Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Water depth inversion based on remote sensing information is one of the research hotspots at present. This manuscript systematically reviews the research on deep learning-based satellite-derived bathymetry (SDB) from 2021 to 2025, summarizes the methodological characteristics by model types, and discusses the data sources, accuracy and existing challenges. With a relatively comprehensive coverage and a clear structure. It is interesting.

After carefully reading the manuscript, it is suggested to revise the following items to improve the manuscript.

The systematic review method is not standardized enough, and there are inconsistencies and insufficient reproducibility in the retrieval and screening processes. Inconsistent expressions of "309 records / 540 records / 130 core papers" appear in the manuscript (Lines 40, 95), and the screening process lacks clear inclusion criteria, exclusion criteria, deduplication strategies, quality control indicators, etc. It is recommended to provide the complete search strategy and screening criteria.
There is a lack of a quantitative comparison and summary table. Table 1 is more like an example list and lacks a quantitative summary that can be used for decision-making. It is recommended to improve Table 1 by introducing relevant indicators to facilitate horizontal comparison.
Some technical expressions contain conceptual and logical flaws and require careful verification. For instance, the description of SAR applied to bathymetry in Section 2.2 is rather vague. SAR-derived water depth estimation typically relies on indirect mechanisms such as wave dispersion and sea surface features, with stringent application conditions. It is recommended to supplement the key assumptions and limitations to avoid giving readers the impression that SAR is widely applicable to SDB.
It is recommended to standardize the technical terms and abbreviations (e.g., SDB, ALB, IOPs, MDD, TDAD, etc.), with their full official designations provided upon the first occurrence.
There are errors in the numbering of figure captions. For example, “Figure 7” is incorrectly labeled as "Figure 3". It is recommended to read through the entire manuscript and correct all such errors.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a recent synthesis (2021–2025) and analyzes 43 models and 130 core studies, with a clear organization into sections covering conventional methods, deep learning architectures, operational challenges, and future directions, and concludes with an integrative discussion supported by a comparative table. The scope is adequately defined, and the review explicitly discusses relevant technical limitations, such as optical complexity, generalization problems, and label quality, in addition to proposing coherent lines of advancement, including multimodal fusion and physics-informed approaches. It is also noteworthy that the search is declared systematic (WOS and Google Scholar) and that the taxonomy by architecture families (CNN, U-Net, MLP, RNN, and Transformers) facilitates reading and conceptual comparison. However, there are still aspects that need to be strengthened before publication: there is an inconsistency in the count of records (309 versus 540) that must be harmonized and explained, and to ensure auditability, it is recommended to incorporate the complete search strings by database, the exact date of consultation, inclusion and exclusion criteria, the duplicate removal procedure, an explicit quality control scheme, and a PRISMA-type diagram. It would also be useful to supplement the table of representative models with a master table for each study (sensors, water type, label source, depth range, metrics, and generalization strategy) and to refine the bibliography, given that inconsistent metadata is observed in some references, for example, in the citation associated with UNCLOS. In this vein, the English could be improved through stylistic editing to increase conciseness, reduce generic formulations, ensure terminological consistency (e.g., “bathymetry retrieval” and “SDB”), and adjust the conclusion to ensure it remains strictly supported by the synthesized evidence.

Comments on the Quality of English Language

Review the entire document

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Comments and Suggestions for Authors

The literature search is restricted to 2021–2025, but satellite-derived bathymetry (SDB) with deep learning has key foundational studies before 2021. Could you explain the rationale for this strict time cutoff and discuss whether omitting these early works limits the comprehensiveness of your review of model evolution and methodological origins?
While you classify deep learning models (CNN, U-Net, MLP, RNN) and their variants, there is a lack of quantitative comparative analysisof their performance (e.g., RMSE, R², computational cost) across consistent datasets/water conditions. Would it be possible to add a unified quantitative benchmark table to clarify the trade-offs between different models in practical SDB applications?
You identify water optical complexity and cross-scene generalization as core challenges, but the discussion of uncertainty quantification (UQ)in deep learning-based SDB is superficial. Could you elaborate on current UQ methods for SDB (e.g., Bayesian neural networks, Monte Carlo dropout) and their limitations in shallow-water bathymetric inversion?
The review mentions multi-source data fusion (optical, LiDAR, SAR, gravity) but does not address practical integration challenges(e.g., spatial/temporal scale mismatches, radiometric calibration differences between sensors, and missing data in fusion pipelines). Could you expand on these technical hurdles and existing mitigation strategies?
You propose physics-informed deep learning as a key future direction, yet the review only briefly introduces a few physics-constrained models (e.g., PI-CNN, PI-RNN). Could you categorize existing physics-informed SDB approaches (e.g., radiative transfer embedding, hydrodynamic constraint, wave mechanics priors) and analyze their relative strengths/weaknesses in different water environments (clear/turbid, shallow/deep)?
The discussion of low-data and transfer learning strategiesis limited, despite the scarcity of in-situ depth labels being a dominant bottleneck. Could you detail state-of-the-art methods (e.g., few-shot learning, domain adaptation, self-supervised pre-training) for SDB and their performance in cross-regional/cross-sensor scenarios?
You mention lightweight models for SDB but do not explore edge deployment and computational efficiency(e.g., model quantization, pruning, real-time inference feasibility). Given the demand for on-site bathymetry mapping (e.g., UAV/USV-based), could you discuss the balance between model accuracy and computational cost for practical SDB applications?
The review focuses on open-ocean and coastal shallow waters but has minimal discussion of inland water bathymetry(e.g., lakes, rivers, reservoirs) which face unique challenges (e.g., high sediment load, aquatic vegetation, complex shorelines). Could you expand on how deep learning models need to be adapted for inland water SDB and identify gaps in current research?

Comments on the Quality of English Language

Comments on the Quality of English Language

The English language used in the review is clear, professional, and academically appropriate for a technical literature survey. The writing maintains a consistent formal tone throughout, with precise terminology (e.g., "uncertainty quantification," "physics-informed deep learning," "radiometric calibration") accurately applied to the satellite-derived bathymetry (SDB) and deep learning domains. Sentence structures are well-organized and logically coherent, avoiding ambiguity even when addressing complex technical concepts—this ensures readability for both specialists and researchers familiar with related fields. Terminology consistency is a notable strength: key terms (e.g., "cross-scene generalization," "in-situ depth labels," "computational cost") are used uniformly across questions, and technical acronyms (e.g., CNN, U-Net, RMSE, UQ) are either widely recognized in the field or clarified contextually. The phrasing of each comment/suggestion is direct yet constructive, adhering to standard academic review conventions without overly complex syntax or jargon. Overall, the English language quality is high and suitable for publication—no grammatical errors, awkward phrasing, or imprecise terminology were identified. The writing effectively conveys technical questions and suggestions in a clear, concise manner that supports productive revision of the review.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Made the changes and followed the suggestions

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Review Comments for the Manuscript A Review on Bathymetric Inversion Research Based on Deep Learning Models and Remote Sensing Images

Comment 1: Grammatical and Syntactical Inconsistencies

Several sentences contain minor grammatical errors, awkward phrasing, or missing prepositions that reduce readability. For example:

Line 241: “On data sources and multimodal fusion, hyperspectral satellites such as PRISMA and EnMAP provide hundreds of narrow spectral bands, enabling finer discrimination of water types and benthic classes.”The preposition “Regarding”is more appropriate than “On” for introducing a topic in academic writing.
Line 326: “In satellite-derived bathymetry, Transformers can effectively capture relationships across spectral bands and over time, thereby improving model performance.”This sentence is a fragment when placed after the CNN model introduction; adding a transitional phrase (e.g., “In addition to CNNs,”) would improve flow and grammatical coherence.
Line 776: “Despite substantial improvements in conventional accuracy metrics (e.g., RMSE and R²), most deep-learning–based SDB models remain deterministic and provide only point estimates of depth without explicit confidence intervals.”The hyphenation of “deep-learning–based”is redundant; the standard form “deep learning-based” is preferred in remote sensing literature.

Comment 2: Inconsistent Nomenclature and Abbreviation Usage

There is inconsistent use of abbreviations and model nomenclature throughout the manuscript, which may confuse readers:

“Satellite-Derived Bathymetry”is abbreviated as SDB(correct) but is occasionally written as “satellite-derived bathymetry (SDB)” with inconsistent capitalization (e.g., Line 192 vs. Line 276). A consistent style (Satellite-Derived Bathymetry (SDB) for first use, SDB thereafter) should be applied.
Model variants such as PI-CNNand PI-RNNare defined, but “physics-informed” is sometimes hyphenated (physics-informed) and sometimes not (physics informed) (e.g., Line 374 vs. Line 551). Standardize hyphenation for all physics-informed model terms.
The abbreviation DoV(Deflection of the Vertical, Line 401) is used without a first-time definition; all non-standard abbreviations must be defined upon their first appearance in the text.

Comment 3: Incomplete Quantitative Benchmark Context in Table 2

Table 2 provides a valuable quantitative comparison of SDB models, but critical contextual details are missing, limiting its utility for practical benchmarking:

The RMSE unitsare listed for most models but are absent for a few (e.g., XGBoost, U-Net + SCNN); uniform unit labeling (meters, m) is required for all entries.
The “Generalization Test (Y/N)”column does not specify the typeof generalization (cross-region, cross-sensor, or cross-water type); adding a parenthetical note (e.g., Y (cross-region), N (no cross-scene testing)) would make this column more informative.
No dataset information(e.g., Sentinel-2, ICESat-2, in-situ survey size) is included for each model; readers cannot assess how training data differences impact reported performance metrics.

Comment 4: Underdeveloped Discussion of Inland Water SDB Data Sources

Section 4.2.3 (Inland Water Bathymetry) thoroughly addresses model adaptation challenges but fails to discuss tailored data sources for inland water bodies (e.g., rivers, lakes, reservoirs). Coastal SDB relies heavily on Sentinel-2/Landsat and ICESat-2, but inland waters require high-resolution data (e.g., UAV multispectral, GF-6/GF-7) and alternative LiDAR datasets due to narrow channels and turbidity. The manuscript should add a brief subsection on inland SDB-specific data sources and their integration with deep learning models to complement the challenge-focused discussion.

Comment 5: Weak Linkage Between Computational Efficiency and Real-World Deployment

Section 4.2.5 (Computational Efficiency and Edge Deployment) discusses model compression techniques (pruning, quantization) but does not establish a clear linkage to real-world SDB deployment scenarios (UAV/USV-based mapping, near-real-time coastal monitoring). The manuscript should:

Add 1–2 examples of lightweight SDB models (e.g., lightweight U-Net, compact CNN) that have been successfully deployed on edge devices (e.g., embedded GPUs on UAVs).
Include a brief discussion of inference latency benchmarks(e.g., frames per second, processing time per km²) for compressed vs. full-capacity models, as latency is a critical metric for on-site bathymetry mapping.
Address gaps in edge deployment (e.g., limited battery power for UAVs, need for offline model inference) to connect computational efficiency to practical operational constraints.

Review Reports

Comments and Suggestions for Authors

Comments on the Quality of English Language

Comment 1: Grammatical and Syntactical Inconsistencies

Comment 2: Inconsistent Nomenclature and Abbreviation Usage

Comment 3: Incomplete Quantitative Benchmark Context in Table 2

Comment 4: Underdeveloped Discussion of Inland Water SDB Data Sources

Comment 5: Weak Linkage Between Computational Efficiency and Real-World Deployment

Comments on the Quality of English Language