Toward Intelligent Underwater Acoustic Systems: Systematic Insights into Channel Estimation and Modulation Methods

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper provides a comprehensive and well-structured systematic literature review on the application of machine learning and deep learning techniques in underwater acoustic communication systems. The research is worth of publication based on minor comments and Suggestions for Improvement:
1. Figure 3 shows the UWA Transceiver. Kindly elaborate in more detail.
2.. Some acronyms e.g., PINNs in Section 6.2 could be spelled out upon first use for clarity.
2. Ensure consistent use of terms like "bit error rate (BER)" vs. "Bit Error Rate (BER)" throughout the paper.
3. In Table 18, "Average Accuracy" and "Average Precision" are introduced without prior explanation in the text. 
4.  Some sentences are lengthy and could be split for readability (e.g., "The adaptive modulation process starts at the transmitter, where the channel encoder adds extra data to protect against errors. The data is then modulated using techniques like BPSK, QPSK, or OFDM [22, 46, 47, 48].").

5. Section 6.2 could be expanded to include a brief discussion on the potential role of federated learning or edge AI in addressing computational constraints for real-time UWA systems.

Author Response

The file is attached. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

(1)In line of 2 on page 1, the word "such as" is repetitive.

(2)Compared with the existing work, what is the key contributions? It should be also listed in the table. or maybe conclude in the following paragraph.

(3)In table 2, the data is how to obtain?

(4)In line 1212, "the section ??", there is a problem.

(5)It is suggested to use some math models for demonstrating the discussed problem.

Author Response

The file is attached. 

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The abstract should go beyond outlining "what was studied" and should explicitly distill core "findings": a comparison of the advantages and disadvantages of current mainstream models across various tasks, the most significant performance bottlenecks, and the most critical research gaps. It is important to clearly communicate the unique value proposition distinguishing this survey from existing works (e.g., whether it represents the first systematic analysis of the transfer potential of ML models across physical layer modules? Does it quantify the gap between simulation and real-world measurement performance?).
2. When elaborating on the inherent challenges of UWA channels (multipath, Doppler, time-varying fading, bandwidth limitation), it is suggested to more closely and concretely link them to the core pain points of physical layer signal processing modules. For example, explaining how strong multipath causes traditional synchronization to fail, thereby driving the need for ML-based channel estimation; explaining how large dynamic Doppler forces adaptive modulation strategies to rely on environmental awareness and real-time decision-making. Avoiding disconnecting technical descriptions from the actual scenarios is recommended.
3. The existing summary of gaps (e.g., real-time performance, energy consumption, model generalizability) is reasonable but could benefit from being further refined with application-specific contexts. For instance: indicating the shortcomings of real-time Doppler compensation models for high-speed mobile platforms (AUV)? Is research lacking on the feasibility of deploying specific models (e.g., DRL) in low-power acoustic modems? In the context of the bandwidth-distance-rate trade-off, is research relatively weak on intelligent resource allocation strategies? Refining the gaps could be more effective in guiding future research directions.
4. When discussing each technique (e.g., CNN/LSTM for channel estimation/modulation recognition, RL for adaptive modulation), it is important to clearly articulate the specific UWA physical layer problem it solves. For example: CNNs in channel estimation primarily target extracting spatial features from multipath structures; LSTMs are recognized for being more adept at capturing the time-varying correlation of channels (addressing time-variability/long-term memory issues), making them suitable for scenarios with severe Doppler spread. The core value of RL (e.g., DQN, A2C) in adaptive modulation is considered to lie in making sequential decisions in complex, unknown, non-stationary channels to approximate optimal strategies.
5. It is important to clearly point out that the performance metrics (BER, SER, Accuracy, MSE, etc.) and evaluation conditions (SNR range, channel model, dataset) used in current research are highly fragmented and lack standardization. It would be beneficial for the survey to advocate for or promote the establishment of standardized virtual or real-measurement evaluation platforms, proposing key recommended metrics and evaluation baselines (classical non-AI algorithms or representative shallow ML models). Integrating scattered performance data (such as Table 18 and similar information) is suggested. Consolidating results for the same type of task (e.g., modulation recognition accuracy of different CNN structures on the same dataset and within the same SNR range) into statistical distribution plots (box plots/violin plots) or radar charts (comprehensively considering dimensions like BER/complexity/latency/generalizability) is recommended. Such visual comparisons can more intuitively reveal trends and anomalies, significantly enhancing the credibility of conclusions and providing deeper insights. Stating only "most models achieve BER of 10^-2" is likely insufficient.
6. Existing studies/surveys often report "highest accuracy/lowest BER values," but critically, there is insufficient assessment of model performance under parameter perturbations, unknown channel conditions, low SNR, and cross-platform testing (acoustic modems in different frequency bands). Emphasis should be placed on the necessity for subsequent research to incorporate robustness testing (e.g., sensitivity to delay/Doppler spread variation?) and on discussing model generalization bottlenecks and potential solutions (e.g., Transfer Learning TL, Domain Adaptation DA).
7. When discussing algorithm limitations (e.g., high computational complexity, strong training data dependency, sensitivity to specific assumptions), these need to be closely linked with the stringent constraints of specific UWA applications. It is crucial to emphasize how these limitations violate the low-power requirements of UWA nodes (especially battery-constrained AUVs/sensors) and hinder real-time physical layer processing (e.g., low-latency channel estimation assisting coherent demodulation). Stating clearly that this leads to drastic performance degradation in real-world, complex, and mutable marine environments (sudden interference, biological noise) (Sim2Real Gap), forming a key obstacle to practical deployment, is important. For embedded platforms, exploring model compression (pruning, quantization, knowledge distillation) is valuable; for low-latency processing, exploring feature engineering optimization, computational graph optimization, hardware acceleration (FPGA/dedicated AI accelerator cores) is suggested. Including model selection guidelines considering the complexity-performance trade-off and expected effectiveness evaluation is recommended.
8. Suggestions regarding the description of specific challenges: When discussing the lack of dynamic adaptability of models under fast time-varying channels, exploring the potential of real-time update techniques such as online learning, incremental learning, and meta-learning could be valuable. Delving deeper into the absence of research on intelligent resource allocation under the bandwidth-distance trade-off and exploring possible directions for RL/game theory in dynamic spectrum allocation, power control, and joint optimization with modulation/coding is beneficial.
9. When summarizing results from different studies, it is preferable to avoid merely listing "effective" or "improved performance." Attention should be paid to pointing out the premises, potential flaws, contradictory results, or reproducibility issues. For example, when different literatures arrive at divergent conclusions for "similar" problems, analyzing whether this stems from differences in test environments or model implementation details could be immensely valuable.
10. It is suggested to add charts displaying model performance distribution (box plots/violin plots) and comprehensive performance/cost comparisons (radar charts). Ensuring all charts have clear, informative titles and legends, with complete axis labeling, is critical. The aim should be that core conclusions are intuitively graspable from the charts, avoiding treating them merely as textual appendages. Simultaneously, pointing out the limitations of evaluation data is important: clearly stating the problem of current research over-relying on simulation data or limited/specific scenario (shallow water, calm sea) real-world data is necessary. Emphasizing the lack of systematic validation in typical deep-sea, high-dynamic, strong-interference environments is also crucial.
11. Restating the main text should be avoided. Clarifying the following is recommended: In which tasks do ML/DL already show significant advantages (e.g., modulation recognition at high SNR)? Which areas still face enormous challenges (e.g., real-time robust adaptive modulation)? What are the core bottlenecks and performance ceilings of mainstream technologies (e.g., high computational load of complex models vs. energy constraints; the gap between simulation performance vs. measured performance)?
12. Unifying terminology is essential: Ensuring consistent use of terms such as: "BER," "ML," "DL" (acronyms acceptable after first full spelling), "Channel Estimation," "Adaptive Modulation," "Modulation Recognition." Carefully distinguishing between "Simulation," "Emulation," "Experiment," "Field test/Measurement" is important.
13. Performing a self-check to ensure citation precision and uniformity is vital: Verifying that all literature citations are accurate and error-free; ensuring citation format is consistent throughout (e.g., IEEE) is necessary. Proofreading the entire text to correct grammatical errors, ambiguities, and redundant expressions, ensuring smooth writing, clear logic, and professional rigor, is strongly recommended.

Author Response

The file is attached. 

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The paper presents a systematic literature review in underwater acoustic communication, focusing on channel estimation, adaptive modulation, and modulation recognition. While the topic is timely and important, especially given recent AI advancements, the manuscript requires some improvements before being suitable for publication.

1. The review falls short in scientific depth. The paper present an extensive collection of metrics such as BER and MSE across different studies, but offer little critical analysis of the trade-offs between techniques, deployment constraints, or comparative strengths. The tables are comprehensive but lack accompanying interpretation that would help readers understand why certain models or approaches perform better under specific UWA conditions. In effect, the paper catalogs research rather than evaluates it critically.
2. The manuscript is overly long, with significant redundancy. Several sections contain boilerplate or textbook-style content that could be condensed or omitted. For example, definitions of terms like BER, MAE, or explanations of general ML concepts do not add value for the intended audience. 
3. The quality of the English writing needs improvement. 
4. The literature coverage, although broad, may miss key works from top-tier conferences and journals such as IEEE JSAC, TCOM, or OCEANS. The paper lacks justification for inclusion and exclusion criteria beyond database names and time ranges.
5. To improve the paper, I recommend the authors significantly condense the manuscript by eliminating redundancies and reducing the emphasis on basic definitions. A concise discussion of current limitations and future research directions in real-world UWA system deployment would be a meaningful addition.

Comments on the Quality of English Language

This paper requires revision to improve clarity, conciseness, and grammatical accuracy. Many sentences are verbose or awkwardly phrased, and terminology is sometimes used inconsistently (e.g., interchanging “ML/DL,” “AI-based,” and “model-driven” without clear distinction). Redundant expressions and vague language (e.g., "enhances reliability" or "improves performance") appear frequently, which weakens the technical rigor. A thorough language edit by a native or professional academic editor is strongly recommended to bring the writing up to publication standards.

Author Response

The file is attached. 

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

no further comment 

Reviewer 4 Report

Comments and Suggestions for Authors

Thank you for the detailed explanation. I am satisfied with the response.