Towards reducing the high cost of parameter sensitivity analysis in hydrologic modelling: a regional parameter sensitivity analysis approach
Abstract. Land surface models have many parameters that have a spatially variable impact on model outputs. In applying these models, sensitivity analysis (SA) is sometimes performed as an initial step to select calibration parameters. As these models are applied on large domains, performing sensitivity analysis across the domain is computationally prohibitive. Here, using a VIC deployment to a large domain as an example, we show that watershed classification based on climatic attributes and vegetation land cover helps to identify the spatial pattern of parameter sensitivity within the domain at a reduced cost. We evaluate the sensitivity of 44 VIC model parameters with regard to streamflow, evapotranspiration and snow water equivalent over 25 basins with a median size of 5078 km2. Basins are clustered based on their climatic and land cover attributes. Performance of transferring parameter sensitivity between basins of the same cluster is evaluated by the F1 score. Results show that two donor basins per cluster are sufficient to correctly identify sensitive parameters in a target basin, with F1 scores ranging between 0.66 (evapotranspiration) to 1 (snow water equivalent). While climatic attributes are sufficient to identify sensitive parameters for streamflow and evapotranspiration, including vegetation class significantly improves skill in identifying sensitive parameters for snow water equivalent. This work reveals that there is opportunity to leverage climate and land cover attributes to greatly increase the efficiency of parameter sensitivity analysis and facilitate more rapid deployment of land surface models over large spatial domains.
Samah Larabi et al.
Status: final response (author comments only)
RC1: 'Comment on hess-2023-21', Anonymous Referee #1, 09 Feb 2023
- AC1: 'Reply on RC1', Samah Larabi, 10 May 2023
- AC3: 'Reply on RC1', Samah Larabi, 10 May 2023
RC2: 'Comment on hess-2023-21', Anonymous Referee #2, 19 Apr 2023
- AC2: 'Reply on RC2', Samah Larabi, 10 May 2023
Samah Larabi et al.
Samah Larabi et al.
Viewed (geographical distribution)
The manuscript ‘Towards reducing the high cost of parameter sensitivity analysis in hydrologic modelling: a regional parameter sensitivity analysis approach’ assesses the spatial pattern of parameter sensitivity and the performance of sensitive parameter transferability over 25 basins of the Pacific Northwest region of North America by using the VIC model. It is notable because a larger suite of 44 parameters parameter considered and the multiple output variables were assessed in the model. Overall, I think that the manuscript is well-written and the topic is attractive. My recommendation is that the paper be published in Hydrology and Earth System Sciences provided the following points are addressed by the authors.
I notice the authors clustered 25 basins and then test the ability of sensitive parameter transferring. To be honest, the number of basins used for clustering and parameter transferring (i.e., regionalization) is insufficient, you see, the basins of cluster #2 are only three in Figure 7. Insufficient samples will cause the classification results to show great randomness, thus impacting the parameter transferring.
The result that two donor basins per cluster are sufficient to correctly identify sensitive parameters in a targeted basin, is doubtful. Have you compared the result to the previous research for parameter regionalization? As far as I know, at least using more than 5 donor basins is credible for parameter regionalization (e.g., Oudin et al., WRR, 2008; Bao et al., JH, 2012).
Have you considered the cross-validation for the evaluation of the transferability of parameter sensitivity? I think cross-validation is a good way to check whether the sensitive parameter regionalization is reliable, i.e., using each of the basins in turn as if it were ungauged (Gou et al., BAMS, 2022).
This research does not involve parameter calibration, so what is the significance of sensitive parameter transferring? Why not transfer the calibrated parameters to ungauged basins directly?
Line 93-Line 98: Check the data ranges again. ‘average annual precipitation over the 25 basins ranges from 448 mm/year to 1666 mm/year.’ But the average annual precipitation of basin ‘BruneauR’ is 337 mm. Meanwhile, you round to one decimal for temperature but keep two decimals for snow index and aridity index, need full text unified. Last, the definition of snow index should be merged with the caption of Table 2, i.e., here gives the calculation method and reference.
Line 172: ‘Arno’ to ‘ARNO’
Line 220: ‘x’ to ‘×’
Line 220: What is the definition of elementary effects (EEs)? Is this value calculated based on the model output? Is the mean of the target modeled output?
Line 239: 75 sets of noninformative/informative parameters are 75 sets of outputs or 75 experiments? Here 75 sets equal to the 25 basins multiply 3 model outputs. I think here cannot be represented as ‘75 sets of noninformative/informative parameters’ because you adopted 44 parameters for each experiment, right?
Line 241: ‘all 75 experiments’ is right.
Table1. The meaning of showing relief?
Table2. The definition and method of the ‘Snow Index’ and ‘Aridity index’ should show in the methods section. There is no point in repeating the emphasis here. And the temperature threshold (i.e., 2°C) for snow index calculation, why? has any references?
Table3. Again, the details of the table header, delete or move to the end of the table by ‘Note’.
Figure 1. The basin ID can be marked.
Figure 2. How to distinguish whether the parameter is invariant-informative? ‘Parameters are considered invariant-informative if the count of basins in which they are informative’, how many informative basins are eligible for invariant-informative? 10 or 15?