the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Key factors for quantitative precipitation nowcasting using ground weather radar data based on deep learning
Daehyeon Han
Yeji Shin
Jungho Im
Juhyun Lee
Abstract. Quantitative precipitation nowcasting (QPN) may assist to mitigate the tremendous socioeconomic harm caused by severe weather. Because of the rapid atmospheric variability, the QPN has been a challenging problem to solve. Recent QPN research has presented data-driven models that make use of deep learning (DL) and ground weather radar. Previous research has mostly concentrated on constructing DL models, while other elements for DL-QPN have received less attention. This research looks at four crucial aspects in the DL-QPN and their impact on predicting performance. The prediction strategy (single, recursive, and multiple predictions), deep learning model (U-Net and convolutional long short term memory; ConvLSTM), input past sequence length (60 and 120 min), and output future sequence length were the four key factors (60 and 120 min). Using weather radar data from South Korea, twelve schemes have been developed to assess the influence of each factor. A long-term operational study for 2018–2020 was conducted, and a summer high rainfall event was examined to investigate the extreme case. In both situations, U-Net outperformed the critical score index (CSI) using a multiple prediction design. While ConvLSTM did not show a definite CSI difference across input sequence length, U-Net performed better with shorter input sequences (i.e., 60 min) than with longer input sequences (i.e., 120 min). The length of future sequences has little influence on model performance. As the lead time increased, all of the DL-QPN schemes showed underestimation and blurry outcomes. U-Net was shown to be significantly reliant on the most recent input time (i.e., 0 previous minutes) in sensitivity analysis, while ConvLSTM was more responsive to multiple time steps. This work may give a modeling technique and contingency plan for future DL-QPN employing weather radar data by explicitly comparing critical elements.
Daehyeon Han et al.
Status: final response (author comments only)
-
RC1: 'Comment on gmd-2022-276', Anonymous Referee #1, 25 Feb 2023
The manuscript titled 'Key factors for quantitative precipitation nowcasting using ground weather radar data based on deep learning' presented a thorough analysis of different schemes to approach precipitation nowcasting problems using deep-learning techniques. The different schemes were tested using ground weather radar data over South Korea. In recent days, multiple works have explored the application of deep-learning algorithms for quantitative precipitation forecasting. Exploring endless options and schemes is necessary to understand better the feasibility of using these methods in an operational scenario. I appreciate the authors' effort in conducting a systematic analysis and discussions. Some parts of the manuscripts are still hard to understand and not very clear.
Major comments:
- Data imbalance: The major problem in precipitation nowcasting is the lack of representation of intense precipitation due to data imbalance. Did the authors try to consider this problem in their analysis?
- Data: Do the authors consider datasets with overlap when training is done? If t1-tn is used as input and tn+1 to tn+m is the forecast, is tn+1 - tn+n used as input for another sample?
- Equations 2 and 3: The Mean Absolute error and mean bias equation are not normalized. Missing 1/n
- Line 271 and Section 4.2.3: Adding a dummy zero variable to input causes sparsity. Adding white noise is a better idea. But, I feel that the entire part (Section 4.2.3) does not add much to the paper. It just lengthens the paper. I will suggest the authors remove that part.
- Table 4 and others: Persistence is not explained previously in the manuscript.
Minor comments:
Figure 2: Why is dBZ converted to rain rate? Why not just train the model for reflectivity values?
Figure 3: It is better to mark the study region on the map.
Line 229: Why is leaky relu not used for U-net and only used for ConvLSTM?
Table 2: Why is SU-120-60 or RU-120-60 not considered for analysis? Please justify this in the text.
Table 4: Is there a reason why the best bias value is not highlighted? Just curious.
Figure 8 and 10: The thresholds should be 5 mm/h. Please check the captions.
Citation: https://doi.org/10.5194/gmd-2022-276-RC1 - RC2: 'Comment on gmd-2022-276', Anonymous Referee #2, 26 Apr 2023
Daehyeon Han et al.
Daehyeon Han et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
445 | 115 | 6 | 566 | 3 | 0 |
- HTML: 445
- PDF: 115
- XML: 6
- Total: 566
- BibTeX: 3
- EndNote: 0
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1