TSRC: A Deep Learning Model for Precipitation Short-Term Forecasting over China Using Radar Echo Data
Round 1
Reviewer 1 Report
General comments:
Deep learning in meteorology is a highly relevant subject and very useful for operational weather forecasting. This research article proposes a new time series based residual convolution deep learning framework using ground radar observations. The authors demonstrate the improvement in real-time forecasting by reducing the smoothing effect and decay of precipitation intensity. The authors improvised the vanilla convolution algorithm in order to preserve the spatial and temporal information. They compared the results with two other algorithms (optical flow and U-Net). Overall, the manuscript is relatively well organized, clearly mentioning the novelty, merits and limitations of this research and results are clearly stated.
Here are few minor comments.
-
Line 350-353, the argue that TSRC algorithm works properly for 1 hr and it’s performance does not work after 2 hr for typhoon case study. The reason is quiet obvious that they did not considered the dynamical parameters of precipitation system as stated in the section 2.3. I have a confusion about the timing of typhoon stage consider. Whether it is the initial stage, severe category, or very severe category ? does the stage impact the results ?
-
Does the TSRC algorithm works for stable convective precipitation system over land region (less movement) ?
-
Why not the authors use Deep forest algorithm which has more advantages than deep learning model ?
Zhi-Hua Zhou and Ji Feng, (2017) “Deep Forest: Towards an Alternative to Deep Neural Networks”
Comments for author File: Comments.pdf
Author Response
please kindly find the attached file: List of Responses (#1).doc
Author Response File: Author Response.doc
Reviewer 2 Report
The manuscript deals with TSRC to improve the accuracy of short-term precipitation forecasting over China. The authors processed the long-term radar echo data for this study and tried to explain the major points of TSRC compared with OF and UNet schemes. However, I think that the manuscript would be accepted after minor revisions.
Comments:
1. Line 128: Authors had better explain the method of data cleaning in the manuscript.
2. Line 130: I would like to recommend that the authors describe the spatial resolution of the data.
3. Lines 246 to 247: I would like to know the categories of reflectivity clearly. For example, ~20 dBZ means 0 to 20 dBZ, ~ 30 dBZ means 20 to 30 dBZ, ~40 dBZ means 30 to 40 dBZ?
4. Line 292 and Line 334: I would like the authors had better remove Case 1 and case 2 in the manuscript and divide section 3 into two sub-sections like 3.2.1 case 1, 3.2.2 case2.
5. Figure 6: The authors divided the reflectivity into 3 categories. Could the authors describe the results with respect to the different categories?
6. Line 362: 2011 to 2021 ?
7. Line 359: I would like to recommend that the authors had better divide Section 4 into two sections. For example, 4. Discussions, 5. Conclusions
Author Response
please kindly find the attached file: List of Responses (#2).doc
Author Response File: Author Response.doc
Reviewer 3 Report
After checking the manuscript, I thought this article presents the application of the DL method in short-term rainfall prediction. I have no other comments on it.
Author Response
We appreciate for your attention and patience.
Reviewer 4 Report
The study proposes TSRC, a new DL-based convolutional neural network for precipitation nowcasting over China with a lead time of 3-hour,in which more contextual information and less uncertain feature is expected to remain in deep networks. Comparison with optical flow model and UNet show the advantage of TSRC in prediction with higher POD, lower FAR, and, smaller MAE. The model can also forecast high-intensity radar echoes even for typhoon rainfall systems. The TSRC seems very potential in precipitation nowcasting. However, before acceptance of the paper, a number of issues should be addressed in the revision:
Major Comments:
1) It's not quite clear why more contextual information and less uncertain feature is remained in deep networks by using this new method and how it affects the evolution of radar echo, especially the intensive echo. The authors should give more explaination on the route for this.
2) How is the new scheme of TSRC is designed and what's the new of it compared with similar methods of convolution network? A flow chart/table in more detail is suggested to be given.
3) How did the TSRC make the best score in all of the statistics, such as probability of detection (POD), false alarm rate (FAR), mean absolute error (MAE)? More case studies should be performed to support the result.
Author Response
please kindly find the attached file: List of Responses (#4).doc
Author Response File: Author Response.doc