In this document, we provide detailed answers to the comments raised by Reviewer 1 on our manuscript "On the Value of Satellite Remote Sensing to Reduce Uncertainties of Regional Simulations of the Colorado River" by Mu Xiao et al., Hydrol. Earth Syst. Sci. Discuss., https://doi.org/10.5194/hess-2022-204-RC1, 2022.

Our answers are reported in italic after the Reviewers’ comments that we have numbered.

General comment:

I enjoyed reading the manuscript. This study combines several remote sensing products to improve the hydrologic model’s physics together with streamflow performance. I only have several concerns regarding the presentation of the work and the framework.

Thanks for working with us to improve this paper. We appreciate the constructive comments and we have provided detailed responses below.

1. L159: USBR dataset needs a reference (url/doi)

The original U.S. Bureau of Reclamation (USBR) webpage that provides the flow data records will be included in the updated version: https://www.usbr.gov/lc/region/g4000/NaturalFlow/provisional.html

The naturalized flow data we obtained from USBR will also be uploaded onto the same Zenodo online archive where we stored the model parameters (it will be mentioned in the “Open Research” section in the updated paper).

2. L239: why monthly and not daily streamflow performance was targeted in calibration? daily water balance is key for hydrologic models. Monthly fit is easier and reducing the value of baseline simulation.

This is a very good point that was mentioned by both Reviewers. The river is heavily regulated and the highest resolution of the naturalized flow records from USBR that is currently available is monthly. This is the main reason why we could not extend our calibration and validation against discharge at a daily scale.

Given the lack of daily streamflow data, we might argue that adding daily remotely sensed products to the model testing phase is even more critical to capture daily dynamics.

3. L248-Fig3: in this section (3.3) I read what has been done but I couldn’t find answers for the question “how”. Framework needs elaboration. Baseline simulation is clear but other steps are not clear.

We acknowledge that we have not provided an in-depth description of the “how” in each step mentioned section 3.3. The reason is that we considered section 3.3 as an overview of the calibration methods, while we preferred reporting the details of the Forcing-adj, Veg-adj, and Snow-adj steps in sections 4.3.1, 4.3.2, and 4.3.3 of the Results, respectively. To address the Reviewer’s comment, we will:

• change the name of Section 3.3 to “Model improvements with remote sensing products: overview of stepwise calibration strategy”;
• mention in section 3.3 that the details of each step are better described in sections 4.3.1, 4.3.2, and 4.3.3 of the Results.

4. Most importantly, model calibration is an exercise of fine tuning of the model parameters. Before calibration a robust sensitivity analysis (SA) must be applied for such sophisticated models with many parameters to reduce the search space. Did authors apply SA in their study?

We did not apply a systematic robust SA, but we adopted a hybrid approach based on the physics and sensitivity of single parameters. Specifically, we first carefully considered the physical equations implemented in the model to simulate the processes related to the observed variable (i.e., land surface temperature, LST, snow cover, SC). These equations are reported in the manuscript Appendix. We then identified the set of key parameters that are (1) involved in the equations, (2) spatially variable, and (3) not derived from any type of direct or indirect observation. We then computed the spatial correlation between these parameters and the pattern of the errors between simulated and remotely sensed LST. The outcomes of these analyses are reported in Figure 7 of the manuscript. For the parameters with the largest correlation, we performed a sensitivity analysis to verify that that parameter importantly affects the simulation of LST, although we did not mention it in the paper. We ultimately focused on the subset of parameters that exhibited the largest sensitivity.

5. The authors followed a stepwise approach but sensitive analysis (sobol’s, LHS O-A-T, Morris etc) may reveal parameter interactions which can be important to consider during calibration. The authors should discuss the implications of parameter interaction in their framework.

We agree with Reviewer 1 that a sensitivity analysis targeting multiple parameters at the same time can ultimately lead to improved performance. However, this type of analysis would be very computationally expensive for a model like VIC and the size of the Colorado River Basin and would most likely require an entirely separate study. Here, our focus is instead to highlight the importance of accounting for spatially variable observations from remote sensors in the calibration process. As discussed in the answer to the previous comment #4, we focused the calibration on a set of parameters identified based on model physics and the impact of these parameters on the spatial variability of the errors between simulated and remotely sensed LST and SC. The calibration was then performed by focusing on one parameter in each step.

That said, we will highlight in the paper that conducting a sensitivity analysis targeting multiple parameters is one of the subsequent research tasks to further advance the use of remotely sensed products for model improvement.

6. It would be good to simultaneously use LST and snow RS data on uncertainty reduction via model calibration.

We agree with the Reviewer that this would be an interesting idea. However, it would require the use of an automatic calibration routine with an optimization function that accounts for both daily LST and monthly SC, which would require significant computational power and would be out of the scope of this study, as for the case of the multiparameter sensitivity analyses (comment #5). To our knowledge, studies that investigated the utility of remotely sensed products in large basins like the Colorado River are still very few; therefore, it is still necessary to separately gain insights into the utility of each remotely sensed dataset, along with the associated parameters and equations. Once this knowledge based on single variables is built, calibration strategies that target multiple variables could be better designed in future work.

7. My biggest concern is about the spatial structure of the selected hydrologic model (VIC) which is a semi-distributed model. In such model parameters get the same value in the same subbasins which inevitably leads to uniform parameter fields and resultant uniform flux maps. One way to avoid this, is using fully distributed models with parameter regionalization tool based on pedo-transfer functions using soil and vegetation properties.

We appreciate the Reviewer’s comment and point out three issues that, hopefully, address this concern.

First, VIC is a macroscale hydrologic model with a gridded domain where most of the parameters and all outputs do vary spatially. For example, all parameters identified in Fig. 7 vary in space. See also the maps in Figs. 1d and 1e that report the vegetation fraction, f_v, and soil depth that are used in the baseline simulation.

Second, the parameters of the baseline simulations are mainly based on the products derived by Bohn and Vivoni (2019), who utilized high-resolution (from 500 m to 1 km) remote sensing products to generate several model parameters in the same grid at 1/16˚(~6-km) resolution used in our manuscript. This is mentioned in Section 3.2 of the manuscript.

Finally, we have more than 15,000 pixels for the Colorado River Basin which allow the spatial variability to be appropriately captured.

8. The authors used bias-sensitive error metrics (rmse, bias) and CC as bias insensitive metric. CC must be used with cautious it can be affected by outliers in the sample. High CC values are not always informing. Instead spatial metrics (SSIM, FSS etc) could be preferred.

We would like to clarify one important issue. As mentioned in lines 255-258: “The first two steps [of the calibration] were guided by metrics quantifying the agreement between simulated and remotely sensed LST, including the correlation coefficient (CC), root mean squared error (RMSE), and Bias (mean LSTV - mean LSTM) between: (1) time series of daily LSTV and LSTM at each grid cell, and (2) daily spatial maps”.

In particular, the maps and metrics shown in Figures 5, 6, 7, and 9 that drove the calibration effort are based on RMSE, CC, and Bias between simulated and observed time series at each pixel. Therefore, for these cases, it is not possible to compute spatial metrics like SSIM, FSS, etc. The only case where we computed metrics between maps is Figure 8, which we used as additional measures of calibration accuracy.

Regarding the role of CC, we fully agree with the Reviewer. When we carried out our analyses, we noted that the CC alone cannot be a good indicator for model evaluation because it is not robust, as pointed out by the Reviewer, and because it is always high (>0.8) even in the baseline (Lines 475-476). Because of this, our model adjustments (Forcing-adj, Veg-Adj, and Snow-adj) are mainly based on either RMSE or Bias of the time series.

To address the concerns of both Reviewers, in the revised manuscript, we computed the Structural Similarity Index Measure (SSIM) and the Spatial Efficiency metric (SPAEF) of long-term LST climatology (see Tables R1 in the attached document). The values of these metrics are in line with the overall trend of RMSE and Bias; the table will be then added to the Supporting Information of the revised manuscript version.

9. Fig6: the readers can be curious why median night time bias for baseline is usually less than other 3 cases.

As shown in Fig. 6, the median bias for nighttime LST in the baseline case is negative. This result was ascribed to the negative bias of the air temperature forcings from the Livneh dataset, which we removed using the PRISM long-term normal products in the “Forcing-adj” calibration step. The resulting median bias for nighttime LST becomes close to 0 or slightly positive in the “Forcing-adj” step, thus improving the simulation (see also Figures 8 and 9). The bias does not change significantly in the other steps because they do not involve changes in parameters that affect nighttime LST. These details are described in Section 4.3.1, where we concluded that “…the Forcing-adj simulations improved Bias, which was reduced in most subbasins” (see Lines 370-371).

10. Fig7 should be better explained. How correlation between parameters is assessed?

The correlation coefficients in Fig. 7 are the Pearson Correlation Coefficient (CC) between two spatial fields in each subbasin. The first spatial field is either T_air or any of the parameters shown in the rows of the heatmaps (e.g., Elevation, Porosity, …, LAI, and f_v), while the second spatial field is either RMSE (heatmaps on left) or the Bias (heatmaps on right) between time series of LST_V and LST_M at each domain pixel. The columns in each heatmap report the correlation coefficient in each subbasin. For example, the first pixel on the top left is showing the correlation coefficient between T_air and Daytime RMSE in the Green subbasin.

In the revised version of the paper, we will update the text to better present this information.

Publisher’s note: this comment is a copy of AC1 and its content was therefore removed.

Publisher’s note: this comment is a copy of AC1 and its content was therefore removed.

In this document, we provide detailed answers to the comments raised by Reviewer 2 on our manuscript "On the Value of Satellite Remote Sensing to Reduce Uncertainties of Regional Simulations of the Colorado River" by Mu Xiao et al., Hydrol. Earth Syst. Sci. Discuss., https://doi.org/10.5194/hess-2022-204-RC1, 2022.

Our answers are reported in italic after the Reviewers’ comments that we have numbered.

General comment:

The authors present a well-written and well-motivated study on the value of remote sensing data to improve hydrological simulations. I enjoyed reading the manuscript and only have a few comments that will require some attention by the authors.

Thanks for reviewing the paper and helping us to improve it. Below, we provide point-to-point responses to the Reviewer’s comments.

1. The authors chose to evaluate the model against monthly runoff. In my point of view this cannot be justified since the remote sensing data are used at daily scale to evaluate the model. In order to achieve a balanced evaluation of the model runoff should also be evaluated at daily scale.

This is a very good point that was mentioned by both Reviewers. The river is heavily regulated and the highest resolution of the naturalized flow records from the US Bureau of Reclamation (USBR) that is currently available is monthly. This is the main reason why we could not extend our calibration and validation against discharge at a daily scale. We also note that other modeling studies of the Colorado River Basin are based on the same monthly dataset.

Given the lack of daily streamflow data, we might argue that adding daily remotely sensed products to the model testing phase is even more critical to capture daily dynamics.

2. More details are required with respect to the “adjustment of the VIC parameters” as presented in section 3.2. Did the authors conduct a manual calibration or were the parameter values estimated via automatic calibration? Please specify.

It is a manual calibration. For the baseline simulation of section 3.2, we manually changed the soil parameters to improve streamflow performance in the “traditional” way. In the revised version of the manuscript, we will specify in Sections 3.2 and 3.3 that the calibration performed in the baseline and the other steps are manual.

3. One of my main concerns relates to how the fit between observed and simulated spatial patterns was assessed. The authors chose to do a grid-wise evaluation of the simulation. I think this makes sense for the forcing adjustment where the temporal dynamics of simulated LST are strongly linked to the forcing data (air Temp). However, when it comes to model parameters, I would suggest to evaluate the model against spatial pattern that are aggregated over time, for example a long-term average annual (or summer) LST map. Evaluating a model at daily scale will always be very much affected by the quality of the forcing data and the model parameters have a limited affect here. Nevertheless, the imprint of the model parameters emerges when aggregating the simulation results over time and quantifying the spatial pattern match (e.g. with help of the SPAEF metric (https://doi.org/10.5194/gmd-11-1873-2018)) instead of the grid-to-grid comparison. Along these lines, the spatial patterns of RMSE and Bias presented in Figure 9 do not show a clear improvement of the model developments. Maybe a spatial pattern oriented evaluated of the long term average LST patterns is more insightful.

Thanks for pointing this out. Reviewer 1 had also a similar comment related to the need to use metrics that quantify the match of spatial patterns. However, there is a detail of our approach that, perhaps, was did not properly explain. As mentioned in lines 255-258: “The first two steps [of the calibration] were guided by metrics quantifying the agreement between simulated and remotely sensed LST, including the correlation coefficient (CC), root mean squared error (RMSE), and Bias (mean LSTV - mean LSTM) between: (1) time series of daily LSTV and LSTM at each grid cell, and (2) daily spatial maps”.

In particular, the maps and metrics shown in Figures 5, 6, 7, and 9 that drove the calibration effort are based on RMSE, CC, and Bias between simulated and observed time series at each pixel. Therefore, for these cases, it is not possible to compute spatial metrics like SPAEF and SSIM (as suggested by Reviewer 1). The only case where we computed metrics between maps is Figure 8, which we used as additional measures of calibration accuracy.

Regarding the option to use long-term averages of LST, we decided not to do so because we wanted to assess the model’s ability to capture dynamics at higher temporal resolutions, especially considering that we do not have streamflow data at a daily scale, as mentioned in the answer to comment #1.

To address this comment, we plan to:

• Clarify that metrics based on pattern-oriented metrics cannot be applied to the maps resulting from pixel-based time series errors.
• Add in the revised manuscript figures showing the changes in the long-term average of LST and SC (Figures R1 and R2 in the attached document), along with a table (Table R1 in the attached document) reporting SPAEF and SSIM that target the ability to match the spatial patterns, as also requested by Reviewer 1. As shown in table R1, the values of these metrics are in line with the overall trend of RMSE and Bias.

4. The authors only present maps of the three selected metrics. For interested readers, observed and simulated maps of the actual variables will provide insightful information. The authors could select single days or long-term averages of LST (night and day) and snow cover to illustrate the catchment characteristics and how the model represents those.

Thanks for the suggestion. As mentioned in the answer to the previous comment, in the revised version we will include maps of long-term averages of LST and SC (Figures R1 and R2 in the attached document).

5. It would be interesting to see how the three steps of model development affect the water balance of the model. The authors only present the simulated runoff of the various simulations, but aggregated numbers of evapotranspiration, runoff, groundwater recharge, etc. would provide relevant alternative information.

This is another good suggestion. In the Revised manuscript, we will add the annual components of the water balance, including precipitation, evapotranspiration, and changes in water storage, so that the readers can have better information about how the water balance changes during our stepwise experiments.