Application of an Encoder–Decoder Model with Attention Mechanism for Trajectory Prediction Based on AIS Data: Case Studies from the Yangtze River of China and the Eastern Coast of the U.S

Round 1

Reviewer 1 Report

Beyond the existing Seq2Seq algorithm, the proposed attention model showed better performance. It has shown that these attention based algorithms outperforms,compared to the LSTM series. 

This paper can be accepted because the attention model has been properly to predict the trajectories. For further study, the algorithm is needed to be applied for various  ship dynamic patterns

The minor edit  is required for publication. 

Author Response

Thank you very much for your consideration and suggestion on our manuscript. We will applied our method into more cases of ship dynamic patterns in future work. We also requested English editing service to polish the language of our manuscript, and attached the certification in revised version. 

Please see the attachments.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper tackles the task of trajectory prediction based on AIS data by exploiting state of the art deep learning models.

I find the task very interesting and I think the method is sound. The novelty of the proposed method is moderate as the developed model is not properly innovative in the field of deep learning but its assemblance and application for the specific objective seems to be original for the field.

I appreciated the ablation study and the explanations on the results of the attention mechanism. 

The main flaw of this study is in its writing. The language used is very confusing and makes hard for the reader to fully follow the discussion. There are language mistakes and inaccuracies in almost every sentence, starting from the abstract. For example, the popular vanishing gradient issue, commonly experienced in deep learning models, has been reported as "gradient disappearance". I recommend a full editing of the manuscript by English language experts. 

I also experienced difficulties in understanding the feature fusion paragraph. I think this technique should be better explained.

The proposed method has been compared with two baseline methods, BPNN and LSTM, but those ones have not been described in the text. I recommend adding a description on their architecture and/or parameters.

I also suggest to better explain which are the predicted values of the proposed model - we can see that they are ship’s navigation longitude and latitude, but are they the ones just one time step ahead of the previous N steps? If this is the case, I can suppose the output is given as input to the proposed model exploiting the recurrent structure, in order to predict a trajectory made of several points ahead in time. That info should be clarified in the text, in the method section or in 4.2.1 subsection.

The related works can be toughened up by adding additional state of the art methods which are working on similar tasks; I recommend adding:

- DANAE: A denoising autoencoder for underwater attitude estimation, Russo et al.

- Deep Learning Methods for Vessel Trajectory Prediction Based on Recurrent Neural Networks, Capobianco et al.

In particular the latest paper has available associated code and proposes a similar network for the same task; for this reason, I recommend to add it to the list of compared methods and to perform some experiments on it.

As mentioned in the previous section, the low quality of the English heavily affects the reading and understanding of your work. I suggest to ask for an extensive english editing by an expert, in order to give your paper a better form and make it clearer for the readers. 

Author Response

Thank you very much for your time and effort on our manuscript. We replied the comments one by one, and completely revised the manuscript according to the comments. We also requested English editing service to polish the language of our manuscript, and attached the certification in revised version.

Please see the attachments.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposes an encoder-decoder model for ship trajectory prediction and collision avoidance using the sequence-to-sequence structure with a self-attention mechanism. This allows the model to capture potential information and correlations present in the AIS data series, which is used for ship navigation. The proposed model is applied to six case studies of ship collision avoidance. The results demonstrate that the mean absolute error of the proposed model is lower compared to classical models such as back-propagation neural network (BPNN) and Long Short-Term Memory (LSTM).

Overall, the paper tackles an engaging problem, presenting diverse case studies and results to evaluate machine learning methods for trajectory prediction and collision avoidance. However, before publication, there are a few key points that require attention:

1. Proofreading: It is essential to carefully proofread the entire manuscript to identify and correct typos and grammatical errors (e.g., caption of Figure 2, "Optimal paramter" in Table 3).

2. Reference Errors: Double-check the references for typos and ensure their accuracy.

3. Clarity in Figures: Improve clarity in Figures 10-11 by changing the line types between LON and LAT plots. Additionally, clarify the legend's usage of "origin" for better understanding.

4. Literature Review: Enhance the literature review in section 2.2 by including encoder-decoder models with uncertainty estimation, like the work by Capobianco et al., "Recurrent Encoder-Decoder Networks for Vessel Trajectory Prediction With Uncertainty Estimation," in IEEE Transactions on Aerospace and Electronic Systems, vol. 59, no. 3, pp. 2554-2565, June 2023.

5. Novelty of the Method: Explicitly describe the novelty of your method concerning the self-attention mechanism. Clarify how your approach differs from prior use of attention mechanisms in encoder-decoder models for trajectory prediction (e.g., reference [30]).

 

1. It is essential to carefully proofread the entire manuscript to identify and correct typos and grammatical errors.

2. Double-check the references for typos and ensure their accuracy.

 

Author Response

Thank you very much for your time and effort on our manuscript. We replied the comments one by one, and completely revised the manuscript according to the comments. We also requested English editing service to polish the language of our manuscript, and attached the certification in revised version.

Please see the attachments.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I would like to compliment with the authors for the extensive editing of their manuscript, with a great improvement of the English. I appreciated a lot the two additional methods they used for comparison and the new paragraphs which better explain some methodological concepts.

I recommend acceptance.