Impacts of spatiotemporal resolutions of precipitation on  flood event simulation based on multimodel structures  – a case study over the Xiang River basin in China

Zhu, Qian; Qin, Xiaodong; Zhou, Dongyang; Yang, Tiantian; Song, Xinyi

doi:https://doi.org/10.5194/hess-28-1665-2024

Articles | Volume 28, issue 7

https://doi.org/10.5194/hess-28-1665-2024

© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/hess-28-1665-2024

© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 28, issue 7

Research article

|

11 Apr 2024

Research article |

| 11 Apr 2024

Impacts of spatiotemporal resolutions of precipitation on flood event simulation based on multimodel structures – a case study over the Xiang River basin in China

Qian Zhu, Xiaodong Qin, Dongyang Zhou, Tiantian Yang, and Xinyi Song

Download

Final revised paper (published on 11 Apr 2024)
Preprint (discussion started on 20 Sep 2022)

Interactive discussion

Status: closed

CC1:
'Comment on hess-2022-301', Xichao Gao, 20 Jan 2023

Meaningful study! could the authors distinguish the sensitivities of these different models to the spatio-temporal resolutions of precipitation? and explain the reasons?

Citation: https://doi.org/10.5194/hess-2022-301-CC1
- AC1: 'Reply on CC1', Qian Zhu, 14 Mar 2023
  
  Thank you for your comment. In our study, the hydrological models are generally more sensitive than the machine learning model, but the sensitivity of the model is also related to the precipitation input. As illustrated in Fig.7 and Fig.8, when the spatiotemporal resolution of CMA changes, DHSVM is the most sensitive one, for the mean NSE of flood events declines from 0.68 to 0.45 when the spatial resolution of precipitation changes from 0.1° to 0.5°. Maybe for CMA-driven DHSVM, the impact of spatial resolution on the capture of precipitation variability during flood event periods can propagate to the flood events simulation.
  But when the spatiotemporal resolution of IMERG changes, SWAT and DHSVM model perform similarly, which is consistent with previous research results (Lobligeois et al. 2014, Huang et al. 2019), where insignificant improvement was reported with higher spatial resolution of observed rainfall. It probably dues to the large catchment area and only the outlet station is used for calibration. Liang et al. (2004) found a critical resolution (1/8° for the VIC model) for a watershed with 1,233 km2, beyond which the spatial resolution shows limited impact on model performance. For our study area (82,375 km2), when the spatial resolution of precipitation changed from 0.1° to 0.5°, small variety is shown in the performance of flood events simulation, which indicates the critical resolution may be larger for large watershed. For HBV, it is not sensitive to changes in temporal resolutions because its simple hydrological model structure.
  For LSTM, higher resolution shows better performance, and similar conclusion is drawn from previous study conducted by Sun et al. (2017), which found that deep learning model performs better with larger datasets.
  
  Reference:
  Huang, Y., Bárdossy, A. and Zhang, K. 2019. Sensitivity of hydrological models to temporal and spatial resolutions of rainfall data. Hydrology and Earth System Sciences, 23(6), 2647-2663.
  Lobligeois, F., et al. 2014. When does higher spatial resolution rainfall information improve streamflow simulation? An evaluation using 3620 flood events. Hydrology and Earth System Sciences, 18(2), 575-594.
  Liang, X., Guo, J. and Leung, L. R. 2004. Assessment of the effects of spatial resolutions on daily water flux simulations. Journal of Hydrology, 298(1-4), 287-310.
  Sun, C., et al. 2017. Revisiting Unreasonable Effectiveness of Data in Deep Learning Era. 2017 Ieee International Conference on Computer Vision (Iccv), 843-852.
  
  Citation: https://doi.org/10.5194/hess-2022-301-AC1
RC1:
'Comment on hess-2022-301', Anonymous Referee #1, 12 Feb 2023

The authors have chosen multiple hydrological models (HBV, SWAT, DHSVM, and LSTM) and applied multiple calibration strategies, separately to all flows and flood flows. Moreover, the models have been applied at multiple spatial and temporal resolutions. The main finding is that calibration with respect to flood flows resulted in better flood flow simulation. LSTM is relatively better in simulating flood flows.
Although the authors have done a lot work and results look interesting, I feel depth is missing and there is not much novelty. It is not surprising that models calibrated using flood flow data more accurately predict flood flows. The authors have considered mean values for comparision, which I think is problematic. Means are unreliable when the distributions are skewed.
There are a lot of results, but the discussions are not doing proper justice -- more in-depth discussions need to be added. For example, it is not sufficiently clear why spatial resolution should matter. The authors can provide more convincing explanations by referring to the concept of time-area diagram.
Model calibration considering parts of discharge time series is not a new idea.
Specific comments:
Lines 20: It is not clear what you mean by "flood event." Also, I am not comfortable with the term "to match continuous streamflow." May be you can write "to match the entire streamflow time series."
How did you select the flood events?
295: Mean NSE may not be a reliable indicator. You should consider median, 75th and 25th percentile NSE.I see 75th NSE falling in case of CMA. The authors need to discuss it.
NSEs in Figure 6: I don’t see any consistent pattern. The results are not discussed properly.
NSEs in Figure 7: Again, I dot see a consistent pattern.
Results and discussions should be put together. It is difficult to follow discussion when results are not immediately available.

Citation: https://doi.org/10.5194/hess-2022-301-RC1
- AC2: 'Reply on RC1', Qian Zhu, 16 Mar 2023
  
  Please refer to the uploaded document for the responses. Thank you very much.
  
  Citation: https://doi.org/10.5194/hess-2022-301-AC2
RC2:
'Comment on hess-2022-301', Anonymous Referee #2, 19 Apr 2023
The authors used the high spatio-temporal resolutions IMERG precipitation products and in-situ observations to force three hydrological modes and a data-driven model to investigate the impacts of spatio-temporal resolutions of precipitation on flood events simulation in a large-scale catchment. The study determines the compatibility between various models and precipitation datasets, which is critical for selecting the appropriate model for precipitation sources in hydrological simulation and prediction, and vice versa.
In general, the manuscript is well organized and well written. The scope of the study fits the journal’s focuses and audience community. The presentations are clear, and the novelty of the work is pointed out. The current manuscript has good potential to be published subject to some minor revisions and a few editing issues. The following suggestions are identified, and the authors shall revise them in their next version towards publication:

Line 66: “the study”-> “they”.

Line 124-127: The structure of these two sentences is suggested to be revised. The conjunction "so" in the beginning of the second sentence may be unclear.

Line 139: “(hereafter CMA)” needs to be put behind “China Meteorological Administration”.

Line 185: the reference “(AghaKouchak et al. 2013)” should be located behind the “HBV model”.

Line 230-233: please pay attention to the format of the variables, such as xt, and t.

Line 268-269: please explain how the eleven historical flood events are selected.

Line 338: “as the resolution get coarser”-> as the resolution is coarser or as the resolution gets coarser.

Line 350-352: “However, the uncertainty of NSE, KGE and BIAS-P values of flood events simulated with IMERG is decreasing as the spatial resolution.” As the spatial resolution what? finer or coarser?

Line 365: in most instances -> in most cases.

Line 407-408: “the same results” means the results are exactly the same, does that what the authors indicate? Otherwise, the same results -> the comparable/similar results or the results are almost the same.

Line 417-418: the calibration strategy II is an effective way for training the LSTM model to obtain the best flood events simulation results -> the calibration strategy II is an effective way to train the LSTM model to obtain the best flood events simulation.

Line 430: performs -> perform.

Line 431: please delete the “results”. And please check the whole manuscript for this issue.

Line 440: larger data set -> larger dataset.

Isn’t the “Fig. 9” shall be colored red to be consistent with other figures?

The colors used in Fig.10 are not so easy to distinguish.

Same issue of Appendix C , and please refer to the comment #15

In summary, I personally think the work is solid and the manuscript has a good potential to be published with some minor editing and polish to be done by the authors.
Citation: https://doi.org/10.5194/hess-2022-301-RC2
- AC3: 'Reply on RC2', Qian Zhu, 24 Apr 2023
  
  Please refer to the attached file for details of our responses. Thanks a lot.
  
  Citation: https://doi.org/10.5194/hess-2022-301-AC3
EC1: 'Comment on hess-2022-301 (Editor's comment)', Dimitri Solomatine, 17 May 2023

The comments of referees are very valid. The authors have shown that they appreciate these and have a plan for the paper revision. I would suggest to give special attention to the genelra comments of Referee 1, more clearly highlighting the novelty of this work.

Citation: https://doi.org/10.5194/hess-2022-301-EC1

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

ED: Publish subject to revisions (further review by editor and referees) (17 May 2023) by Dimitri Solomatine

AR by Qian Zhu on behalf of the Authors (06 Jun 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (17 Jun 2023) by Dimitri Solomatine

RR by Anonymous Referee #2 (24 Jun 2023)

RR by Anonymous Referee #1 (05 Sep 2023)

Suggestions for revision or reasons for rejection

The authors have applied multiple hydrological models using multiple precipitation datasets at various temporal and spatial resolutions to predict flood peaks. Clearly, a lot of hard work has been put into this research. The results are interesting. However, certain parts of their methodology are quite unclear to me, which is why I am unable to provide a fair opinion. I am also a bit disappointed by the lack of effort in providing insights, which would make their study complete. My specific comments are given below.

3.2.2. It is not clear whether you have considered peak discharge only or all the data points of the flood hydrographs. If the former is true, the number of data points is very small for any meaningful calibration. The term 'flood event' has not been explicitly defined, which gives rise to additional confusion.

3.3. It is not clear if Eq. (1) uses only flood peaks or all the data points in the time series for computing NSE. If the former is the case, the metric is not reliable since there are not many data points considered by the author.

Eq. (2): Bias is not typically presented in this way. Again, how reliable is the equation when there are so few data points?

4.1. The results are not very surprising since you have calibrated the model for flood peaks only. I would be surprised if you also show an improvement in overall NSE (i.e., NSE considering all the data points).

4.2. The effects of precipitation data type on model performance are quite informative. However, no proper explanation is provided in the discussion section, which makes the analysis incomplete. Line 460 is unclear. What do you mean by error propagation? Please explain instead of merely citing another paper.

4.3. The results look interesting. Again, no explanation is provided. For example, why does 0.25-degree data give the best 75th NSE and the worst 25th NSE for HBV (Figure 7a)?

The abstract says LSTM is outperforming other models. Figure 6 says HBV is better than LSTM.

Line 90: The term 'physically based' is typically used for hydrological models based on hydrodynamic equations. The models you are referring to are typically called conceptual models. This is just a semantic issue though.

Hide

ED: Publish subject to revisions (further review by editor and referees) (10 Sep 2023) by Dimitri Solomatine

AR by Qian Zhu on behalf of the Authors (19 Sep 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (01 Oct 2023) by Dimitri Solomatine

RR by Anonymous Referee #1 (03 Nov 2023)

Suggestions for revision or reasons for rejection

The authors have made some effort to address my earlier comments, which I appreciate. However, some of the major concerns still remain unaddressed.

Earlier I wrote “on. The term 'flood event' has not been explicitly defined.” The concern remains unaddressed. The authors write “… 17, where eleven historical flood events are selected with flood peak exceeding the threshold of 8,600 m3/s in this study.” Again, what is an event? Where does it start and where it ends? It appears the authors follow some subjective criteria to select the events, which are not elaborated.

I am not comfortable with the overly simplistic conclusion that LSTM is better than HBV. The results provide a much more nuanced picture. Figure 6 NSE plots: HBV is more consistent across the data products compare to LSTM. For instance, LSTM’s 25th percentile is much lower compared to HBV’s for IMERG-L. A statement like “LSTM has a higher likelihood of success” would be much more acceptable.

I reiterate my earlier statement: there is no surprise that model performance improved for flood events after calibrating exclusively for flood events. It is widely acknowledged that calibration with respect to a specific objective function leads to its improvement. As I said earlier, I would be surprised to see overall NSE (NSE for the whole time series) improving after calibration with respect to flood events.

As I mentioned earlier, there are numerous results but proportionally less discussion. For instance, no explanation is provided for why a 0.25-degree resolution appears to perform well for NSE-CMA-SWAT (apologies for the earlier typo) but not for NSE-CMA-LSTM. Similarly, a 0.5-degree resolution seems to work for KGE-CMA-SWAT but not for KGE-CMA-LSTM. The authors have predominantly presented the results without a thorough critical analysis, which is the point I am emphasizing. Once again, I am not suggesting that the study is irrelevant, but I believe that additional effort is needed to enhance the paper's overall appeal.

Mean NSE does not make a lot of sense if the distribution is skewed (which is very likely).

Hide

ED: Publish subject to revisions (further review by editor and referees) (22 Nov 2023) by Dimitri Solomatine

AR by Qian Zhu on behalf of the Authors (08 Dec 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (07 Jan 2024) by Dimitri Solomatine

RR by Anonymous Referee #1 (21 Feb 2024)

ED: Publish as is (22 Feb 2024) by Dimitri Solomatine

AR by Qian Zhu on behalf of the Authors (23 Feb 2024) Author's response Manuscript

Short summary

Input data, model and calibration strategy can affect the accuracy of flood event simulation and prediction. Satellite-based precipitation with different spatiotemporal resolutions is an important input source. Data-driven models are sometimes proven to be more accurate than hydrological models. Event-based calibration and conventional strategy are two options adopted for flood simulation. This study targets the three concerns for accurate flood event simulation and prediction.