Review of Ji et al., CLGAN: A GAN-based video prediction model for precipitation nowcasting

Ji et al. present a method combining a generative adversarial network (GAN) with a U-Net and recurrent LSTM cells as the generator, for a high spatiotemporal resolution prediction of precipitation up to two hours ahead. The CLGAN model introduced in this work is compared against a set of baseline models and competing deep-learning based models with a comprehensive set of evaluation criteria. The work is well presented and the evaluation appears comprehensive, and is good for the scope of GMD. I only have minor comments before recommending this manuscript for publication.

General comments:

1.    The authors use an “ablation study” (as defined in the abstract), or more specifically a sensitivity analysis on the weight of the GAN component, to assess the importance of the newly added GAN component for generating forecasts that pass the visual test and are more skillful in capturing the statistical properties of observed precipitation. This is a very useful contribution, although I feel that too many different technical terms are used in different places of the manuscript for describing this process (L8-10; L53-57; L73-74; L331-L342). I recommend the authors start with more general terms (e.g., “weight of the GAN-component”) then use the terms reconstruction los and adversarial loss after they are introduced in L189 eq. (3). Using it too early in the introduction may confuse readers.

2.    Through looking at the code archived on Zenodo there appears to be code from a Git repository. Besides providing the Zenodo DOI could you also provide the GitLab repository link as well? This way potential users can follow the developments of the code and look at the README and documentation easier.

Specific comments:

1.    L73-74: As in general comment 1, try to avoid terms like “adversarial loss” before introducing it. I also suggest being more specific about “sheds light on the interaction between the generator and the discriminator”, e.g., the role of the GAN-component in generating forecasts with closer statistical properties of the observed precipitation.

2.    L161-162: The authors apply stronger weighting on higher precipitation rates to optimize towards heavy precipitation events. How much of an effect (if any) does this have on the precipitation of lighter precipitation events?

3.    L372-374: It is mentioned that CLGAN can provide probabilistic forecasting by adding random noises. There is also mention of using ensemble forecasts to quantify forecast uncertainty in the introduction (L49-51). Is this part of future work or shown in this work? It was not very clear and I had to look if probabilistic forecasts were made in this work.

4.    Figure 1(c). Is L^G supposed to be L^GAN here, to be consistent with text?

5.    Figure 6. For ease to read please label the subfigures (a)-(e) with subtitles (Observation, Persistence, DenseRotation, …)

The paper presents an interesting new method for nowcasting of precipitation, as well as a new precipitation dataset which could be useful for future machine learning applications. The work is thorough and well thought out, and would be suitable for publication in GMD after a few modifications and clarification listed below:

Major Comments

• There are a couple of instances where the paper would benefit from an additional grammar check. This is especially noticeable in the abstract, the first few paragraphs of the introduction and the conclusion. I have highlighted some specific examples in my minor comments, but I would advise the authors to check the prepositions they use throughout the paper.
• I find it odd that in Figure 3, the models cannot outperform the persistence forecast. Surely the ConvLSTM should be able to at least match the persistence as in effect one of its inputs is the persistence forecast. Do you have any intuition as to why you cannot beat persistence?
• In Figure 3c) and 3d) there is a very large degradation in skill at the 20 minute lead time for the ConvLSTM. As far as I can tell this is not discussed in the work and I think it needs to be discussed as it is quite a stark difference.
• Lines 286-288: It would be nice to see some discussion here as to why CLGAN is superior in terms of dichotomous forecast scores but not for RMSE. What attributes does it have or do other models not have, which help here?
• Line 329: You say that CLGAN is doing better than ConvLSTM in the difference plot but it is different to see this in the eyeball norm. It would be useful to have some metrics even if it is just RMSE. Also please clarify what you mean by “fine cells”
• Line 331: The topic change here is very confusing as I thought you were still talking about the case study. Maybe add a sub-seciton title
• Lines 370-371 “It shows…”: This sentence is very unclear. Have you already tried adding additional predictors? If so please provide a reference to this work. If this is future work then the sentence needs to be rephrased because currently it reads like this is a conclusion of the paper
• Lines 371-372 “In addition…”: Please explain further what you mean by expanding to probabilistic forecasting by adding random noise and how you could do this with your model.

Minor Comments

• Lines 6-8: This sentence is very difficult to understand and needs to be rephrased
• Lines 73-74: The last highlight is very difficult to understand and needs to be rephrased
• Lines 100-102: This sentence needs to be rephrased
• Lines 120-121: This sentence needs to be rephrased
• Section 3.1.4: It would be helpful to put references to figure 1 in this section because it is very difficult to follow the CLGAN structure without references to the figure
• Line 242: You are missing the index i in the expression and it is very confusing to have the forecast expression with the observation expression in brackets. You should separate out the two expressions.
• Figure 4: Please clarify in the text and figure caption what the box and whiskers represent
• Figure 5: Add units to the legend and mention what the legend is in the caption
• Figure 6: It is interesting that ConvLSTM does better at longer lead times here given that it does worse at longer lead times in the metrics. Do you have any intuition on this? Is it just a quirk of the case study you chose?
• Line 344: This line needs to be rephrased
• Line 358: More potential than what?
• Section 5: It would be interesting to have a comment about how you think your model would perform at longer lead times (say 6hrs). Would you still see such good results?