acceptodds
Withdrawn from ICLR 2027

Distributed Hydrological Modeling in the Feature Space

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate forecasting of river discharge and floods is very challenging. River dynamics are affected by storage, meteorological forcing, and flow propagation at different spatial and temporal scales. Forecasting thus requires a framework that considers the upstream-to-downstream flow through river networks across grid cells and catchments. This modeling is known in hydrology as distributed modeling and routing. Existing deep learning approaches either ignore this topology, operate on lumped catchments, or route predicted physical quantities through a separate graph or physical routing model. We instead introduce feature-space routing: a topology-aware state-space operator embedded directly in the forecasting dynamics. At every forecast step, the operator gathers latent states from upstream grid cells and causally updates the downstream state according to the known river network. This preserves the physical connectivity of the river system while allowing the propagated state itself to be learned end-to-end and allows the model to predict river discharge considering both local dynamics and neighboring upstream contributions. To address uncertainty and provide probabilistic forecasts, we minimize the fair continuous ranked probability score (fCRPS) as a training objective. Our experiments on the European Flood Awareness System (EFAS) and observational data for river discharge forecasting demonstrate that encoding the physical structure of river networks explicitly in the feature space substantially improves the forecasting skill, particularly in an ungauged setting. This enabled us to achieve state-of-the-art results on both reanalysis and observational data and to be able to forecast maps of river discharge at arcminute and -hourly resolution up to days lead time.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mamba, Earth science, routing, distributed modeling, hydrology, flood forecasting, river discharge

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B2LNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a state space operator for river discharge and flood forecasting that combines the following: (1) a state-space model inspired by Mamba-2; (2) a feature-space routing mechanism to account for the river structure; (3) a loss function based on fair continuous ranked probability score (CRPS); and (4) a warmup scheme to reduce distribution shifts between the encoder and forecasting engine. The experiments compare the proposed approach against multiple alternatives (DRUM, ED-LSTM, ME-LSTM, RiverMamba, and IFAS) using a benchmark combining EFAS reanalysis data and GRDC and CAMELS station data. The results show that the proposed approach outperforms the baselines in terms of accuracy metrics for both reanalysis data and gauged observations.

Critical strengths

S1) The paper addresses a relevant problem: predicting river discharge (or streamflow)

S2) The idea of combining LLD information as network topology for state-space modeling in river networks seems novel.

S3) The CRPS loss and the warmup procedure are interesting and useful extensions over RiverMamba.

Critical weaknesses

W1) Methodological novelty needs clarification: many ideas in the paper have been explored by previous work and are not properly contextualized such as other graph-based extensions of SSD, LLD-based graph construction for streamflow forecasting models, and CRPS loss for hydrological modeling. The main novelty, in my opinion, is the particular use of LLD with SSD. However, this contribution is only compared against GraphMamba and some simple alternatives and even then the gains over summation and mean seem minor.

W2) Experimental setting: the comparison with RiverMamba and the other baselines is a bit confusing because they use different metrics. I would be much simpler to apply the RiverMamba sequential encoding of the river network in Table 4.c. Other graph encodings for state-space models should also be considered, even if they haven't been applied specifically to streamflow prediction.

W3) It is unclear if DRUM, ED-LSTM, and ME-LSTM leverage the same input data and the warmup procedure as the proposed solution. The warmup procedure, in particular, seems to be quite effective but it is not a core contribution of the paper.

W4) The paper uses EFAS as training and, ground-truth, and baseline. It should be clear whether all these streamflow values are generated the same way and if they also leverage the same information as the proposed solution (i.e., is this a limitation in the model or its inputs).

W5) The figures in the paper are not friendly to human readers due to small fonts.

Questions for authors

Q1) What in the proposed method is novel more broadly, novel only in the context of streamflow forecasting, and novel only in combination with state-space modeling?

Q2) How does the routing procedure proposed compare agains the one used by RiverMamba? Can this be inferred through the ablation study or is it a missing experiment?

Q3) Are the gains achieved by the routing procedure and shown in Table 4.c significant? Do they support the main hypothesis in the paper?

Q4) Is the comparison against DRUM, ED-LSTM, and ME-LSTM fair in terms of inputs and the warmup procedure?

Q5) Can the EFAS baseline leverage all the inputs used by the proposed method? If not, how would the approaches compare against each other if constrained to the same inputs?

2 comments

Thank you for your time and for highlighting the relevance of the addressed problem, the novelty of combining Bi-SSD with LDD for river routing, the state-of-the-art performance against multiple baselines, and the usefulness of probabilistic forecasting with CRPS for distributed modeling in hydrology.

(W1) Methodological novelty needs clarification.

Thank you for the note. We would like to clarify more about the novelty aspects of our work:

  • The main contribution is doing river routing based on SSD and LDD along the river network in the feature space. This was not explored before. Deep learning approaches in general are lumped models, use routing schemes that are differentiable but still based on physics, use GNN, or vision modeling without an explicit routing operator. Our method keeps a learned latent state on every 1-arcminute cell and rolls it forward over lead time. It combines a routing operator with Bi-SSD. The Bi-SSD exchanges information along the domain and the routing propagates and constrains the signal to the LDD topology. Therefore, the main contribution is compared to other approaches in hydrological modeling. We compared our routing approach to time-series models (ED-LSTM, ME-LSTM, and DRUM), physics-based routing (EFAS), differentiable physics-based routing (Muskingum and Hayami in response to #2 Reviewer zLt6), GNN routing operators (GraphSAGE, GAT, and GGNN in response to #2 Reviewer zLt6), Graph-SSM sorting (GraphMamba), and a vision model without explicit routing (RiverMamba).
  • Built upon the proposed methodology for routing, we introduced the first operational autoregressive deep learning model that forecasts probabilistic river discharge up to 10 days and 1-arcminute resolution over Europe.

There seems to be a misunderstanding about Table 4c and we are happy to clarify. Summation and mean in Table 4c are different variants to aggregate upstream features and run by our routing operater in the end. They represent an ablation for how we aggregate upstreams before running the routing with SSD. Sum and mean aggregation pool the upstream features into one value before the SSD. We kept all upstream (last row Table 4c) so the model learns the aggregation in an adaptive way while achieving better performance. Note that summation (first row Table 4c) doesn't perform well on obs CAMELS and given all upstreams to SSD without a mean aggregation achieves slightly better results.

We will clarify this in the paper.

(Q1) What in the proposed method is novel more broadly, novel only in the context of streamflow forecasting, and novel only in combination with state-space modeling?

Please see our response to W1. Our main contribution is a novel way to do feature-space routing and forecasting in distributed hydrological modeling. The routing operator is not restricted to state-space modeling and works also in a combination with standard local-runoff models as we showed in our response to #2 Reviewer zLt6.

While in this work we addressed flood forecasting, we believe that the method is not limited to streamflow forecasting or specific input variables. For instance, many processes in land surface/sub-surface modeling require lateral flow (i.e., horizontal movement of water) and our model is promising for more general tasks.

(W2) Experimental setting: the comparison with RiverMamba and the other baselines is a bit confusing because they use different metrics.

Thank you for this comment. RiverMamba was compared only over the Mid-Europe domain. It is a deterministic model with no nowcasting and is expensive to train over the full domain, so we give it its own comparison. We compare to it in Appendix Sec. E in more detail using the same metrics. We agree that this could be confusing and we will make it clearer in the revision.

(W2) It would be much simpler to apply the RiverMamba sequential encoding of the river network in Table 4.c.

(Q2) How does the routing procedure proposed compare against the one used by RiverMamba? Can this be inferred through the ablation study or is it a missing experiment?

Thank you for the note. We are happy to clarify. RiverMamba is a vision model that scans the domain with Bi-SSM to do the forecast but without explicit routing. The sequential encoding of the river network is done in RiverMamba with space-filling curves. The equivalent experiments to RiverMamba are in Table 3 (first row) and Table 4a, representing a model without routing (rows 1 and 3). The full comparison to RiverMamba is in Appendix Sec. E. We will make this clearer in the revision. Thanks.

(W2) Other graph encodings for state-space models should also be considered, even if they haven't been applied specifically to streamflow prediction.

Thank you for the suggestion. We already compared with a Graph-SSM method: Wang et al., 2024 (https://arxiv.org/abs/2402.00789) which orders all nodes by feature magnitudes. Another common Graph-SSM is GMN by Behrouz and Hashemi, 2024 (https://arxiv.org/abs/2402.08678) which uses subgraph tokenization, but there is no official code to test it (see https://github.com/GraphMamba/GMN). We will discuss and mention more Graph-SSM methods in the revision.

In our work, we want to propagate information along the river network as a fixed topology. Our aim is not to introduce a general Graph-SSM method but to use Graph-SSM for river routing in the feature space.

(Q3) Are the gains achieved by the routing procedure and shown in Table 4.c significant? Do they support the main hypothesis in the paper?

Thanks for this comment. Our main research hypothesis is that "encoding the physical structure of river networks explicitly in the feature space improves the forecasting skill.". This is supported in Tables 2, 3, and 4a (with routing vs without routing), qualitative results in Figs. 1 and 7, and through additional experiments in Appendix Sec. B-C. The gain of adding routing is significant.

(W3) It is unclear if DRUM, ED-LSTM, and ME-LSTM leverage the same input data and the warmup procedure as the proposed solution. The warmup procedure, in particular, seems to be quite effective but it is not a core contribution of the paper.

(Q4) Is the comparison against DRUM, ED-LSTM, and ME-LSTM fair in terms of inputs and the warmup procedure?

We thank the reviewer for pointing out to this ambiguity. All baselines (DRUM, ED-LSTM, and ME-LSTM) use the same input data as our model but as time series (the initial condition is given as 6-hourly input for LSTMs so they have even more memory), and they are evaluated under the same protocol of training/testing. The comparison is fair. We will state this explicitly in the paper.

The warmup is not added to the baselines because they are equivalent by design. Our model is autoregressive and a warmup is needed to initialize its first rollout state. ED-LSTM and ME-LSTM are encoder-decoder models following Nearing et al. (2024) and Cohen et al. (2026). Their forecast state is initialized by a hindcast LSTM that runs over the previous days before the forecast time and is then given to the forecast LSTM (e.g., in ED-LSTM a learned handoff network gives the final state to the forecast LSTM). DRUM conditions on the same hindcast days. This hindcast is built from the same data as our warmup. We agree that the warmup is not a core contribution and we did not claim that. We described it because it is useful for the autoregressive design in our model.

We will state this explicitly in the revised paper.

(W4) The paper uses EFAS as training and, ground-truth, and baseline. It should be clear whether all these streamflow values are generated the same way and whether they also leverage the same information as the proposed solution (i.e., is this a limitation in the model or its inputs).

(Q5) Can the EFAS baseline leverage all the inputs used by the proposed method? If not, how would the approaches compare against each other if constrained to the same inputs?

Thank you for the note. EFAS has two parts: EFAS-forecast as a baseline and EFAS-reanalysis as a ground truth. The way how we used it in the paper is as follows:

  • EFAS-reanalysis is used as a benchmark to compare different baselines because it simulates river discharge everywhere and it is a skillful hydrological system. EFAS forecast is excluded here because the ground truth reanalysis is generated by the same model.
  • EFAS forecast is used as a physics-based baseline and benchmarked against observational datasets i.e., GRDC and CAMELS data.

We did not change its input and we did not run the model ourselves. We will make this clearer in the revision to avoid confusion.

(W5) The figures in the paper are not friendly to human readers due to small fonts.

We will increase the font in the revision. Please note that we reduced the size of figures due to page limit in the main paper.

Thank you again for your time in reviewing the paper. We hope that the response has resolved all your concerns. Please let us know if there are still any open issues.

Reviewer_zLt6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to design a data-driven river routing model to support flood forecasting. The motivation is that deep-learning hydrologic models usually operate at the lumped scale. The few distributed studies apply a separate physical or neural-network routing step on top of a local rainfall-runoff model.

This study builds on the RiverMamba architecture and couples local and routing processes in one State Space Model (SSM). It keeps RiverMamba's bidirectional SSM encoder over space-filling curves and adds a routing module inside each 6-hourly forecasting step. For each grid cell, the module scans the cell's upstream neighbours from the LISFLOOD drainage-direction map, ordered by upstream area, and adds the result to the cell's latent features. The model runs at 1 arcmin (about 5.2 million land cells) for lead times up to 10 days, and is trained with the fair CRPS to produce ensembles.

The model is evaluated over the European EFAS domain against the EFAS reanalysis and gauge observations (GRDC and the European CAMELS datasets). It is compared with recent data-driven forecasters (DRUM, ED-LSTM, ME-LSTM, RiverMamba) and with the EFAS operational forecasts. Results show better hydrograph matching (R2 and KGE) and probabilistic skill (CRPS) than the lumped baselines. Ablations show improved performance over the same model without routing, and over graph-neural-network aggregators (GraphSAGE, GAT) in place of the SSM.

The main contribution is a topology-aware feature-aggregation operator integrated into RiverMamba's forecasting architecture.

Critical strengths
  1. Useful model design.

    • The paper formulates river routing as a topological sequence-modeling problem that an SSM can handle efficiently, with linear complexity in the number of nodes. This partly preserves the physical structure of routing by explicitly processing upstream and downstream states.
    • The results provide evidence that explicit routing helps this architecture: the same model without routing is clearly worse. They also support the choice of aggregator, since the SSM outperforms GraphSAGE and GAT.
    • The benefit is largest in a spatial-holdout ("ungauged") test. There, adding routing raises the average R2 from about −0.5 to +0.2 and makes KGE positive in all upstream-area classes (Table 14).
  2. Operational scale and evaluation.

    • The model forecasts at 1 arcmin, every 6 hours, up to 10 days, over about 5.2 million cells. It uses forcing available at operational latency (IFS-HRES forecasts).
    • It is compared against the EFAS operational forecasts on observations. On GRDC stations in the test period, its ensemble mean reaches a KGE of 0.55, against 0.38 for ME-LSTM and 0.40 for the EFAS control forecast.
Critical weaknesses
  1. Insufficient benchmark.

    • The study aims to show the superiority of a new SSM-based routing method within distributed hydrologic modeling. However, most comparisons are against lumped data-driven models with no routing at all (DRUM, ED-LSTM, ME-LSTM). The only routing comparisons are internal swaps of the aggregator for GraphSAGE and GAT.
    • Many physical or hybrid routing methods are structurally simple and have established performance at large scales. The authors review several: graph-network routing (Mosaffa et al., 2026; Scholz et al., 2025), differentiable Muskingum–Cunge routing (Bindas et al., 2024; Song et al., 2025) and impulse-response routing (Hascoet et al., 2026). It remains open whether feature-space SSM routing improves on simpler established approaches.
    • Published discharge products are not used as references. Google's model output (Nearing et al., 2024) is publicly released, while the ME-LSTM baseline only re-implements its architecture. Upon simple searches, several other more recent routed global products are available, such as Yang et al. (2025), which the authors cite. Since the evaluation window is the authors' choice, a comparison at overlapping stations and periods would strengthen the evaluation, or the authors should explain why it is not possible.
    • The comparison with RiverMamba, the direct predecessor, is mixed. RiverMamba was retrained only over Mid-Europe. There, the deterministic RiverMamba has higher R2 and KGE than the deterministic version of the proposed model at GRDC-split stations (test KGE 0.69 vs 0.66, Table 19). The proposed model's advantage is mainly in flood-detection F1.
    • Less central, but also missing:
      • Flood detection is scored mainly at frequent thresholds. The main comparisons emphasize F1 averaged over the 1.5- to 20-year return periods, and a 1.5-year exceedance is a frequent high flow, not an extreme flood. Per-threshold results for the main baselines are not reported, and peak magnitude and timing are not evaluated.
      • Using the fair CRPS in this distributed setting may be new, but the experiments do not isolate the effect of the objective. Comparing deterministic and probabilistic versions changes both the predictive formulation and the loss. A comparison with another probabilistic objective on the same backbone would show whether the fair CRPS itself helps.
  2. Limited insight into the model architecture.

    • More hydrological justification is needed for why this SSM-based design is preferred, beyond the performance metrics. Is there mathematical intuition from the structure of river routing that makes an SSM suitable? For example, does the approximate linearity of routing favour an SSM over nonlinear gated GNNs?
    • The routing module exchanges information only with immediate upstream neighbours at each 6-hour step, while a flood wave can travel many 1.5-km cells in that interval. The bidirectional SSM also mixes information over long distances, and Table 8 shows it remains important. It is unclear how much realistic propagation the routing operator itself learns, as opposed to reproducing patterns in the physically routed EFAS reanalysis used for pretraining.

References

  • Bindas et al. (2024). doi:10.1029/2023WR035337
  • Hascoet et al. (2026). doi:10.1029/2025JH000760
  • Jia et al. (2021). doi:10.1137/1.9781611976700.69
  • Mosaffa et al. (2026). doi:10.5194/hess-30-2079-2026
  • Moshe et al. (2020). arXiv:2007.00595
  • Nearing et al. (2024). doi:10.1038/s41586-024-07145-1
  • Scholz et al. (2025). doi:10.5194/hess-29-6257-2025
  • Song et al. (2025). doi:10.1029/2024WR038928
  • Yang et al. (2025). doi:10.1029/2024WR039764
Questions for authors
  1. Is it appropriate to call the method a new routing scheme? It propagates feature states rather than physical quantities, so it is not a plug-and-play module for standard local rainfall-runoff models outside the RiverMamba architecture. In other words, is this study an incremental extension of RiverMamba or a new general routing scheme?
  2. A simple search suggests using CRPS as a training objective is not new in geophysics or hydrology. What does the fair CRPS add in this distributed setting, beyond the effect of making the model probabilistic?
  3. The SSM topological operator is said to have low computational cost. Table 11 reports training and inference times for the backbone against RiverMamba, but not for the routing module itself. Can the authors report the cost of the routing module per rollout step, compared with the GNN alternatives?
4 comments

Thank you for your time and the detailed description of the paper. Your thoughtful feedback was really helpful. We appreciate that you have found the model design useful and supported by ablation studies, the clear benefit in an ungauged setting, and the operational scale and evaluation.

Benchmark.

Thank you for the suggestion. We would like first to clarify few things. In our work, we focus on fully distributed modeling where we do routing in the feature space between grid cells at high spatio-temporal resolution and at a continental scale. Timeseries baselines like LSTM with no spatial information were used as a no-routing models in our work. Routing with deep learning has been explored in semi-distributed settings, between sub-basins (e.g., Moshe et al., 2020, Jia et al., 2021, Scholz et al., 2025, Mosaffa et al., 2026).

Mosaffa et al., 2026, for instance, route with graph attention, which corresponds to our GAT variant (Table 4b). Scholz et al., 2025 use a gated graph neural network. We added a GGNN routing variant with the rest of the model unchanged. These compare the operators within the same fully distributed model. Although both GNNs benefit from our Bi-SSD.

ModuleKGE | KGE | F1 ↑
GGNN0.720 | 0.729 | 0.298
SSM0.740 | 0.747 | 0.339

Physics-based and differentiable routing methods route discharge, not latent features, and need a runoff model with routing parameters for every sub-catchment. Plugging them into our fully distributed model would change the model design, not just its routing operator (they are not plug-and-play modules in our setting). In the fully distributed, an established physical routing baseline is EFAS itself, which uses kinematic wave on the same 1-arcminute graph at continental scale and we have included this baseline in our paper.

However, we agree with you that a comparison with other hybrid and physics-based routing approaches would strengthen the evaluation. We therefore compare them in a standard semi-distributed setting over Europe (graph on sub-catchments). Our method can be used here as a general routing method. The experiment isolates the routing operator: the local model is the same, only the routing changes, and we compare against established physics-based routing. For the experiment, we used our processed dataset which has (3463 gauges with 1538 of them are upstream gauges). The dataset combines GRDC and CAMELS stations as a connected graph on the EFAS river network. The local lumped model for rainfall-runoff is the deterministic ED-LSTM and it is shared, we change only the routing between gauges. We train on 1993-2018, validate on 2019, and test on 2020-2024. We used the same input data as in our main fully distributed model with 365 days of hindcast and forecast up to 10 days of observational river discharge including nowcasting without pretraining on reanalysis.

All gauges (3463)Gauges with upstreams (1538)
RoutingKGE ↑R2 ↑F1@(1.5–20) ↑F1@(1.5–500) ↑KGE ↑R2 ↑F1@(1.5–20) ↑F1@(1.5–500) ↑
-0.671±0.0150.647±0.0020.116±0.0140.065±0.0080.688±0.0130.655±0.0040.121±0.0160.068±0.009
Muskingum (DiffRoute)0.684±0.0090.672±0.0020.116±0.0060.065±0.0040.722±0.0060.718±0.0020.138±0.0060.077±0.004
Hayami (DiffRoute)0.684±0.0040.676±0.0020.114±0.0060.063±0.0030.722±0.0020.720±0.0010.136±0.0060.076±0.003
SSM (Ours)0.688±0.0160.677±0.0020.122±0.0160.068±0.0090.724±0.0210.716±0.0030.139±0.0230.078±0.013

SSM routing routes the hidden state at each step using Eq. (6) as in the paper. Muskingum and Hayami were implemented using DiffRoute by Hascoet et al., 2026, (https://doi.org/10.5281/zenodo.17888293). As can be seen, adding the SSM routing clearly outperforms the baseline without routing, especially on sub-catchments with upstream. The operator learns realistic propagation without pretraining on reanalysis. It also on par with or slightly better than the Diffroute baselines. The benefit of feature-space routing is that we can use it in a fully-distributed modeling framework with deep learning without estimating any physical quantities.

We will add this comparison to the revised paper and discuss the related work more.

Published discharge products are not used as references.

Thank you for the suggestion.

GRADES-hydroDL by Yang et al., 2025 is a simulation, not a forecast. The model there is forced by observed data. So the data can be used as a reference e.g., similar to EFAS-reanalysis but not as a forecasting system baseline.

Regarding the published forecast by Nearing et al., 2024. In our work, we focus on fully distributed modeling which has different aims (fully distributed vs lumped modeling), resolution (6-hourly vs daily, global vs Europe) and evaluation protocol than Nearing et al., 2024. We compared to their model under our experimental setting since it is easier to apply timeseries models than distributed models. Our task is to forecast high-resolution river discharge at a grid. Google’s model is lumped: it scores high at gauges but cannot forecast at every grid cell, so it is not the same framework.

Finally, please also not that, testing our model under their protocol is also difficult to be addressed in the short period of rebuttal. We will explain this in the revision.

The comparison with RiverMamba.

Thanks for the comment. RiverMamba was retrained on the same Mid-Europe domain because of its computational costs. It was retrained with same inputs and lead times as our model.

We do not agree that the advantage is mainly in flood detection and we are happy to explain why we do not think this is the case.

  • The comparison in Table 19 refers to EFAS reanalysis split at 480 GRDC stations, where the target is the EFAS-Reanalysis. Against observed discharge at the CAMELS stations, the deterministic version of our model has KGE=0.676 and R2=0.622 which are better than RiverMamba with KGE=0.626 and R2=0.543. Over the full domain, our model doubles flood-detection across val/test splits.
  • Our new model is faster (see Table 16).
  • Our model has less parameters (2.3 M against 4.4 M).
  • RiverMamba needs one forecasting block per lead step. It can only forecast with a fixed horizon window. Ours is autoregressive.
  • RiverMamba does not provide nowcasting at lead time t=0t=0. Ours does.
  • RiverMamba is a deterministic, ours can generate consistent ensemble forecast.

However, we are happy to address your concern. For this, we run an additional evaluation over the full domain on reanalysis for the same deterministic models. This include 280820 grid cells (much larger sample evaluation compared to Table 19). The results is below:

Results on EFAS Reanalysis (all grid cells).

ModelVal R2 ↑Val KGE ↑Val F1@(1.5-500) ↑Test R2 ↑Test KGE ↑Test F1@(1.5-500) ↑
DetermRiverMamba0.3910.5090.0990.3320.4880.087
DetermOurs0.5250.5980.1400.4520.5740.124

Ours with routing has better forecast which is also consistent with our experiments in Appendix Sec. B-D. It has better performance than RiverMamba in R2 on 71.2 % of the points in the domain and in KGE it is better on 66.4 % of the points.

We will add this additional evaluation to the paper. Thanks for your suggestion.

Less central, but also missing: Flood detection is scored mainly at frequent thresholds.

Thank you for the suggestion. Our paper already includes extensive analysis and a comprehensive appendix. To maintain conciseness, we initially omitted some additional figures and details, but we would be happy to include them in the revised manuscript.

Please also note, some experts argue that international disaster response organizations usually base their action plans on more frequent events as mentioned by Nearing et al., 2026 (https://doi.org/10.1016/j.hydroa.2026.100222) and especially with climate change, frequent floods become more relevant.

More hydrological justification is needed for why this SSM-based design is preferred, beyond the performance metrics. Is there mathematical intuition from the structure of river routing that makes an SSM suitable? For example, does the approximate linearity of routing favour an SSM over nonlinear gated GNNs?

Thanks for the note. Our motivation was that linear hydrological routing schemes are actually state space models e.g., the Muskingum scheme, as well as linear reservoirs, which are linear recurrences. However, Routing is only approximately linear in reality. Muskingum-Cunge addresses this with flow-dependent coefficients. The selective SSM in Mamba has the same structure: the recurrence is linear in the state, while Δ\Delta, B and C depend on the input. It is therefore a linear time-varying system, close to the routing schemes used in hydrology. Also, because the state update is linear, it is associative. So the full domain can be computed with parallel scan. Non-linearity is added in the Mamba block (see Fig. 4). A gated GNN, in contrast, applies a nonlinear transform at every connection. Mamba inductive bias matches the structure of routing intuitively, and our comparison to GNN (Table 4b) and to GGNN (see rebuttal) supports that empirically.

We will make the motivation clear in the revision. Thanks.

The routing module exchanges information only with immediate upstream neighbours at each 6-hour step, while a flood wave can travel many 1.5-km cells in that interval. The bidirectional SSM also mixes information over long distances, and Table 8 shows it remains important. It is unclear how much realistic propagation the routing operator itself learns, as opposed to reproducing patterns in the physically routed EFAS reanalysis used for pretraining.

Thanks you for the note and we are happy to address your point. Our experiment on semi-distributed modeling verifies that the SSM routing learns realistic propagation between sub-catchments without pretraining on EFAS reanalysis.

However, we would also like to elaborate more on Bi-SSD and routing. The routing module can also work alone and it is essential to learn routing on the river network in a distributed setting (Compare Table 4a third row against Table 8 first row).

We will discuss this in the revised paper. Thanks for the note.

Is it appropriate to call the method a new routing scheme? It propagates feature states rather than physical quantities, so it is not a plug-and-play module for standard local rainfall-runoff models outside the RiverMamba architecture. In other words, is this study an incremental extension of RiverMamba or a new general routing scheme?

Thank you for the question. Yes, the proposed routing method can be seen as a general method and can be used as a plug-in in standard local rainfall-runoff models. It is not restricted to RiverMamba architecture. In our work, we focus on fully distributed modeling and RiverMamba is the state-of-the-art so we used it as a baseline and tested our method on it. This is for sure not an incremental extension of RiverMamba. Please see our experiment on semi-distributed (in the rebuttal above). It further verifies that the proposed routing scheme is general although it was designed mainly for fully distributed modeling with deep learning.

A simple search suggests using CRPS as a training objective is not new in geophysics or hydrology. What does the fair CRPS add in this distributed setting, beyond the effect of making the model probabilistic?

CRPS objective in a distributed setting has two main benefits beyond providing just a probabilistic forecast:

  • It has one step inference in comparison to standard denoising diffusion/flow-matching objectives. This is very important for fully-distributed modeling at a large scale.
  • It can model the spatio-temporal joint structure. See our response to Reviewer #skY8 (the Variogram Scores (VS)). CRPS ens members preserve the spatio-temporal consistency in our probabilistic setting because with CRPS, we share the noise across space and lead time.

We didn't claim to be the first one to use CRPS. The CRPS loss is currently used in state-of-the-art weather forecasting (Alet et al., 2025, Lang et al., 2026). Only recently, few studies used it in a lumped setting for deep learning in hydrology e.g., Karan Ruparell et al., 2026. Our approach is optimized with fCRPS in a fully spatially distributed modeling framework at a large-scale.

Less central, but also missing: Using the fair CRPS in this distributed setting may be new, but the experiments do not isolate the effect of the objective. Comparing deterministic and probabilistic versions changes both the predictive formulation and the loss. A comparison with another probabilistic objective on the same backbone would show whether the fair CRPS itself helps.

As suggested. We compared the proposed fCRPS loss to another probabilistic loss in a distributed setting (CMAL as in Klotz et al., 2022, https://doi.org/10.5194/hess-26-1673-2022). The results are shown for the test years 2024-2025.

LossGRDC MAE ↓GRDC Spread-skill (calibrated=1)CAMELS MAE ↓CAMELS Spread-skill (calibrated=1)
CMAL6.521.662.021.70
fCRPS4.190.781.350.81

We will clarify this in the revised paper. Thanks for the suggestion.

The SSM topological operator is said to have low computational cost. Table 11 reports training and inference times for the backbone against RiverMamba, but not for the routing module itself. Can the authors report the cost of the routing module per rollout step, compared with the GNN alternatives?

Thank you for the note. On the H200 in bf16, the SSM routing takes about 5.7 ms of a 22.4 ms forecasting step (280820 grid cells) and about 9.8 ms of a 43.3 ms forecasting step in fp32. It is slower than the GNN variants but more accurate (see Table 4b in the paper). The table below shows the routing share of the compute in one forecasting step of the full model over the Mid-Europe domain (280820 grid cells). We report the forward pass (fwd) and the forward and backward passes (fwd + bwd) on an H200 GPU.

Routing moduleShare of forward (%)Share of forward + backward (%)
GraphSAGE (mean)9.8 ± 0.28.0 ± 0.1
GraphSAGE (sum)9.0 ± 0.27.5 ± 0.1
GAT12.5 ± 0.212.5 ± 0.1
GGNN15.0 ± 0.214.5 ± 0.1
SSM25.5 ± 0.228.9 ± 0.1

Thank you again for your time and the thoughtful reviewing. We hope that the response has resolved all your concerns. Please let us know if there are still any open issues.

Reviewer_skY8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces feature-space routing for distributed river-discharge forecasting. At each autoregressive step, a Mamba2/SSD operator aggregates upstream latent features according to a prescribed drainage graph, with neighbors ordered by upstream area. The model combines meteorological observations and forecasts with static spatial embeddings, and uses noise injection and fair CRPS training to generate ensembles. It is pretrained on EFAS reanalysis and fine-tuned separately on GRDC and CAMELS observations. Experiments report improved discharge and flood-threshold prediction, including spatial-holdout evaluations, at one-arcminute spatial resolution and six-hour intervals up to ten days ahead.

Critical strengths

The routing module provides a useful, computationally practical way to integrate known upstream–downstream connectivity into a dense forecasting model. The controlled routing/warmup ablation supports its contribution: with warmup fixed, adding routing increases reanalysis KGE from 0.687 to 0.740 and F1 from 0.162 to 0.339 (Table 4). Comparisons with GNN aggregation and the topology-only binary-mask control further strengthen this evidence.

The evaluation extends beyond reproducing simulated discharge to real gauge observations, spatial holdouts, and operational EFAS comparisons. The reported gains are substantial: Table 5 gives test KGE of 0.5520 versus 0.3964 for the EFAS control run and 0.3843 for ME-LSTM. The extensive ablations and station-level analyses make this a potentially valuable applied contribution.

Critical weaknesses

The operational evaluation does not yet establish that all inputs were available at forecast issuance. Appendix O.9.1 averages AlphaEarth embeddings over 2017–2024, incorporating observations from the validation period and the first test year. These embeddings summarize annual surface conditions, so treating them as static does not remove their temporal information. Appendix K also explicitly ignores IMERG’s four-hour latency and suggests a truncated daily average for operational use, without establishing that the reported evaluation used that construction. A forecast-time-consistent rerun is needed to quantify the effect; the present results do not establish the claimed operational performance.

Numerical reporting inconsistencies weaken confidence in the empirical conclusions. Tables 27 and 29 repeat nearly all validation RMSE, R², and KGE entries exactly for GRDC and CAMELS, despite describing separate datasets and fine-tuning procedures. Table 5 also reports different validation R²/KGE values from Table 27 for the same stated years without explaining the evaluation-mask difference. Additionally, the evaluation fCRPS formula in Eq. (16) incorrectly repeats forecast–observation differences in the ensemble-spread term, unlike Eq. (9). These may be reporting errors, but corrected tables, explicit evaluation masks, and confirmation of the implemented metric are necessary.

Questions for authors

Operational clarification: What exact observation windows and availability times were used for IMERG and IFS inputs at each forecast issuance? How do the results change when static embeddings exclude validation/test years and all dynamic inputs respect their actual publication latency?

Results verification: Please reconcile Tables 5, 27, and 29, including station counts, date ranges, lead-time aggregation, and missing-data masks. Is Eq. (16) solely a typographical error, and was evaluation performed after inverse-transforming discharge to physical units?

Probabilistic validation: Can you provide reliability or coverage diagnostics for flood-exceedance probabilities and a test of spatial/temporal dependence, such as joint exceedances or a variogram score? Marginal CRPS and shared noise do not by themselves establish the claimed joint-distribution accuracy. The EFAS ensemble comparison also covers only June–August 2024; does the advantage persist across flood seasons?

Routing implementation: How are upstream edges crossing the 200,000-point partitions handled during training and inference? If the routing SSD includes a causal convolution, is its state also reset at segment boundaries? These details determine whether the packed implementation preserves the intended drainage connectivity.

3 comments

Thank you for your careful review and the helpful feedback. We highly appreciate your time and we are glad that you have found the new routing approach useful, the extensive ablations support its valuable contribution and the state-of-the-art performance compared to several baselines.

The operational evaluation.

Thank you for this note. As written in Appendix Sec. K and shown in Fig. 80, we considered the latency for each input data. CPC has 2 days of latency, operational IFS (forecast and analysis) is available in near real-time, and EMO data has 2 days of latency. Only that we ignored the 4 hours latency for IMERG to compute the daily mean.

However, we are happy to address your concern. We run an additional sensitivity study where we account for 4 hours latency of IMERG and average AlphaEarth for 2017-2021 as suggested. Inputs are changed at inference and the sensitivity is computed over the Mid-Europe domain for 4 days forecast. Table below shows the evaluation on EFAS reanalysis from 2 January to 1 July 2025. KGE and R2 are medians over stations and averaged over lead times. As can be seen, the effect is insignificant. Note that AlphaEarth embedding varies in space much more than in time.

ModelKGE (GRDC | CAMELS splits)R2 (GRDC | CAMELS splits)
baseline0.5856 | 0.57920.4543 | 0.4395
ΔKGE (GRDC | CAMELS splits)ΔR2 (GRDC | CAMELS splits)
IMERG with 4 hours latency−0.0005 | −0.0008−0.0016 | +0.0019
AlphaEarth 2017-2021−0.0059 | −0.0058−0.0072 | −0.0070
both−0.0064 | −0.0064−0.0084 | −0.0081

We also repeated the same test for the fine-tuning on GRDC observations. The changes across GRDC stations in this period for the 3 settings range from 0.000 to +0.012 in KGE (the baseline is KGE=0.6432) and from −0.0003 to +0.0130 in R2 (the baseline is R2=0.6246).

We will make a clarification about the sensitivity of operational run in the paper.

Numerical reporting inconsistencies.

Thank you for the careful reading. Yes, the evaluation was performed after inverse-transforming discharge to physical units m3/s−1m^3/s^{-1}. We fixed the typo in Table 27. The validation set in Table 27 was copied from CAMELS obs by mistake.

Yes, Eq. (16) is solely a typographical error. We fixed the typo in Eq. (16) for CRPS. The spread term measures the distance between ensemble members (similar to Eq. (9)) and it was evaluated correctly.

fCRPS=1P∑p=1P[1M∑m=1M∣Xppred−Xpobs∣−12M(M−1)∑m,m′=1M∣Xpm−Xpm′∣]\text{fCRPS} = \frac{1}{P} \sum_{p=1}^{P} \left[ \frac{1}{M} \sum_{m=1}^{M} \left| X_p^{\text{pred}} - X_p^{\text{obs}} \right| - \frac{1}{2M(M-1)} \sum_{m,m'=1}^{M} \left| X_p^{m} - X_p^{m'} \right| \right]

Probabilistic validation.

Thanks for the suggestion. In appendix Sec. H and G, we showed qualitatively the spatial (Sec. H) and temporal (Sec. G) consistency of the ensemble forecast.

As requested, we additional computed the Variogram Score (VS) based on Eq. (3) from Scheuerer and Hamill, 2015 (https://doi.org/10.1175/MWR-D-14-00269.1). For the experiment, we used an order p=0.5 and set weight w=1. We run two main analyses over Europe for the years 2022-2025 for both GRDC and CAMELS observations:

Temporal consistency. Here we check the ensemble across lead days up to 10 days forecast. For each station, one pair is two consecutive days. The VS score is computed for the generated ensemble members as a baseline (Ours) and for shuffled ensemble members (Shuffled Ens) to remove dependence. The difference (ΔVS) gives the change of our ensemble relative to the shuffled one. A lower VS score is better.

PeriodGRDC (1277 stations): OursShuffled EnsΔVSCAMELS (715 stations): OursShuffled EnsΔVS
Dec, Jan, Feb0.1240.201−38 %0.1390.184−24 %
Mar, Apr, May0.0910.152−40 %0.0970.128−24 %
Jun, Jul, Aug0.0570.090−36 %0.0710.083−15 %
Sep, Oct, Nov0.0890.140−36 %0.1260.157−20 %

As can be seen the generated ensemble is consistent across lead time.

Spatial consistency. Here we check the ensemble across space along the river network. One connected pair is two stations which are connected along the river network i.e., located along the same river. Unconnected pairs don't have connection in the river network. We compute VS scores within different distances starting from short to long distance. Similar to the temporal consistency check, we compare to shuffled ens members as a baseline which has similar CRPS but without dependency.

Each cell: Ours | Shuffled (ΔVS).

DistanceGRDC ConnectedGRDC UnconnectedCAMELS ConnectedCAMELS Unconnected
0–25 km0.068 | 0.131 (−48 %)0.085 | 0.121 (−30 %)0.064 | 0.105 (−39 %)0.094 | 0.110 (−15 %)
25–50 km0.075 | 0.115 (−35 %)0.092 | 0.120 (−23 %)0.074 | 0.095 (−22 %)0.102 | 0.112 (−9 %)
50–100 km0.082 | 0.107 (−24 %)0.103 | 0.121 (−15 %)0.088 | 0.093 (−5 %)0.111 | 0.113 (−2 %)
100–200 km0.096 | 0.109 (−13 %)0.113 | 0.122 (−7 %)0.090 | 0.088 (+2 %)0.121 | 0.119 (+2 %)
200–400 km0.096 | 0.103 (−7 %)0.120 | 0.123 (−2 %)0.077 | 0.074 (+4 %)0.124 | 0.122 (+1 %)
400–800 km0.094 | 0.099 (−5 %)0.123 | 0.123 (0 %)–0.117 | 0.115 (+1 %)

For the GRDC river network, the consistency along rivers persists up to 400–800 km, while it vanishes beyond 200 km for unconnected pairs. The ensemble thus represents dependence that follows the river network. Note that CAMELS catchments are smaller that GRDC so the gain along rivers vanishes after 50–100 km distance.

As can be seen the generated ensemble is also consistent across space and river network.

We will include this suggested experiment in the appendix in the revised paper.

The EFAS ensemble comparison also covers only June–August 2024; does the advantage persist across flood seasons?

Thanks for the comment. Note that the June-August 2024 includes a major flood event i.e., Augsburg flood in Germany. The processed data is 3.2 TB for June–August 2024 alone. However, we additionally downloaded and processed January 2024 (1.1 TB) and repeated the comparison for January in the winter season. As can be seen, the advantage persists also for January in Winter 2024.

Results on gauged GRDC Observations. Test set (2024-01-01 to 2024-01-31).

ModelRMSE ↓R2 ↑KGE ↑CRPS ↓CRPS mae ↓CRPS spread ↑
ProbEFAS (mean of 51 ens members)8.0788−1.13140.07195.92617.05400.7704
ProbOurs (mean of 12 ens members)5.9317−0.35690.40802.67796.16483.4245

Routing implementation.

Thank you for the note. For the ablation studies over mid Europe, the whole domain with 280820 points is processed at once in both training and inference. So no partitioning inside the domain is happening. Only upstream points outside the domain are not considered for routing. For the experiments on the full domain with 5.2 million points, the domain is split into sub-domains of 200000 points each and the upstream neighbors in another sub-domain are treated as absent. So only points at the edges are affected. Training and inference use the same split and the model sees the same graph for both.

Regarding the packed implementation, yes, our implementation keeps exactly the flow connectivity as given by the local drainage direction graph. The routing module forms one small sequence per downstream point and concatenates these small sequences into one long sequence. Every token along the sequence carries the id of its small sequence via seq_idx similar to masking in flash-attention. We used the fused kernel from (Tri Dao and Albert Gu, 2024, https://arxiv.org/abs/2405.21060) which applies the index on causal convolution and on the SSD scan. The causal convolution masks out tokens that come from a different segment and the SSD scan resets the hidden state at the beginning of each new segment.

We will make this clearer in the method section and add a comment on how the causal convolution also respects the boundary.

Thank you again for your reviewing our paper. We hope that our response has resolved all your concerns. Please let us know if there are still any open issues.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall response

We appreciate the reviewers' thoughtful comments and valuable feedback. We are glad that reviewers found our proposed feature-space routing approach novel and valuable (skY8, zLt6, B2LN), the clear benefit in an ungauged setting (zLt6), the scale and evaluation (skY8, zLt6), the probabilistic forecast in a distributed hydrological setting (B2LN), the relevance of the addressed problem (B2LN), and sound ablations (skY8, zLt6). Typically, the reviewers also raised some concerns. We addressed all of them in the rebuttal. We hope our response has sufficiently addressed everyone's concerns.

However, while working on the rebuttal, we have discovered a bug in the evaluation of the LSTMs baselines. Unfortunately, this affected all results for LSTM baselines. While we have corrected the numbers and figures in the current revised version, we want to improve our manuscript further and be more sure about it. We have therefore decided to withdraw the submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.