Review of: “Exploiting radar polarimetry for nowcasting thunderstorm hazards using deep learning”.

In this paper, the authors investigate to what extent additional polarimetric variables form dual-polarization radars can improve nowcasting of thunderstorm hazards. To do this, they add the polarimetric variables and quality information as additional inputs to the existing deep learning-based nowcasting framework by one of the authors. The novelty lies in (1) the use of not just the CAPPI dual-pol variables, but the volumetric data from multiple altitudes that are combined using a weighting function, and (2) the forecasting of multiple hazards, that is, lightning and hail, in addition to (heavy) precipitation.

The authors find that the impact of polarimetric variables and quality information in addition to radar reflectivity is not unambiguously positive. While improving heavy precipitation forecasts for the next hour, the hail forecasts are slightly degraded by adding polarimetric and quality information. The authors suggest that this is potentially due to overfitting as the multiple information sources may contain redundancies.

General comments

The authors present relevant and new scientific results, which fit well within the scope of the journal.  

Overall, the level of is English is good but the clarity and flow of the text can be improved, for example the following sentence is unclear: “... due to high demand in computation time results, are not available in real time.” or L122": "The main difference between the predicted thunderstorm hazards is that heavy precipitation is trained ... “. Some formulations are ambiguous or confusing and should be improved e.g. L174 “The average loss of lightning”.

I missed the information on how the train-test split was done (randomly, different time periods?). Did you make sure to include all kinds of events (and non-events) in both? Please provide some more details on the learning process (learning rate, stopping criterion etc.) which are crucial pieces of information for this kind of research. 

I would also suggest that the authors better motivate the limited lead time of one hour. Is this enough to act upon? Or is it motivated by an inherent predictability limit to the phenomena you are trying to forecast? Moreover is there a reason why rain gauges are not used as an input, or for validation of heavy rainfall?

It would be useful for the authors to also discuss the added value of their DL-based methodology as compared to more traditional methods in terms of computational performance. How long does it take to run this 1-hour nowcast?

Specific comments

L22: The sentence “the initial state of the atmosphere in NWP assimilation is based on previous model predictions rather than the latest available observations, which makes it less suited for accurately predicting the time and location of convective storms”: appears to stem from some misconception. Indeed, the previous forecast is used as a first guess or background field, but this is in fact combined with the latest observations to create an initial state in the data assimilation process. The main reasons why nowcasting is important is because of the faster computational time which allows a higher update frequency, and because a purely (or mostly) observation-driven system will (by construction) be closer to observations for the shortest lead times.

L37: The rainfall fields on which these nowcasting systems are based typically already make use of dual polarization variables to estimate the rain rate, clutter etc. What you mean is that you explicitly add polarimetric variables in the nowcasting scheme.

L58: "dataset from Leinonen et al. (2022b)”- can you provide a bit more detail here (which variable? Radar rainfall dataset?)

L71: “Weather radar (R) observations”- please be more specific. I suppose you mean reflectivity, but there are many more weather radar observations (e.g. radial velocity…)

L93: “After hyper-parameter turning, a value of -0.5 for β was selected, as Wolfensberger et al. (2021) found that this resulted in the best parameter value” - it’s unclear to me whether the value selection was the result of hyperparameter tuning or just taking the best value from literature? Also, “this resulted in the best …”should probably read “this was the best…”

L94: Next, the data was transformed by first normalizing it by bringing the mean close to 1 -> How was this "brought close to 1", why not simply set it to 1? 

L127-128: "The focal loss is an adaptation of the CE and focuses more on the difficult cases" - this is not informative, please be more specific or leave out the last part of this sentence. 

L129: "We trained each possible combination of data sources..." - you cannot train data (or sources, or variables), you train a model. 

L130: trained only 3 times: is it enough?

L134: Shapley value, which distributes the total score among its predictors... -> In game theory it is used the score this way, but there is no "score" in the picture here, please rephrase.

L138: You convert the probabilities to binary values, why? Since both the network and POH algorithms output a probability, why not use a metric that quantifies PDF overlap/mismatch?

L139: What do you mean by “predictability”?

Throughout the manuscript, “radar data” is used to refer to reflectivity data. This is confusing, since polarimetric data also comes from radars. Please change this throughout the manuscript (e.g. L189 “radar is the most important source…”-> “radar reflectivity data is …”) also in the captions (e.g. Fig. 3) 

L200: Please specify which thresholds you mean (probability? intensity?)

L213: “resulting in penalization”-> what kind? Can you be more specific (double penalty?)

L215: <FSS for> RPQ and R

Typographic, grammar and other small corrections

Abstract and L14: humans -> human lives

L21: NWP <models> often have

L33: In the recent years -> in recent years

L63: Operational<ly> available products

L186: Lie more apart -> Lie further apart

L229: stresses out -> confirms, shows

L232: can not -> cannot

The paper investigates the potential enhancement of nowcasting accuracy for hail, lightning, and heavy rainfall over Switzerland by incorporating dual-polarization radar data as supplementary input. Lead times of up to 60 minutes were considered in this analysis. The nowcasting framework employed a preexisting convolutional neural network. The findings indicate that, in comparison to conventional radar data, the inclusion of polarimetric variables enhances the nowcasting performance for heavy rainfall and lightning prediction, although no significant improvement is observed for hail.

While the paper is commendably well-written and organized, certain sections suffer from limited clarity and detailed information, thereby posing challenges for readers attempting to grasp the complete essence of the content. I recommend the paper for publication EGUsphere if the minor revision points listed below are taken into account.

Minor revisions:

• Some parts of the paper are difficult to understand without consulting the papers by Leinonen et al. (2002a, b, 2023). Of course, it is not useful to repeat all the details from those studies, but including the essentials would be very helpful for the reader.
• Abstract: The first five sentences are more of an introduction than an abstract; consider shortening this part and adding some more details about your specific work.
• L18-19: NWP models are useful not only for stratiform precipitation, but also for convection. In particular, high-resolution EPS and rapidly updated cycle (RUC) models are quite good at predicting convection.
• P1, last paragraph: It is unclear what is meant by “These models…are not available in real time.” Besides, the general statements about NWP models do not hold true for RUC)and ensembles EPS forecasts.  
• L54: but also heavy rainfall? Figure 2 shows the result for nowcasting precipitation based on dual-pol variables.
• Introduction: Can you describe the objectives of your study in more details? Only one sentence (L48) is too short.
• First paragraph of Section 2: A period of 5 months is very short. The affiliation of the authors suggests that they have direct access to the data. So why didn't you consider a longer period? In any case, at least in the conclusions I would expect a discussion of the reliability of the results given the short time period. Finally, please state the training period of the model and at least briefly state the data used for the model.
• L64 and Figures 1, 2: “…maximum range of observations is 246 km…” is unclear. Do you mean the study area (shown in Figures 1 and 2)? Where is the location? But why didn't you use the whole radar range? You should also explain that your study area is different from the one used by Leinonen et al. (2022). Perhaps an additional figure would help.
• L105: Where does the 8 km distance come from? Have you performed sensitivity tests with variable distance?
• L112: The classes for precipitation totals are rather coarse. Can you comment on this?
• Section 3.1: Could you add some more (mathematical + theoretical) details of the model used, so that a reader not familiar with CNN can get the gist?
• Section 3.2: For the interpretation of the Tables and because the Shapley score in not well known, it would be very helpful to give the range of values and their interpretation.
• L138: A threshold of 50% for POH makes sense, but it would be very interesting to see how the results would change if the probability were higher (note that several studies have found a POD of ~30% for a POH of 50%, which means that POH = 50% means <40% really hail on the ground).
• Figures 1 and 2: Please insert the units of the color bars; it does not make sense to show ZDR in logarithmic units (values can also be negative)
• L183: It’s very interesting that the skill for heavy rain is increased when using polarimetric parameters, but not for hail. Are there any meteorological reasons for that (you may speculate a bit)?
• Section 4.3. Again, I miss some interpretation of the results (try to give answers or speculate about the why of the results).
• L200: Can you specify the different thresholds considered here?
• L209: “…time and space scales of the target variables” not sure on this. Lightning and hail have smaller spatial and temporal scales compared to precipitation. So I would expect a higher skill for precipitation compared to the other two.
• Conclusions: This section is rather short. Consider expanding it with more substance.

Questions/Edits/Typos:

L13: Not the convective storms can turn into flash floods, but the associated heavy rainfall

L15: reformulate “…by these weather phenomena”; in this sentence, these refers to flash floods (object of the last sentence); but as you know, hail in Switzerland causes the largest economic losses.

L20 “…time results…” delete results

L21: NWP models

L24-25: nowcasting is simply warning or immediate rather than early warning

L31: “…to take the life cycle of convective cells with growth and dissipation …”

L44 and others: the term “hazard” represents the potential for harm of a certain phenomena. As this is unclear in your manuscript (no information about hail size, rain intensity), I would suggest to replace “hazard” by “phenomena” when used in conjunction with hail or precip

L48: delete will

L49: specify “model” e.g., convolutional neuronal network model

L53: I doubt whether you really retrieve relevant information about microphysics; from the dual-pol radar you can get information about hydrometeors and their characteristics and not about physics (o.k., the latter could be true, but requires complex post-processing which is not mentioned in the paper)

L59 and others: be consistent in the use of “data”: either singular or plural, but do not mix.

L71: I would be more specific here as weather radar observations may also include polarimetric variables

L73: is maximum echo maximum reflectivity? And what is meant by maximum, CAPPI or the maximum in overlapping areas?

L87: either use (plural) or used

Eq 1: Check the dimensions of the equation. Are VIS and w dimensionless?

L92: what is “slope of the exponential”?

L120: I suggest to refer here directly to the target values lightning, POH, and CombiPrecip

L130: again, what are training and application period?

L137: “ground truth” for hail is weird given the fact that POH is obtained from an integral bulk (reflectivity) only, measured aloft, and does not consider horizontal drifting between the height of the radar signal and the ground

L146: “…is an imbalance…”

Figures 1/2: can you write a few words about the weather situation on that day? Please indicated the date.

Figure 3: A continuous color scheme for the colorbar makes no sense here.

L202: “…lead times the skill of the…” “…while for hail the values drop…”

L205: indicats; nowcast --> predict

L209: “…than for lightning…”

L2010: Reference not in brackets

L2012: lightning (plural does not exist); PR, AUC, and CSI

L216: “…while for hail it we find…” delete "it"