Wang presented the framework of a deep convolutional neural network to forecast haze events in Beijing and Shanghai. He utilized the trained neural network model and cluster analysis to categorize the regional meteorological regimes associated with haze and non-haze events, focusing on the technical aspects, such as the design of the model. The methods on the combination of a convolutional network and cluster analysis in this paper could be of high interest to the community, and it would be useful to better understand what the neural network model learns and why it misses the haze events. However, discussions of new scientific findings and more in-depth interpretations are not sufficient. For instance, generally, there is no significant difference in features between clusters (e.g., the four clusters of TP in Beijing look similar), and the author did not explain the physical meanings of the clusters. Thus, it is hard to see the necessity of cluster analysis without significant differences in clusters and explanations of their physical meanings. Although the methods and ideas are innovative, the manuscript is not well-written, mainly that the structure is a bit challenging for readers without sufficient background to follow. The methods could be helpful to the community once the author clarifies and addresses the important issues and includes more scientific discussions about the findings.

Specific comments:

1. Page 3, line 95-102: This paragraph is confusing. The first part of the paragraph says that U-net can be used when weather patterns associated with the targeted outcome are known or irrelevant to the task, but the second part mentions that when the environmental conditions related to the targeted outcome are yet known, U-net is not applicable. The two paragraphs are conflicting and confusing. Additionally, the sentences are convoluted and difficult to follow. I suggest rewriting the whole paragraph.

2. Page 4, line 130-132: Are the model structure and parameters the same for the two models (Beijing and Shanghai)? If so, shouldn’t the models’ parameters be optimized separately for the two regions to obtain the best performance for each of them?

3. Page 6, line 187-189: Do you include a testing set (a holdout dataset that has never been used in the training process)? From your results, I think the validation set in this paper is used to tune the hyperparameters of the model and monitor potential overfitting that occurs when training accuracy decreases but validation accuracy increases. An evaluation for a testing set is recommended to assess the performance of the trained model.

4. Page 6, line 197-198: Please include the description of class-weight and batch normalization in the appendix or model description and how they help the imbalance-data issue.

5. Page 7-8, line 246-249: It is worth checking the maps of features for the TP, FP, and FN for April and May verse the TP, FP, and FN for other months to see their differences in the weather and hydrological features.

6. Page 4-9, Section 2 and 3: The subsections of “Kernel size and CNN performance” and “Reducing the number of input features” could be moved to Section 2, as these two subsections are more related to the model architecture and design. Additionally, the author shows the validation results of reducing the number of input features before introducing the methods of reducing the number of features. The two sections together make it difficult for readers to follow. I suggest reorganizing these two sections and the subsections.  

7. Page 8, Figure 4: The unit/axis label of top figure?

8. Page 10, line 322: Please specify what VAE stands for.

9. Page 11, Section 4: The feature patterns of the clusters for the TP, FN, and FP are very similar with slight differences in certain features. Generally, I don’t see the point of conducting cluster analysis because (1) there is no significant difference between clusters, (2) the author did not explain the physical meanings of each cluster (if based on the slight differences), and (3) the author focused on the differences between TP and FN when explaining the missing haze events, which can be done by simply comparing averaged feature maps of TP and FN without cluster analysis.

10. Page 11, line 349-354: It seems like the feature patterns of the four clusters for the TP in Beijing cases are very similar (also the clusters for FN and FP). Why did you choose four clusters? Could you justify the purpose of using cluster analysis, given that there is no significant difference between the clusters?

11. Page 12, line 380: If the results of Shanghai are similar to Beijing and there is no regional characteristic for Shanghai, I suggest removing the results of Shanghai from the paper. Or you could add more discussions and highlight the common or different characteristics shown in the two regions.

12. Page 12, Figure 7: (1) Please include the color bar. The title of the features and cluster labels are hard to see; please enlarge them. It will also be helpful to add a contour map or add a point/shape to label Beijing on the map. (2) As mentioned before, there is no significant difference between clusters, and it is hard to read the key messages from 117 plots, especially that they are all super small. I wonder whether it is necessary to demonstrate the results of all the clusters in the main text? Is it possible to only show the results of the major cluster (the cluster with the largest number) and move the results of all clusters to the supplement? A similar issue also is shown in Figures 8 and 9.

13. Page 12, line 386: For the results of unnormalized format, I wonder whether there is a trend shown in the features and the haze events have seasonality or not. If so, how would the trend/seasonality affect the clustered features? 

14. Page 11, line 349-354 and Page 13, line 398-399: Do the four clusters represent different regimes/scenarios of haze events based on their differences shown in DTCV, SW1, and SW2?

15. Page 13, line 400 and line 402: It should be Figure S7 and Figure S8?

Interactive comment on “Forecasting and Identifying the Meteorological and Hydrological Conditions Favoring the Occurrence of Severe Hazes in Beijing and Shanghai using Deep Learning” by C. Wang

General comments

This paper presents a Deep Learning based research of severe air pollution events in Southeast China, namely “haze events”. The technological backbone of this research is a deep Convolutional Neural Network (CNN) that is commonly used in advanced image processing algorithms and applications but recently introduced to the climate and weather prediction communities in the private sector. The author well justifies the model choice and clearly elaborates on its technical structure. The model performance on the two selected regions of interest (Beijing and Shanghai) shows impressive accuracy when predicting severe haze events, as defined by the author. An additional classification analysis extended from that same model could also reveal the significance of each meteorological feature for severe haze event occurrence, contributing to the optimization of any future CNN analyses for similar purposes and raising new insights for further physical discussion. This paper introduces new techniques to the field of weather, climate, and environmental prediction, leveraging the growingly available computing power. Furthermore, this work joins the cutting edge of spatiotemporal physical models, recently shifting from deterministic into statistic methods. While most of the relevant research in the private sector is already implementing similar weather and climate prediction approaches, I see great promise and importance in introducing such work to our academic community.

Therefore, I recommend this paper to be published in Atmospheric Chemistry and Physics, with reservation to the author’s response to my comments below, hoping that my notes could help the author improve his paper.

Specific comments

1. The selected threshold for defining a severe haze event in the 2-class training is set to days whose surface visibility decreases below the 25^th While the value of that percentile varies in time and space, I suggest elaborating more on selecting that exact percentile threshold.
2. Please elaborate on the input data and its possible effects on the results. How many ground stations are used for the analysis in each city? What are the potential limitations that are resulted from the ERA5 spatial resolution of 0.25 degrees?
3. Please include, possibly as a supplemental, some technical details regarding the CNN analysis. What were the data and computational volumes and costs? What kind of computation platform was used, how long each training session take? What were optimization and approximation procedures implemented? Such information could assist future researchers when planning their analyses and also provide the scientific community with a technological benchmark for comparison with future projects.

Technical corrections

1. Please keep consistency in number representation along the manuscript (e.g. “11,376” in P. 4 Line 132 vs. “14975” in P. 11 Line 338.
2. Please keep consistency in technical terminology along the manuscript (e.g., “class-1” vs. “class 1”, etc.).
3. Please specify the unit following physical quantities (e.g. “…heights at 500 (Z500) and 850 (Z850) hPa” in P. 6 Line 188).
4. Please follow a consistent terminology for classes 0 and 1 in the 2-class analysis (e.g., “non-haze events” and “severe haze events”).
5. 2 Line 38: Please change “event” to “events”.
6. 2 Line 39: Please change “has” to “has”.
7. 4 Line 119: Please change “Introduction” to “introduction”.
8. 5 Line 158: Please add punctuation marks where necessary.
9. 5 Line 176: I suggest replacing or dropping the words “longitude-latitude” that are already self-embedded in surface map objects.
10. 6 Line 214: Please change “metrics” to “metric”.
11. 6 Lines 219-220: Please clarify that sentence.
12. 7 Line 225: Please correct a typo “class0weight”
13. 8 Line 267: Please modify to “there are many hyper-parameters in HazeNet that need …”.
14. 12 Line 398: Please change “soli” to “soil”.
15. 8 – caption (Lines 405-408): Please specify the explicit variable descriptions for better readability.