Experimental Study of Cloud-to-Ground Lightning Nowcasting with Multisource Data Based on a Video Prediction Method

Round 1

Reviewer 1 Report

The manuscript illustrates a short-term lightning forecasting model, namely the Convolutional Long and Short-term Memory Lightning Forecast Net (CLSTM-LFN). Given the importance of lightning and the increasing use of AI methods for forecasting environmental parameters, this kind of studies are welcome and worth to be published. Results are quite encouraging, although conclusions reached by the article are common to other forecasting methods (decreasing performance and less importance of observations) with increasing forecasting time. Conclusions are drafted considering one season. If there are not obstacles in obtaining more lightning data or WRF runs, a longer test set (it is shorter than the training set) would allow to draft more solid conclusions. The manuscript is correctly structured and quite well understandable, although several parts of the manuscript deserve attention. However, I have several comments that authors should take into consideration in preparing the revised manuscript.

1. I found the Title and abstract, using the words like “multisource”, “radar echoes” to be a bit misleading. Actually, used are a number of selected WRF parameters (instead of methods based on single WRF parameter) from the WRF implementation used and lightning data from National Lightning Monitoring Network of China. CAPE and “radar” are not observations, but WRF output. Please make it clear in the abstract and in the text.
2. From Line 44: Relatively recent studies on radars have demonstrated that dual-pol radars outperform single polarization radars in identifying convection characteristics and the relation between lightning and the presence of graupel. The introduction mentions mainly old studies based on reflectivity measurements. Other experimental GNSS studies consider water vapor as a parameter for lightning prediction. I think the objective of this introduction is to select variables from WRF that are relevant to lightning prediction. If I am correct, I invite authors to make it explicit in the Introduction. Otherwise, introduction lacks important citations on lightning predictions.
3. Lightning Data: historical lightning observations are mentioned in the manuscript and t is not clear what they are. They deserve a clearer explanation in section 2.1.
4. From line 231: The description of the implementation of the method through like “In the input variables to the neural network model, the dimension of the lightning frequency is [3, 256, 256, 1] ….” sounds are a bit difficult to follow. Please try to reword the section for increased clarity.
5. Section 3.5: The authors should explain the reason of choice of the alternative methods listed here. Again, “radar_echoes” are not actual radar observations, but WRF output, and historical lightning observations should be better defined. The description of CLSTM-LFN-O is not clear to me: is WRF excluded both from train and forecast (in other words, the neural network runs only with observations of lightnings, as opposed to CLSTM-LFN-W that uses only WRF variables?)
6. Line 365: “..observed within a grid range of ±2”. What is the unit for the grid range ?
7. Line 380: Please fix “Tabel”. Please explain why different threshold (I assume “threshold” indicates the number of lightnings, but it should be specified in the table). I suggest authors to comment more on Table 2, for example, about the other prediction methods and the other indices.
8. Line 412: It is strange that lightnings are predicted for regions in which precipitation is absent (e.g. lightning spots at around 118°E, 33°N) or wrongly forecasted by WRF. On the contrary, text says “The CLSTM-LFN forecasts of lightning are closer to the observed results due to the indication of the Rmax prediction...”
9. Line 416: What is the meaning of “… the small distribution of historical lightning observations input to the neural network model, “? Lightnings are not frequent in regions considered, so that the neural network tends to limit their occurrence there ?
10. Line 491: similarly, there is a discrepancy between CAPE and Rmax.
11. Section 5. I do not understand the rationale of applying “disordering” to the different variables. Maybe authors could explain in the text what “disordering” is.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this study, the authors developed a lightning nowcasting model called CLSTM-LFN. It provides a stronger ability in the 0-3-h lightning forecast, relative to the traditional methods based on lightning parameterization or the model with only a single data source being input. They also found that the contribution of the WRF predication products increases with time, and in the products, the CAPE play the most prominent role. The article is generally well organized and written. The study provides a reference method for improving lightning early warning and forecasting. I suggest it can be published in this journal after a minor revision.

 

Comments and suggestions:

 

1. Some abbreviations were not fully written when they first appeared, such as PR92, LPI, R_max, CAPE and so on. Please check the text through the manuscript for improvement.

 

2. There are two study areas that are not included in the introduction section, they should be added. The first is the lightning prediction (or simulation) based on the numerical models directly coupled with electrification and discharge parameterization schemes. You can search the contributions from Liangtao Xu (Chinese), A. O. Fierro (American) and so on. The second is the previous studies that are very similar to this study in data usage. You can search the contributions from Yangli-ao Geng, Tianyang Lin and so on. They also developed the AI models to predict the lightning activity based on the observational lighting and WRF products.

 

3. 100-101: Should here be a “is” after “trajctory”?

 

4. Section 2.1: Please make it clear that what the National Lightning Monitoring Network observed were cloud-to-ground lightning flashes.

 

5. 138: I believe the longitude and latitude of the center point of the simulated area were given incorrectly.

 

6. 145-149: This paragraph focus on the study area. But it is separated from the description of the simulated area of the WRF (137-140). I recommend putting the contents in lines 145-149 and 137-140 together.

 

7. 165-166: “Thirteen variables from WRF model prediction products, such as maximum vertical velocity and CAPE, are selected in this study.”

 

8. 197-200: It is not to increase the lightning density, but increase the number of grids with lightning. Does this match the actual situation? Another approach is to randomly select the grids without lightning so that the number of grids is close to that of the grids with lightning. Why is this method not adopted? After using the method in the article, what is the ratio of the number of grid points with and without lightning? The intuitive feeling is that the grids without lightning are still far more than the grids with lightning.

 

9. 205-213: The meaning of this part is not very clear. The authors said their output is the number of lightning flashes. So, during the model training stage, was the one of input parameters the lightning frequency in the grid? If so, I believe that, for a specific lightning frequency value, the number of grids labelled with this value should be very small, especially relative to the number of grids without lightning; then, how to ensure the enough sample for the model training? Furthermore, if the lightning frequency is referenced in the model, the original lightning location data should be processed, such as grouping return strokes (location records) into flashes at least; this is not introduced in section 2. On the other hand, in the section of case study, the authors didn’t exhibit the lightning frequency, but the presence and absence of lightning output by the model (i.e., 1 and 0). These may confuse the readers.

 

10. 217: Please clarify what the sample number is for, the hour time? weather process?

 

11. 300: “SHI” should be “Shi”.

 

12. 4.2: Case study.

 

13. In the section of case study, please also provide the FAR and TS.

 

14. 498: should “calculate” be “calculating”?

 

15. 551 and other places including Summary and Conclusion: The prominence of CAPE is relatively clear at 3h. I recommend to describe this point more precisely.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study presents a method of combining lightning observations and numerical weather (forecasts) model output data of the WRF model for improving the hourly prediction of lightning occurrence within the next 3 h. The authors combine observations and models using a deep learning approach and by fusing previously established methods. The implemented approach is applied to lightning occurrence over China, where the algorithm is trained using data of two months (June, July) in 2020. The algorithm is then tested and evaluated using observations of August 2020. Different model output variables are briefly assessed concerning their importance for the lightning prediction accuracy.

The concept of the approach seems generally worth the publication, especially because the test of the applied model provides promising results and obviously better performance scores than other described methods. Nevertheless, the final classification of the surplus or novelty of the authors’ approach is not easy to judge on, as the introduction and especially the discussion of the results fall a bit short in this context. Moreover, I found it sometimes difficult to follow the authors’ arguments due to stylistic errors, such as repetitive words and phrases within one or two subsequent sentences. Some of my specific and technical comments address these, too, but I have a couple of additional comments and suggestions to improve the manuscript. Although these probably require a major revision, I encourage the authors to address (the majority of) the comments carefully as I believe that their study will then be worth the publication in “Remote Sensing”.

Please find all detailed comments and suggestions in the enclosed PDF.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors have done a good job with their revision of the manuscript. The overall quality and the discussion have been improved, as well as the introduction and analyses.

I therefore recommend accepting the manuscript for publication after a few minor phrasing revisions, for example:

line 33: "emergency [due to] local strong storms"

line 35-37: "whether there is conducive" - This sentence/grammar does still require rephrasing.

line 128: "factors affecting" -> "parameters affecting"