Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios

Courvoisier, Fabien; Kraft, Basil; Haddad, Yann Yasser; Zappa, Massimiliano; Gudmundsson, Lukas

doi:10.5194/egusphere-2025-6058

Preprints

https://doi.org/10.5194/egusphere-2025-6058

Preprints

12 Dec 2025

| 12 Dec 2025

Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios

Fabien Courvoisier, Basil Kraft, Yann Yasser Haddad, Massimiliano Zappa, and Lukas Gudmundsson

Abstract. Climate change is intensifying the global water cycle, altering both mean runoff and extremes, and strengthening the need for reliable hydrological projections to support adaptation. Traditionally, such projections have relied on process-based models. More recently, machine learning models, and in particular Long Short-Term Memory (LSTM) networks, have shown strong skill in predicting and reconstructing runoff from observations, raising interest in their use for hydrological projections. However, their ability to provide stable and physically credible results when forced with future climates beyond their training domain remains largely unexplored. Here we evaluate this question in Switzerland, a region strongly exposed to warming due to its alpine environment and glacier influence. An LSTM trained on observed meteorological and discharge data is driven with CH2018 climate and glacier projections for 1981–2100, and benchmarked against Hydro-CH2018 simulations from the process-based model PREVAH under identical forcings. Results show that the LSTM reproduces key hydrological signals closely – wetter winters, drier summers, and elevation-dependent trends – consistently across catchments and climate chains. Divergences are most pronounced in alpine and glacier-fed catchments, where runoff dynamics are more complex, yet the main governing patterns are captured. The largest limitation arises for extremes, where the LSTM underestimates peak flows, consistent with previously reported saturation effects. Overall, this study demonstrates that LSTMs can deliver robust mean-flow projections and trends comparable to a process-based benchmark, while highlighting persistent challenges in representing hydrological extremes.

Received: 07 Dec 2025 – Discussion started: 12 Dec 2025

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

Preprint (PDF, 27026 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (27026 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

16 Sep 2026

Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios

Fabien Courvoisier, Basil Kraft, Yann Yasser Haddad, Massimiliano Zappa, and Lukas Gudmundsson

Hydrol. Earth Syst. Sci., 30, 5873–5900, https://doi.org/10.5194/hess-30-5873-2026,https://doi.org/10.5194/hess-30-5873-2026, 2026

Short summary

Fabien Courvoisier, Basil Kraft, Yann Yasser Haddad, Massimiliano Zappa, and Lukas Gudmundsson

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2025-6058', Anonymous Referee #1, 20 Jan 2026

Dear Authors,
I have read your manuscript with great interest. The study is relevant and well-conducted, and the manuscript is written coherently. The findings address the suitability of LSTMs for runoff projections under future climate scenarios systematically.

I have no major remarks regarding the implementation of the methods or the presentation of the results and their discussion.

Please find attached a short summary and my general and minor comments.

Best regards,

Sanika Baste

Citation: https://doi.org/10.5194/egusphere-2025-6058-RC1
- AC1: 'Reply on RC1', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authors.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC1
RC2:
'Comment on egusphere-2025-6058', Anonymous Referee #2, 24 Jan 2026

The study by Courvoisier et al. compares an LSTM architecture with a process-based model regarding runoff behaviour across Switzerland - for historical as well as future (extreme) climate change projections.
The paper addresses an important topic and I generally like how the authors approached the problem, presented and discussed the results. I enjoyed reading the work, particularly the discussion and interpretation.

However, from a process-based perspective and for explaining differences between the models, I am missing some crucial information. And I think a few additions to to the results could help with the interpretation. While I feel the paper has the quality to be accepted in HESS, I think a few of my comments require this to be a 'major revision'.

I hope the authors find my comments useful to improve the manuscript.

Comments to individual sections:
Section 2.1: Suggest to better explain in the text what the differently labelled catchments in figure 1 are used for. e.g. unclear what "second part of our analysis" means in the caption at this stage.

It is unclear why 96 minimally impacted catchments were chosen and then projected on 307 - you introduce a bias already (e.g. missing out on reservoir impacts to name one obvious point)?
Section 2.5: The glaciers play a crucial role in the future runoff predictions, I think. Could you add some more detail here. E.g. which/how many of the basins have no glacier extent in summer by end of century and is this consistently applied in both LSTM and PREVAH?
Section 2.6 and 3.1: You compare these two models, and try to attribute differences - but I don't know where the models actually differ. I strongly suggest to give an overview (a table with a side-by-side comparison would work nicely) of the diverging input data and model structure of PREVAH vs the LSTM - observed meteo, cc data (and meteo parameters used), cal-val-test approach, glacier data and simulation approach in PREVAH, spatial representation, observational vs projection catchments, you train the LSTM on natural catchments - what about reservoirs in the projection catchments - are those included in PREVAH...?

In this sense, I don't see a value in section 3.1 without mentioning PREVAH. I think it is more important to understand the differences than explaining the LSTM in detail alone (E.g. I'd be interested in how PREVAH simulates glacier, snow, ET which is needed to understand the differences presented (e.g. in section 4.5)).
Section 4.3/4.4/4.7 I like the maps. However, I wonder how a scatter plot LSTM vs PREVAH Runoff (could add the corresponding r²) would look like next to the maps (one dot=one catchment, x=LSTM runoff, y=PREVAH runoff, color of dot could be represented by observational, projection, selected catchments or the elevation band color or the ecoregion ... whatever you find most appropriate). This would give a better sense of the actual differences between the models and allow more intuitive diagnostics for why you see the differences.

Major comments:

l.132ff I'd like to understand the bias adjustment and cc data better: Which "Swiss gridded observations" - are those the same that you forced the model with (chapter 2.2)? Did you conduct the QM yourself or does CH2018 provide that? I suggest to discuss the implications if the bias adjustment was done on a meteorology dataset different to the forcing dataset. Also, how did you prepare the cc data for the models: Spatial aggregation->Bias adjustment. Or Bias adjustment->spatial aggregation?
l.220ff I like this cross validation methodology. I assume, per fold, you have 12 catchments that are validation, 12 test and 72 train. Is it correct that you took for valid: year 2016, 2017, 2018 of the 12 validation catchments; year 2019-2024 of the 12 testing catchments and the remaining years of the 72 training catchment for train? But what do the grey shades in Fig 2ab mean - do they correspond? what is the light grey in b that is not part of a?
l.242/369ff as far as I understand your LSTM model architecture is the same as Kraft et al. 2024 - with the only difference that you added dynamic glacier data. I'd have expected the LSTM would perform slightly better with this additional information. Do you have a (short) reasoning why it didn't (including Kraft et al's LSTM in the suggested table might help to explain this - see comment to 2.6)?
l.339 I am confused by this. Yes, fig 8 shows a stronger decline for PREVAH vs the LSTM, but I can't see that in fig 7. For 7b,c,d,f I even see the opposite.
l.347ff, 413ff and l.529ff I think the key question here is: Why is this happening and which model is closer to reality? I know that you cannot answer this question as it's the future. This likely goes beyond your scope, but I wonder if it would help (you could add your thoughts to the discussion):

1. when comparing to observed data. I assume these future events are significantly outside your training data? If that wouldn't be too extreme, you could look at the test data of the LSTM (section 3.5) and extract the maximum/minimum precip periods and check the same events for PREVAH, and evaluate which model is closer to obs (this could help for your discussion in chapter 5.7)?

2. to attribute this to PREVAH's 'knowledge' of the driving processes (such as the additional climate data it gets to calculate ET) different constraints (or no constraints) for the glacier extent, the internal mass balance constraint that the LSTM doesn't have (could the application of a mass-conserving LSTM improve the situation)?
l.584 You earlier mention that this interpretation of low-flow performance requires caution and I don't see this adequately evaluated in your paper to merit mentioning this in the short conclusion.

Minor comments:
l.93 when does the data end/what was your cutoff? Mention the spatial resolution.
l.96 PRISM and SYMAP - reference and spell out on first use
l.102 how did you spatially average in detail? Extract entire cells or did you 'split' cells? If the catchment area is small and the grids large, you can introduce errors particularly in mountainous terrain.
l.116 Can you give a rational why you used topsoil information only?
l.126 suggest to add that it is based on the CMIP5 framework
l.129 you earlier write that the resolution was 2km - how is the product downscaled from the 12-50km CORDEX resolution to 2km?
l.148 Suggest to make it clearer whether you did any glacier simulations of if this was Brunner et al. 2019b work.
l.150 I think you mean section 3.2
l.185 in section 2.5 you cite Brunner et al. 2019b as the source for the glacier data
l.201 I don't see the 24 tested alternatives in Kraft et al. 2025 section 3.6 - seems you employed the pre-print version of their approach?
l.251 fig 4: suggest to add the number of catchments per boxplot (n=x) in the caption.
l.261 I don't think DJF is a period where much melt dynamics is going on in Switzerland and JJA is not really low flow in most of the catchments (you mention this yourself in l.271-272)
l.262 add that this is about the annual panel
l.307 I think the LSTM surpasses PREVAH by 2060, or?
l.327 fig 8 caption "(mm period−1 vs 1991–2020) for 2071-2100" I think you mean something like 2071-2100 - 1991–2020?
l.476-482 the language here suggests we are still in results. Suggest to rephrase.
l.539 deeper deeper
l.576 whether
l.715, l.623 provide links to final paper?

Citation: https://doi.org/10.5194/egusphere-2025-6058-RC2
- AC2: 'Reply on RC2', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authorss.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC2
RC3:
'Review of egusphere-2025-6058', Anonymous Referee #3, 16 Feb 2026
General comments:

This paper investigates the question if LSTMs can be used for climate change impact prediction or rather: are they a valuable alternative to process-based hydrological models for climate change impact predictions in alpine environments? Or even more precise: are LSTM a valuable alternative to the selected process-based model (PREVAH)? Previous work has asked this question for other case studies and other models. From the paper as it is currently framed, it is unclear if this is simply a state-of-the-art case study or if it goes beyond the state-of-the-art. This having said, case studies can well desserve publication in HESS but given that the methods have all been used / have been developed in previous work, I think that the paper could make clearer what it contributes to the literature.

More importantly: as far as I see the reference simulations with the selected PREVAH model are problematic as reference simulations: these simulations are calibrated for some of the catchments; but for many other catchments, the model is somehow regionalized (we do not have information on this in the paper). I do not think that this is good practice: How can such a sample of two very different populations be used as a reference to evaluate a purely data-driven approach against? Furthermore, we see from Figure 3 that for many catchments, the reference model has an NSE value below 0.7 and even below 0.5: even if I cannot relate entirely to what 0.7 means for these catchments, 0.5 at daily time step is for sure a really bad performance for this climate, so these catchments should not serve as reference to compare the LSTM against. Furthermore, given the presented performance comparison, we simply know that the LSTM has a similar performance distribution. This is not informative: what if the LSTM does well for the the ones that PREVAH does not well and vice versa? If the focus of the paper is on seasonal signals only, we should at least expect a sensoring of the catchments / regions for which the seasonal signal is not well reproduced by a) by PREVAH, b) by LSTM, c) by both (which is very interesting!).

Furthermore: we should not forget that we talk about streamflow simulations here but we do not see any actual simulations, not even for seasonal simulations: I think that it is of outmost importance to reconnect the data-based methods to actual data and hydrology. Otherwise we do not learn much about how well LSTMs perform for the selected hydroclimatic region.

From a model development perspective, information is missing: how well does the PREVAH model, forced with climate data (rather than with observed precip & temperature) reproduce observed streamflow? This is a serious lack of information: the models are trained with observed data and then run with climate data; the divergence between the models for the future periode might be rooted at least partly in the divergence of their simulations for the observed period.

Furthermore, the training input data set contains a gridded spatial rainfall product, which can be assumed to show very different quality for different catchments in an alpine environment. Do the climate scenarios have the same spatial resolution? And are the gridded climate scenarios debiased to the spatial rainfall product?

Otherwise, each of the models (PREVAH and LSTM) learns how to deal with the observation-based data set (i.e. how to translate it into observed streamflow), but then makes its own errors when applying what it has learned to a rather different data set. In this setting, what can you learn from such a comparison framework?

Let's imagine: Model A is better than model B for the observed period if fed with observed data, but worse than B if fed with climate data. What do we learn from this? That model A was overfitted? Or that the climate data is not good? If now model A is close to model B for the future period, what does this tells us? That by chance, they both agree? I know that this is not the objective of the paper but given that it contributes to the question of how to use LSTM for climate change impact studies, these questions should be discussed.

Next: the hydrological process model certainly received PET as input for training. How was dealt with this for climate change simulations? And does the LSTM receive PET ? And if not, is this not unfair?

Next: The LSTM is optimized with the mean squared error (MSE) between normalized simulated

and observed discharge. This is certainly very different from the optimisation criterion used for the process model (MSE is rarely used). Why was not a similar / the same criterion used? And what can you learn from two models that are optimized based on different criteria? Besides: was the PREVAH model calibrated with some global optimisation method?

I believe that this paper needs to better frame the analysis framework, improve is A compared to B and what conclusions can be drawn in its methods section. Furthermore, we need more details on the model set up.

Detailed comments:

Abstract and elsewhere: what means "stable results" and how do you measure "physically credible"?

Intro: what do you mean by "These results highlight their robustness and scalability"? Does any of the references explicitely assess the robustness of the models and if so, how is this defined in a hydrological modelling context? What is meant by scalability, what paper discusses this and how is it defined in a hydrological modelling context? Similar what is meant by "achieve strong generalization performance" in a hydrological context? The introduction has to be much more hydrology-specific.

Intro: what do you mean by "Despite their success in present-day forecasting,"? Is this paper about present-day forecasting (forecasting = predicting current state based on previous state) or is it about continuous simulation (prediction) of hydrological states based on driver inputs (i.e. without feeding in the previous hydrological state)?

The abstract states "Divergences are most pronounced in alpine and glacier-fed catchments, where runoff dynamics are more complex, yet the main governing patterns are captured. ": I would argue that the mean governing patterns in these environments can be capture by feeding only temperature into the LSTM; should this be toned down?

Section 2: "The dataset includes most of the variables used in Kraft et al. (2024) and combines those typically employed to force the spatially distributed model PREVAH" - can this be more precise: did PREVAH use similar data or might part of the divergence come from different reference data for parameter estimation?

Section 2: do the same dynamic glacier fractions feed into the LSTM and into the PREVAH simulations or do they receive different input? This is unclear in my view.

Section 4: what is CH-RUN? This should be clearer somewhere

What is "dynamical stability of projections", should this be defined in the methods?

The paper mentions that PREVAH is widely used; the model is however only unsed in Switzerland (given the references)
Citation: https://doi.org/10.5194/egusphere-2025-6058-RC3
- AC3: 'Reply on RC3', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authors.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC3

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

ED: Reconsider after major revisions (further review by editor and referees) (20 Mar 2026) by Ralf Loritz

AR by Fabien Courvoisier on behalf of the Authors (01 Jun 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (11 Jun 2026) by Ralf Loritz

RR by Anonymous Referee #2 (12 Jul 2026)

RR by Anonymous Referee #3 (10 Aug 2026)

Suggestions for revision or reasons for rejection

First of all, I would like to apologise for the delay of this re-review and like to thank the authors for the revised version. Unfortunately, the authors’ response does not contain a detailed response on how the manuscript was revised, the response is mostly written in future tense, and I had to check myself what changes have been made to the manuscript.

Overall, the new version has improved, the scope and details have become clearer and section 3.7 is very valuable to understand differences between modelling frameworks. From reading the track changed version, I understand that my comments have been addressed even if this was not clear from the responses. Despite the above (and I acknowledge that you made clear that the process-based model is not the truth), I come back to my critic that we do not know enough about the process-based model:

First of all, some of my comments were answered by saying that “this is how it is done in large sample hydrology”. I understand that large sample hydrology tends to (uncritically) re-use existing simulation results but I do think that this is only justified if the original simulations are well documented (which seems not to be the case here).

If a new modelling study uses an existing one as reference data set, this reference data set should be critically assessed by someone. This step seems to have been omitted so far. I went to the cited papers and noticed that none of them critically discusses the simulation quality of the most recent PREVAH generated data set:

- Viviroli et al., 2009b presents the model and briefly some past applications
- Viviroli et al. 2009a presents a regionalisation approach for flood estimation, with a regionalization technique that focuses on standard conditions and on flood conditions. The results are presented for flood-focused calibration (“It is important to note that all regionalisations were computed on the basis of the flood-calibrated parameter sets ») for 49 representative test catchments; in other words, it does not present the regionalization for regime studies or water balance studies. The model is at resolution 500 m x 500 m.
- Brunner et al., 2019b (in the response it is stated: “the PREVAH simulations used here originate from the Hydro-CH2018 dataset (Brunner et al., 2019b), where the modelling setup is described in detail »): the paper focuses on extreme flow regimes; Brunner et al do not discuss how well the model performs. Instead, it refers to another paper for calibration: “For the calibration of the model parameters, meteorological and discharge time series from 140 mesoscale catchments covering different runoff regimes were used.(..) More details on the calibration and validation procedures can be found in Köplin et al. (2010).»
- Köplin et al. (2010) is a climate change study (the work of Viviroli et al. 2009 is not); it refers back to Viviroli, 2009a,b,c for model calibration and states that it also uses the parameter set for standard conditions. (“The tuneable parameters are calibrated for standard and flood conditions, (..). According to the objectives of our study, both parameter sets (standard and flood calibration) are used for further investigations. ».
- The work of Köplin also used a 500 m x 500 m resolution. They probably used very different input data than is available to date (in particular climate data but probably also meteo station or gridded data). Accordingly, the performance of their model driven with new data (as done in Brunner et al., 2019b) would need to be checked. As far as I see, we have no details on the Brunner et al., 2019b data set and I am not sure that we have access to it (did not check). I guess we would need a data availability statement for this.
- In fact, the work of Brunner et al., 2019b uses a 200 m gridded version (“A gridded version of the model at a spatial resolution of 200 m was set up for Switzerland (Speich et al., 2015)”; Question: did you use a 500 m x 500m version as stated in the paper or a 200 m x 200 m set up?
- Going to the Speich et al. 2015 paper reveals yet another reference: “The gridded values were calculated using the rainfall-runoff model PREVAH (Viviroli et al., 2009) in its spatially explicit version, as applied by Schattan et al. (2013)”
- going to Schattan et al. 2013 refers to Zappa et al., 2012, a German paper for which it is unclear if it presents model calibration and performance; but Schattan says that they used the Köplin parameters (“Calibration parameters for the investigated regions were obtained from a regionalized set of parameters (Köplin et al., 2010). »);
- To this point, it is not clear what calibrated version Brunner et al., 2019b used. What I guess is that they used the parameters of Köplin that were regionalized for a 500 m x 500 m grid but applied to a 200 m x 200 m grid.
- The later work of Brunner et al. 2019b explicitly did not do any evaluation of the simulations obtained from this “old” calibrated model, driven with new data (any calibrated model has to be recalibrated for new data). The paper simply states: «The hydrological model has been calibrated using observed meteorological data, but will subsequently be fed with meteorological data simulated by a set of GCM–RCM combinations. It is assumed that the parameter set derived in the calibration procedure will still produce reliable results (..)”.
In conclusion: Brunner et al., 2019 used parameters calibrated for a model set up for a different resolution and with different input data. One has to dig into many papers to understand where the set-up actually comes from (Zappa et al. 2012?); I do not think that this is a good reference data set to compare a new model against. At the very least, this should be critically discussed and the reader should be enabled to understand where the set-up comes from and where to download the data.
In general, and as discussed earlier: if we compare models, we need to be able to understand where differences potentially come from. At the very least, we need relevant references to understand the process-based model.
What is important to note here is that the above work should have been done by earlier authors. It is unfortunate that this strong critic “falls back” on the current first author but it is perhaps time to interrupt the chain of papers that re-use older simulations without giving some more details on them.

Furthermore:
Comment 37 (your numbering) : as far as I see, the MS does not show simulated streamflow time series anywhere; we cannot appreciate if the LSTM produces times series that ressemble streamflow in the corresponding hydroclimates; even if other, similar papers do not present streamflow time series: it is an important question to ask: do we want data-driven models to produce time series are only statistics? No need to add a figure at this stage but perhaps simply say that you checked the time series and that they look like streamflow?

References :
- Köplin, N., Viviroli, D., Schädler, B., and Weingartner, R.: How does climate change affect mesoscale catchments in Switzerland? – A framework for a comprehensive assessment, Adv. Geosci., 27, 111–119, https://doi.org/10.5194/adgeo-27-111-2010, 2010.
- Schattan, P., Zappa, M., Lischke, H., Bernhard, L., Thürig, E., and Diekkrüger, B., 2013. An approach for transient consideration of forest change in hydrological impact studies. In: Climate and Land Surface Changes in Hydrology, Proceedings of H01, IAHS-IAPSO-IASPEI Assembly, Gothenburg, Sweden, July (IAHS Publ. 359). pp. 311–319.
- Speich, M. J., Bernhard, L., Teuling, A. J., and Zappa, M.: Application of bivariate mapping for hydrological classification and analysis of temporal change and scale effects in Switzerland, J. Hydrol., 523, 804–821, https://doi.org/10.1016/j.jhydrol.2015.01.086, 2015.
- Zappa, M., Bernhard, L., Fundel, F. & Jörg-Hess, S. (2012) Vorhersage und Szenarien von Schnee- und Wasserressourcen im Alpenraum. In: WSL (ed.): Alpine Schnee- und Wasserresourcen gestern, heute, morgen. Forum für Wissen 2012, 19–27.

Hide

ED: Publish subject to technical corrections (17 Aug 2026) by Ralf Loritz

AR by Fabien Courvoisier on behalf of the Authors (24 Aug 2026) Manuscript

Post-review adjustments

AA – Author's adjustment | EA – Editor approval

AA by Fabien Courvoisier on behalf of the Authors (14 Sep 2026) Author's adjustment Manuscript

EA: Adjustments approved (14 Sep 2026) by Ralf Loritz

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2025-6058', Anonymous Referee #1, 20 Jan 2026

Dear Authors,
I have read your manuscript with great interest. The study is relevant and well-conducted, and the manuscript is written coherently. The findings address the suitability of LSTMs for runoff projections under future climate scenarios systematically.

I have no major remarks regarding the implementation of the methods or the presentation of the results and their discussion.

Please find attached a short summary and my general and minor comments.

Best regards,

Sanika Baste

Citation: https://doi.org/10.5194/egusphere-2025-6058-RC1
- AC1: 'Reply on RC1', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authors.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC1
RC2:
'Comment on egusphere-2025-6058', Anonymous Referee #2, 24 Jan 2026

The study by Courvoisier et al. compares an LSTM architecture with a process-based model regarding runoff behaviour across Switzerland - for historical as well as future (extreme) climate change projections.
The paper addresses an important topic and I generally like how the authors approached the problem, presented and discussed the results. I enjoyed reading the work, particularly the discussion and interpretation.

However, from a process-based perspective and for explaining differences between the models, I am missing some crucial information. And I think a few additions to to the results could help with the interpretation. While I feel the paper has the quality to be accepted in HESS, I think a few of my comments require this to be a 'major revision'.

I hope the authors find my comments useful to improve the manuscript.

Comments to individual sections:
Section 2.1: Suggest to better explain in the text what the differently labelled catchments in figure 1 are used for. e.g. unclear what "second part of our analysis" means in the caption at this stage.

It is unclear why 96 minimally impacted catchments were chosen and then projected on 307 - you introduce a bias already (e.g. missing out on reservoir impacts to name one obvious point)?
Section 2.5: The glaciers play a crucial role in the future runoff predictions, I think. Could you add some more detail here. E.g. which/how many of the basins have no glacier extent in summer by end of century and is this consistently applied in both LSTM and PREVAH?
Section 2.6 and 3.1: You compare these two models, and try to attribute differences - but I don't know where the models actually differ. I strongly suggest to give an overview (a table with a side-by-side comparison would work nicely) of the diverging input data and model structure of PREVAH vs the LSTM - observed meteo, cc data (and meteo parameters used), cal-val-test approach, glacier data and simulation approach in PREVAH, spatial representation, observational vs projection catchments, you train the LSTM on natural catchments - what about reservoirs in the projection catchments - are those included in PREVAH...?

In this sense, I don't see a value in section 3.1 without mentioning PREVAH. I think it is more important to understand the differences than explaining the LSTM in detail alone (E.g. I'd be interested in how PREVAH simulates glacier, snow, ET which is needed to understand the differences presented (e.g. in section 4.5)).
Section 4.3/4.4/4.7 I like the maps. However, I wonder how a scatter plot LSTM vs PREVAH Runoff (could add the corresponding r²) would look like next to the maps (one dot=one catchment, x=LSTM runoff, y=PREVAH runoff, color of dot could be represented by observational, projection, selected catchments or the elevation band color or the ecoregion ... whatever you find most appropriate). This would give a better sense of the actual differences between the models and allow more intuitive diagnostics for why you see the differences.

Major comments:

l.132ff I'd like to understand the bias adjustment and cc data better: Which "Swiss gridded observations" - are those the same that you forced the model with (chapter 2.2)? Did you conduct the QM yourself or does CH2018 provide that? I suggest to discuss the implications if the bias adjustment was done on a meteorology dataset different to the forcing dataset. Also, how did you prepare the cc data for the models: Spatial aggregation->Bias adjustment. Or Bias adjustment->spatial aggregation?
l.220ff I like this cross validation methodology. I assume, per fold, you have 12 catchments that are validation, 12 test and 72 train. Is it correct that you took for valid: year 2016, 2017, 2018 of the 12 validation catchments; year 2019-2024 of the 12 testing catchments and the remaining years of the 72 training catchment for train? But what do the grey shades in Fig 2ab mean - do they correspond? what is the light grey in b that is not part of a?
l.242/369ff as far as I understand your LSTM model architecture is the same as Kraft et al. 2024 - with the only difference that you added dynamic glacier data. I'd have expected the LSTM would perform slightly better with this additional information. Do you have a (short) reasoning why it didn't (including Kraft et al's LSTM in the suggested table might help to explain this - see comment to 2.6)?
l.339 I am confused by this. Yes, fig 8 shows a stronger decline for PREVAH vs the LSTM, but I can't see that in fig 7. For 7b,c,d,f I even see the opposite.
l.347ff, 413ff and l.529ff I think the key question here is: Why is this happening and which model is closer to reality? I know that you cannot answer this question as it's the future. This likely goes beyond your scope, but I wonder if it would help (you could add your thoughts to the discussion):

1. when comparing to observed data. I assume these future events are significantly outside your training data? If that wouldn't be too extreme, you could look at the test data of the LSTM (section 3.5) and extract the maximum/minimum precip periods and check the same events for PREVAH, and evaluate which model is closer to obs (this could help for your discussion in chapter 5.7)?

2. to attribute this to PREVAH's 'knowledge' of the driving processes (such as the additional climate data it gets to calculate ET) different constraints (or no constraints) for the glacier extent, the internal mass balance constraint that the LSTM doesn't have (could the application of a mass-conserving LSTM improve the situation)?
l.584 You earlier mention that this interpretation of low-flow performance requires caution and I don't see this adequately evaluated in your paper to merit mentioning this in the short conclusion.

Minor comments:
l.93 when does the data end/what was your cutoff? Mention the spatial resolution.
l.96 PRISM and SYMAP - reference and spell out on first use
l.102 how did you spatially average in detail? Extract entire cells or did you 'split' cells? If the catchment area is small and the grids large, you can introduce errors particularly in mountainous terrain.
l.116 Can you give a rational why you used topsoil information only?
l.126 suggest to add that it is based on the CMIP5 framework
l.129 you earlier write that the resolution was 2km - how is the product downscaled from the 12-50km CORDEX resolution to 2km?
l.148 Suggest to make it clearer whether you did any glacier simulations of if this was Brunner et al. 2019b work.
l.150 I think you mean section 3.2
l.185 in section 2.5 you cite Brunner et al. 2019b as the source for the glacier data
l.201 I don't see the 24 tested alternatives in Kraft et al. 2025 section 3.6 - seems you employed the pre-print version of their approach?
l.251 fig 4: suggest to add the number of catchments per boxplot (n=x) in the caption.
l.261 I don't think DJF is a period where much melt dynamics is going on in Switzerland and JJA is not really low flow in most of the catchments (you mention this yourself in l.271-272)
l.262 add that this is about the annual panel
l.307 I think the LSTM surpasses PREVAH by 2060, or?
l.327 fig 8 caption "(mm period−1 vs 1991–2020) for 2071-2100" I think you mean something like 2071-2100 - 1991–2020?
l.476-482 the language here suggests we are still in results. Suggest to rephrase.
l.539 deeper deeper
l.576 whether
l.715, l.623 provide links to final paper?

Citation: https://doi.org/10.5194/egusphere-2025-6058-RC2
- AC2: 'Reply on RC2', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authorss.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC2
RC3:
'Review of egusphere-2025-6058', Anonymous Referee #3, 16 Feb 2026
General comments:

This paper investigates the question if LSTMs can be used for climate change impact prediction or rather: are they a valuable alternative to process-based hydrological models for climate change impact predictions in alpine environments? Or even more precise: are LSTM a valuable alternative to the selected process-based model (PREVAH)? Previous work has asked this question for other case studies and other models. From the paper as it is currently framed, it is unclear if this is simply a state-of-the-art case study or if it goes beyond the state-of-the-art. This having said, case studies can well desserve publication in HESS but given that the methods have all been used / have been developed in previous work, I think that the paper could make clearer what it contributes to the literature.

More importantly: as far as I see the reference simulations with the selected PREVAH model are problematic as reference simulations: these simulations are calibrated for some of the catchments; but for many other catchments, the model is somehow regionalized (we do not have information on this in the paper). I do not think that this is good practice: How can such a sample of two very different populations be used as a reference to evaluate a purely data-driven approach against? Furthermore, we see from Figure 3 that for many catchments, the reference model has an NSE value below 0.7 and even below 0.5: even if I cannot relate entirely to what 0.7 means for these catchments, 0.5 at daily time step is for sure a really bad performance for this climate, so these catchments should not serve as reference to compare the LSTM against. Furthermore, given the presented performance comparison, we simply know that the LSTM has a similar performance distribution. This is not informative: what if the LSTM does well for the the ones that PREVAH does not well and vice versa? If the focus of the paper is on seasonal signals only, we should at least expect a sensoring of the catchments / regions for which the seasonal signal is not well reproduced by a) by PREVAH, b) by LSTM, c) by both (which is very interesting!).

Furthermore: we should not forget that we talk about streamflow simulations here but we do not see any actual simulations, not even for seasonal simulations: I think that it is of outmost importance to reconnect the data-based methods to actual data and hydrology. Otherwise we do not learn much about how well LSTMs perform for the selected hydroclimatic region.

From a model development perspective, information is missing: how well does the PREVAH model, forced with climate data (rather than with observed precip & temperature) reproduce observed streamflow? This is a serious lack of information: the models are trained with observed data and then run with climate data; the divergence between the models for the future periode might be rooted at least partly in the divergence of their simulations for the observed period.

Furthermore, the training input data set contains a gridded spatial rainfall product, which can be assumed to show very different quality for different catchments in an alpine environment. Do the climate scenarios have the same spatial resolution? And are the gridded climate scenarios debiased to the spatial rainfall product?

Otherwise, each of the models (PREVAH and LSTM) learns how to deal with the observation-based data set (i.e. how to translate it into observed streamflow), but then makes its own errors when applying what it has learned to a rather different data set. In this setting, what can you learn from such a comparison framework?

Let's imagine: Model A is better than model B for the observed period if fed with observed data, but worse than B if fed with climate data. What do we learn from this? That model A was overfitted? Or that the climate data is not good? If now model A is close to model B for the future period, what does this tells us? That by chance, they both agree? I know that this is not the objective of the paper but given that it contributes to the question of how to use LSTM for climate change impact studies, these questions should be discussed.

Next: the hydrological process model certainly received PET as input for training. How was dealt with this for climate change simulations? And does the LSTM receive PET ? And if not, is this not unfair?

Next: The LSTM is optimized with the mean squared error (MSE) between normalized simulated

and observed discharge. This is certainly very different from the optimisation criterion used for the process model (MSE is rarely used). Why was not a similar / the same criterion used? And what can you learn from two models that are optimized based on different criteria? Besides: was the PREVAH model calibrated with some global optimisation method?

I believe that this paper needs to better frame the analysis framework, improve is A compared to B and what conclusions can be drawn in its methods section. Furthermore, we need more details on the model set up.

Detailed comments:

Abstract and elsewhere: what means "stable results" and how do you measure "physically credible"?

Intro: what do you mean by "These results highlight their robustness and scalability"? Does any of the references explicitely assess the robustness of the models and if so, how is this defined in a hydrological modelling context? What is meant by scalability, what paper discusses this and how is it defined in a hydrological modelling context? Similar what is meant by "achieve strong generalization performance" in a hydrological context? The introduction has to be much more hydrology-specific.

Intro: what do you mean by "Despite their success in present-day forecasting,"? Is this paper about present-day forecasting (forecasting = predicting current state based on previous state) or is it about continuous simulation (prediction) of hydrological states based on driver inputs (i.e. without feeding in the previous hydrological state)?

The abstract states "Divergences are most pronounced in alpine and glacier-fed catchments, where runoff dynamics are more complex, yet the main governing patterns are captured. ": I would argue that the mean governing patterns in these environments can be capture by feeding only temperature into the LSTM; should this be toned down?

Section 2: "The dataset includes most of the variables used in Kraft et al. (2024) and combines those typically employed to force the spatially distributed model PREVAH" - can this be more precise: did PREVAH use similar data or might part of the divergence come from different reference data for parameter estimation?

Section 2: do the same dynamic glacier fractions feed into the LSTM and into the PREVAH simulations or do they receive different input? This is unclear in my view.

Section 4: what is CH-RUN? This should be clearer somewhere

What is "dynamical stability of projections", should this be defined in the methods?

The paper mentions that PREVAH is widely used; the model is however only unsed in Switzerland (given the references)
Citation: https://doi.org/10.5194/egusphere-2025-6058-RC3
- AC3: 'Reply on RC3', Fabien Courvoisier, 15 Mar 2026
  
  Dear Editor and Reviewers,
  We thank the reviewers for their constructive comments. Our detailed responses to all referee comments (RC1–RC3) are provided in the attached PDFs.
  The responses were prepared on behalf of all co-authors.
  Best regards,
  
  Fabien Courvoisier (on behalf of all co-authors)
  
  Citation: https://doi.org/10.5194/egusphere-2025-6058-AC3

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

ED: Reconsider after major revisions (further review by editor and referees) (20 Mar 2026) by Ralf Loritz

AR by Fabien Courvoisier on behalf of the Authors (01 Jun 2026) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (11 Jun 2026) by Ralf Loritz

RR by Anonymous Referee #2 (12 Jul 2026)

RR by Anonymous Referee #3 (10 Aug 2026)

Suggestions for revision or reasons for rejection

Hide

ED: Publish subject to technical corrections (17 Aug 2026) by Ralf Loritz

AR by Fabien Courvoisier on behalf of the Authors (24 Aug 2026) Manuscript

Post-review adjustments

AA – Author's adjustment | EA – Editor approval

AA by Fabien Courvoisier on behalf of the Authors (14 Sep 2026) Author's adjustment Manuscript

EA: Adjustments approved (14 Sep 2026) by Ralf Loritz

Journal article(s) based on this preprint

16 Sep 2026

Assessing the stability of LSTM runoff projections in Switzerland under climate scenarios

Fabien Courvoisier, Basil Kraft, Yann Yasser Haddad, Massimiliano Zappa, and Lukas Gudmundsson

Hydrol. Earth Syst. Sci., 30, 5873–5900, https://doi.org/10.5194/hess-30-5873-2026,https://doi.org/10.5194/hess-30-5873-2026, 2026

Short summary

Fabien Courvoisier, Basil Kraft, Yann Yasser Haddad, Massimiliano Zappa, and Lukas Gudmundsson

Viewed

Total article views: 3,685 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
2,131	1,344	210	3,685	216	242

HTML: 2,131
PDF: 1,344
XML: 210
Total: 3,685
BibTeX: 216
EndNote: 242

Views and downloads (calculated since 12 Dec 2025)

Month	HTML	PDF	XML	Total
Dec 2025	740	385	45	1,170
Jan 2026	685	390	50	1,125
Feb 2026	232	161	52	445
Mar 2026	316	240	55	611
Apr 2026	53	55	0	108
May 2026	27	62	1	90
Jun 2026	7	9	0	16
Jul 2026	15	13	1	29
Aug 2026	19	12	2	33
Sep 2026	37	17	4	58

Cumulative views and downloads (calculated since 12 Dec 2025)

Month	HTML	PDF	XML	Total
Dec 2025	740	385	45	1,170
Jan 2026	685	390	50	1,125
Feb 2026	232	161	52	445
Mar 2026	316	240	55	611
Apr 2026	53	55	0	108
May 2026	27	62	1	90
Jun 2026	7	9	0	16
Jul 2026	15	13	1	29
Aug 2026	19	12	2	33
Sep 2026	37	17	4	58

Viewed (geographical distribution)

Total article views: 3,642 (including HTML, PDF, and XML) Thereof 3,642 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 22 Sep 2026

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (27026 KB)
Metadata XML

Short summary

Climate change strongly modifies runoff in alpine regions like Switzerland, altering water supply and affecting both human sectors and natural ecosystems. We test whether a model that learns only from past observations can credibly project future runoff under climate scenarios. It captures major expected shifts and reveals known limits for extremes. Its speed and ability to generalize across catchments make it a promising tool for assessing climate impacts on water resources.


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