the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Calibrating macro-scale hydrological models in poorly gauged and heavily regulated basins
Dung Trung Vu
Thanh Duc Dang
Francesca Pianosi
Stefano Galelli
Abstract. The calibration of macro-scale hydrological models is often challenged by the lack of adequate observations of river discharge and infrastructure operations. This modelling backdrop creates a number of potential pitfalls for model calibration, potentially affecting the reliability of hydrological models. Here, we introduce a novel numerical framework conceived to explore and overcome these pitfalls. Our framework consists of VIC-Res (a macro-scale model setup for the Upper Mekong River Basin) and a hydraulic model used to infer discharge time series from satellite data. Using these two models and Global Sensitivity Analysis, we show the existence of a strong relationship between the parameterization of the hydraulic model and the performance of VIC-Res – a co-dependence that emerges for a variety of performance metrics we considered. Using the results provided by the sensitivity analysis, we propose an approach for breaking this co-dependence and informing the hydrological model calibration, which we finally carry out with the aid of a multi-objective optimization algorithm. The approach used in this study could integrate multiple remote-sensed observations and is readily transferable to other basins.
- Preprint
(7520 KB) - Metadata XML
-
Supplement
(906 KB) - BibTeX
- EndNote
Dung Trung Vu et al.
Status: closed
-
RC1: 'Comment on hess-2023-35', Anonymous Referee #1, 13 Mar 2023
Overall, the manuscript could be very helpful in providing a guideline for a calibrating model where almost no in-situ data is available. The authors discussed the in-depth methodology of their proposed calibration, providing a detailed analysis of the impact of parameter tuning on the performance metrics of the model. The authors also provided an analysis of the co-dependence of the hydrological and hydraulic model parameterizations and techniques to break the co-dependency. The central argument presented in the manuscript is using satellite data to infer river discharge against which the model-derived river discharge will be compared and the model recalibrated to achieve the desired accuracy. Sattelite data itself can cause a wide range of uncertainty in the calculation of river cross-section, water surface slope, and so on, especially in the Upper Mekong river basin, which is so complex in topography. The satellite data (used as a proxy of observation for calibration purposes) is prone to uncertainty that eventually impact the parameter tuning. So there's a need to strengthen the discussion by providing a detailed discussion on the uncertainty in the river discharge estimation from satellite data, without which the calibration framework could be questionable. My recommendation is for a major revision with the specific comments and concerns listed below:
1. In the abstract, the authors mentioned that their approach could be readily transferable to another basin. However, the authors did not provide any convincing discussion on how the same framework can be used for another basin. For example, what cautions should other researchers follow when applying the same technique to a highly complex basin with rugged terrain or complex topography? As the estimation of river cross-section and water surface slope could be challenging/more uncertain in some other basin.
2. In section 3.3.1 authors discussed the calibrated model parameters and presented the calibration outcomes later. However, one would expect to see a discussion of the calibration of the most sensitive parameters.
3. As the discussion is so central to the simultaneous calibration of the using RS Discharge (from satellite data). Thus an uncertainty analysis of the estimation of the river cross-section or uncertainty in the RS Discharge from satellite data is extremely necessary. It directly impacts the RS Discharge estimation against which VIC-RES discharge is compared to calculate performance matrices. Thus uncertainty in RS Discharge can substantially impact the calibration process and parameter tuning, fundamentally questioning the Novel technique the authors suggested in that manuscript. I recommend a discussion on uncertainty in RS Discharge estimation and how it may affect the calibration process. Although the author provided some insights in section 4.2.3
4. In section 3.3.2- The authors discussed that they used multiple performance matrices to cover a different aspect of modeling accuracy. However, the use of KGE as a performance metric is also suggested, as it considers bias, correlation, and variability.
5. In Figure 8: could you discuss why there is less variability in 2009-2012 and after that, there is considerable variability, particularly in the low flows?
6. In Figure 9: the timing of the peak is missed in some of the years, e.g., 2007. You can just add a discussion on the sensitivity of different parameter tuning in capturing timing/seasonality or peak. Or which is the most sensitive parameter?
7. Also. How can the hydrological response unit's resolution or size impact calibration? It can impact the calibration substantially. For example, the Lancang river basin is so narrow and elongated. Thus, the use of a coarse hydrological response unit of Coarse-resolution may accurately impact the identification of river grid cells.
8. Figure 2 and 3 can be merged together. Having two figures does not add much value to the discussion.
9. In section 3.2.1: The authors used a regression technique (sixth-degree polynomial) to fit the data point best. However, the author said that is best works for the natural condition of rivers. AS MANUSCRIPT TITLE, the authors mentioned "Heavily Regulated Basin." One would like to know the author’s novel technique for heavily regulated basins. In the suggested numerical framework, I think I do not see any strong linkage of the reservoir operation (like heavy regulation) with the calibration/parameter tuning. Or could you provide an explanation of how your technique is mainly applicable to the heavily regulated basin? Or it may be more justifiable to say Novel calibration technique for the poorly gauged basin.Citation: https://doi.org/10.5194/hess-2023-35-RC1 - AC1: 'Reply on RC1', Dung Trung Vu, 03 May 2023
-
RC2: 'Comment on hess-2023-35', Andrea Galletti, 17 Mar 2023
The authors propose an interesting framework for exploiting satellite data for the calibration of large-scale hydrological model. The steps undertaken to translate satellite-derived information into are explained in thorough detail. Then, the co-dependence between the hydrological model's performance and the calibration of the hydraulic model based on remote-sensed data is analyzed via Global Sensitivity Analysis. Finally, the authors adopt the satellite-derived information to calibrate VIC-Res and validate the predictions against available ground observation.
The concept is promising, however it would benefit from some considerations on its general applicability were drawn. Furthermore, the densely-flowing work sometimes overshadows the connection between the overall proposed framework and the particular tool/result being explained. I recommend acceptance of the paper, provided that the authors address or clarify the following points (minor revision):
1. The paper addresses the issue of calibrating macro-scale models in ungauged catchments. The introduction seems to implicitly assume that every large-scale hydrological model is always calibrated, and that this is most times done against available streamflow time series. However, several large scale hydrological models do not actually undergo case-specific calibration. While a calibrated model can provide results that are generally more reliable or realistic, this might not be clear to every reader, or someone could disagree. I recommend addressing the importance and effectiveness of calibration at the beginning (lines up to 15).
2. The subsequent paragraph (line 17 onwards) could also benefit from a more thorough introduction to models' calibration and what is needed. In particular, it is not clearly explained why and how the presence of hydropower infrastructures for which operations are not known can create pitfalls for model calibration, nor why could it be at all important to include such factors in a calibration of hydrological model's parameters.
3. Some references and examples should be provided to support the sentence "Yet, this approach may still partially rely on in-situ data": this sentence opens up to the second class of RS-based works and basically justifies the study, therefore it needs to be backed by clear examples.
4. Following point #1, I feel like a non-trivial question is whether it is possible and helpful to calibrate large scale hydrological models in ungauged catchments. This opens up to research questions I, II and III, helping to place them in a context that is broader than a mere numerical exercise.
5. The choice of the 2009-2018 time window is not motivated as of line 84. I suspect the decision was taken backwards for compatibility with data coverage and quality (Figure 3). If this is the case, it should be stated clearly and the corresponding statements (line 96 and caption of fig 3) adjusted. If the choice was driven by something else, it should be pointed out.
6. I am dubious about the reliability of interpolating (linearly) daily reservoir operations between two monthly values. Hydropower often operates at the daily scale (or lower). The configuration of cascading dams might help masking errors in this procedure and their repercussions on the performance metrics. Did the authors compare their results with a fully-natural setup in order to evaluate the impacts of the reconstructed reservoir operation?
7. The estimation of the river cross section represents an important step in the framework, influencing the outcomes of the rating curve and all subsequent analyses. With the river width ranging from 400m to 200m depending on water level, and with DEM and Landsat images having a resolution of 30m, one could think that the combination of cells taken as the cross section is not univoque. Exploring multiple alternatives for the chosen cross section could help assessing the variability of this extrapolation (if at all present). Otherwise, a plane view or schematic of the chosen pixels would help understanding the decision taken. Furthermore, river bed might be subjected to vertical evolution (i.e., erosion or sediment accumulation) which is hardly grasped by this methodology since it relies on a static DEM. Performing a similar extrapolation (of the river cross section) e.g.,at location 812 which is closer to Chiang Saen and then comparing the so-obtained discharge values with ground observations would have provided a rather solid, yet not exhaustive, base for any assumption made at Virtual Station, reinforcing the implicit validation described at lines 267-269.
8. Line 338 onwards: the selected 25% is said to be the best according to all metrics: does this mean that the (i assume normalized, as defined in the methods) metrics were averaged and the best 25% averages were taken? This concept could be made a bit clearer. Furthermore, it is unclear what determined the choice of the 12 and 58 stations presented later in this section.
9. Line 359: Phrasing here is a bit misleading, and a more thorough introduction could be useful: GSA was used to determine the co-dependence between n and the model's performance, while the potential bias intrinsic to the co-dependence was not really assessed, although one could reasonably suppose that it is present. Providing further evidence (references) that joint estimation of discharge and parameters leads to a biased calibration could help strenghten the need to break the co-dependence.
TECHNICAL CORRECTIONS:
Line 227: is 2008-2018 correct? it seems it should be 2009
Line 297: the value in parentheses should be 0.045
Line 382-383: do not contribute to (reducing) modelling uncertainty?
Line 393: One often recurring... one should be removed?
Citation: https://doi.org/10.5194/hess-2023-35-RC2 - AC2: 'Reply on RC2', Dung Trung Vu, 03 May 2023
Status: closed
-
RC1: 'Comment on hess-2023-35', Anonymous Referee #1, 13 Mar 2023
Overall, the manuscript could be very helpful in providing a guideline for a calibrating model where almost no in-situ data is available. The authors discussed the in-depth methodology of their proposed calibration, providing a detailed analysis of the impact of parameter tuning on the performance metrics of the model. The authors also provided an analysis of the co-dependence of the hydrological and hydraulic model parameterizations and techniques to break the co-dependency. The central argument presented in the manuscript is using satellite data to infer river discharge against which the model-derived river discharge will be compared and the model recalibrated to achieve the desired accuracy. Sattelite data itself can cause a wide range of uncertainty in the calculation of river cross-section, water surface slope, and so on, especially in the Upper Mekong river basin, which is so complex in topography. The satellite data (used as a proxy of observation for calibration purposes) is prone to uncertainty that eventually impact the parameter tuning. So there's a need to strengthen the discussion by providing a detailed discussion on the uncertainty in the river discharge estimation from satellite data, without which the calibration framework could be questionable. My recommendation is for a major revision with the specific comments and concerns listed below:
1. In the abstract, the authors mentioned that their approach could be readily transferable to another basin. However, the authors did not provide any convincing discussion on how the same framework can be used for another basin. For example, what cautions should other researchers follow when applying the same technique to a highly complex basin with rugged terrain or complex topography? As the estimation of river cross-section and water surface slope could be challenging/more uncertain in some other basin.
2. In section 3.3.1 authors discussed the calibrated model parameters and presented the calibration outcomes later. However, one would expect to see a discussion of the calibration of the most sensitive parameters.
3. As the discussion is so central to the simultaneous calibration of the using RS Discharge (from satellite data). Thus an uncertainty analysis of the estimation of the river cross-section or uncertainty in the RS Discharge from satellite data is extremely necessary. It directly impacts the RS Discharge estimation against which VIC-RES discharge is compared to calculate performance matrices. Thus uncertainty in RS Discharge can substantially impact the calibration process and parameter tuning, fundamentally questioning the Novel technique the authors suggested in that manuscript. I recommend a discussion on uncertainty in RS Discharge estimation and how it may affect the calibration process. Although the author provided some insights in section 4.2.3
4. In section 3.3.2- The authors discussed that they used multiple performance matrices to cover a different aspect of modeling accuracy. However, the use of KGE as a performance metric is also suggested, as it considers bias, correlation, and variability.
5. In Figure 8: could you discuss why there is less variability in 2009-2012 and after that, there is considerable variability, particularly in the low flows?
6. In Figure 9: the timing of the peak is missed in some of the years, e.g., 2007. You can just add a discussion on the sensitivity of different parameter tuning in capturing timing/seasonality or peak. Or which is the most sensitive parameter?
7. Also. How can the hydrological response unit's resolution or size impact calibration? It can impact the calibration substantially. For example, the Lancang river basin is so narrow and elongated. Thus, the use of a coarse hydrological response unit of Coarse-resolution may accurately impact the identification of river grid cells.
8. Figure 2 and 3 can be merged together. Having two figures does not add much value to the discussion.
9. In section 3.2.1: The authors used a regression technique (sixth-degree polynomial) to fit the data point best. However, the author said that is best works for the natural condition of rivers. AS MANUSCRIPT TITLE, the authors mentioned "Heavily Regulated Basin." One would like to know the author’s novel technique for heavily regulated basins. In the suggested numerical framework, I think I do not see any strong linkage of the reservoir operation (like heavy regulation) with the calibration/parameter tuning. Or could you provide an explanation of how your technique is mainly applicable to the heavily regulated basin? Or it may be more justifiable to say Novel calibration technique for the poorly gauged basin.Citation: https://doi.org/10.5194/hess-2023-35-RC1 - AC1: 'Reply on RC1', Dung Trung Vu, 03 May 2023
-
RC2: 'Comment on hess-2023-35', Andrea Galletti, 17 Mar 2023
The authors propose an interesting framework for exploiting satellite data for the calibration of large-scale hydrological model. The steps undertaken to translate satellite-derived information into are explained in thorough detail. Then, the co-dependence between the hydrological model's performance and the calibration of the hydraulic model based on remote-sensed data is analyzed via Global Sensitivity Analysis. Finally, the authors adopt the satellite-derived information to calibrate VIC-Res and validate the predictions against available ground observation.
The concept is promising, however it would benefit from some considerations on its general applicability were drawn. Furthermore, the densely-flowing work sometimes overshadows the connection between the overall proposed framework and the particular tool/result being explained. I recommend acceptance of the paper, provided that the authors address or clarify the following points (minor revision):
1. The paper addresses the issue of calibrating macro-scale models in ungauged catchments. The introduction seems to implicitly assume that every large-scale hydrological model is always calibrated, and that this is most times done against available streamflow time series. However, several large scale hydrological models do not actually undergo case-specific calibration. While a calibrated model can provide results that are generally more reliable or realistic, this might not be clear to every reader, or someone could disagree. I recommend addressing the importance and effectiveness of calibration at the beginning (lines up to 15).
2. The subsequent paragraph (line 17 onwards) could also benefit from a more thorough introduction to models' calibration and what is needed. In particular, it is not clearly explained why and how the presence of hydropower infrastructures for which operations are not known can create pitfalls for model calibration, nor why could it be at all important to include such factors in a calibration of hydrological model's parameters.
3. Some references and examples should be provided to support the sentence "Yet, this approach may still partially rely on in-situ data": this sentence opens up to the second class of RS-based works and basically justifies the study, therefore it needs to be backed by clear examples.
4. Following point #1, I feel like a non-trivial question is whether it is possible and helpful to calibrate large scale hydrological models in ungauged catchments. This opens up to research questions I, II and III, helping to place them in a context that is broader than a mere numerical exercise.
5. The choice of the 2009-2018 time window is not motivated as of line 84. I suspect the decision was taken backwards for compatibility with data coverage and quality (Figure 3). If this is the case, it should be stated clearly and the corresponding statements (line 96 and caption of fig 3) adjusted. If the choice was driven by something else, it should be pointed out.
6. I am dubious about the reliability of interpolating (linearly) daily reservoir operations between two monthly values. Hydropower often operates at the daily scale (or lower). The configuration of cascading dams might help masking errors in this procedure and their repercussions on the performance metrics. Did the authors compare their results with a fully-natural setup in order to evaluate the impacts of the reconstructed reservoir operation?
7. The estimation of the river cross section represents an important step in the framework, influencing the outcomes of the rating curve and all subsequent analyses. With the river width ranging from 400m to 200m depending on water level, and with DEM and Landsat images having a resolution of 30m, one could think that the combination of cells taken as the cross section is not univoque. Exploring multiple alternatives for the chosen cross section could help assessing the variability of this extrapolation (if at all present). Otherwise, a plane view or schematic of the chosen pixels would help understanding the decision taken. Furthermore, river bed might be subjected to vertical evolution (i.e., erosion or sediment accumulation) which is hardly grasped by this methodology since it relies on a static DEM. Performing a similar extrapolation (of the river cross section) e.g.,at location 812 which is closer to Chiang Saen and then comparing the so-obtained discharge values with ground observations would have provided a rather solid, yet not exhaustive, base for any assumption made at Virtual Station, reinforcing the implicit validation described at lines 267-269.
8. Line 338 onwards: the selected 25% is said to be the best according to all metrics: does this mean that the (i assume normalized, as defined in the methods) metrics were averaged and the best 25% averages were taken? This concept could be made a bit clearer. Furthermore, it is unclear what determined the choice of the 12 and 58 stations presented later in this section.
9. Line 359: Phrasing here is a bit misleading, and a more thorough introduction could be useful: GSA was used to determine the co-dependence between n and the model's performance, while the potential bias intrinsic to the co-dependence was not really assessed, although one could reasonably suppose that it is present. Providing further evidence (references) that joint estimation of discharge and parameters leads to a biased calibration could help strenghten the need to break the co-dependence.
TECHNICAL CORRECTIONS:
Line 227: is 2008-2018 correct? it seems it should be 2009
Line 297: the value in parentheses should be 0.045
Line 382-383: do not contribute to (reducing) modelling uncertainty?
Line 393: One often recurring... one should be removed?
Citation: https://doi.org/10.5194/hess-2023-35-RC2 - AC2: 'Reply on RC2', Dung Trung Vu, 03 May 2023
Dung Trung Vu et al.
Dung Trung Vu et al.
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 694 | 197 | 19 | 910 | 47 | 7 | 6 |
- HTML: 694
- PDF: 197
- XML: 19
- Total: 910
- Supplement: 47
- BibTeX: 7
- EndNote: 6
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1