Summary

This paper describes a method for assessing the likelihood of ICI derived from geostationary satellite observations and derived products, coupled with cloud water content estimates derived from the CloudSat/CALIPSO-based DARDAR product.  ICI and HIWC often occurs within deep convection, though it has also been observed within mid-latitude frontal cloud bands, and represents a significant hazard to aviation.  Machine learning identified the most important metrics for diagnosing ICI that are combined to estimate a (daytime only) ICI/HIWC likelihood for the MSG SEVIRI imager, that has been validated with a subset of DARDAR not used in model training.  Performance is fairly comparable to existing methods, though with slightly weaker validation stats. The product is then validated with DARDAR and compared with a small sample of European ICI events encountered by in-service Lufthansa aircraft.

The authors have a clear understanding of the ICI/HIWC hazard, existing satellite-based methods from the literature focused on diagnosing ICI/HIWC, machine learning best practices, and the most appropriate geostationary parameters for diagnosing this hazard.  The paper is clear and well written.  My concerns with the paper begin with the exclusive focus over Europe.  Convection over Europe is relatively infrequent compared with Africa or the tropical Atlantic and typically weaker (with warmer tops) than these regions. Figure 3 shows several African/Atlantic Lufthansa ICI events that could have been studied, which if included would increase confidence in the method’s global applicability.  DARDAR likely viewed many intense storms over Africa as well that would serve as excellent training for the model.  Another concern is the fact that there is no attempt to try to develop/validate a model that can operate at night. Aside from tau and visible reflectance, all other parameters are available at night. I would like to see a night-time product demonstration.  Third, there was an international HIWC/HAIC field campaign based in Cayenne, French Guyana in 2015 that was within the SEVIRI field of view.  IWC data was collected at 5 sec intervals from 2 aircraft which would be an extremely robust dataset for model validation.  The authors should explore this data as it has been a number of years since collection and the data should be freely available by now. I have a number of other more minor comments/concerns listed below

Given that the paper and its writing are of high quality, but due to the significant concerns mentioned above, I say the paper is acceptable but with major revisions to address these concerns.

Specific Minor Comments/Concerns

Sections 2.1.1 and 2.1.2, the differences between CIPS and APICS, and the ramifications of these differences on the analysis are not very clear.  I see both produce optical depth but you use the optical depth from one model for water cloud and the other for ice cloud.  Additionally it is not explained why you are not using the cloud product data operationally generated by EUMETSAT which would make your method more easy to apply by others in the community. 

Section 4.1, I don’t think you need to use paper space to define very commonly used validation metrics.  You could simply cite the Wilks meteorological statistics book and move one Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed. International Geophysics Series, Vol. 100, Academic Press, 648 pp.

Validation stats in general, it would be interesting to see the validation applied to > 1.0 g m^-3 data in addition to > 0.5 as the higher value is likely to be more consequential for aircraft.

Figure 7 and many other mapped data figures (i.e. Figure 10), the mapped product is very hard to see details of.  For Fig 7, I recommend you make the map much larger and place below the curtain plot data. For Fig 10, consider enlarging the graphics as I cannot see details when printed out on paper. 

Figure A.3, there is an extremely odd look to the HIWC product with a discontinuity at 49.3 N latitude.  What is the reason for this?  Figure A.7 has an odd diagonal discontinuity too. 

All Figures in Appendix, what is the purpose of plotting the wind information on the maps?  It seems like an unnecessary detail that adds clutter to the map.

General Comments

In this paper, Aricò et al. present a method for detecting high ice water content (HIWC) conditions associated with aviation ice crystal icing (ICI) events using geostationary satellite data. They apply a machine learning approach to derive a set of input variables from Meteosat Second Generation (MSG) SEVIRI derived products. Training of a random forest is accomplished using CloudSat radar and lidar (DARDAR) products. Finally, the authors obtained a database comprised of in-service ICI events which they have used to asses performance of their detection method.

The paper is clearly written and well-organized. It begins with a comprehensive description of the ICI threat to aviation and a review of previously published HIWC detection methods. It provides thorough descriptions of the various data sets used. Explanations of the random forest approach, why it was chosen, and how it was applied to this problem are detailed and clear. Results presented are meaningful, and conclusions are well-justified. Significantly, the authors have demonstrated the need for a dedicated HIWC detection product by showing displacement between their HIWC Mask and detected convective cells. Overall, this is a very good paper that needs only minor revision.

Specific points:

1. If I understand correctly, the method is limited to cruise altitudes, but it’s not clear why. ICI engine events have occurred during ascents and descents, so the hazard is not limited to cruise altitudes.
2. While I understand why the method is limited to daytime, given its reliance on products derived from visible channels, I think it would be appropriate to at least discuss how you might develop a corresponding nighttime method. 
3. It’s difficult to see the symbols in certain figures (e.g., Figs 1, 7, 9, 10). The images and symbols could be enlarged and/or the symbol color could have better contrast with the background image.
4. Verification against in-service ICI events is very important, but the Lufthansa database apparently only includes air data system (ADS) events (e.g., TAT anomalies), not engine events. I assume the authors did not have access to the latter. Some discussion of the relationship between ADS events and engine events would bolster the significance of your results.
5. In the box and whisker plots shown in Fig. 6, some of the variables on the vertical axis only have outlier points, i.e., no box and whiskers. Could you explain how this should be interpreted?