Joint Retrieval of Multiple Species of Ice Hydrometeor Parameters from Millimeter and Submillimeter Wave Brightness Temperature Based on Convolutional Neural Networks

Round 1

Reviewer 1 Report

In this manuscript, several convolutional neural networks (CNNs) to estimate ice-related parameters from submillimeter-wave passive satellite observations are presented.  The neural networks used in the manuscript are advanced and the overall methodology is sound, but there are several major deficiencies that make the manuscript not suitable for publication in the current form.  Specifically,

1)    No details regarding the microphysical scheme used in the WRF simulations are provided.  Historically, numerical models tended to produce excessive amounts of large precipitation-sized ice particles (Lang et al. 2011), and although some of these biases have been continuously reduced, they are still likely to exist and be positive.  It is thus surprising that although the amount of ice produce by WRF in Figure 2 is larger than that from the ATMS retrievals, the associated scattering signal is weaker. 

2)    No details on the calculation of the electromagnetic scattering properties of ice particles are provided.  The fact that the brightness temperatures calculated from the ATMS products exhibit more scattering may be a result of biased electromagnetic scattering calculations, and not a WRF deficiency.  In fact, the distribution of ATMS graupel in Fig. 3 is suspicious, with significant amounts of graupel above 400 mb (which is about 7.0 km above the ground).  Graupel forms when supercooled water droplets (at a temperature below 273.15) freeze onto a snow crystal, a process called riming.  Consequently, most of the graupel production occurs near the freezing level (where the largest amounts of supercooled water exist), and, although it may be transported vertically by convective updrafts, graupel tends to peak near the freezing level (as correctly reproduced by WRF) and not above 400 mb as in the ATMS retrievals.

3)    The manuscript does not provide any information on the size of the training dataset. The validation dataset consists of eight cases.  However, the size of the training dataset is, by all indications, rather small, as the number of tropical cyclones in a 6-year period cannot be very large.  The authors mention (without explanations) the use of rotation in extending the dataset, but CNNs are rotation invariant, and, therefore, rotations are not likely to significantly add to the original data.  Given that CNN process entire scenes and no individual pixels individually, it is very likely that the training data set is small and potentially insufficient to train a robust model.

4)    The merging of WRF and ATMS data is unusual and likely to introduce artifacts in the retrievals.  Specifically, numerical weather model simulations of real events may be characterized by large phase-errors (with predicted features being significantly displaced or distorted relative to reality).  Replacing hydrometeor fields by other estimates, but potentially at significantly different locations, could cause serious artifacts.  For example, the predicted position of eyewall may be misplaced by 100 km.  A naïve replacement of the model graupel may result in significant rain at the surface with no graupel aloft, and significant amounts of graupel with no rain beneath.

Minor Comments:

1)    Ice, snow and graupel represent only one phase, ice.  They constitute multiple species of hydrometeors differentiated by size, density, shape, etc.. But the phase is the same.

2)    Figure 3.  What does serial number means here (and everywhere where serial number is used)? The serial number concept is usually used in manufacturing to differentiate a product from another one produced in the same factory and identical in appearance (but different for practical purposes).  What is the meaning of the x-axis variable? What are the units in this and the previous figure?

3)    Citations do not follow the correct protocol.  Fuzhong Weng is Weng and Grody, Pingyi Dong is Dong et al., etc.

References

 Lang, S. E., W. Tao, X. Zeng, and Y. Li, 2011: Reducing the Biases in Simulated Radar Reflectivities from a Bulk Microphysics Scheme: Tropical Convective Systems. J. Atmos. Sci., 68, 2306–2320, https://doi.org/10.1175/JAS-D-10-05000.1.

The use of English is appropriate.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

General comments:

The manuscript entitled ‘Joint Retrieval of Multiphase Ice Hydrometeor Parameters from millimeter and submillimeter wave brightness temperature Based on Convolutional Neural Networks’ by Chen et al. presents a method for retrieving ice hydrometeor content through millimeter and sub-millimeter wave brightness temperature. The work is interesting, but modifications are needed before formal publication.

Specific comments:

The details of the WRF model need to be elucidated, including the region, grids, resolutions, and particularly the parameterization schemes, especially the cloud microphysics parameterization scheme.

The verification of simulated brightness temperature in Section 2.1 is insufficient. Drawing conclusions and deciding on methods based solely on a single case is highly unrigorous; additional statistical information needs to be provided.

Why focus only on the tropical cyclone process? What about ice clouds at other times? While reading the paper, there is no mention of tropical cyclones in the Abstract and Introduction, so it is confusing when the method selectively screens only for tropical cyclone processes.

Was the choice of using WRF intended to provide information on ice and snow because ATMS can only provide information on graupel? The paper also mentions significant differences between WRF simulations and real conditions. In light of this, why was WRF still chosen as the basis for dataset creation? The paper does not provide an explanation or clarification. Additionally, why not manually vary the content of the three types of ice hydrometeors to establish the dataset?

Sensitivity experiments need to be conducted to demonstrate the sensitivity of the established CNN model to brightness temperature observation errors.

In the results comparison section, it is desired to include more intuitive scatter plots for a visual comparison and presentation of the results.

Minor comments / typos:

The citation format used in the paper needs to be carefully revised (e.g., line 73, 76, and 78.).

Line 74, the definition of Dme should be introduced.

Figure 3, what is the meaning of the label of x-axis, ‘Contour of the serial number’?

The text in Figure 15 is too small to read.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have addressed all my comments, and I recommend publication of the manuscript.

Reviewer 3 Report

I appreciate the efforts made by the authors to clarify a few confusing points and improve the manuscript based on comments/suggestions. 

I believe the manuscript is now of much better quality and is suitable for publication as is.