A Branched Convolutional Neural Network for Forecasting the Occurrence of Hazes in Paris Using Meteorological Maps with Different Characteristic Spatial Scales

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the author presented a deep learning approach based on a branched CNN to forecast LVEs in the Paris region, using meteorological and hydrological data from the ERA5 reanalysis dataset. The model processes input data over a 96*128 grid, covering a large part of Western Europe, and predicts visibility conditions based on observations from CDG. The approach is evaluated using both classification and regression tasks, with the model demonstrating improved performance compared to previous methods, particularly in reducing false positives. The branched CNN architecture allows the model to capture both large and small-scale meteorological patterns effectively, and the paper reports successful results when validated on test data from 2021 to 2023.

 

Some comments and questions are as follows.

 

1. The description of the model’s output in the paper is not very clear. Is the visibility observation at CDG considered as the representative point for the entire Paris region? Then are the CDG data used as the model’s output? In other words, is the model generating a single output value? If that’s the case, does the input really need to cover such a large area (96*128 grid)? Could a smaller region serve as the input with little to no degradation in the model’s performance? It would be beneficial to conduct experiments to verify this.

 

2. The paper uses the 25th percentile (P25), corresponding to visibility of 7.89 km, as the threshold for distinguishing low visibility events. While this choice covers most or even all the low visibility cases, more attention should arguably be given to extreme cases of very low visibility. The evaluation of the model focuses on P15 and P25, but would it be possible to provide a more granular analysis, for example, by examining P5 or even P1 thresholds? This would help assess the model’s performance in predicting more extreme visibility conditions.

 

3. The model is based on a CNN architecture, which may appear somewhat outdated in the deep learning field. Moreover, the paper only compares the proposed model to the author’s prior works. It would strengthen the paper to compare the model with a broader range of approaches to highlight its advantages. For example, comparisons with classic methods or more recent models, such as attention mechanisms, could provide a more comprehensive evaluation.

 

4. In Figure 11, only the results of the proposed model are given, and the results of the comparison algorithm are not reflected, so the actual effect of the proposed algorithm cannot be compared and analyzed. It is recommended to supplement the results of the comparison algorithms.

 

5. The model uses ERA5 reanalysis data, but ERA5 typically has a few days' delay before the data becomes publicly available. This could limit the real-time applicability of the proposed model in scenarios that require immediate predictions. Will this delay impact the use of the model in real-time applications?

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Originality and Contribution

The paper deals with the development of a two-branch convolutional neural network (CNN) for the prediction of low-visibility events (LVE) in the Paris region. The use of CNNs in the field of deep learning and atmospheric phenomena is not a new concept, but using a two-branch structure in a region with complex weather conditions such as Paris is one of the original contributions of the paper. In contrast to previous applications of HazeNet, it is emphasised how the two-branch architecture performs better in maintaining different spatial scales. However, the literature section of the paper could explain in more depth why this new method was chosen specifically for Paris. The performance results in different cities could have been compared.

2. Methodological robustness

The method section of the paper clearly explains how the two-branch structure of the CNN architecture was designed. Training sets and validation sets are carefully separated, and parameter optimisation is done appropriately. The use of a large data set, especially for the years 1975-2023 and 1975-2023, increases the reliability of the method. However, since most of the data used in the training of this model is historical, it is questionable how much the model can adapt to current weather conditions and changing climatic trends. In addition, it could have been explained in more detail why different CNN structures were needed to improve the accuracy of the validation results. The advantages of the two-branch structure are not only limited to performance data, but in which scenarios it provides superiority could be discussed in more depth.

3. Spelling and Structure

The article has a coherent structure and conforms to the norms of scientific writing. There are clear transitions between sections, and the text flow is smooth. However, some deficiencies can be observed in terms of the visual clarity of graphs and tables. For example, the comprehensibility of the colour scales used in figures 5 and 6 could be improved and could be more descriptive. Also, excessive use of technical terms in some sections may make it difficult to understand for general readers.

4. Interpretability of Results

In the results section, it is stated that the model was able to accurately predict low visibility events 70 percent of the time. Although this is a very impressive result, it should have been clearer under which specific conditions or data sets this success was achieved. For example, it is not specified which weather parameters have the greatest impact on the model's performance. How much difference the results make compared to general weather forecasting approaches could be explained in more detail.

5. General Contribution

This study contributes to studies on the use of deep learning in weather forecasting models. In particular, the conclusion that the two-branch architecture better recognises different spatial scales is an important finding. However, whether the model has been tested in other geographical regions and its comparative performance against other models should have been addressed more comprehensively.

Overview:

The paper fills an important gap between deep learning and weather forecasting applications. However, a detailed discussion of the methodology and more comprehensive analyses of the results would help the reader to better grasp the broad scope of the work.

While there are important scientific contributions in the paper, I believe that improvements and additional clarifications are needed in some areas. In particular, corrections and improvements can be requested in the following areas:

Clarity and Extension of the Method: The advantages of the two-branch CNN architecture should be explained in more detail. Also, why the model is applied specifically to the Paris region and what kind of performance it can give in different geographical regions should be discussed in more depth.

Elaboration of Results: It is important to elaborate more on which weather parameters affect the performance of the model and the validation results. It can be more clearly demonstrated how much difference the results make compared to general weather forecasting approaches.

Improving Graphs and Tables: Especially the colour scales and explanations in the graphs should be made clearer, and the reader should be able to understand the data better.

 

 

Comments on the Quality of English Language

The English language used in the article is generally in accordance with academic norms, but the use of language and expression in some sections could be made more fluent and understandable. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for your replies to my review report. The editable content has been improved in the paper, and some explanations have also been provided. Additionally, I have a few suggestions:

Table 4 presents the results and analysis for the years 2021-2023, along with detailed explanations. P5 only has 3 samples, making it difficult to directly compare the performance of various algorithms.

Since Table 1 provides detailed statistics for the years 1975-2019, where P5 has 878 samples and the fog samples total 878-251=627, even if the paper does not use cross-validation and only performs a single split, it should still present the performance metrics for P5, P10, P15, and P20 on the validation set in Table 3 and Figure 9. Otherwise, what is the significance of the statistics for each percentile in Table 1?

Author Response

Comment of the Reviewer 1 (in Italia bold font):

Thank you for your replies to my review report. The editable content has been improved in the paper, and some explanations have also been provided. Additionally, I have a few suggestions:

Table 4 presents the results and analysis for the years 2021-2023, along with detailed explanations. P5 only has 3 samples, making it difficult to directly compare the performance of various algorithms.

Since Table 1 provides detailed statistics for the years 1975-2019, where P5 has 878 samples and the fog samples total 878-251=627, even if the paper does not use cross-validation and only performs a single split, it should still present the performance metrics for P5, P10, P15, and P20 on the validation set in Table 3 and Figure 9. Otherwise, what is the significance of the statistics for each percentile in Table 1?

Response to the Reviewer's Comment:

Thank you for providing the additional suggestion! The validation performance scores, including accuracy, precision, recall, and F1 scores have now been added in the revised Table 4 (note also that the new table now adopts a number format with 3 rather than 2 digits after the decimal to benefit ranking), along with corresponding discussions as: “The performances of various versions of HazeNet in validation clearly indicate that the two branched architectures have both delivered better F1 scores in regression-classification mode for events with vis. equal or lower than the 25^th, 20^th, and 15^th percentiles than the non-branched version, while the latter performed better for the events equal of lower than the 10^th and 5^th percentiles of the long-term observations (Fig. 9 and Table 3). The evolution of regression-classification scores with the probability of targeted LVEs shown in Table 3 demonstrate that the accuracy of all machines increases with the lowering of probability, indicating the effect of increasing imbalance (i.e., ratio of non-targeted to targeted events). On the other hand, F1 score decreases with probability, implying the forecasting skill of the machines decreases in low probability cases than in higher ones”.

For Figure 9, an attempt to add more sets of bars appears to make the figure too crowded without offering more information than the updated Table 4 does, thus I decide to let Figure 9 to stay in its current format.