acceptodds
Under review as a conference paper at ICLR 2027

A Model-Agnostic Physics-Guided Adapter for Few-Shot Transfer of Coastal Flood Prediction Models to Unseen Regions

Abstract

Deep learning (DL) surrogates can produce high-resolution coastal flood maps orders of magnitude faster than physics-based hydrodynamic simulators, yet transferring them to new coastal regions remains costly, since generating target-region data for fine-tuning typically requires numerous time-consuming simulations. To tackle this bottleneck, we introduce the Physics Adapter (PA), a compact, architecture-agnostic adaptation interface that enables efficient few-shot transfer of flood prediction models across diverse coastal regions. PA predicts peak water level through a differentiable wet/dry response that compares terrain elevation against a learned water level, and blends this physics-structured prediction with a data-driven branch through a learned gate. Unlike physics-informed formulations, PA imposes no PDE-residual or conservation losses and instead exploits elevation as an architectural inductive bias, adding a negligible number of trainable parameters. We integrate PA into 12 heterogeneous models, spanning graph, convolutional, Transformer, state-space, depth-foundation and diffusion models, and evaluate them on two coastal regions with markedly distinct geometries, topographies, and shoreline protection configurations. The performance of PA is benchmarked against a no-physics baseline, full fine-tuning, and standard parameter-efficient fine-tuning (PEFT) methods, considering both within-region generalization to unseen sea level rise (SLR) values and between-region transfer. In low-shot regime (K=3), and averaged over all backbones and transfer settings, adding PA reduces root mean square error (RMSE) by 11.5% when only the output head is adapted on a frozen backbone, by 15.4% when combined with PEFT methods, and by 22.9% under full fine-tuning, compared to matched configurations without PA. Taken together, the findings of this work offer practitioners a concrete recipe for extending DL-based coastal flood predictors to new, data-scarce regions, thereby advancing scalable AI support for coastal adaptation planning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Coastal Flood Prediction, Few Shot Learning, Sea Level Rise, Physics-guided Adapter, Hydrodynamic Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g7VGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses transfer learning and how to make pretrained coastal flood prediction models adapted to new regions with few shot learning, and whether this can be achieved in a model-agnostic way. For this, the paper proposes a model-agnostic post-hoc module (physics adapter (PA)) that takes elevation and data driven features from pretrained model and outputs the predicted water level. The transferability with the proposed method was tested in two regions (Abu Dhabi and San Francisco) and the model-agnostic aspect was tested with different backbones. The experiments showed on these two regions that the proposed PA improves RMSE under different scenarios.

Critical strengths
  • The evaluation regarding the model-agnostic aspect is convincing.
  • I really enjoyed reading the main part of the paper, because it is clear and easy to follow.
Critical weaknesses

I think this paper is leaning to reject:

  • The main argument of the paper is that the proposed PA helps with transferability e.g., L016-019: “we introduce the Physics Adapter (PA), a compact, architecture-agnostic adaptation interface that enables efficient few-shot transfer of flood prediction models across diverse coastal regions” and L116-117: “... including across geographically and hydrodynamically distinct coastlines.” I think this is over-claim and not supported by the results. The model is only tested on two regions, and the proposed PA helps improving RMSE and not helping or solving the transferability issue. In addition, no flood detection metrics were reported to support the main argument, RMSE is not suitable metric for flood detection (see Q3).
  • The experiments do not provide any clue of the contribution of the proposed method compared to existing approaches. There is no comparison to other transfer learning methods for coastal floods. Table 1 looks like an ablation study rather than a benchmarking of the proposed module (see learned surrogates for flood prediction in related works). Also, there are no ablation studies about the components of the proposed physics adapter e.g., some ablation should include a comparison to location-aware adaptive normalization where affine transformation is learned inside the model like AdaNorm, also the whole PA can be replaced by an AdaNorm or different ways to inject elevation as conditioning. Otherwise, the physics adapter is not justified.
Questions for authors
  • The first question that comes to my mind is why the elevation is injected two times in the model! The backbone uses elevation already so maybe the way such static information is injected in each backbone is not good.
  • Is elevation enough to achieve transferability and isn’t this contradicting L60-62: “... domain shift arises in different forms: geographic transfer changes terrain, protection geometry, landcover structure, and flood response, whereas SLR transfer affects the forcing within the same region. An effective adaptation mechanism must handle both from only a handful of labeled examples.”?
  • Can you report transfer learning results like tables 6 and 7 in appendix? In flood we are interested mainly on flood detection not RMSE. Different models can have different systematic biases and important are metrics for flood detection like F1.
  • Line 252-253: why would the model collapse to data-driven branch if the physics branch is useful? What is the motivation?
  • The model does not work in zero-shot learning. Do you have any thoughts why?
  • The model works as an emulator of Delft3D rather than estimating the observed water level to test real-world transferability.

Things to improve the paper that did not impact the score:

  • L284-286: “Geographic transfer changes the region (rs≠rtr_s \neq r_t), whereas SLR transfer keeps the region fixed and changes only the forcing”, how the forcing was changed? Do you do extrapolation in time?
  • Format page 6: the bottom margin is against the guideline format.
  • I would make figure 2 (b) do multiplication of gate gig_i with both branches then some the two branches so it becomes clearer from the figure without reading the text.
3 comments
AuthorsOpenReview ↗

## Response to Reviewer g7VG (1/3)

Thank you for the careful review, and for the positive comments on the clarity of the paper and on the model-agnostic evaluation. All changes are marked in blue in the revised PDF.

W1. Over-claiming, only two regions, and no flood-detection metric.

  • Claims revised. We rewrote the sentences you quoted. The abstract (L016–019) now describes the PA as an interface "for few-shot transfer of coastal flood surrogates across regions and sea levels". The related-work sentence (L116–117) now says "under the tested geographic and sea-level shifts". The contributions and the conclusion state the gains "in RMSE", and the limitations state that all targets are simulated.
  • Third region added. We added Dubai, which is never seen in training, with the source models and hyperparameters unchanged (Sec. 5.4, Table 2a, Appendix H). With 10 Dubai simulations, FT+PA from SF reaches 0.102 m and a flooded-class F1 of 0.953, against 0.134 m and 0.936 for FT. With only the head adapted, the PA more than halves the RMSE (0.22 against 0.50 m).
  • Flood-detection metrics added. Thank you for this point. F1 is now reported throughout the paper (see Q3 below).

W2. Comparison with other transfer methods, and ablations.

  • Existing flood-transfer methods. The few studies that transfer flood surrogates either fine-tune the model on target data (Xu et al., 2025), which is our FT regime, or use a domain-adaptation scheme built for one neural operator (Taghizadeh et al., 2026), which does not carry over to our 12 architectures. We now also compare with LP-FT and with surgical fine-tuning of the first or last block, each with and without the PA (Appendix I, Table 36). Adding the PA raises F1 for all seven transfer methods in both settings and lowers RMSE in 9 of the 14 cases.
  • Elevation conditioning, including AdaNorm. As you suggested, we replaced the whole PA with (i) a head on features with elevation concatenated, (ii) FiLM conditioned on elevation, and (iii) an elevation-conditioned adaptive normalization at all 17 normalization layers of the decoder, similar to LOAN/AdaNorm. Each was trained from the source stage (Sec. 5.4, Table 2b; Appendix I, Table 35, Figure 21).
Head (VM-UNet, mean of 4 settings)Adapted paramsHead only, K=3 RMSE / F1Whole model, K=5 RMSE / F1
PA2940.127 / 0.9470.090 / 0.983
Elevation concatenated1,3510.201 / 0.9110.091 / 0.976
Elevation FiLM4,1050.200 / 0.9290.090 / 0.975
AdaNorm in the decoder52,4570.191 / 0.8810.091 / 0.977
No adapter46,9450.169 / 0.9340.094 / 0.969

Findings. With only the head adapted, the PA is clearly the best, with far fewer parameters. When the whole model is adapted, all designs are within 0.002 m, and the PA keeps the highest F1. The ablation of the parts of the PA is in our response to Reviewer H7Zj and in Table 2b.

AuthorsOpenReview ↗

## Response to Reviewer g7VG (2/3)

Q3. Flood detection (F1) and transfer tables like Tables 6 and 7.

We agree that flood detection matters. In our setting it is a sub-task of the PWL regression, since a location is flooded when its predicted PWL exceeds 0.05 m, so F1 follows from the same predictions. We focused on PWL because protection planning needs water levels and not only the wet/dry status, for example to size defenses.

Changes made. Detection is now reported in full.

  • F1 next to RMSE in Table 1 and Figures 2b and 2d, and the F1 gain from the PA in Table 11.
  • Transfer tables in the format of Tables 6 and 7, for every setting at K=10 and averaged over K (Appendix G.4, Tables 12–31), with flooded, dry and mean F1 in Tables 19–21 and 29–31.
  • Precision and recall of the flooded class by backbone family (Appendix G.5, Table 32), and the definitions in Appendix F.

Findings. All models detect floods well, so F1 leaves much less room for improvement than RMSE. F1 is 0.91 to 0.99 in domain and 0.93 to 0.97 after transfer at K=10 (Table 1). The relative gain from the PA is therefore smaller in F1 than in RMSE. Over all settings at K=3 it is 4.5 to 8.2% in F1 against 33 to 46% in RMSE, both computed from the averaged metrics (Table 11). The gain is concentrated where models fail without the PA (Table 32, K=3, cross-region, full fine-tuning).

Backbone familyRMSE (m), FT → FT+PAF1, FT → FT+PA
Graph and operator1.309 → 0.3940.454 → 0.918
Dense0.214 → 0.1930.916 → 0.907

This is the rise in F1 "from 0.45 to 0.92 at K=3" stated in the abstract, and it comes from the graph and operator backbones. Without the PA, they predict flooding at many dry locations (precision 0.28 to 0.38), and the PA raises their precision to 0.70 to 0.82. For the dense backbones, F1 is already about 0.9. The PA lowers their RMSE, while F1 changes by −0.01 under full fine-tuning and by −0.05 to −0.07 with PEFT and head-only adaptation, with a higher recall and a lower precision. We report this openly in Sec. 5.2, Appendix G.5 and the conclusion. In Dubai, FT+PA has the highest F1 at every K with SF as the source (Table 2a).

Q1. Why is elevation given twice?

The backbone does receive elevation as an input, but inside the network it is normalized and mixed with land cover, protection status and SLR, so its scale in meters is lost. The PA uses the raw DEM in meters to compare it with a predicted water level (Eq. 3). The ablation shows that this matters. With elevation only through the backbone (data branch only), the RMSE is 0.262 m against 0.182 m for the PA (cross-region, K=3, head only, Table 35). Change made. This explanation is added to Sec. 3.3.

Q2. Is elevation enough for transfer?

No, and the PA does not rely on elevation alone. Only the physics-guided branch uses the raw DEM, and even there the wet/dry threshold and water level are predicted from the backbone features, which carry land cover, protection status and SLR. The data branch uses all features, and the gate decides per location how much to trust each branch. The ablation confirms that elevation alone is not enough, since the physics-guided branch alone gives 0.269 m against 0.182 m for the full PA (Table 35).

AuthorsOpenReview ↗

## Response to Reviewer g7VG (3/3)

Q4. Why would the gate collapse onto the data branch?

The data branch sees all the features and can fit the training labels of the source region on its own. Without a penalty, the gate drifts towards it during source training, and the elevation structure that helps in a new region is lost. The penalty only discourages this collapse and does not force the same gate everywhere. The ablation supports this. Without the penalty, the PA gives 0.289 m against 0.182 m across regions and 0.182 m against 0.072 m under SLR transfer (K=3), and the gate of the AD source model falls well below that of the SF model and varies strongly between seeds (Appendix I, Figure 22). Change made. This motivation is added to Sec. 3.4.

Q5. Why does the model not work zero-shot?

The reference water level η\eta is learned on the source region. Without any target data, the PA compares the target terrain with the source water level, which is wrong for a new region or sea level. One target simulation is enough to correct this, and from K=1 on every PA regime has a lower RMSE than every regime without it (Figure 2a). This is discussed in Sec. 5.2 ("Zero-shot behavior") and in the limitations. On VM-UNet with only the head adapted, about three target simulations are needed (Appendix I).

Q6. The model emulates Delft3D rather than observed water levels.

We agree. This is inherent to the task. Protection scenarios are planning options that have not been built, so their flooding cannot be observed, and the surrogate replaces the simulator when screening them. The AD hydrodynamic model was validated over a three-month period at 194 locations in the Gulf (Hassan et al., 2026, supplement).

Change made. The limitations now state that all targets are simulated and not observed.

Minor points (changes made)

  • Forcing (L284–286). SLR transfer changes the sea level in the simulations (0.5, 1.0 and 1.5 m). There is no extrapolation in time. This is now stated in Sec. 3.4.
  • Page 6 margin. Fixed.
  • Adapter figure. Figure 1(b) is redrawn. Each branch is multiplied by its gate weight (gig_i and 1−gi1-g_i), and the two products are summed, as in Eq. (5).
Reviewer_7XZ1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a physics adapter (PA) for flood few-shot coastal flood prediction. More specifically, the idea is to be able to transfer a model trained in one region to another one by training only the adapter based on a small number of examples in the target region. The proposed adapter operates on the representations generated by any ML-based flood predictor and applies gated fusion to combine the data-driven prediction with an elevation conditioned term. The adapter is trained jointly with the flood prediction model in the source dataset and is fine tuned based on target examples. Other parameters can also be adapted based on the target examples depending on the adaptation regime. The experiments are based on Deft3D simulations for two locations (SF and Abu Dhabi), 12 flood prediction models, and two sea rise levels. Results, based on R^2 and RMSE, show that the proposed adapter improves adaptation across models and adaptation regimes whenever at least on target example is available (1-shot).

Critical strengths

S1) The experimental results show that the proposed physics adapter improves few-shot learning in multiple settings.

S2) The paper provides a comprehensive appendix with additional details and experiments

S3) The paper addresses a relevant problem, which is the few-shot adaptation of coastal flooding models

Critical weaknesses

W1) The contributions of the paper are quite limited. The physics adapter is a simple (linear) combination of a well-known parameterized inundation rule and the flood model predictions. One could train an ensemble of models with different levels of complexity, including the inundation rule applied, and then use the few-shot examples to decide which model to apply globally or at each location. The best combination of the model could also be identified based on the examples. The proposed approach is small modification of this idea and it is unclear how it advances the field, even in the coastal flooding domain.

W2) The experiments are based on only two regions. It is difficult to generalize the conclusions in the paper without a more comprehensive benchmark. In particular, the paper seems to rely on fixed hyperparameters (Equation 6) that might not generalize across regions.

W3) The error after adaptation seems particularly high across the methods. One could argue that such results have limited use and that training the model on target simulations is still the best approach in practice, even if it takes more resources.

Questions for authors

Q1) How does the proposed approach compare against simpler adaptation methods that select between an ensemble of models with different levels of complexity and uses the target examples to select or combine them?

Q2) How can the claims in the paper be supported using such a limited number of locations (2)?

Q3) Why are the adaptation results still useful in practice? What relevant decisions can be made using them despite their error being significantly higher than what one would expect from target training?

2 comments
AuthorsOpenReview ↗

## Response to Reviewer 7XZ1 (1/2)

Thank you for the detailed review. We ran the ensemble baseline you suggested and added a third region.

W1 / Q1. Comparison with an ensemble that selects or combines models.

We agree this is a natural baseline.

Experiments done. We implemented it (Appendix I, Table 35, summary in Table 2b). The pool holds the two members that the PA itself combines, the network without the PA and the inundation rule b 1[zi<a]b\,\mathbb{1}[z_i<a], with aa and bb fitted on the KK supports. We test the rule alone, selecting one member for the whole region or per location, blending with one weight for the region or per location, and stacking. All use the same support draws and test sets as our K-shot regimes. RMSE (m) on VM-UNet with only the head adapted (Table 35):

MethodCross-region K=1K=3K=10SLR K=1K=3K=10
PA0.3610.1820.1410.2230.0720.066
Network without PA0.2340.2030.1900.1540.1350.125
Inundation rule only0.5490.4950.4920.4630.4350.398
Selection, region0.2750.2030.1900.1540.1350.125
Selection, per location0.2450.2010.1740.1260.1030.096
Blend, region0.3040.2010.1880.2720.1330.123
Stacking0.3040.2010.1880.2720.1330.123

Findings. Per-location selection is the strongest ensemble and is better than the PA at K=1. From K=3 on, the PA is clearly better in both settings, and the ensembles with regional weights also have a lower F1 across regions (0.77 to 0.81).

Difference between the two approaches. An ensemble chooses among fixed models using the few target examples, so the rule keeps two global parameters and the networks are not changed. In the PA, the water level, the wet/dry threshold and the gate are predicted per location from the backbone features and trained together with the backbone. The PA can therefore be adapted with any regime (head only, PEFT or full fine-tuning) and plugs into 12 different architectures unchanged, with 294 parameters on the dense backbones. It also keeps or improves in-domain accuracy for 10 of the 12 backbones (Sec. 5.1).

W2 / Q2. Only two regions, and fixed hyperparameters (Eq. 6).

Experiments done. We added Dubai as a third region that is never seen in training. We reused the SF and AD source models and every hyperparameter, including λg=0.1\lambda_g = 0.1 in Eq. (6), with no tuning for Dubai (Sec. 5.4, Appendix H). The results follow the same pattern as in AD and SF (Table 2a, source SF, RMSE in m / flooded-class F1).

RegimeK=1K=3K=5K=10
FT+PA0.440 / 0.5730.250 / 0.7830.188 / 0.8540.102 / 0.953
FT0.538 / 0.3730.322 / 0.6880.214 / 0.8060.134 / 0.936
PA (head only)0.461 / 0.3170.264 / 0.3200.223 / 0.3310.217 / 0.362
NPA (head only)0.512 / 0.3070.493 / 0.3140.493 / 0.3150.496 / 0.315

With AD as the source, FT+PA is ahead from K=5 on. The three regions differ strongly. AD lies on the open Gulf with wind waves and run-up and has 17 OLUs, SF is a sheltered bay with 30 OLUs, and Dubai has 13 OLUs at a third sea level. The SLR settings add a change in forcing within one region.

Changes made. We have narrowed the claims to "the tested geographic and sea-level shifts", and the limitations state that the third region was tested with one backbone.

AuthorsOpenReview ↗

## Response to Reviewer 7XZ1 (2/2)

W3 / Q3. Error after adaptation, and training on the target region instead.

Experiments done. We tested this directly in Dubai by training VM-UNet without the PA on Dubai alone, with 12 extra simulations for early stopping (Table 2a, Appendix H, Tables 33–34, Figure 19). One Dubai simulation takes about 38 h, or 850 CPU-hours, estimated from the costs reported for AD and SF by Hassan et al. (2026).

ApproachDubai simulationsSimulation timeRMSE (m)Flooded-class F1
FT+PA from SF, K=335 days0.2500.783
FT+PA from SF, K=101016 days0.1020.953
Target-only1930 days0.4580.265
Target-only3048 days0.0880.592
Target-only4775 days0.0720.781
All configurations (2132^{13})8,19236 years––

Findings. The reviewer is right that training on the target region gives a lower RMSE once enough simulations exist. However, it needs about three times as many simulations to match the RMSE of transfer at K=10, its F1 never exceeds 0.805, and with all 47 simulations it overpredicts the flooded area by 59%, against 4 to 6% for FT+PA. With 213=8,1922^{13}=8{,}192 possible protection configurations in Dubai (2172^{17} in AD and 2302^{30} in SF), only a small fraction can ever be simulated, which is exactly the few-shot setting we study.

On the size of the errors, the averages in the paper are dominated by the graph and operator backbones. At K=10 the dense backbones reach an RMSE of 0.08 to 0.12 m and an F1 of 0.93 to 0.97 (Table 1). Training only the adapter (294 parameters on the dense backbones) already gives a lower RMSE than full fine-tuning without it at every K ≥ 1 (Figure 2a).

Practical use of the adapted models

  • Screening protection scenarios. With ten Dubai simulations, FT+PA ranks the 15 test scenarios by flooded area almost exactly as the simulator does (Spearman ρ=0.971\rho=0.971, Table 33). A planner can therefore rank thousands of configurations and send only the most promising ones to the simulator.
  • Flood extent. At K=10 the predicted flooded area in Dubai is within 4 to 6%, and the cross-region F1 of VM-UNet with FT+PA is 0.950 (Table 19).
  • Planning the simulation budget. Figure 19 and Table 33 show the accuracy reached with 1, 3, 5 and 10 target simulations, which tells a user how many simulations a new region needs for a given accuracy. For reference, K=10 is 3.5% of the 285 SF scenarios and 7% of the 142 AD scenarios used for source training.

Final design decisions for a specific defense would still be checked with the full simulator. The adapted surrogate is meant to make the search before that step affordable.

Reviewer_H7ZjOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Physics Adapter (PA), a lightweight and model-agnostic framework for few-shot transfer of coastal flood prediction models to unseen regions. PA incorporates terrain elevation as a physical inductive bias through a differentiable wet/dry response and combines it with a data-driven prediction branch using a learned gate. The adapter can be integrated with different backbone architectures and parameter-efficient fine-tuning methods without introducing PDE-based constraints. The authors evaluate PA on 12 diverse backbones across two coastal regions, considering both cross-region transfer and sea-level-rise transfer. Experimental results demonstrate consistent improvements over matched adaptation methods without PA, particularly in low-shot settings.

Critical strengths

The paper addresses an important challenge in transferring coastal flood prediction models to data-scarce regions. The proposed Physics Adapter introduces a lightweight elevation-based inductive bias that can be incorporated into different backbone architectures and combined with existing PEFT methods. The overall design is intuitive and practically relevant.

The authors evaluate the approach across 12 diverse backbones, multiple adaptation strategies, and both geographic and sea-level-rise transfer settings. The experiments include different few-shot budgets, multiple training seeds, repeated support-set sampling, and qualitative visualizations. The consistent improvements across model families provide encouraging evidence for the general applicability of the proposed adapter.

Critical weaknesses

While the experiments demonstrate the effectiveness of the overall adapter, the individual contributions of the elevation-based wet/dry response and learned gating mechanism are not fully analyzed in the main paper. Further analysis would help clarify which components contribute most to the observed improvements.

Questions for authors

Could the authors further clarify the individual contributions of the physics-guided wet/dry response and the learned gating mechanism?

1 comment
AuthorsOpenReview ↗

## Response to Reviewer H7Zj

Thank you for the positive assessment and for the clear question. We have added the ablation you asked for.

Experiments done. For VM-UNet we trained four variants of the adapter, each from the source stage, and adapted them in the same way as the PA (Sec. 5.4, Table 2b; full results in Appendix I, Table 35, Figures 20 and 22).

  • Physics-guided branch only (gi=1g_i = 1).
  • Data branch only (gi=0g_i = 0), so elevation reaches the output only through the backbone input.
  • Fixed gate (gi=0.95g_i = 0.95 at every location).
  • The PA without the gate penalty (λg=0\lambda_g = 0).

Results (Table 2b, averaged over the four transfer settings)

HeadHead only, K=3 RMSE (m)F1Whole model, K=5 RMSE (m)F1
Full PA0.1270.9470.0900.983
Physics-guided branch only0.2190.9150.1070.967
Data branch only0.2180.9400.0930.975
Fixed gate0.2420.9130.0950.966
No adapter (NPA / FT)0.1690.9340.0940.969

Findings

  • Neither branch is enough on its own. With only the head adapted, each branch alone is about 0.09 m worse than the full PA. The data branch sees elevation only through the backbone input, so its gap to the PA (0.08 m across regions at K=3) is what the wet/dry response on raw elevation adds.
  • The gate must be learned per location. A fixed gate is worse than either branch (0.242 m), so the gain comes from choosing between the branches location by location.
  • The gate penalty matters. Without it, the same 294 parameters give 0.289 m instead of 0.182 m across regions, and 0.182 m instead of 0.072 m under SLR transfer (K=3, Table 35). The gate of the source model also drifts towards the data branch and varies between seeds (Figure 22).
  • When the whole model is adapted, the differences become smaller, but the full PA still has the lowest RMSE and the highest F1.

Limitations reported. We also state the cases where the PA is not best. With a single target scenario (K=1) and only the head adapted, the PA is not the best head, and under SLR transfer with the whole model adapted all variants are within 0.02 m of each other, with the PA behind most of them (0.058 m against 0.050 to 0.055 m, Appendix I).

Comments on the paper

1 comment
AuthorsOpenReview ↗

## General response to all reviewers

We thank the three reviewers for their careful reading and helpful comments. We also apologize for posting our responses late in the discussion period. The new experiments (a third region, ablations of the adapter and additional baselines) were extensive and took longer than we expected, and we preferred to report complete results rather than partial ones. We have revised the paper substantially, and all changes are marked in blue in the revised PDF. The main changes are listed below.

  1. A third, completely unseen region (Dubai). Dubai has 13 OLUs and 62 simulated protection scenarios at 0.75 m SLR, and it is used only as a transfer target. We reused the SF and AD source models and all hyperparameters unchanged, including λg\lambda_g in Eq. (6). With 10 Dubai simulations, FT+PA from SF reaches an RMSE of 0.102 m and a flooded-class F1 of 0.953, against 0.134 m and 0.936 for FT without the PA (Sec. 5.4, Table 2a, Appendix H, Tables 33–34, Figure 19).
  2. Flood-detection metrics. F1 is now reported next to RMSE throughout the paper. It is the mean of the flooded-class and dry-class F1, with a location counted as flooded when its PWL exceeds 0.05 m. See Table 1, Figures 2b and 2d, Tables 9–11, the transfer tables 19–21 and 29–31 (flooded, dry and mean F1), and precision and recall by backbone family in Table 32 (Appendix G.5). Definitions are in Appendix F.
  3. Ablations of the adapter. We remove or fix each part of the PA (physics-guided branch only, data branch only, fixed gate, no gate penalty) and compare it with three other ways of conditioning on elevation (concatenation, FiLM, and AdaNorm-style normalization in the decoder). See Sec. 5.4, Table 2b, Appendix I, Table 35 and Figures 20–22.
  4. More baselines. Support-set ensembles of the network and an inundation rule (selection and blending, regional and per location, and stacking), and two further transfer methods, LP-FT and surgical fine-tuning, each with and without the PA (Appendix I, Tables 35–36).
  5. Transfer against training on the target region alone, together with the simulation cost of each option (Table 2a, Table 34, Figure 19).
  6. Claims revised to match the results. The abstract, introduction, related work and conclusion now state the gains "in RMSE" and "under the tested geographic and sea-level shifts", and the limitations state that all targets are simulated.
  7. Figure 1(b) redrawn to show the two gated products and their sum, as in Eq. (5).

Choice of backbone for the new experiments. Because of the limited time of the rebuttal period, all new experiments (Dubai, ablations and additional baselines) use VM-UNet. It is the strongest of the 12 backbones in our benchmark. It has the lowest RMSE and the highest F1 in domain (0.054 m and 0.986), and the lowest RMSE and the highest F1 after transfer at K=10, averaged over the four settings (0.080 m and 0.973, Table 1). It also has the lowest RMSE across regions and under SLR transfer separately (Tables 16 and 17). The limitations state that the third region was tested with this one backbone.

We answer each reviewer's points individually below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.