see attached file

see attached file.

The paper, “Ice water path retrievals from Meteosat-9 using quantile regression neural networks,” develops a new approach toward estimating cloud ice water path during any time of the day using a machine learning method. The technique is trained with matched SEVIRI pixels and DARDAR profiles using a quantile regression neural network that permits an estimate of the uncertainty for each retrieval. Two versions are applied, a single pixel method (MLP) and a single pixel plus surrounding pixel data, a convolutional neural net (CNN). The latter was found to be superior to the former in that the average uncertainty was reduced, although the CNN tends to “smear” the IWP signal across neighboring pixels. Three input datasets were tested: VISIR (daytime) using all but one SEVIRI channel, an infrared (IR) only method using all non-solar SEVIRI channel, and a subset IR case using only two channels to simulate an historical Meteosat imager. All were trained using DARDAR data and their results were compared to DARDAR data taken within the same time frame as the training set. The VISIR performed best, but is limited to daytime and would require training for various solar angles that are not available for the DARDAR. The IR input produced quite acceptable results that are consistent in relative terms with the daytime portions of diurnal cycles of IWP determined from other passive sensor methods. The IR subset input shows less skill but provides information that is not obtainable with more physically based retrievals. This approach shows promise for improving the estimation of IWP at all times of day.

I recommend publication with a few revisions.

1. It would help in section 2.3 to use the same terminology in the text and Figure 1 description of the collocation. “Cell” is only mentioned in the caption, not in the text. SEVIRI “pixel” is used in the text. Also, the caption should note the units used for lat and lon, as degree is the usual unit. Is it correct to assume that the SEVIRI pixels used for given image were at least 16 km apart?

2. The units of the statistical parameters in Figure 5 and 7 are given in kg m^-3. That is good. But, the mean DARDAR value should be noted for each plot, or the values given in percent of the mean DARDAR value in the text.

3. Line 280: By stating that “CLAAS has been thoroughly validated” suggests that the CLAAS IWP values agreed well with actual IWP measurements. The cited studies showed that CLAAS agreed well with similar passive remote sensing techniques, but not particularly well with DARDAR, the “ground truth” used here. For the most part, the CLAAS values are significantly lower than their DARDAR counterparts, as indicated later in the discussion. As the results shown in the citations vary, and none arise from an actual comparison with any in situ data, “thoroughly validated” is a bit of an overstatement. I would suggest rewriting this line, so that it is no surprise to find the CLAAS mean running below that from the CNN in Fig. 9.

4. Figure 10. This plot is difficult to examine closely. I think it would be easier to compare the two methods by putting them on the same graph with the two scales, and maybe only using 4 months instead of all twelve, just to illustrate the relative consistency.

5. Mean mass height (Zm) and mean mass size (Dm) are derived with somewhat mixed results. It would be helpful if the authors could remind us of the importance of these parameters.