|This manuscript describes an effort to combine satellite observations and model simulations of volcanic ash to derive a best estimate of the ash concentration evolution as a function of 3D space and time. The Eyjafjallajökull eruption of 2010 is used as a test case. The ash transport model LOTOS-EUROS is used, along with ash loading estimates from the SEVERI satellite instrument. An ensemble data assimilation (DA) framework is used, with ensemble members branched with different plume heights, representing the uncertainty in this parameter. The results of the DA appear to produce much better results than the ensemble mean of the pure model simulations.|
DA is a powerful tool for forecasting, and it is increasingly being explored as a method to be used in forecasting the evolution of volcanic ash plumes. The subject matter is of relevance to ACP.
However, the manuscript does not adequately describe a significant advance in the development of volcanic ash DA. DA has already been shown to improve the forecasting of volcanic ash transport (e.g., Fu et al., 2015, 2016). The main contribution of the paper appears to be the introduction of a satellite observational operator (SOO), which is not convincingly needed or effective. The authors also claim to quantify the duration of forecast improvement brought about by DA, but the single case study explored here, especially in the manner presented, is not suited to any general conclusions about the timescale of volcanic ash forecast skill.
The authors state that the satellite observations “are often two-dimensional (2D), and cannot easily be combined with a three-dimensional (3D) volcanic ash model”, which motivates their development of a satellite observational operator (SOO). In fact, this issue is common to many DA applications using satellite measurements. Satellite instruments using nadir viewing geometry measure 2D fields. The standard technique in DA is to apply an observational operator (H) to the model output, to be able to compare model and satellite data in measurement space. Satellite observational operators (SOO) are sometimes used to transform observations closer to model space before assimilation. Using an SOO simplifies some aspects of the DA, but does so at the cost of including some set of assumptions in the SOO. In the present case, I find the motivation for the use of a SOO to be lacking (a mismatch between 2D and 3D fields is not sufficient).
Furthermore, the assumptions that are included into this SOO appear to be questionable. The authors argue that volcanic ash clouds have thicknesses between 0.5 and 3.0 km, although they quote studies having measured even smaller thicknesses. Given the relative large range, it is not clear why even smaller thicknesses are not included in the range. Then they calculate a range of possible concentrations by dividing the satellite retrieved ash loading (in g/km^2) by thicknesses between 0.5 and 3.0 km. They take the mean and SD of these concentrations as their best guess (and uncertainty in) ash concentration. This seems like a complicated way of assuming a mean thickness of 1.75 km. Then the authors focus on the ash concentration at cloud top, which appears to be the quantity which is compared to the model output within the DA (although this is not sufficiently explained). There is no real justification for why the focus is restricted to the concentration at cloud top, and this process seems like it discards a significant amount of useful information (e.g., the total mass loading).
Most of these assumptions and complications would be avoided if measurements and model results were compared in measurement space rather than model space, as is the standard in DA. This would simply require the observation operator H to take the vertical integral of the modeled ash profile.
Fig 1a shows ash mass loadings from SEVERI, with values between 0 and 5 g/m^2 (although values could be higher). Taking 3 g/m^2 as a typical value, and under the assumption that cloud thicknesses are between 0.5 and 3 km thick (and that concentrations are uniform within the cloud), one would expect typical concentrations between 1-6 mg/m^3. The SOO extracted measurements in Fig 3a are around an order of magnitude less than this, with most values falling between 0.1-0.6 mg/m^3. The rough concentrations estimated here (1-6 mg/m^3) are also more in line with the model results shown in Fig 4. Given the apparent simplicity of the SOO, it should be possible for the reader to gain a quantitative feel for the magnitudes of the concentrations involved. At present, this is not possible. If the SOO is more complicated than I have understood it, then the description needs fair amount more detail.
There is only a small amount of evidence shown to support the conclusion that “satellite data assimilation can force the volcanic ash state to match the satellite observations, and that it improves the forecast of the ash state.” Fig 5 shows a snapshot of ash loadings at a single time, comparing the SEVERI measurements with results from the model with and without assimilation. While the assimilation does seem to correct a rather large bias in the pure model output, it’s not clear that the DA result is better in all locations: while the values over the Netherlands may be more realistic, it appears that the DA results over the west coast of Norway may be underestimated by the DA system. Whether the authors have chosen this single comparison randomly, or if this is a best- or worse-case is not stated. One would expect some overall measure of skill with respect to time and space for such a comparison, in order to gauge the impact of DA.
Apparently, DA has already been shown to improve the forecasting of volcanic ash transport (Fu et al., 2015, 2016). The present study claims to quantify the duration of such improvement. Fig 6 shows a comparison of DA forecast results with in situ airplane measurements. Again, the DA initiation helps to correct for a major over estimation by the model without DA. The authors conclude from this comparison that with initial conditions taken from DA, the “effective duration of the improved regional volcanic ash forecasts is about a half day”, although this is a “conservative” estimate, which is based only on this single case, and is in fact limited by the 15 h duration of the airplane measurements. It’s not clear why the continuous satellite measurements were not used to quantify the duration of the improved forecast. The duration of the skill improvement also likely has much to do with the assumptions regarding the noise added to the plume height estimates. A convincing assessment of the skill duration would require a much more thorough and detailed analysis.
P1, L1: “Data assimilation” is very general, the focus on ash forecasting should be clearer.
P1, L5: It doesn’t appear that cloud thickness *data* is being included in the DA. Data-based assumptions, maybe.
P2, l4: A number of eruption source parameters are mentioned here, but only plume height is included in the DA ensemble generation. Some discussion would be nice at the end of the paper on the potential impact of uncertainties in the other ESPs.
P2, l10: Actually, you would hope that the VATDM was fairly accurate, and the DA was used because of unknowns in the ESPs.
P2, l11: DA does more than provide the initial conditions.
P2, l17: It’s not clear if these measurements are specific to Eyjafjallajökull, or more general.
P2, l25: should mention that the satellite is geostationary here.
P2, l26: Earth’s
P3, l6: Following from general comment, sparseness of data is not a real impediment to its use in DA, in fact, one could argue that sparse data is exactly what DA is useful for!
P3, l9: This discussion seems to assume that ash clouds have very well defined edges, but could there not be cases where the ash concentration decays smoothly over some vertical range? How is the thickness defined in such cases?
P3,l10: “hundreds of meters”
P3, l16: “Kasatochi”?
P3, l16: Personal communication should probably list from whom the information is from.
P3, l21: Whether thin ash clouds are less of a concern to aviation or not does not necessarily make them less of a concern for a data assimilation system.
P4, l12: “As a parameter…” The meaning of this sentence is hard to understand.
P4, l15: Also unclear, in the previous paragraph a number of data are described as SEVIRI products.
P5, l25: “ML_blue” is not an intuitive terminology, something physically motivated would be much better.
P6, Eq. 7: Where does this formula come from? Usually, uncorrelated errors add in quadrature. On line 13, a “conditional probability relation” is mentioned, but this does not make Eq. 7 easier to understand.
P7, l12: This description of the plume height needs much more detail to be understandable.
P8, l6: “… which is mainly due to lack of sedimentation processes” scared me quite a bit, but Fu et al. (2016) only state that the model contains sedimentation processes, but not coagulation, evaporation and resuspension. I guess assumptions about the size distribution are also important here.
P8, l7: Using DA to correct for model overestimation is a rather unphysical “solution”. It would be more satisfying to improve the physics of the model if in fact this is the source of the over-estimation.
P8, l26: This is not a very convincing validation of the SOO.
P8, l27: Is there a reason why SEVIRI mass loading retrieval error and the standard deviation of the mass loadings should be similar? The reason is not apparent.
P9, l3: I don’t see why satellite measurements wouldn’t be better to test the duration of improvements to forecasts from DA. That satellite measurements have “big uncertainties” seems a strange argument, given that the preceding portions of the paper have used satellite measurements for the DA. While I don’t mean to play down the importance of in situ measurements, in this case, the airplane flight track sampled only the very edges of the old ash plume, with concentrations orders of magnitude less than the peak values and much lower than the threshold for airplane safety. Therefore, the accuracy of the forecasts at the times and places of the in situ data used here seem to be of little importance to the question of how good the model is at predicting ash concentrations dangerous to airplanes.