Multiple-objective calibration of a conceptual hydrological model using satellite data of snow cover, soil moisture and limited streamflow observations

Khalil, Asma; Tong, Rui; Majid, Zahra; Széles, Borbála; Bertola, Miriam; Parajka, Juraj

doi:10.5194/egusphere-2026-2880

Preprints

https://doi.org/10.5194/egusphere-2026-2880

Preprints

18 Jun 2026

| 18 Jun 2026

Status: this preprint is open for discussion and under review for Hydrology and Earth System Sciences (HESS).

Multiple-objective calibration of a conceptual hydrological model using satellite data of snow cover, soil moisture and limited streamflow observations

Asma Khalil, Rui Tong, Zahra Majid, Borbála Széles, Miriam Bertola, and Juraj Parajka

Abstract. One way to improve hydrological predictions in data-sparse regions is to assimilate satellite data of water cycle components into the calibration of hydrological models. This study evaluates the value of combining satellite snow cover (MODIS) and soil moisture (ASCAT) data with limited streamflow observations to improve hydrological model calibration at the regional scale. The study compares model performance of eleven calibration variants that differ in (a) whether they use satellite data only, (b) the type and temporal distribution of streamflow observations used, and (c) whether satellite and streamflow data are combined. The streamflow sampling strategies cover two scenarios: regularly spaced observations distributed over multiple years (one value per season or per month), and event-based strategies that mimic a single short-term gauging campaign in the wettest, driest, or average year (a peak-flow event with recession plus six bimonthly background samples). The analysis is performed for 213 catchments in Austria, grouped into 119 alpine and 94 lowland catchments. The results show that calibration to satellite data only provides reliable runoff simulations primarily in lower-elevation, drier, and more agricultural lowland catchments. In alpine catchments, adding any limited streamflow data substantially improves model efficiency. The combination of monthly streamflow observations and satellite data (Vmo+sat) results in the best overall runoff performance in lowland catchments, with a median validation runoff efficiency of 0.67. In alpine catchments, event-based streamflow-only strategies achieve median validation runoff efficiencies of 0.69–0.71, close to the regular monthly variant (Vmo, median 0.75) and substantially better than satellite-only calibration (Vsat, 0.26). In lowland catchments, Vmo+sat (0.67) outperforms all event-based variants. Adding satellite data to any streamflow-based variant reduces median snow-cover errors in alpine catchments by approximately a factor of five and consistently improves the simulated soil moisture, although it can reduce runoff efficiency in alpine catchments compared to streamflow-only calibration. These results support the practical value of short, targeted gauging campaigns combined with satellite remote sensing for hydrological modeling in data-sparse regions.

Received: 20 May 2026 – Discussion started: 18 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Asma Khalil, Rui Tong, Zahra Majid, Borbála Széles, Miriam Bertola, and Juraj Parajka

Status: open (until 29 Aug 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-2880', Anonymous Referee #1, 17 Aug 2026 reply
General Comments:
In this preprint, authors aim to address relevant and distinct research objectives with clear added value on top of their methodological predecessor [Tong et al. (2021)], specially by introducing event-based calibration experiments & limited flow observations combining with satellite snow-cover & soil moisture data. Comparison of results between alpine and lowland/hilly catchments from simulation of large sample data of Austria is indeed important strength of this study.
However, some of the main conclusions are difficult to interpret casually as experimental design does not isolate the effect of individual factors. Particularly, effect of adding satellite data cannot be separated from the effect of simultaneous reduction of weights assigned to the streamflow objective. Observation numbers and timing vary concurrently across sampling strategies; event campaigns are defined using retrospective information and uncertainty assessment is missing for supporting several small performance difference interpretations. Notably, most interesting regional patterns revealed remain inadequately explained with hydrological process-based understandings.
This manuscript has therefore high potential but needs careful major revision to enrich interpretation before consideration for publication. Below are some specific comments that Authors should consider to address and improve it for publication:
Specific Comments:
Introduction
The introduction argues legitimate knowledge gaps on sparse discharge observations and satellite data ingested calibration. However, stated objectives are framed like analytical tasks only rather than as clear scientific research questions to answer grounded on previous literature. Inclusion of literature supported research questions would make the manuscript insightful & well justified.

Authors should clearly distinguish what is new addition in this study than Tong et al. (2021). Previous study already investigated satellite data informed calibration and effect of MODIS/ASCAT on internal states.

The motivation behind the expected fundamentally different responses between alpine and lowland/hilly catchments should also be established more explicitly.

Study Area & Data
The physical/empirical basis behind 900 m threshold for separating alpine and lowland catchments is not clearly established. Manuscript should establish that conclusions are robust to this classification.

The statement in lines 96-98 regarding mean annual precipitation in lowlands is 400 mm/year should be clarified with respect to Table 1, where mean annual precipitation for lowlands is stated as 999.5 mm/year.

Section 2.3.3 should describe ASCAT-DIREX product more precisely. The role of Senitnel-1 in processing should be clearly differentiated from the source of SWI as it was utilized in Tong et al. (2021) for downscaling/resampling. Authors should clarify soil moisture retrieval and what is represented by 500 m SWI.

Methods
Clustering
Choice of 4 clusters are supported by elbow method only, while 4^th cluster contains only 5 catchments. Robustness of this clustering process should be presented more convincingly. Authors should provide silhouette score/gap statistics or other cluster stability assessment and test sensitivity to scaling.

Authors should clarify implications of absolute monthly runoff differences in clustering process, as these differences are inherently influenced by runoff magnitude and hydro climatic condition.

In addition, Vsat validation error defined clustering and then clusters are used to evaluate how other strategies could reduce those errors. It implies this clustering analysis is mainly diagnostic, and therefore, authors should be cautious for interpretation in worst performing clusters.

Interestingly, cluster membership is assigned based on observed validation period runoff errors, which ideally would not be available for real ungauged catchments. Therefore, their direct application to ungauged catchments is limited, despite the broader context framing of this study is around spare data and ungauged conditions. It should be explicitly taken into consideration during interpretation and discussions.

Calibration and Optimization
DEoptim is used for calibration & optimization, but manuscript provides insufficient information on population size, number of iterations. Stopping criteria, random seeds or repeated runs. These should be reported with high importance, as several conclusions depend on differences in validation efficiency within very small range.

15 model parameters are calibrated using only 12 discharge observations. Authors should address parameter identifiability/equifinality and demonstrate that argued performance is not strongly dependent on optimization variability.

Authors should state NSE/NSElog calculation approach explicitly. It is not clear if it was calculated from 12 observations or taking complete record. It should be also clarified that how validation performance was evaluated?

Approximately, 120 streamflow values are used for Vmo, 40 for Vse, and only 12 for event-based variants. Therefore, effects of observation number and sampling strategy cannot be separated. Comparison with same observation size is needed to separate sampling-strategy effects from observation-number effects. Table 3 should also report the number of streamflow observations used for each variant.

Another major concern is that ‘streamflow-only versus +sa’ comparison introduces satellite constraints but reduces relative runoff weight from 1 to 0.33. Therefore, again the effect of adding satellite data cannot be separated from the effect of reducing the streamflow objective weight. Consequently, the decline in alpine runoff performance cannot be attributed uniquely to satellite constraints or an inherent multi-objective trade-off. It is very important as the three objective components do not necessarily have comparable numerical scales. Therefore, interpretation discussed throughout Sections 4 and 5 should reflect this identification limitation unless its effect is explicitly evaluated. I would encourage you to evaluate its explicit effect.

Event based strategy relies on retrospective knowledge of campaign year and annual peak event, therefore, its practical application as a field gauging strategy should be discussed with caution.

Please justify the changes of snow objectives (thresholds for cloud cover, SWE and SCA) in this study than previous Tong et al. (2021), as this manuscript compares current performance with previous framework.

Soil moisture has strong seasonal cycles. Please clarify how much of raw Pearson correlation metric reflects calibration sensitive soil moisture variability rather than shared seasonality

Results
Results are overall clearly presented, but several ‘outperforming’ conclusions are made based on narrow range of differences or improvements without conducting statistical tests or uncertainty assessment. For instance, lowland runoff has been reported as improved in more than 50% catchments with median change within 0.1-0.3. Authors did not indicate if these small changes exceed optimization variability or statistical uncertainty. Similarly, very small changes are considered as evidence of preferencing campaign years. Authors should consider uncertainty assessment with appropriate tests for such cases, for instance, pair-wise bootstrap confidence intervals, or Wilcoxon signed rank test or other appropriate tests fitting the purposes.

Reported monthly Vsat errors -520 mm/month (cluster 3) or -1260 mm/month (cluster 4) seems unusually large and highly concentrated in one month only than the reported mean annual precipitation which ranges from 1000-1500 mm/year. How this could be explained in terms of annual water balance? Please verify that units or aggregation are correct or revisit the results.

Manuscript states “significantly underestimate” in Section 4.1 without reporting a statistical significance test. “Significantly” should not be used in this statistical sense unless it is supported formally.

Discussions
All major quantitative results should be reported first in the result section and then should be discussed. For instance, large snow error improvement is discussed only under discussion section which were not reported under Section 4 or in Table 6.

Sections 5.1-5.3 in general restate the overall results and performance more than the underlying hydro-climatic mechanism behind those. Authors should discuss the reasons behind the improvement or degradation of results. For instance, why does adding satellite information slightly improve or preserve runoff in lowland catchments but considerably degrade it in many alpine catchments? Referring this an “inherent multi-objective trade-off” is insufficient because several hydro-climatic mechanisms could dominate this pattern. Please revisit whole discussion sections accordingly. Authors are encouraged to consider examining parameter changes or water balance components to support the discussion.

Apparently, recommendations to prefer dry or average years over wet years are not consistently supported across regions. Therefore, general recommendations regarding this in Section 5.2 and Conclusion should be explicitly linked to its supporting conditions only.

The statement that Vmo+sat is the “best” lowland strategy should preferably be described as the best overall performance balance, since it is not numerically best for every individual output.

Term “reliable” for satellite only runoff should be reconsidered, as median validation values are approximately 0.12 in lowlands and 0.26 in alpine catchments, respectively.

Limitations
Current limitation section seems too brief with respect to the methodological issues that affect interpretation. Authors should also explicitly acknowledge other important aspects deemed necessary to highlight after revision based on above feedback. For instance, hindsight based event selection, potential influence of seasonality, pseudo-ungauged nature of this experiment, or limited generalization beyond Austrian catchments and so on.

Conclusions
Overall, formal objectives of this study are broadly fulfilled at the performance comparison level. But broader practical goal of providing robust guidance for ungauged or data-sparse catchments is only partially achieved, as the campaign experiments are dependent on hindsight, variable observation samples, and unquantified optimization uncertainty. Therefore, authors are encouraged to address those to support its conclusions about relative model behavior under controlled sparse-data experiments strongly.

Minor/Editorial Comments:
Please include the number of streamflow observations associated with each calibration variant in Table 3

Under Section 4.1, please report the alpine/lowland composition of each cluster.

Clipping the snow-error distributions at 0.6 obscures part of the distributions in Figures 5 & 8. Please correct this plot.

Final sentence appears grammatically incomplete for Figure 5

Please correct the term “snow cover error OS” in Figure 8

Presentation quality of all figures needs to be improved.

Wording “short term gauging campaign” may be misleading because the event strategy includes additional bimonthly measurements distributed across the year.

Please check consistency of parameter ranges and notation relative to Tong et al. (2021), especially where direct performance comparisons are made.

Double check first appearance of abbreviations/expansion and then abbreviations in subsequent sections.

Please check reference style as per HESS guidelines.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-2880-RC1
- AC1: 'Reply on RC1', Asma Khalil, 24 Aug 2026 reply
  
  Dear Editor and Reviewers,
  We sincerely thank the Editor and the Reviewers for their careful evaluation of our manuscript and for their constructive and valuable comments. We have carefully considered all the comments and have revised the manuscript accordingly. We believe that these revisions have substantially improved the quality and clarity of the manuscript.
  In the attached file, we provide a point-by-point response to the reviewer’s comments. The reviewer’s comments are reproduced in Italics, followed by our responses. All changes made in the revised manuscript are in Bold for ease of reference.
  We hope that the revised manuscript and our responses adequately address the concerns raised by the reviewers, and we appreciate the opportunity to improve our work.
  Sincerely,
  
  Asma Khalil and all the Co-authors
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2026-2880-AC1
RC2:
'Comment on egusphere-2026-2880', Valentina Premier, 18 Aug 2026 reply

This work presents different calibration strategies for a conceptual hydrological model in the context of data scarcity across catchments with different hydroclimatic characteristics, including both Alpine and lowland catchments in Austria. The results suggest that incorporating remotely sensed information on snow cover and soil moisture into the calibration leads to only limited improvements in model performance, particularly for the lowland catchments.
The paper is well written, and the different sections are clearly structured and presented. The topic is relevant, and I find the proposed approach particularly interesting as a potential strategy for improving hydrological modelling in poorly monitored catchments.
These findings are broadly consistent with observations from our recent study (https://hess.copernicus.org/articles/30/1189/2026/), where we found that constraining the snow module with remote sensing data—through post-calibration replacement of the snowmelt coefficient—resulted in slightly improved or largely unchanged streamflow performance in the target catchments. Although streamflow performance did not substantially improve, we argued that the additional information could nevertheless improve the internal consistency of the model by providing a more realistic representation of processes and states such as snow cover and soil moisture.
An interesting difference, however, is that in our study the snowmelt parameter was replaced after calibration, while the remaining model parameters were left unchanged. We therefore hypothesized that a complete recalibration, allowing the other parameters to adjust to the additional constraint on the snow module, could potentially lead to improved streamflow performance. In the present study, however, the results appear to show that the inclusion of additional remotely sensed information can actually lead to a deterioration in streamflow performance, particularly for the Alpine catchments, where one might expect snow processes and therefore snow-cover information to play a comparatively important role.
I think this result deserves further discussion. In particular, the authors could better explain why constraining the model with snow-cover information does not translate into improved streamflow simulations in the more snow-influenced catchments, and why it may even lead to poorer performance. Could this arise from the multi-objective calibration strategy itself (used methodology, objective function definition, parameter range selection)? It would also be useful to investigate whether the response is related to the actual degree of snow influence of each catchment (e.g., fraction of precipitation falling as snow, snow-cover duration, or relative importance of snowmelt to runoff), rather than simply the Alpine/lowland classification. In fact, I would expect an improvement especially in Alpine catchments where snow is more dominant and in-situ measurements are sparser and hence, the use of satellite data should be an added value.
I am also wondering about the statement that the internal consistency is improved, which should be justified by the results shown in Figs. 5 and 6. I agree with the statement, but the results look a bit obvious. Of course, if the calibration function directly considers those metrics, adding the satellite observations and therefore those additional constraints will improve Osm and Osc. I have the impression that the robustness is evaluated using the same metrics that are included in the calibration, without showing a clear improvement in terms of streamflow prediction.
Specific comments:
L61/L83 The focus on high-flow events is understandable, but is it also justified in a context of drought conditions? In this context, I wonder how well low flows are represented by the model and by the different calibration strategies.
Table 1: Mean annual precipitation includes both liquid and solid precipitation. Given the importance of snow processes for some of the investigated catchments, it would be informative to provide an indication of the partitioning between rainfall and snowfall (e.g., fraction of annual precipitation falling as snow). This would also help the reader assess the degree to which each catchment can be considered snow influenced and would support the later interpretation of the results.
L115 For the interpretation of the results, and particularly in relation to the statement in Sect. 5.1 that poorer model performance may be associated with the sparse network of in situ stations used to interpolate the meteorological forcing, I suggest providing information on the number and/or density of available stations for each catchment. This could be included in an Appendix. More importantly, I suggest investigating this hypothesis more explicitly: is there evidence that catchments with poorer model performance are indeed characterized by a sparser station network? Such an analysis would provide stronger support for the interpretation proposed in Sect. 5.1.
L121 Why are you not using newer version 6.1?
L125 I am not an expert of this product, but so far as I know soil moisture products can be quite uncertain. Please provide the uncertainty of this product and discuss if this can also influence results.
Sec. 3.1 Is clustering just a trick to aggregate and visualize the results?
Table 2 How are the ranges of the parameters chosen? For DDF, for example, we allowed for a larger range. Do you think that constraining the range can also lead to a non-optimal calibration? Are the parameters calibrated for each variant falling within the ranges or at the bounds? In my opinion, it would also be interesting to provide information on how much these parameters vary from one variant to another and which of them are more sensitive to the different variants.
L167 How is a pixel of snow identified? MOD10A1 is a fractional snow product. Do you set a hard threshold and convert it to binary snow/snow free information?
L171 Having as calibration objective function something that counts for the Pearson correlation coefficient, could lead to results that are highly biased in terms of Soil Moisture Content (but still correlated)?
Table 3: It is not clear whether the satellite products are also considered as one observation per month/year or whether you consider all available satellite observations. It would also be meaningful to provide information on how many satellite observations are actually used for both snow and soil moisture, considering limitations such as cloud obstruction and the number of available acquisitions.
Fig 2 Could the bad performance of cluster 4 also be due to the fact that a small number of catchments are within that cluster?
Figure 3 Don’t you consider forest coverage as an indicator? And the soil type/effects of groundwater recharge? Is the hydrological regime also influencing the results? Catchments with higher/lower Q perform differently? Snow cover duration?
Fig 10 There is a strong underestimation for clusters 3 and 4 during the months when the snowmelt peak occurs (right?). Are these catchments characterized by a strong snow influence? Is snowmelt not fully explaining the observed discharge during this period? Are there also strong storm/rainfall events occurring at the same time? How does the hydrograph look during this period?
Sec 5.4 needs to be expanded and better discussed.

Reply

Citation: https://doi.org/10.5194/egusphere-2026-2880-RC2
- AC2: 'Reply on RC2', Asma Khalil, 24 Aug 2026 reply
  
  Dear Editor and Reviewers,
  We sincerely thank the Editor and the Reviewers for their careful evaluation of our manuscript and for their constructive and valuable comments. We have carefully considered all the comments and have revised the manuscript accordingly. We believe that these revisions have substantially improved the quality and clarity of the manuscript.
  In the attached file, we provide a point-by-point response to the reviewer’s comments. The reviewer’s comments are reproduced in Italics, followed by our responses. All changes made in the revised manuscript are indicated in Bold for ease of reference.
  We hope that the revised manuscript and our responses adequately address the concerns raised by the reviewers, and we appreciate the opportunity to improve our work.
  Sincerely,
  
  Asma Khalil and all the Co-authors
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2026-2880-AC2

Asma Khalil, Rui Tong, Zahra Majid, Borbála Széles, Miriam Bertola, and Juraj Parajka

Viewed

Total article views: 78 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
46	25	7	78	5	3

HTML: 46
PDF: 25
XML: 7
Total: 78
BibTeX: 5
EndNote: 3

Views and downloads (calculated since 18 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	1	0	1
Jul 2026	30	16	5	51
Aug 2026	15	9	2	26

Cumulative views and downloads (calculated since 18 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	1	0	1
Jul 2026	30	16	5	51
Aug 2026	15	9	2	26

Viewed (geographical distribution)

Total article views: 53 (including HTML, PDF, and XML) Thereof 53 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 24 Aug 2026

Short summary

Study signifies how the use of limited streamflow records with satellite data shows promising approach to improve hydrological predictions in data-sparse regions. Results show combining satellite observations with even a small number of targeted river flow measurements can support reliable hydrologic modeling in data limited catchments. Further, this approach can enhance research dynamics in regions with no or limited observed data availability by using consistent datasets of satellite products.


Total:	0
HTML:	0
PDF:	0
XML:	0