Fusion of Rain Radar Images and Wind Forecasts in a Deep Learning Model Applied to Rain Nowcasting

Round 1

Reviewer 1 Report

This paper introduced a precipitation nowcasting approach based on AI U-Net. I think it is a good move by including more relevant weather variables in image-based deep-learning precipitation nowcasting. The authors demonstrated an improvement of precipitation nowcasting by adding two surface wind components into the learning algorithm. I have the following significant concerns and some minor suggestions that need to be addressed carefully in the next revision.      

1) The authors chose to do 30 and 60 minutes rain only. Due to the fast change of natural rain and quick degrading of the forecast performance, I suggest working prediction of every 5 minutes from 5 - 60 minutes, then compare the performance degrading with an increase of the forecast lengths of different schemes.

2) For the classes, I do not understand why the authors chose to use inclusive classification (Table 1), but not the bins or intervals (Table 2). I do not think "inclusive classification" is independent.

3) The result in Table 6 is critical information for this work. However, there is no definition of metrics "Mean" and "Std". Therefore it is hard to understand the conclusions based on the Table. 

4). Two examples shown in Figures 10 and 11 do not support the conclusions, and inconsistent with the data Table 6. Both examples are for scenarios of weakening convention. In both cases, I do not see the U-NET result is any better than the optical flow, especially for Class 3.

5). For validating the forecast result, TS and CSI are more proper than Mean and Std.     

6). The authors use all days for learning. I think the clear sky days should be excluded. null-rain images only add unnecessary uncertainties to the training, especially when U and V are included.     

Minor suggestions:

1) Page 1, the authors' affiliations should be spelled out.

2) Line 60, remove "and can prove difficult to outperform".

3) Throughout the paper (such as table 1, Figure 4, and elsewhere) - you are dealing with 5-minute rain accumulation. You can use the unit of "mm" or mm/5-minutes for the accumulated rain, but it is wrong to use "mm/h" for it.

4) Line 194, remove one "training".

5) Line 198, "... and multiply them", how?  

Author Response

see attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see separate file.

Comments for author File: Comments.pdf

Author Response

see attached file

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper presents a new deep learning model trained on rain radar images and wind velocity produced by a weather forecast model which outperforms the traditional optical flow method on moderate and higher rain events forecasts. The method is noval and expected to be verified for use in operational forecast. The paper can be accepted based on revision according to the following comments: 

1) In 3.1.2, the classification of the precipitation scale is not quite appropriate. for example, all of the thresholds 0.1 mm/h, 1mm/h, 2.5mm/h are for small to medium rainfall. I suggest that 10mm/h, 25mm/h, 50mm/h should be considered at least for this concern.

2) the comments 1) might help to improve the problem in Line 391, Page 18:"unable to predict the details at a small scale resulting in very smooth borders...". THe reason is that small scale rainfall are usually relate to intensive convection which leads to intensive rainfall. 

3) A sensitivy experiment is suggested to be performed on the optional (with/without) use of wind data from numerical model (not only compare to that of optical flow).   This should be a major point for the novelty of the paper. 

4) For evaluation of the new method, the Threat Score (TS) /Bias (usually used in meteorological verifications) are suggested to be used.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf