A Deep Learning-Based Trajectory and Collision Prediction Framework for Safe Urban Air Mobility

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study proposes a trajectory and collision prediction framework for urban air mobility (UAM) based on an LSTM–Attention deep learning model. Using simulated data, the model achieves high-precision 10-second trajectory predictions and multi-level collision risk assessments, aiming to enhance safety and real-time responsiveness in future low-altitude urban transportation systems.

1. Reliance on Simulated Data Without Real-World Validation. The current model is entirely based on synthetic data generated by Korean Air’s Virtual Trajectory Generator (VTG), with no incorporation of real-world flight records. It is recommended to include empirical trajectories or conduct cross-validation to improve the model's practical applicability.

2. Limited Evaluation of Model Generalizability. Although tested on 23 unseen flight paths, all data originates from the same simulation system and geographic region (Seoul). To better assess robustness and generalizability, the dataset should be expanded to include diverse environments, such as international cities.

3. Narrow Scope of Baseline Comparisons. The study compares only seven deep learning models (e.g., GRU, LSTM). It is recommended to include more advanced architectures like Transformers or Graph Neural Networks to more convincingly highlight the advantages of the proposed method.

4. Collision Risk Classification Lacks Empirical Justification. While the four-level risk system (Collision, Warning, Caution, Safe) is logical, the thresholds used (e.g., 15 m/7.5 m) are not sufficiently justified by industry standards or dynamic aircraft characteristics. Sources or sensitivity analyses should be provided.

5. No Integration of Collision Avoidance Mechanisms. The framework detects risk based solely on TCA (Time-to-Closest Approach) but does not suggest any evasive strategies. Incorporating an automated avoidance module or proposing emergency action plans would improve the system’s practical utility.

6. Over-Reliance on RMSE as a Performance Metric. RMSE is the only main evaluation metric used. It is advisable to add complementary indicators for trajectory prediction (e.g., MAE, DTW) and collision classification (e.g., F1-score, precision-recall curves) to provide a more comprehensive evaluation.

7. Real-Time Performance Testing Lacks Deployment Context.The latency evaluation is conducted solely on an NVIDIA RTX 6000 GPU, which does not reflect actual edge computing conditions. Testing on embedded platforms like Jetson or onboard processors would provide more realistic insight.

8. Limited Model Interpretability and Attention Visualization. Although the model includes an attention mechanism, the paper does not visualize or analyze which time steps are most influential. Adding attention heatmaps or interpretability studies would enhance transparency and user trust.

9. No Clear Integration Path with Existing Air Traffic Systems. The deployment of UAM systems must interface with ground-based ATC or autonomous traffic management. The framework should include architectural design or communication interface suggestions for system integration.

10. Excessive Length and Redundant Descriptions. Some sections, especially those explaining LSTM and Attention mechanisms, are overly detailed. Streamlining these parts and focusing more on the novel contributions and practical implications would strengthen the paper.

This paper addresses a timely and important topic with a well-structured framework and solid experimental foundation. However, to reach publishable standards, key aspects such as data authenticity, generalization testing, real-world deployment feasibility, and interpretability need further reinforcement. Enhancing the study with empirical data, integrated avoidance strategies, and broader evaluation metrics is strongly recommended.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

While the authors acknowledge that they do not consider critical environment variables such as wind speed and turbulence, they also do not consider issues such as obstacles.

This must also be acknowledged.

Figure 1 appears to be an AI generated figure. Has this been acknowledged in the manuscript? There is also a citation in the figure caption. Have the authors taken the figure from reference number 4?

Is figure 7 mentioned/referenced in-text? Please double check

I believe some other figures and tables are also not mentioned in-text. Please check for all tables and figures.

Critical statements in-text need references.

For example, the paragraph on wind and turbulence modeling needs references to studies such as

Phadke, A.; Medrano, F.A.; Chu, T.; Sekharan, C.N.; Starek, M.J. Modeling Wind and Obstacle Disturbances for Effective Performance Observations and Analysis of Resilience in UAV Swarms. Aerospace 2024, 11, 237. https://doi.org/10.3390/aerospace11030237

and

Neto, Euclides C. Pinto, et al. "Trajectory-based urban air mobility (UAM) operations simulator (TUS)." arXiv preprint arXiv:1908.08651 (2019).

Finally the study appears to address trajectory prediction using DL for a single vehicle. How would the framework perform in situations where multiple vehicles have such prediction implementations running on-board?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Modifications have been made but the following issues still exist:

1. The manuscript is lengthy, and certain sections contain overlapping content. Consolidation and tighter editing of repeated descriptions would improve clarity and readability.
2. Although the system supports multi-vehicle prediction and collision assessment, no dynamic multi-agent scenario is presented. Including a case study showing prediction and risk escalation over time in a dense operational setting would strengthen the experimental evidence.
3. The model is currently validated entirely on synthetic data generated by the VTG simulator. While high-fidelity, it lacks real-world verification. Incorporating real UAM or helicopter flight trajectories, or performing cross-validation with external datasets, would enhance the model’s practical applicability and robustness. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I have some minor comments.

The abstract is at 92 words and should be expanded to the common convention of 200-25o words. More details can be included here that better summarize the study.

Figure 2 bottom section has small text and a lot of space around the text. Why not increase size of text to improve readability and visibility?

Same issue for Figure 3. Even at 150% zoom, the text is barely visible. Figure 8 can be similarly improved as well.

I have no further comments. The authors have addressed all my previous concerns and the paper can be accepted.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

The modifications have been completed as required and can be accepted after completing the modifications according to the editorial department's requirements.