The effect of satellite-derived surface soil moisture and leaf area index land data assimilation on streamflow simulations over France

Fairbairn, David; Barbu, Alina Lavinia; Napoly, Adrien; Albergel, Clément; Mahfouf, Jean-François; Calvet, Jean-Christophe

doi:https://doi.org/10.5194/hess-21-2015-2017

Articles | Volume 21, issue 4

https://doi.org/10.5194/hess-21-2015-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

https://doi.org/10.5194/hess-21-2015-2017

© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 3.0 License.

Articles | Volume 21, issue 4

Research article

|

13 Apr 2017

Research article |

| 13 Apr 2017

The effect of satellite-derived surface soil moisture and leaf area index land data assimilation on streamflow simulations over France

David Fairbairn, Alina Lavinia Barbu, Adrien Napoly, Clément Albergel, Jean-François Mahfouf, and Jean-Christophe Calvet

Download

Final revised paper (published on 13 Apr 2017)
Supplement to the final revised paper
Preprint (discussion started on 09 May 2016)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of HESS article 2016-195', Anonymous Referee #1, 10 Jun 2016
- AC1: 'Response to Reviewer 1', Jean-Christophe Calvet, 26 Jul 2016
RC2: 'Review of “Integrated validation of assimilating satellite derived observations over France using a hydrological model”', Anonymous Referee #2, 21 Jun 2016
- AC2: 'Response to Reviewer 2', Jean-Christophe Calvet, 26 Jul 2016
SC1: 'Comment on comparing observations and model simulations for anthropogenically influenced catchments', Jean-Philippe Vidal, 04 Jul 2016
- AC3: 'Response to Jean-Philippe Vidal', Jean-Christophe Calvet, 26 Jul 2016

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (29 Jul 2016) by Wolfgang Wagner

AR by Jean-Christophe Calvet on behalf of the Authors (04 Aug 2016) Author's response Manuscript

ED: Referee Nomination & Report Request started (12 Aug 2016) by Wolfgang Wagner

RR by Anonymous Referee #3 (11 Oct 2016)

Suggestions for revision or reasons for rejection

Integrated validation of assimilating satellite derived observations over France using a hydrological model.

HESS, Fairbairn et al. 2016

The paper discusses the assimilation of ASCAT soil moisture retrievals and LAI into the integrated SAFRAN-ISBA-MODCOU modeling system over France, using the SEKF.

The topic is suitable for HESS, but from a data assimilation point of view, the paper is rather weak. I also noticed that several recommendations from an earlier review were not included and the text can be much improved.

Main comments on data assimilation:

- The introduction sounds like a twisted excuse to not follow recent advances in land surface data assimilation and to get away with a suboptimal system. Please acknowledge the true state-of-the-art:
a) P.2, L29: one (?) study found no advantage of 2D Kalman filtering over 1D Kalman filtering. Maybe. Yet, very many other studies use 2D/3D Kalman filtering and that is the only correct way of doing Kalman filtering if we deal with different spatial resolutions.
b) The SEKF may be a preferred method at some operational centers, but in most other centers, there is a push towards the EnKF.

- Why did the authors continue to use the SEKF without (or w/ minimal) alterations and then run into the same problems as already reported in Draper et al. (2011)? One of the conclusions is that an EnKF and assimilation in a slightly deeper surface soil layer may alleviate the problems that are experienced in this paper: trying out these recommendations would be food for a paper, but rerunning the same problems, is not so nice.

- Would the EnKF really help, as suggested in the abstract? The introduction says that the EnKF and SEKF produce results with similar accuracies. (p.3 L13).

- The model is run at 8 km, the ASCAT data are at 25 km, the LAI data are at 1 km resolution. Data assimilation should take care of these spatial discrepancies, especially to downscale the coarser data to the finer resolution. A priori interpolation just does not make sense: this adds unnecessary errors and the subsequent 1D assimilation wrongly assumes that spatially independent observations are assimilated, while in reality there is a perfect oversampling with perfect spatial error correlations. The latter is especially a problem when (p.7, L26) the “SEKF analysis is calculated independently for each patch, with the same obs used for all patches in de grid-box.”

If the spatial errors are discarded, then the observation error variance should at least be increased – which is done, but far too little to make any difference (i.e. from 0.050 to 0.055 m3/m3). In addition, the observation error should be adjusted in line with the rescaling and be spatially variable if the same obs is used for different vegetation classes. Yet, there is no linkage between obs errors and vegetation class in this paper.

- P.6, L28: How is it possible to assume zero (error, I assume) covariances between LAI and WG2, and at the same time derive meaningful Jacobians that calculate e.g. dLAI/dWG2. Isn’t this a basic contradiction? Please explain.

- Why calculate dLAI/dLAI by adding perturbations? Should this not be simply =1? Why not?

- P.4, L26: Why is soil moisture rescaled, whereas LAI is not? The KF assumes unbiased innovations in either case: does this work out in the end?

- P.6, L30: why is WG1 not part of the state vector?

- P.7, L2: what do you mean by a 24-hour assimilation window? A filter is used, not a smoother. Data are assimilated every 3 days for soil moisture and every 10 days for LAI.

- In short, the paper needs a thorough and in depth acknowledgement of all the violated assumptions in the KF: no spatial error correlations where perfect error correlations are present (and consequently, an exaggerated impact of the DA); either a relationship or none between LAI and WG2 in the setup of the background errors and the Jacobians, the choice for the observation error variance, etc.

I would suggest to validate the optimality of the data assimilation system (e.g. white innovations in space and time?) and hopefully it comes out “just fine”: in that case, the paper could perhaps be considered for publication, otherwise it becomes hard to justify.

Other:

- The title is not very representative: “Integrated validation” sets high expectations, we do expect advanced validation methods, or at least more than just only streamflow (i.e. maybe include turbulent fluxes, groundwater, in situ soil moisture,…). I would rephrase it as something like “The effect of [assimilation] on streamflow estimates”. Secondly, the authors responded to one of the reviewers that he/she was inaccurate about referring to assimilation in a hydrological model – yet, that is exactly what the title says.

- P.2, L10: observation network --> observation coverage (already mentioned in earlier review, and I agree that this needs to change)

- Throughout, please check all sentences. There are too many textual issues (in an already reviewed manuscript!) to list, but here are some examples:

* P.2, L9: “… instruments… are subject to retrieval errors.” --> The *data* are subject to retrieval errors. And besides: we could assimilate raw radiance or backscatter observations and circumvent ‘retrieval’ errors (but errors would be elsewhere).
* p.4, L2: “… integrated validation of the soil moisture fluxes.” What is a soil moisture flux? Did you mean groundwater recharge, runoff, river discharge, …?
* p.9, L13 : SIM is not a “tool to validate”. SIM provides model simulations which can be validated using in situ observations, and using some specific validation metrics.

- p.4, L17: Draper et al. (2011) already assimilated SSM into SIM, but they did not perform the validation against streamflow, which is the whole goal of the current paper. To be complete, this paper should include a validation of these older results (Draper et al., 2011) against streamflow.

- P.12, section 3.2: Are all the WG1 and LAI metrics simple internal checks? I.e. comparison against data that are assimilated into the model? Please repeat that here. It would make much more sense to validate against independent soil moisture observations. What is the statistical significance level when it is stated that “LDAS1 significantly improves the fit…”

- What is the model integration time step? Since the assimilation is done at 9:00 UTC, whereas the data are taken at 9:30 UTC, I assume that there must be a very long model time step (one hour or more?). Could the long model integration time step be another cause of inferior model performance at some times?

Hide

RR by Anonymous Referee #4 (30 Nov 2016)

Suggestions for revision or reasons for rejection

General comments

1. I understand how correcting the minimum LAI for an important land cover type (grassland) in the model has the desired effect of reducing streamflow overestimations. However, what is the scientific merit of replacing the minimum LAI simulation by the minimum of satellite observations? Firstly, it would be more justifiable to address these LAI shortcomings in the modeling itself, rather than replacing the values that are undesired. Alternatively, external observations can be used to correct the model LAI, but in that case, I believe it should not be restricted to only the LAI minimum for grassland. Only correcting this feature and not any other features is quite arbitrarily. Looking at Figure 3, it seems that, besides underestimating observations over grassland in winter, the model LAI is significantly larger than the observations over deciduous forest (almost the entire year) and over C3 crops in summer and fall. The latter discrepancies seem to be equally important (if not more) as the underestimation over grassland. Furthermore, looking at Table 3, correcting the grassland minimum LAI even significantly increases the average bias against the LAI observations, which feels like a counter-intuitive approach.

Therefore, I recommend the authors to perform an additional experiment, with complete rescaling of the model LIA to the climatology of the LAI observations, and to re-evaluate the simulations of discharge, or to provide extensive scientific proof (e.g. based on other LAI products like MODIS) that the winter grasslands are really the main or only concern.

2. I have some concerns regarding the choice of not correcting biases in LAI prior to the assimilation. I can see that the purpose of this experiment is to verify whether the assimilation can mediate the bias in LAI. Although being appealing, data assimilation is theoretically not designed to correct model bias. Again, as mentioned in the previous comment, such bias should actually be resolved in the model itself, or by rescaling the observations. I believe the same principle applies for LAI as for soil moisture, for which biases were appropriately removed in this study. The assimilation of biased LAI may not have had the effected that was hoped for, because of the small Jacobians in winter, but it could for instance also have disturbed the model behavior in summer, when larger Jacobians as well as large biases in LAI over crops were present. I believe this is an important issue that is not addressed in the paper. Moreover, the assimilation of (biased) LAI may also impact the climatology of the soil moisture simulations, which then again becomes biased with respect to the (previously bias-corrected) ASCAT retrievals. This could potentially be another important reason why the soil moisture assimilation, in combination with LAI assimilation, wets the lower layer and fails to improve discharge simulations.

I suggest the authors to include an assimilation run for which the bias in LAI between model and observations is removed a priori, e.g. by rescaling the model LAI to the observations, cfr. comment 1 above. The impact of the summer bias over cropland (in combination with large Jacobians), and of deciduous forest bias need to be better addressed. Also, the impact of LAI assimilation on soil moisture bias between simulations and ASCAT retrievals requires further investigation. I believe the bias-correction of the soil moisture observations should best be done with respect to the LAI assimilation experiment.

Specific comments

1. P2.L30-P3.L22: After a comment from the previous review, parts of the introduction have been rewritten to more strongly motivate the use of an SEFK relative to an ENKF. However, I feel like the text is rewritten as to actually recommend the SEFK over the ENKF. I don’t believe there is really a need to defend the use of the SEFK so strongly, almost in a confrontational way with respect to the ENKF. Each has its own advantages. I would suggest to rewrite some of the phrases, making it less of a confrontation between both methods. It also feels a bit strange that the authors commend the SEKF in the introduction, whereas in the conclusion, they recommend future use of an ENKF.

Similarly, the authors mention the study by Gruber et al. (2015) in P2.L29 to defend the choice for a 1D Kalman filter. I don’t have any problem with using a 1D filter and mentioning this paper as a support, but suggest to make a more cautious statement. Theoretically one would expect a better performance with 2D systems in case of coarser-resolution observations, as they decrease the representativeness error between model forecasts and observations. As it stands, it feels like a general statement that advises the use of a 1D filter, with which I do not agree.

2. P6.L9-11: The processing of the ASCAT data needs a little bit more explanation. How are data interpolated to the 8-km grid? Which temperature threshold was used for filtering frozen areas (zero Celsius or larger)? Have open water fractions or snow been dealt with? Why is altitude used instead of the topographic slope, or the topographic complexity flag in the ASCAT product? A plateau at high altitude seems more preferable to me than a terrain with strong slopes at low elevation, considering the backscattering mechanisms at hand. Adjusting these processing issues could be a way forward for getting improved results in future studies.

3. P7.L26: The authors assimilate the same (aggregated) LAI observation for all grid patches, which may for instance contain a forest patch and grassland patch? Is there any benefit of doing so? I would think this may cause very large increments that could potentially destabilize the model? Updating the aggregated grid cell LAI seems both more theoretically correct and computationally efficient to me. Please comment, and potentially modify the approach.

4. P25.Table2: The assimilation of LAI only seems to flip the sign of the bias against observations, i.e. from +0.11 to -0.08. How is this possible? If the assimilation correctly balances forecast and observation errors (set equal for LAI), it should provide a result that is in-between the observations and simulations? Please comment.

5. P25.Table2: Also, the impact of the LAI assimilation seems to be very large. Is it potentially overdone? This is in large contrast with the results of the soil moisture assimilation, for which Table 3 shows almost negligible impact. Could you please comment? I was also wondering if the impacts on surface soil moisture are comparable to those observed by Draper et al. (2011)?

Hide

ED: Reconsider after major revisions (further review by Editor and Referees) (01 Dec 2016) by Wolfgang Wagner

AR by Jean-Christophe Calvet on behalf of the Authors (02 Mar 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (09 Mar 2017) by Wolfgang Wagner

RR by Anonymous Referee #4 (13 Mar 2017)

Suggestions for revision or reasons for rejection

Overall, I am pleased with the modifications to the manuscript and responses to the review comments. I believe the manuscript has improved in quality.

There are still a few items that I would like to bring to your attention, which you could consider for adjusting some parts of the manuscript:

1. I do not entirely agree with the statement of the authors in the reply to comment 3 from referee 1 and comment 1 from referee 2:
a. Biases in SSM can show (systematic) variability over time. For instance, saturating and particularly drying rates are often different between models and observations, which could be due to deficiencies in the model, rather than in the parameters (wilting point and porosity). Applying ‘seasonal’ CDF-matching could therefore eliminate part of the soil moisture information contained in the observations, potentially reducing the impact and benefits of the assimilation.
b. Regarding LAI, a linear rescaling (as referred to in the reply to the comments) would not impact the timing of the seasonal cycle, unless it is applied separately over each season (as for the SSM bias-correction). Therefore, even with bias-correction (but depending on the method applied), it should still be possible to correct for the model phase error by the assimilation. But at least, such bias-correction would converge the magnitude (and variability) of the LAI observations and simulations.
If the authors do not want to eliminate the climatological information (seasonal cycle) from the LAI observations, this is achievable by not performing bias-correction over each season separately. If the idea is to be consistent with SSM bias-correction, then the latter should probably also conserve the seasonal dynamics (which is not the case with seasonal CDF-matching).

2. I suggest the authors add at least a caveat that the LAI innovations (Fig. S1.8) are quite strongly correlated over time, which is not optimal.

3. In general, I think it is important to clearly stress that the reasons for not obtaining improvements in discharge simulations may be due to model deficiencies and the particular setup of the assimilation system, for not to discourage future research efforts on the joint assimilation of SSM and LAI, which I believe is a promising way forward.

Hide

ED: Publish subject to minor revisions (further review by Editor) (14 Mar 2017) by Wolfgang Wagner

AR by Jean-Christophe Calvet on behalf of the Authors (20 Mar 2017) Author's response Manuscript

ED: Publish as is (22 Mar 2017) by Wolfgang Wagner

AR by Jean-Christophe Calvet on behalf of the Authors (22 Mar 2017)

Short summary

This study assesses the impact on river discharge simulations over France of assimilating ASCAT-derived surface soil moisture (SSM) and leaf area index (LAI) observations into the ISBA land surface model. Wintertime LAI has a notable impact on river discharge. SSM assimilation degrades river discharge simulations. This is caused by limitations in the simplified versions of the Kalman filter and ISBA model used in this study. Implementing an observation operator for ASCAT is needed.