Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors, please find attached my small contribution 

Comments for author File: Comments.pdf

Author Response

We have provided all the point-by-point responses in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Goal: The study asks whether the segmentation architecture or input representation is contributing more to the accuracy of flood-event surface-water mapping. Using a CNN encoder named GeoHybridNet, they compare it under matched training with a U-Net, an Attention U-Net, a parameter-matched wide U-Net, and a SegFormer, with 3 seeds.

Major Concerns:

1. The positive class is all surface water, permanent plus flood (lines 302–308). IoU is therefore dominated by rivers and lakes that NDWI separates trivially, which likely inflates the apparent importance of NDWI. Flood-specific conclusions, and the title, require either separating flood from permanent water using a pre-event water mask (e.g., JRC Global Surface Water) or reframing the claims as surface-water mapping.
2. The central question ("input representation matters more than architecture") is not helped by the design of the experiment. The architecture and input are never varied jointly, so their contributions cannot be compared or checked for interaction. The authors acknowledge this in Section 4.4 (lines 799–801), but the abstract, highlights, and conclusions state the result without that qualification. A 3 (architecture) × 3 (raw bands / NDWI only / full) design with at least five seeds is needed to support the claim.
3. NDWI is a deterministic function of B3 and B8, which are already in the S2 input, yet the optical-only model reaches only 50.5% IoU versus 75.9% for NDWI alone. They should also test simple conditioning controls such as a B3−B8 difference channel. Without these, the claim that "431 chips are insufficient to induce from data alone" (lines 546–548) is unsupported.
4. The only formal test is a single Wilcoxon signed-rank test (GeoHybridNet vs U-Net, seed 42, n = 83, p = 0.93; lines 489–493). It is used for the conclusions about all five models. It does not correct for multiple comparisons, and the chips are spatially clustered by event and therefore not independent.

Summary: The paper asks a worthwhile question and is candid about null results. As written, however, its main claim (input matters an order of magnitude more than architecture) is unverified and has an under-tuned training exercise. Thus, the intellectual contribution is not clear. I urge the authors to perform a major revision and clarify these issues.

Author Response

Please see the attached file with detailed, point-by-point responses.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Overall, this manuscript presents a clear research question, reasonable experimental design, and well-structured writing. However, several issues remain as follows.

In the section on study area overview, line 278 mentions using 10 flood events, while line 291 refers to 11 flood events—an inconsistency. It is recommended to specify the types of each flood event (riverine flooding, flash floods, coastal/pluvial) and provide corresponding notes on SAR/optical data availability.

Section 2.3 should clarify whether NDWI was included in per-channel standardization and whether statistical metrics were derived solely from the training set.

Section 2.4 should elaborate on the rationale behind Transformer hyperparameter settings, including why a 1/16 resolution was adopted, and include a sensitivity analysis of parameters.

Section 3.3 states that input modality ablation "was now measured over three seeds," whereas Section 4.3 claims "the modality ablation was run on a single seed"—a direct contradiction. Please verify and unify the descriptions across the main text, Table 4, and the limitations section, clearly explaining the source of ±std values.

It is recommended to add a legend to Figure 13 to improve readability and comprehension for readers.

In the conclusion, the statement "input representation and honest evaluation—not architecture—govern flood-mapping accuracy" should be qualified as "on Sen1Floods11 at this benchmark scale." Additionally, boundary conditions should be clarified: whether the findings hold for larger training scales, multi-temporal inputs, urban inundation, or coastal flooding.

In the discussion section, the claim in Section 4.3 that "the modality ablation was run on a single seed" contradicts the main text and requires correction. Furthermore, it is suggested to report the range or coefficient of variation of IoU across different folds to make variance more intuitive.

Author Response

Please see the attached file with detailed, point-by-point responses.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Accepted.

Author Response

We thank the reviewer for accepting our revised manuscript.