Deep Learning-Based Source Localization with Interference Striation of a Towed Horizontal Line Array
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript has proposed a wideband target localization algorithm using deep learning for the towed horizontal array in deep sea, which is a current concern in the field of underwater acoustic and has high scientific and application values. Due to my professional background not being in the field of deep learning, I primarily examine the following issues and provide suggestions from the perspective of traditional underwater acoustic signal processing in this paper.
- Line 46, "Direction on Arrival" should be written as "Direction of Arrival".
- In the Introduction, the current difficulties and challenges associated with utilizing interference striation for target localization have not been adequately discussed. Furthermore, the advantages of employing deep learning methods based on interference striation characteristics for target localization are not comprehensively elaborated, which results in the innovation points of this paper being insufficiently highlighted.
- 2.2.3 The analysis results of the vertical array data seem not to be closely related to the manuscript.
- Formula (25) does not explain the parameter ω. Line 293, "arrival range" seems to be incorrect.
- Line 299, Figure 16 should be written as Figure 8.
- This manuscript demonstrates the effectiveness of the proposed algorithm for underwater target localization using simulated data. Comparative analysis reveals that the estimation accuracy of the proposed algorithm surpasses that of various existing deep learning algorithms. To further validate its practical significance, it is recommended to conduct additional simulation experiments or refer to relevant literature in order to illustrate the advantages of the proposed algorithm over traditional sonar signal processing methods in terms of target localization performance.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsManuscript Review: "Deep Learning Based Source Localization with Interference Striation of a Towed Horizontal Line Array"
This paper proposes a novel deep learning model, MoELocNet, for underwater acoustic source localization using a Towed Horizontal Line Array. The model integrates a mixture-of-experts (MoE) structure with a multi-task learning approach, enabling simultaneous estimation of range and depth from a single snapshot. It is trained and validated using deep-sea environmental data generated by the Bellhop simulator, and is particularly noteworthy for achieving high performance in the acoustic shadow zone (MAE: 0.029 km for range, 0.072 m for depth). However, the paper presents the following some issues.
Minors.
- The most significant limitation is that the experiments were conducted solely under overly idealized (theoretical) conditions. The proposed model was trained and evaluated based on a static and simplified environment simulated with Bellhop, assuming a flat seafloor, fixed sound speed profile, and a perfectly aligned array. As a result, common real-world conditions in the marine environment are not reflected at all. These include: (1) array deformation or tilting due to towing, (2) complex acoustic phenomena such as seafloor undulation and internal waves, and (3) actual marine noise and various SNR conditions. Although the authors verify theoretical patterns using some real measurement data from the Indian Ocean, these data were not used directly for model training or evaluation. Thus, the practical applicability of the model is highly limited. To address this, at minimum, the paper should include a discussion on possible performance degradation in real environments or reflect on the model’s generalizability.
- MoELocNet is designed such that different experts learn to capture diverse interference patterns through its MoE structure, but the paper does not provide sufficient interpretation of which expert is selected for which input or how the model responds to ambiguous or noisy data. In current domains such as maritime defense, search and rescue, where high reliability is required, interpretability of deep learning models is becoming increasingly important. Therefore, expert selection visualization (e.g., expert activation maps) and sensitivity analysis on prediction changes under input variation are needed.
- Although the paper shows superior performance compared to models like ResNet18, Swin Transformer, and ViT, it lacks fundamental analysis on how much the MoE structure itself contributes to the improvement. If the study included comparisons with simpler CNN or LSTM-based models, quantitative comparisons with traditional acoustic methods like Matched Field Processing (MFP), and ablation studies to analyze the contribution of each key component (e.g., MoE, multi-task loss, expert router), the contribution of the proposed structure could have been more clearly demonstrated. As models become more complex, the role of each internal component must be clearly examined, and without such experiments, the claimed superiority relies only on empirical results.
4. The paper emphasizes that the model shows excellent performance in the shadow zone, but its accuracy appears to drop significantly in other acoustic zones such as the convergence zone and the direct arrival zone. However, the paper does not present any analysis or compensatory strategy regarding this performance drop. In real applications, a single system must maintain consistent performance under various oceanic conditions. This may require zone-specific transfer learning, data augmentation that accounts for interference pattern distortion, or a hybrid structure combining physical and deep learning models. Clearly stating the practical operating range and limitations of MoELocNet would strengthen the practical utility and research direction of this work.
Comments on the Quality of English LanguageIt might be helpful to consult a professional English editing company that specializes in academic or scientific manuscripts.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript has been greatly improved. I have no further comments.