GAN-argcPredNet v2.0: A Radar Echo Extrapolation Model based on Spatiotemporal Process Intensification

By

Kun Zheng, Qiya Tan, Huihua Ruan, Jinbiao Zhang, Cong Luo, Siyu Tang, Yunlei Yi, Yugang Tian, and Jianmei Cheng

This paper presents deep learning-based precipitation nowcasting algorithms. In particular, they employ GAN-based network with several attention mechanisms to tackle the well-known problem in machine learning-based precipitation nowcasting, i.e., weakening of rain as the forecast time increases. The results presented in this manuscript are very promising. However, the quality of writing is poor at this moment. Scientific logics are not well considered, and the sentences are redundant. There are so many figures, but the descriptions and discussions on these figures are very limited. Most of the figures are difficult to read. Therefore, I recommend significant revisions before rendering a decision.

Major comment:

1. The structure of the manuscript is not well organized and redundant. For example, the last two sentences of the introduction section (i.e., Section 1) mention about the experimental setup and the results. They should be in the dedicated sections for the "experimental setup" and "results." The first paragraph of Section 2 also repeats the same thing as that in Section 1.

2. Line 126, "STIC-Prediction generator reduces information loss...": This is not proven yet up to this line in the manuscript, nor supported by previous studies (indeed, no citation here). Therefore, this is just your working hypothesis at this moment. If you intend to prove that STIC-Prediction generator reduces information loss, it should be in the results section, not here. Besides, you need to define a metric for information loss. Otherwise, you cannot prove it quantitatively. I would recommend writing as follows: "STIC-Prediction generator is designed to reduce information loss and...". Other sentences in the same paragraph have the same problem. Please clearly separate introduction, experimental setup, result, and discussion.

3. Line 153-157: This is the same as the previous comment. This is your working hypothesis until you present evidence for that.

4. Line 166, "By using hard_sigmoid as the activation function, the training speed is accelerated.": This is interesting, but not supported by evidence. You may accelerate the training near the origin, where the gradient is non-zero, while you will decelerate the training where the gradient gets exactly zero.

5. Page 8-10: In general, descriptions lack important information, such as the number of channels, padding, stride, batch size, etc. All these hyperparameters affect the model performance. Comparing different models without knowing setup of each model does not really make sense. At least, you should provide details of all the models you used in this paper. If it is too big, you can put them in the appendix or supporting information.

6. Line 219-221: It is a little bit strange to conclude before showing results.

7. Section 4.1-4.3: They are the experimental setup. I would recommend to separate the experimental setup from the result section.

8. Equation 15: This is somewhat strange. Usually, radar reflectivity is converted to rain rate with the equation as follows (e.g., https://glossary.ametsoc.org/wiki/Z-r_relation):

Z = a * R^b.

Then, the radar reflectivity is expressed in decibel (dBZ, https://glossary.ametsoc.org/wiki/Dbz):

dBZ = 10 * log10(Z).

Although Eq 15 is the same as that in Shi et al. (2017), your definition uses log, not log10. If that is the case, your rain rate may be wrong. Please check your code and reprocess data if necessary. Despite the wrong definition, the values in Table 1 seem correct. In addition, the citation just before this equation (Watters et al. 2021) seems misplaced. I do not understand why this is cited here.

9. Line 261, "From Fig. 7, 8, 9, 10, and 11": Although you placed 5 figures with 4 panels each in the manuscript, you just describe them in a single line. This is not acceptable. Please add meaningful descriptions and discussion for them, or please consider reducing the figures.

10. Figures 7-11, 13-17: These figures have 4-5 lines each, but they are not clear, and they are drawn with similar colors (many of them are in blueish colors). Please improve their quality. However, before improving them, please consider summarizing them more concisely.

11. Figure 12: You highlighted an intense rain area near the center of the domain, which is predicted well by the proposed method. Meanwhile, the rain area on the bottom right corner over-intensifies in the proposed method. The rain area near the top right in the domain goes out of the domain, which is different from the ground truth. In other methods, this rain area is located at the right place. I would say that the proposed method has pros and cons. Please discuss the results more carefully. Not just saying that you make a great system, but a more scientific consideration is needed. I imagine that the over-intensification near the bottom right may be related to the use of attention mechanism.

Minor comment:

1. Line 20, "intensification of echo feature sequence": It is not clear what you intensified. If you always increase the extracted features by CNN, I do not think you can make a good prediction. I guess you meant the attention mechanism, but this sentence did not make sense.

2. Line 45, "Existing deep learning models, ...": It is impossible to prove non-existence, so I would recommend adding "to the knowledge of authors."

3. Line 77, "these traditional methods fail to utilize": Most of the traditional methods are semi-process-based, so they do not intend to utilize historical data directly. Therefore, it is not fair to use the phrase "fail to."

4. Line 130, "where H, W and C denote": "T" and "l" are missing.

5. Equation 9: \sigma before \phi_21 (there are two) must be \gamma.

6. Table 4, the values for 5 mm/h, FAR: CS-GAN and GAN-argcPredNetv2.0 show the same number within significant digits. Therefore, both should be presented with bold.

7. Line 339, "However, there are further improvements on the basis of current accuracy.": I could not understand the sentence. Please consider rephrasing.

8. Line 341, "hardware conditions": What does this mean?

This article presented an application study of GAN-argcPredNet v2.0 on precipitation nowcasting. The problem setup, data processing, and neural network training procedures are correct, but the result evaluations and wording can be improved. I recommend that major revisions are required before the paper could be accepted for publication. More specific comments regarding the manuscript are included below.

Major comments:

• The written quality of this manuscript is poor. I recommend re-structuring the result section.
• Same as comment 1, more detailed descriptions for Fig. 7-11 need be included. In addition, in Figure 12, a general conclusion that the extrapolation of the new method is superior to other methods was given, the reason behind the results obtained in the manuscript should be provided. The false prediction showed in the lower right corner of the figure should be mentioned and explained.

• Precipitation nowcasting is generally defined as the prediction within 0-2 hours, but in the manuscript, the extrapolation results for a longer time (such as 1h, or one hour later) are not mentioned and presented. please state this in the discussion part (that this work only focuses on the 45-minute prediction?)

• The description of input and output parameters of model training is too brief. Although the code is provided, more detailed model parameters should be listed.

• The conclusion section of this manuscript is too brief. As a neural network-based study, many key concerns were not discussed. I recommend proposing a separate discussion section to summarize evaluation results, comparisons with other works.