Intelligent Reconstruction of Radar Composite Reflectivity Based on Satellite Observations and Deep Learning

Round 1

Reviewer 1 Report

I liked the work and I consider that this topic needs to be more explored.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors try to reconstruct the radar composite reflectivity(CREF) based on satellite observations and deep learning. The research is important for early warning of severe weather in areas that suffers from poor radar coverage. However, some problems should be addressed before publication. The comments are as follows:

1) In the Section 2, since both Figure 1 and Figure 2 contain DEM information, I will suggest that the authors can merger these two figures into one figure.

2)Please provide more information for the loss function of the model in the page 10.

3)In the data processing, what is the dimension of the dataset for Model (VIS) and Model (IR)?  Please provide more details.

4)The language should be undergone strict review by a native English speaker.

In this paper, the authors try to reconstruct the radar composite reflectivity(CREF) based on satellite observations and deep learning. The research is important for early warning of severe weather in areas that suffers from poor radar coverage. However, some problems should be addressed before publication. The comments are as follows:

1) In the Section 2, since both Figure 1 and Figure 2 contain DEM information, I will suggest that the authors can merger these two figures into one figure.

2)Please provide more information for the loss function of the model in the page 10.

3)In the data processing, what is the dimension of the dataset for Model (VIS) and Model (IR)?  Please provide more details.

4)The language should be undergone strict review by a native English speaker.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors,

the first parts of the manuscript are well written. It was a pleasure to read this parts. The aim of the work addresses a very important topic. So far so good, but, the results of your study do not support your conclusions or provide a proof for a succesfull training.

In my opinion there are clear indicators of overfitting.  The statistical model fits the training data (the discussed 3 events) well enough. But there is no proof that the algorithm works well on data it has not yet seen. Comparison with GPM is no proof that it works well, as GPM as well as other satellite based precipitation data sets comes with a very high uncertainty. That exactly is the reason to try AI to improve it, hence, GPM can not be used to proof the improvement. Further, it is only one case and no statistics is provided. Thus, even if GPM would be perfect, it coud match by chance.

Backround Overfitting: The IR channels provide not enought patterns or feature information for a learning process that can reproduce radar structures so well when applied to unknown test cases (to a lesser amount this is also true for VIS). The correlation between the temperature of the cloud top and the precipitation rate is not good (e.g. as a comparison of satellite based products with radar clearly shows). Also AI can not learn out of nothing. It can reproduce the training data set, but I doubt that the patterns would still match well enough when applied to test data sets that were not part of the training. 

I regret, but for theses reasons I recommended to reject the manuscript. However, the manuscript might be worth to considered again after thorough evaluation of the method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Thank you for your replies and the clarifications.

I think it would be worth to add more information on the question on how the network can learn radar patterns from IR, although IR sees only the cloud top temperature. One reason might be the higher spatial resolution compared to e.g. MSG. Another the use of BDTs at this at least "normalizes" the temperatures.

Also there might be a difference in the learning process and the quality of the net between stratiform and convective rain. Within this scope, it is not clear how many stratiform rain events are part of the training and testing sets, please clarify and discuss. 

Finally, do you consider to release the software, so that it can be used by others ? This would be great and would increase the transparency of the work.

best regards

Author Response

Please see the attachment.

Author Response File: Author Response.pdf