Mastin et al. 2009; Witham et al. 2012). In Figs. 1e–g (3-hourly from 2232 to 0432 UTC), the ash cloud moves away from the volcano in an approximately west-southwestward direction. By 0732 UTC 14 February 2014 (Fig. 1h), the cloud has become more difficult to identify, most likely because of a combination of dispersion and the presence of water/ice clouds. Therefore, the brightness temperature observations shown in Fig. 1 will be used to evaluate model forecasts only up to 0432 UTC 14 February 2014.
A time series of 11-μm BTs observed by an MTSAT over Kelut, Indonesia, from 1632 UTC 13 Feb to 0732 UTC 14 Feb 2014, with dashed-line polygons manually added to identify the feature of interest. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
The second set of verifying data consists of the fields of brightness temperature difference (BTD) between the 11- and 12-μm channels (Prata 1989; Wen and Rose 1994; Prata and Grant 2001) and the probability of ash, as generated by the Geostationary Cloud Algorithm Testbed (GEOCAT) system (Pavolonis et al. 2013) run at the BoM. These two fields are plotted together to produce a binary field based on the presence of BTDs ranging from 0° to −10°C, and ash probabilities ranging from 20% to 100%. The areas covered by these fields have been captured using solid-line polygons, as shown in Fig. 2. These polygons will be used later to evaluate model forecasts and will be displayed consistently using solid lines. The four times shown by Figs. 2a–d, 6 h apart, are selected because they correspond approximately to the four forecast times that will be discussed in the following sections (6-, 12-, 18-, and 24-h forecasts, with respective times and dates of 2200 UTC 13 February and 0400, 1000, and 1600 UTC 14 February 2014).
Shaded areas based on the combined presence of BTDs ranging from 0° to −10°C and ash probabilities ranging from 20% to 100%, as detected by the GEOCAT system, outlined by solid-line polygons, at approximately (a) 6, (b) 12, (c) 18, and (d) 24 h from the start of the eruption. Dotted-line polygons represent manual analyses produced by VAAC forecasters. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
The third set of information used to evaluate the model forecasts consists of manual analyses, based on satellite observations and other available information including ground-based observations and pilot reports, produced by VAAC forecasters in routine volcanic ash advisories. Each area judged to potentially contain hazardous concentrations of volcanic ash was identified and outlined by defining a polygon. These polygons are illustrated using dotted lines for each of the four times displayed in Fig. 2.
All three sets of verifying data show that the ash cloud moves away from the volcano to the west-southwest. The manual analyses show that the area affected by the volcanic ash cloud increases over time. During the first few hours after the eruption began, the BT data also showed some growth in the area covered by the ash cloud (Fig. 1), but over time the cloud became increasingly difficult to identify manually. The areas covered by the BTD and ash probability fields are relatively patchy, as shown by the shaded areas within the solid-line polygons (Fig. 2), and do not indicate consistent expansion over time of the areas affected by the ash cloud. However, the areas of ash identified by these fields (solid lines) are almost completely located within the areas covered by the manual analyses (dotted lines).
4. Forecasts of the volcanic ash cloud
In this section, forecasts are presented of the volcanic ash cloud at 6, 12, 18, and 24 h following the start of the Kelut eruption. Following information on the pollutant source and wind fields, and a discussion concerning the choice of field and the computation of probabilities, the ensemble forecasts are compared with deterministic (single member) forecasts, and also with observations. Following this, the sensitivity of the forecast field to the selected threshold used in calculating the probabilities is considered.
a. Definition of pollutant source
The longitude, latitude, date, time, duration, and plume height noted in section 3 are used to define the pollutant source in every one of the 24 dispersion model simulations. A single vertically uniform column of erupted mass is released, similar to past work by Witham et al. (2007), Webley et al. (2009), and Webster et al. (2012). This column extends from the surface to a height of 19 km. The total amount of volcanic ash released into the model’s atmosphere during the 4-h eruption is 1 unit of mass (M), released uniformly in time over the 4-h period.
b. Wind fields
Interpretation of the forecasts presented below will benefit from information on the wind fields at different heights in the atmosphere. Forecast results are produced from the dispersion model in layers each of 5 km thickness up to 25 km above sea level (ASL). Results from the meteorological control forecast at 0000 UTC 14 February 2014 are used here to provide a general illustration of the winds within each layer (Fig. 3), although, of course, these fields vary between members of the ensemble. During, and for at least 24 h after the eruption, the winds in the 10–15-km layer were approximately easterly in the vicinity of the volcano, with northeasterly and northerly winds far to the southwest, near 90°–100°E (Fig. 3a). In the 5–10-km layer, winds were approximately easterly to the north of the volcano, relatively weak southerly to the south, and relatively weak westerly far to the southwest (Fig. 3b). Below this layer, from 0 to 5 km ASL, winds were southwesterly in the vicinity of the volcano (Fig. 3c).
Horizontal wind fields predicted by the control forecast at 0000 UTC 14 Feb 2014 at (a) 200 hPa (~12.4 km), (b) 400 hPa (~7.5 km), and (c) 750 hPa (~2.5 km). Shaded regions represent wind speeds above 10 m s−1. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
c. Concentration and atmospheric column load
The concentration of the volcanic ash pollutant predicted by the dispersion model is specific to a selected layer of the atmosphere. The field of concentration (mass per volume; M m−3) is therefore useful when investigating a particular atmospheric layer of interest, such as a layer that contains aircraft cruise altitudes. Alternatively, multiple layers may be considered together by weighting the concentration in each layer by the layer’s depth to produce a total atmospheric column load (mass per area; M m−2). It is useful to consider column load when examining the dispersion model results because the visualization of this field shows all locations where volcanic ash is predicted to be present irrespective of altitude. Additionally, images obtained from satellites are based on the view of the atmosphere from space, not on the view of a specific layer of the atmosphere. In this section, the dispersion model forecasts are examined based on column load, while in section 5 there is some discussion concerning concentrations.
d. Probability of volcanic ash
Each member of the ensemble produces a forecast of the column load. These forecasts from the 24 members are combined to produce a probability forecast. Each spatial grid point of the column load field produced by each member is evaluated relative to a selected threshold value. Points where the column load exceeds the threshold are categorized as “true,” otherwise they are “false.” The total number of forecasts that are found to be true for a given grid point is divided by the total number of members needed to produce a probability for that grid point. In the current work, probabilities are represented as percentages.
As there are 24 members in the ensemble, a probability of approximately 4.2% at a given grid point means that only one member predicted a column load above the threshold at that location. The prediction from that one member is not necessarily a poor forecast. It may even be the best forecast of all the members. However, the low percentage means that the ensemble of forecasts indicates that there is a low likelihood of a column load above the threshold occurring at that particular location.
Another point to note concerns the interpretation of the probability field relative to the column load. While in some cases there might be some coincidence between the appearance of a column load field from a single forecast and the probability field from the ensemble of forecasts, it is important to remember that low probabilities, for example, do not represent low values of column load. In the approach used here to compute the probabilities, values of column load less than the threshold have been neglected for all members, and are not considered as part of the calculation.
e. Ensemble forecast
Ensemble forecast (EF) probabilities of column loads above a threshold of 10−19 M m−2 are shown in Fig. 4 for forecast hours 6, 12, 18, and 24. Alternative values of the threshold are considered in section 4h. Also shown are observational data discussed in section 3, with BT polygons (dashed), BTD and observed probability of ash (BTDP) polygons (solid), and VAAC hazard area analyses shown by dotted polygons. The location of the volcano is shown by the black dot.
Probabilities (%) of column load above a selected threshold of 10−19 M m−2, at forecast hours (a) 6, (b) 12, (c) 18, and (d) 24 based on the 24-member ensemble. Dashed (BT), solid (BTDP), and dotted (VAAC analyses) polygons represent observations presented in Figs. 1 and 2. Locations A–D are discussed in the text. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
Shaded areas, representing probabilities greater than 5%, increase over time in Figs. 4a–d. Probabilities shown range from greater than 90% down to 5%. The cloud of volcanic ash disperses mainly to the west-southwest of the volcano over the first 12 h, but also with some movement to the northeast. From 12 to 24 h, the western arm of the cloud shows some dispersal southward between approximately 95° and 100°E, consistent with the description of the wind field in section 4b. Highest probabilities (>90%) are found within the core of the cloud, showing that there is a high level of consensus between members of the ensemble for these locations.
At 6 h, most of the cloud, and particularly the red core, is contained within the dotted- and solid-line polygons. However, the cloud reaches farther to the west beyond the boundaries of these polygons. The core of this arm of the cloud at 6 h does, however, coincide with the BT polygon, showing that this is a very good forecast. At 12 h, the cloud has moved and is dispersed farther to the west. While the dotted- and solid-line polygons also show some expansion to the west, the EF has placed most of the area containing the highest probabilities outside of the areas covered by these two polygons. However, the BT polygon (dashed) validates the accuracy of the 12-h EF, similar to what was found for the 6-h forecast.
Validation of the 18- and 24-h forecasts depends on the dotted- and solid-line polygons because the BT signal could not be identified clearly after about 12 h, as noted in section 3. Moderate to high probabilities (>50%) within the extensive western and southwestern regions of the cloud at forecast times of 18 and 24 h are located within the dotted polygon analyzed by the VAAC. The forecast cloud appears to match the size and general shape of the VAAC polygons very well. The solid-line polygons are located within the forecast cloud, coinciding with a range of probabilities ranging from 5% to more than 90%.
f. Comparison of ensemble and deterministic forecasts
A separate deterministic model forecast could be used for comparison with the results produced by the DEPS. However, it is convenient to use the ensemble’s own control (nonperturbed) member simulation for this purpose because it has a configuration that is identical to all members of the DEPS. By using the control forecast, factors such as initialization and resolution, which may differ between the DEPS and a separate deterministic model, do not need to be considered, which allows for a clear comparison between the deterministic control forecast (CF) and the EF. Note that the CF is not used to assess the accuracy of the dispersion model, but rather it provides a single deterministic simulation to compare with the EF.
The CF is shown in Fig. 5 for the same forecast hours as presented above for the EF. For consistency with the EFs, the CFs are presented for points where the column loads exceed the same threshold of 10−19 M m−2. As this is a single member, the probabilities are therefore either 0% or 100%, with the latter shown by gray shading in Fig. 5.
As in Fig. 4 but for the forecast of column load above a selected threshold of 10−19 M m−2.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
On the broad scale, the dispersion of the volcanic ash cloud predicted by the CF shows some similarities with the EF, with expansion of the cloud to the west and southwest and some additional dispersion to the northeast. However, differences do exist between the EF and the CF. At 6 h (Fig. 5a), the plume has moved and dispersed to the west of the volcano, and its western edge is collocated with the BT polygon (dashed). The plume also covers approximately half of the BTDP polygon (solid line). A similar level of agreement between the observations and shaded areas is predicted by the EF at this time. Later, at forecast hours 12, 18, and 24, the shaded areas predicted by the EF correspond very well to the BTDP polygons, but relatively less agreement results from the CF, as shown by the unshaded portions of these polygons (Figs. 5b–d).
Some locations where the EFs predict probabilities above 5% correspond to predictions by the CFs of zero 0% probability. Some examples of these locations are identified in Figs. 4 and 5 by A–D for forecast hours 12, 18, and 24. At location A, the probabilities are no higher than 20%, based on predictions by 4–5 members. Although this is a relatively low probability compared with the core of the cloud where values exceed 90%, the low consensus at A does not mean that these 4–5 forecasts are incorrect. In the vicinity of location B, the probabilities reach up to approximately 40%, due to predictions by 9–10 members. Similarly, C is also located on the western edge of the cloud, where probabilities are up to 30%. These three locations show that the EFs predict column loads above the selected threshold at locations not identified by the CFs. Given the highly hazardous nature of volcanic ash (Heinold et al. 2012), information such as this has the potential to be valuable to forecasters. It is interesting to consider point D because it identifies a location within the VAAC polygon where probabilities are up to approximately 40%. In this example, the EF shows good agreement with the analyzed polygon, but the location is not identified by the CF. Although A–D identify some areas where there is relatively low consensus between members of the DEPS, it is important to note that they are not necessarily poor forecasts and at least provide some additional guidance not available from the CF.
The largest differences between CFs and EFs may be expected over periods of greater than a day or two, as the spread of the ensemble members continues to increase. Here, differences between the CFs and EFs are evident by 6 h, with differences increasing up to 24 h. This is interpreted as a positive characteristic of the DEPS, for several reasons. First, a DEPS must be able to produce a range of results beyond that produced by a deterministic forecast. Second, forecasts produced by the VAAC cover periods up to 18–24 h, so it is beneficial to utilize a DEPS that generates an adequate spread over this period. Third, the possible lack of spread in the tropics, noted in section 2a, did not appear to eventuate in this case.
Although this is a preliminary investigation, the results presented here suggest that it may be useful in the future to automate the production of forecast polygons at the VAAC based on these probability fields. For example, in Fig. 4d the western edge of the VAAC polygon and the 40% probability shading are approximately collocated, and the shape of the polygon and the shape of the cloud containing probabilities above approximately 40% are similar. Additionally, in Fig. 4b the EF predicted an area of high probabilities that coincided very well with the BT polygon, but that was located to the west of the VAAC polygon. In this latter example, an automated generation of the VAAC polygon based on the DEPS would have produced a more accurate result based on a forecast than that based on the VAAC analysis.
g. Selecting the threshold: Probability
The level of probability that a user may consider appropriate to interpret the EF is considered here. In the future, maps of probabilities, similar to those presented in Fig. 4, will be produced from the DEPS for operational forecasting. These maps provide information on the probability of volcanic ash column loads above a selected threshold, derived from the ensemble of forecasts that represents uncertainties in the initial conditions and model formulations. Using these maps as guidance to produce a manual forecast by identifying hazardous areas based on locations of high probabilities may produce a forecast with reduced errors. However, for an individual event, the hazardous area predicted by the most accurate individual member forecast will not necessarily correspond to the area containing the highest probabilities. Depending on the context, it may not be suitable for a forecaster to aim solely at producing a forecast with small errors. Within the context of high risk, such as the potential for aircraft engines to be disabled as a result of encountering airborne volcanic ash, it may be useful for a forecaster to produce a forecast that includes the identification of areas where the probability is relatively low. This is consistent with the work of Richardson (2000), who stated that users with relatively large potential losses will benefit from taking action even when the forecast probability is low.
h. Sensitivity to threshold: Column load
In assessing the probability based on all members of the DEPS, each individual member forecast is assessed for exceedance of a selected threshold value of the atmospheric column load. The results shown in Figs. 4d and 6 demonstrate the sensitivity of the 24-h probability forecast to column load thresholds ranging from 10−19 to 10−11 M m−2. The probabilities shown in Figs. 4d, 6a, and 6b are very similar in terms of both the location of the edge of the volcanic ash cloud and the distribution of probabilities within the cloud. These examples show that in this case the probabilities do not vary over the wide range of thresholds from 10−19 to 10−15 M m−2.
Probabilities (%) of column load for the 24-h forecast, based on the 24-member ensemble, for thresholds (a) 10−16, (b) 10−15, (c) 10−14, (d) 10−13, (e) 10−12, and (f) 10−11 M m−2. Solid (BTDP) and dotted (VAAC analyses) polygons represent observations presented in Fig. 2. Locations E–G are discussed in the text. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
However, when the threshold is increased to 10−14 M m−2 (Fig. 6c), there are noticeable changes in the probabilities relative to those discussed above. While differences in the location of the edge of the volcanic ash cloud are very small, peak values within the cloud are reduced. The area containing high probabilities in the southern part of the cloud (to the east of E) has been reduced, while probabilities within the extensive west–east-oriented area located near 7°S (identified by F) have mostly dropped from 90%–100% to 70%–90%.
The probabilities decrease further when the threshold level is increased to 10−13 M m−2 (Fig. 6d). Changes in the location of the edge of the cloud are again small, while reductions in the probabilities within the cloud are more noticeable. The extensive west–east region (F) of probabilities with values above 90% seen in Fig. 6b has been replaced by probabilities in Fig. 6d of approximately 20%–60%, while in contrast, the region of high probability (>90%) near E maintains a relatively high probability (>70%) when the threshold is increased. These comparisons show that distributions of probabilities do not vary uniformly throughout the cloud when the threshold is varied.
An increase in the threshold to 10−12 M m−2 (Fig. 6e) has a larger impact on the area of the cloud than is seen in the examples discussed above. The area has decreased, particularly in the western parts of the cloud. Probabilities located within the dotted polygon have mostly fallen to less than 40%. The highest probabilities are now found in the eastern part of the cloud, in the vicinity of G, with peak values of approximately 50%, although there are some very small areas exceeding 50%. When the threshold is increased further to 10−11 M m−2 (Fig. 6f), probabilities do not exceed 5%, apart from some small areas near G. In Figs. 6e and 6f, the highest probabilities are found near G, while in previous frames (Figs. 6a–d), the highest probabilities occurred near E or F.
These results demonstrate that the forecast probabilities are sensitive to the choice of threshold. However, this sensitivity depends on the range of thresholds considered, with values from 10−19 to 10−15 M m−2 having little impact on the probabilities in this case, while varying the threshold from 10−15 to 10−11 M m−2 has a large impact.
5. Presentation of information from individual members of the ensemble
A large amount of information is produced from the DEPS. In section 4, for example, probabilities were presented based on combined forecasts of column load from all 24 members of the DEPS (Fig. 4). However, in addition, concentrations could have been used to compute probabilities within multiple layers of the atmosphere. Also, including a range of thresholds further increases the amount of information generated from the DEPS (Fig. 6). Alternatively, fields such as column load and concentration might be examined from individual member forecasts, which would greatly increase the amount of information extracted from the DEPS. As with all EPSs, there is a need to provide useful information to the forecaster, or other end user, without presenting so much information that it becomes unmanageable to interpret. In this section, the aim is to provide an additional layer of information without obscuring the original display of probabilities (Fig. 4) and without generating a large number of additional diagrams.
a. Column load maxima
Locations of maximum ash column load from each ensemble member forecast have been identified for each of the four forecast times discussed previously. So as to present this additional information without obscuring the probabilities, these locations are displayed using a black ring for each member forecast, the center of which corresponds to the location of the maximum (Fig. 7). There are three points to note concerning these maximum points. First, the method of displaying these points appears to succeed in providing information about individual member forecasts without obscuring the shaded probabilities. Second, at 6 h these points are located in a small bunch close to the volcano, but as the forecast time increases from 6 to 24 h, the points disperse to cover a wider area. Third, the maxima do not correspond well with areas displaying highest probabilities. This is particularly obvious at forecast hours 18 and 24 (Figs. 7c,d), where most probabilities above 90% are located to the west and southwest of the volcano, while the maxima are found relatively much closer to, and north of, the volcano. The lack of correspondence between locations of the maxima and the highest probabilities is investigated next.
As in Fig. 4, but with the addition of locations of column load maxima from individual ensemble member forecasts, shown by black rings.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
b. Concentration maxima
Probabilities of ash concentrations above a selected threshold (10−19 M m−3) are computed for four 5-km-thick layers of the atmosphere up to 20 km ASL based on 24-h forecasts produced by the DEPS (Fig. 8). Note that although individual layers of the atmosphere are presented here, the probabilities within each layer depend on the representation of uncertainties in the NWP system, as discussed in sections 1 and 2, while uncertainties due to the definition of the height of the pollutant source are not considered in the current work. Also shown in Fig. 8, using black rings, are locations of maximum ash concentration from each individual ensemble member forecast. The two lowest layers of the atmosphere (0–5 and 5–10 km ASL) contain both probabilities of at least 5%, a value that represents at least one member, and maxima in the vicinity of the volcano (Figs. 8a,b). In contrast, in the higher layers (10–15 and 15–20 km ASL), the ash is located away from the volcano to the west and southwest (Figs. 8c,d). The distributions of ash throughout these layers are consistent with the wind fields within these layers described in section 4b.
A 24-h forecast of probabilities (%) of concentration above a selected threshold of 10−19 M m−3, based on the 24-member ensemble, within model atmosphere layers (a) 0–5, (b) 5–10, (c) 10–15, and (d) 15–20 km ASL. Locations of concentration maxima from individual ensemble member forecasts are shown by black rings. Solid and dotted polygons represent observations presented in Fig. 2. The dashed-line box and H are discussed in the text. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
As column load is computed as an aggregate of layer concentrations weighted by layer depth, concentration maxima in the vicinity of the volcano in the lower two layers (Figs. 8a,b) have, in this case, led to a column load field (Fig. 7d) in which maxima are found at similar locations. The concentration fields and their maxima provide only a partial explanation for the question of the lack of correspondence between locations of maxima and highest probabilities. In fact, the concentration data (Fig. 8) also exhibit a lack of correspondence between locations of maxima and highest probabilities. This is investigated further using the 10–15-km layer (Fig. 8c) as an example. This is of interest because typical aircraft cruise altitudes are located within this layer.
There are three areas of interest in Fig. 8c. First, the area containing the highest probabilities is outlined by the dashed-line box. Within this box are seven maxima. This means that 17 maxima are located outside the box, demonstrating the lack of correspondence between maxima locations and highest probabilities. Second, 13 of these maxima are found in the northeast area of the cloud, corresponding to probabilities of approximately 20%–60%. One may expect that some maxima would be located in this area as a result of the transport of ash from the source by winds in this layer (Fig. 3a). Third, four maxima are located in the northwest area of the cloud (indicated by H in Fig. 8c). These are of interest because the probabilities at this location are relatively low (<30%) and the maxima are located near the edge of the cloud, relatively isolated from other maxima.
c. Individual member forecasts
The individual member forecasts of concentration that were used to compute the probabilities in Fig. 8c are presented in Fig. 9. The location of the maximum concentration is indicated by the black ring in each frame. For reference, and comparison between members and against the probability map, the same box is displayed as in Fig. 8c. The four members displaying maxima in the northwest (near H in Fig. 8c) are numbers 5, 11, 12, and 21 in Fig. 9. Within the context of the entire collection of member forecasts in Fig. 9, these four forecasts do not appear to be unusual or extreme, and they may be accepted as valuable members of the ensemble.
Individual member (frames numbered 0–23) 24-h forecasts of concentration in the 10–15-km layer, with values above 10−16 M m−3 shown in red (otherwise blue). The dashed-line box is discussed in the text. The location of the maximum concentration is shown for each individual ensemble member forecast by a black ring.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
A characteristic of all member forecasts in Fig. 9 is the positioning of part of the cloud within the box. Another prominent feature is the relatively narrow branch connected to the cloud from the east. This narrow branch contains concentrations above 10−16 M m−3 in almost all member forecasts, while the concentrations within the box are much less consistent. Based on these features, one might expect that the highest probabilities (of concentrations above 10−19 M m−3) in Fig. 8c would be located in the branch to the northeast of the box, rather than within the box. How is it possible for the box to contain higher probabilities than the branch to the east? The answer is that the location of the narrow branch is highly variable between member forecasts (Fig. 9), which prevents detection of this feature at a consistent location and results in a relatively low consensus between members and a low probability. In contrast, there are enough occurrences of the cloud within the box in member forecasts to produce a higher probability. The 13 maxima found to the northeast of the box in Fig. 8c correspond to the maxima located within the narrow branch of 13 member forecasts in Fig. 9. The inclusion of the black rings to represent individual member forecast maxima along with the shaded probability information (Fig. 8c) is therefore very valuable because these maxima identify the narrow branch feature that is prominent (Fig. 9) but would otherwise not be displayed using probabilities alone. The fact that the highest probabilities and the maxima are found at different locations should not be interpreted as an inconsistency. Instead, the probabilities and maxima should be seen as complementary, and potentially valuable, sets of information.
d. Sensitivity to threshold: Concentration
Probabilities of the occurrence of ash concentrations above a chosen threshold vary depending on the value of that threshold. The fields presented in Figs. 8c and 10 demonstrate the sensitivity of the probability forecast to ash concentration thresholds ranging from 10−19 to 10−16 M m−3. Increasing the threshold from 10−19 to 10−18 M m−3 has a minor effect near the edge of the cloud, with some small but noticeable changes near the center of the cloud in the dashed-line box where probabilities at some points decrease by approximately 10%. The impact due to a further increase to 10−17 M m−3 is similar, with largest changes in the inner region of the cloud where probabilities are highest.
Probabilities (%) of concentration for the 24-h forecast, based on the 24-member ensemble, for thresholds (a) 10−18, (b) 10−17, and (c) 10−16 M m−3. Locations of concentration maxima from individual ensemble member forecasts are shown by black rings. Solid and dotted polygons represent observations presented in Fig. 2. The dashed-line box is discussed in the text. The location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
Over the range of thresholds from 10−19 to 10−17 M m−3, the area within the cloud containing probabilities above 40% is a fairly consistent feature, approximately filling the dashed-line box. When the threshold is increased further to 10−16 M m−3, probabilities above 40% are found mainly in a much smaller area near the northeast corner of the box. Similarly, while the area of the entire cloud does not change much for thresholds ranging from 10−19 to 10−17 M m−3, there is a more noticeable reduction in the area when the threshold is increased from 10−17 to 10−16 M m−3. As the threshold is increased and the cloud covers less area, there is increased correspondence between locations of maxima and the remaining areas of cloud, defined by relatively low probabilities.
e. Ensemble mean forecast
In addition to computing probabilities or examining individual member forecasts, the ensemble member forecasts may be averaged to produce an ensemble mean forecast, which is essentially a quasi-deterministic consensus forecast (Van den Dool and Rukhovets 1994; Ebert 2001). It is well known that ensemble mean forecasts verify well, but can provide poor guidance for individual high-impact events (Van den Dool and Rukhovets 1994; Ebert 2001; Palmer 2012; Ancell 2013). For the case investigated here, the ensemble mean 24-h forecast of ash concentration in the 10–15-km layer (Fig. 11a) may be compared with both the probability forecast (Figs. 8c and 10) and individual member forecasts (Fig. 9).
Ensemble mean 24-h forecasts of (a) the concentration (M m−3) in the 10–15-km layer, with concentration maxima from individual ensemble member forecasts shown by black rings, and (b) the column load (M m−2), also with maxima from individual members. The dashed-line box is discussed in the text and the location of Kelut is shown by the black dot.
Citation: Journal of Applied Meteorology and Climatology 55, 1; 10.1175/JAMC-D-15-0079.1
The areas covered by the respective ash clouds in Figs. 8c and 11a are similar, and there is some indication that the location of the highest mean concentrations in Fig. 11a, in the northeast corner of the dashed-line box, corresponds to the region of highest probabilities above 70% in Fig. 8c. The probability field computed based on an increased threshold (Fig. 10c) also shows that there are similarities between the ensemble mean and probability fields, such as the respective locations of mean concentrations above 10−16 M m−3 and highest probabilities above approximately 40%.
For comparing the ensemble mean concentration with individual member forecasts, red shading representing values above 10−16 M m−3 has been applied consistently in both Figs. 9 and 11a. The area containing mean concentrations above this level is smaller than the corresponding areas produced by every individual member forecast. Although the approximately west–east swath of red shading produced by the ensemble mean forecast may indicate the presence of the narrow branch discussed in section 5c, it is much less prominent than the features displayed by the individual member forecasts in Fig. 9. Another difference between the ensemble mean and the individual member forecasts is the area covered by the respective clouds. The ensemble mean cloud covers a larger area than any individual member cloud because of the compositing of all member forecast clouds in the process of computing the mean. Although the mean area is larger, the features within the cloud have been effectively smoothed by the process of averaging, thereby removing details that may be very useful to forecasters.
The areas covered by the ensemble mean 24-h forecast of column load (Fig. 11b) and by the probability field (Fig. 6) are similar. However, within the cloud, the highest mean loads are found east of approximately 107°E, while the highest probabilities are found west of this longitude (Figs. 6a–d). The use of a higher threshold in computing the probabilities (Fig. 6e) results in the highest probabilities (>40%) being located east of 107°E, a spatial distribution that is similar to that of the highest mean column loads (>10−12 M m−2). In this case the locations of the maxima correspond well with the locations of the highest mean column loads and with the probability field computed using a threshold of 10−12 M m−2. The locations of the concentration maxima within different layers of the atmosphere (black rings in Fig. 8) show that the column load maxima are found east of approximately 107°E in the vicinity of the volcano as a result of the distributions of volcanic ash in the lowest two layers of the atmosphere (0–10 km ASL).
One of the main aims of producing composited fields, such as probability and the ensemble mean, is to present ensemble forecast information concisely. However, in the case investigated here, although the presentation is concise, the use of these fields underrepresents areas of relatively high concentrations of volcanic ash predicted by individual member forecasts. The use of the complementary display of locations of maxima from individual member forecasts is a simple method that can be employed to identify areas within composited fields where individual member forecast data points may be underrepresented, while still maintaining a fairly concise presentation of the ensemble forecast information. In addition to composited fields, it is important that forecasters be provided with information, such as complementary fields and individual member forecasts, to ensure that ensemble forecasts can be interpreted correctly so as to maximize the benefits of using an ensemble prediction system.
6. Conclusions
A meteorological EPS that represents uncertainties in both initial conditions and model formulations has been coupled with a modified version of HYSPLIT. Uncertainties due to the definition of the pollutant source have not been considered in the current work. The 24 EPS member forecasts are used to define meteorological conditions for 24 individual HYSPLIT forecasts. The coupled system has been used to forecast the dispersion of the volcanic ash cloud produced by the 13 February 2014 eruption of Kelut, Indonesia.
The 6- and 12-h forecasts from both the deterministic (control member) and coupled EPS show very good qualitative agreement with satellite observations of BT, BTD, and ash probability provided by GEOCAT, confirming the ability of HYSPLIT to accurately simulate the dispersion of pollutants. By forecast hours 18 and 24, observations appear to be less coherent, so they are complemented by manual analyses from the Australian VAAC. At these forecast hours the coupled EPS forecast shows better qualitative agreement with the observations than does the deterministic forecast.
The coupled system also provides additional information in terms of probability, identifying the levels of consensus between member forecasts. Although the probability field presents information concisely, experiments here have shown that it is very important to also consider results from individual member forecasts in order to identify features that may be undervalued in the probability field. For example, an area of high ash concentration that was forecast by most of the members was not particularly evident in the probability field because the location of this feature was highly variable between member forecasts. This feature was also not well represented by the ensemble mean forecast. A large amount of information is available from the coupled EPS, and needs to be presented concisely to allow interpretation by forecasters working under operational time constraints, but care should be taken to avoid neglecting vital data.
A range of thresholds has been used in computing probability fields. As these thresholds affect the appearance of the probability fields, it would be useful to provide forecasters with the ability to adjust the threshold while perusing probability fields. It is also important to provide access to individual member forecasts because the examination of these forecasts can be helpful in interpreting probability fields. There is an additional requirement to select the level of percentage to use in assessing probability fields. Within the context of high risk, such as the potential for the failure of aircraft engines as a result of encountering airborne volcanic ash (Mastin et al. 2009), it may be useful to include consideration of areas where the computed probabilities are relatively low.
The meteorological EPS used here, based on MOGREPS, is run daily at the BoM to produce forecasts over a global domain. In the future, the implementation of a higher-resolution version of this system is expected to improve the accuracy of the coupled EPS. It is also hoped that the issue of the relatively small initial spread of the ensemble in the tropics will be addressed. It may also be useful to consider increasing the number of member forecasts in the EPS beyond 24. Possibly a more important point to consider in the future is the representation of uncertainties based on characteristics of the volcanic source. These may include, for example, variations in the height of the cloud column above the volcano and variations in the physical properties of volcanic ash particles.
Acknowledgments
The authors thank Elizabeth Ebert and Rod Potts for providing valuable comments on the manuscript. We are also thank the reviewers for their positive comments.
REFERENCES
Ancell, B. C., 2013: Nonlinear characteristics of ensemble perturbation evolution and their application to forecasting high-impact events. Wea. Forecasting, 28, 1353–1365, doi:10.1175/WAF-D-12-00090.1.
Baumann, K., and A. Stohl, 1997: Validation of a long-range trajectory model using gas balloon tracks from the Gordon Bennett Cup 95. J. Appl. Meteor., 36, 711–720, doi:10.1175/1520-0450-36.6.711.
Bishop, C. H., B. J. Etherton, and S. J. Majumdar, 2001: Adaptive sampling with the ensemble transform Kalman filter. Part I: Theoretical aspects. Mon. Wea. Rev., 129, 420–436, doi:10.1175/1520-0493(2001)129<0420:ASWTET>2.0.CO;2.
BoM, 2012: APS1 upgrade of the ACCESS-G Numerical Weather Prediction system. Bureau of Meteorology NMOC Operations Bull. 93, 29 pp. [Available online at http://www.bom.gov.au/australia/charts/bulletins/apob93.pdf.]
Bowler, N. E., A. Arribas, K. R. Mylne, K. B. Robertson, and S. E. Beare, 2008: The MOGREPS short-range ensemble prediction system. Quart. J. Roy. Meteor. Soc., 134, 703–722, doi:10.1002/qj.234.
Buizza, R., and T. N. Palmer, 1995: The singular-vector structure of the atmospheric general circulation. J. Atmos. Sci., 52, 1434–1456, doi:10.1175/1520-0469(1995)052<1434:TSVSOT>2.0.CO;2.
Dacre, H. F., and Coauthors, 2011: Evaluating the structure and magnitude of the ash plume during the initial phase of the 2010 Eyjafjallajökull eruption using lidar observations and NAME simulations. J. Geophys. Res., 116, D00U03, doi:10.1029/2011JD015608.
Dare, R. A., 2015: Sedimentation of volcanic ash in the HYSPLIT dispersion model. CAWCR Tech. Rep. 079, 26 pp. [Available online at http://www.cawcr.gov.au/publications/technicalreports/CTR_079.pdf.]
Devenish, B. J., P. N. Francis, B. T. Johnson, R. S. J. Sparks, and D. J. Thomson, 2012: Sensitivity analysis of dispersion modeling of volcanic ash from Eyjafjallajökull in May 2010. J. Geophys. Res., 117, D00U21, doi:10.1029/2011JD016782.
Draxler, R. R., 2003: Evaluation of an ensemble dispersion calculation. J. Appl. Meteor., 42, 308–317, doi:10.1175/1520-0450(2003)042<0308:EOAEDC>2.0.CO;2.
Draxler, R. R., and G. D. Hess, 1997: Description of the HYSPLIT_4 modeling system. NOAA Tech. Memo. ERL ARL-224, 24 pp. [NTIS PB98-116593.]
Draxler, R. R., and G. D. Hess, 1998: An overview of the HYSPLIT_4 modeling system for trajectories, dispersion, and deposition. Aust. Meteor. Mag., 47, 295–308.
Ebert, E. E., 2001: Ability of a poor man’s ensemble to predict the probability and distribution of precipitation. Mon. Wea. Rev., 129, 2461–2480, doi:10.1175/1520-0493(2001)129<2461:AOAPMS>2.0.CO;2.
Galmarini, S., F. Bonnardot, A. Jones, S. Potempski, L. Robertsen, and M. Martet, 2010: Multi-model vs. EPS-based ensemble atmospheric dispersion simulations: A quantitative assessment on the ETEX-1 tracer experiment case. Atmos. Environ., 44, 3558–3567, doi:10.1016/j.atmosenv.2010.06.003.
Ganser, G. H., 1993: A rational approach to drag prediction of spherical and nonspherical particles. Powder Technol., 77, 143–152, doi:10.1016/0032-5910(93)80051-B.
Heffter, J. L., and B. J. B. Stunder, 1993: Volcanic Ash Forecast Transport and Dispersion (VAFTAD) Model. Wea. Forecasting, 8, 533–541, doi:10.1175/1520-0434(1993)008<0533:VAFTAD>2.0.CO;2.
Heinold, B., I. Tegen, R. Wolke, A. Ansmann, I. Mattis, A. Minikin, U. Schumann, and B. Weinzierl, 2012: Simulations of the 2010 Eyjafjallajökull volcanic ash dispersal over Europe using COSMO-MUSCAT. Atmos. Environ., 48, 195–204, doi:10.1016/j.atmosenv.2011.05.021.
Hobbs, P. V., L. F. Radke, J. H. Lyons, R. J. Ferek, D. J. Coffman, and T. J. Casadevall, 1991: Airborne measurements of particle and gas emissions from the 1990 volcanic eruptions of Mount Redoubt. J. Geophys. Res., 96, 18 735–18 752, doi:10.1029/91JD01635.
Kristiansen, N. I., A. J. Prata, A. Stohl, and S. A. Carn, 2015: Stratospheric volcanic ash emissions from the 13 February 2014 Kelut eruption. Geophys. Res. Lett., 42, 588–596, doi:10.1002/2014GL062307.
Langmann, B., A. Folch, M. Hensch, and V. Matthias, 2012: Volcanic ash over Europe during the eruption of Eyjafjallajökull on Iceland, April-May 2010. Atmos. Environ., 48, 1–8, doi:10.1016/j.atmosenv.2011.03.054.
Mastin, L. G., and Coauthors, 2009: A multidisciplinary effort to assign realistic source parameters to models of volcanic ash-cloud transport and dispersion during eruptions. J. Volcanol. Geotherm. Res., 186, 10–21, doi:10.1016/j.jvolgeores.2009.01.008.
Mccarthy, E. B., G. J. S. Bluth, I. M. Watson, and A. Tupper, 2008: Detection and analysis of the volcanic clouds associated with the 18 and 28 August 2000 eruptions of Miyakejima volcano, Japan. Int. J. Remote Sens., 29, 6597–6620, doi:10.1080/01431160802168400.
Merrill, J. T., R. Bleck, and L. Avila, 1985: Modeling atmospheric transport to the Marshall Islands. J. Geophys. Res., 90, 12 927–12 936, doi:10.1029/JD090iD07p12927.
Molteni, F., R. Buizza, T. N. Palmer, and T. Petroliagis, 1996: The ECMWF Ensemble Prediction System: Methodology and validation. Quart. J. Roy. Meteor. Soc., 122, 73–119, doi:10.1002/qj.49712252905.
O’Kane, T. J., M. J. Naughton, and Y. Xiao, 2008: The Australian Community Climate and Earth System Simulator Global and Regional Ensemble prediction scheme. ANZIAM J., 50, C385–C398. [Available online at http://journal.austms.org.au/ojs/index.php/ANZIAMJ/article/view/1421.]
Palmer, T. N., 2012: Towards the probabilistic Earth-system simulator: A vision for the future of climate and weather prediction. Quart. J. Roy. Meteor. Soc., 138, 841–861, doi:10.1002/qj.1923.
Pavolonis, M. J., A. K. Heidinger, and J. Sieglaff, 2013: Automated retrievals of volcanic ash and dust properties from upwelling measurements. J. Geophys. Res., 118, 1436–1458, doi:10.1002/jgrd.50173.
Prata, A. J., 1989: Infrared radiative-transfer calculations for volcanic ash clouds. Geophys. Res. Lett., 16, 1293–1296, doi:10.1029/GL016i011p01293.
Prata, A. J., and I. Grant, 2001: Retrieval of microphysical and morphological properties of volcanic ash plumes from satellite data: Application to Mt. Ruapehu, New Zealand. Quart. J. Roy. Meteor. Soc., 127, 2153–2179, doi:10.1002/qj.49712757615.
Richardson, D. S., 2000: Skill and relative economic value of the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 126, 649–667, doi:10.1002/qj.49712656313.
Rose, W. I., and A. J. Durant, 2009: El Chichón volcano, 4 April 1982: Volcanic cloud history and fine ash fallout. Nat. Hazards, 51, 363–374, doi:10.1007/s11069-008-9283-x.
Scheele, M. P., and P. C. Siegmund, 2001: Estimating errors in trajectory forecasts using ensemble predictions. J. Appl. Meteor., 40, 1223–1232, doi:10.1175/1520-0450(2001)040<1223:EEITFU>2.0.CO;2.
Sears, T. M., G. E. Thomas, E. Carboni, A. J. A. Smith, and R. G. Grainger, 2013: SO2 as a possible proxy for volcanic ash in aviation hazard avoidance. J. Geophys. Res., 118, 5698–5709, doi:10.1002/jgrd.50505.
Straume, A. G., 2001: A more extensive investigation of the use of ensemble forecasts for dispersion model evaluation. J. Appl. Meteor., 40, 425–445, doi:10.1175/1520-0450(2001)040<0425:AMEIOT>2.0.CO;2.
Straume, A. G., E. N. Koffi, and K. Nodop, 1998: Dispersion modelling using ensemble forecasts compared to ETEX measurements. J. Appl. Meteor., 37, 1444–1456, doi:10.1175/1520-0450(1998)037<1444:DMUEFC>2.0.CO;2.
Stunder, B. J. B., and J. L. Heffter, 2004: Volcanic ash dispersion modeling research at NOAA Air Resources Laboratory. Proc. Second Int. Conf. on Volcanic Ash and Aviation Safety, Alexandria, VA, Office of the Federal Coordinator for Meteorological Services and Supporting Research, 105–110.
Stunder, B. J. B., J. L. Heffter, and R. R. Draxler, 2007: Airborne volcanic ash forecast area reliability. Wea. Forecasting, 22, 1132–1139, doi:10.1175/WAF1042.1.
Tupper, A., I. Itikarai, M. Richards, F. Prata, S. Carn, and D. Rosenfeld, 2007: Facing the challenges of the International Airways Volcano Watch: The 2004/05 eruptions of Manam, Papua New Guinea. Wea. Forecasting, 22, 175–191, doi:10.1175/WAF974.1.
Turnbull, K., and Coauthors, 2012: A case study of observations of volcanic ash from the Eyjafjallajökull eruption: 1. In situ airborne observations. J. Geophys. Res., 117, D00U12, doi:10.1029/2011JD016688.
Van den Dool, H. M., and L. Rukhovets, 1994: On the weights for an ensemble-averaged 6–10-day forecast. Wea. Forecasting, 9, 457–465, doi:10.1175/1520-0434(1994)009<0457:OTWFAE>2.0.CO;2.
Wadell, H., 1932: Volume, shape, and roundness of rock particles. J. Geol., 40, 443–451, doi:10.1086/623964.
Warner, T. T., R.-S. Sheu, J. F. Bowers, R. I. Sykes, G. C. Dodd, and D. S. Henn, 2002: Ensemble simulations with coupled atmospheric dynamic and dispersion models: Illustrating uncertainties in dosage simulations. J. Appl. Meteor., 41, 488–504, doi:10.1175/1520-0450(2002)041<0488:ESWCAD>2.0.CO;2.
Webley, P. W., B. J. B. Stunder, and K. G. Dean, 2009: Preliminary sensitivity study of eruption source parameters for operational volcanic ash cloud transport and dispersion models—A case study of the August 1992 eruption of the Crater Peak vent, Mount Spurr, Alaska. J. Volcanol. Geotherm. Res., 186, 108–119, doi:10.1016/j.jvolgeores.2009.02.012.
Webley, P. W., T. Steensen, M. Stuefer, G. Grell, S. Freitas, and M. Pavolonis, 2012: Analyzing the Eyjafjallajokull 2010 eruption using satellite remote sensing, lidar and WRF-Chem dispersion and tracking model. J. Geophys. Res., 117, D00U26, doi:10.1029/2011JD016817.
Webster, H. N., and Coauthors, 2012: Operational prediction of ash concentrations in the distal volcanic cloud from the 2010 Eyjafjallajokull eruption. J. Geophys. Res., 117, D00U08, doi:10.1029/2011JD016790.
Weigel, A. P., M. A. Liniger, and C. Appenzeller, 2008: Can multi-model combination really enhance the prediction skill of probabilistic ensemble forecasts? Quart. J. Roy. Meteor. Soc., 134, 241–260, doi:10.1002/qj.210.
Wen, S. M., and W. I. Rose, 1994: Retrieval of sizes and total masses of particles in volcanic clouds using AVHRR bands 4 and 5. J. Geophys. Res., 99, 5421–5431, doi:10.1029/93JD03340.
Witham, C. S., M. C. Hort, R. Potts, R. Servranckx, P. Husson, and F. Bonnardot, 2007: Comparison of VAAC atmospheric dispersion models using the 1 November 2004 Grimsvotn eruption. Meteor. Appl., 14, 27–38, doi:10.1002/met.3.
Witham, C. S., H. Webster, M. Hort, A. Jones, and D. Thomson, 2012: Modelling concentrations of volcanic ash encountered by aircraft in past eruptions. Atmos. Environ., 48, 219–229, doi:10.1016/j.atmosenv.2011.06.073.
