Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript proposes a multi-module fusion framework, DDA-SIM-ATT-CatBoost. This model is validated using real-world operational data from Shenzhen Bao'an International Airport. The overall approach offers some engineering insights, and experimental results indicate a hit rate of 74.57% within a ±120 s interval. However, there are still issues with the current manuscript:

1) There is an obvious error in the chapter numbering. The next main chapter title after "1.Introduction" is "1. Influencing Factors of Departure Flight Taxi-Out Time and Their Correlation Analysis", please revise.

3) In section 1.2, the definitions of departure queue x₃ and arrival queue x₄ rely on the takeoff and landing information of other flights within the taxiing interval [AOBT(i), ATOT(i)]. In real-world prediction scenarios, the future ATOT may not be not known at the AOBT moment, so the future information within this window is not available at the prediction. The author needs to clarify the prediction horizon and demonstrate that all input features are available at that time point.

4) It is necessary to clarify whether data augmentation occurs before or after the data training/testing split, and to ensure that there is no cross-set mixing of augmented samples with original samples.

5) It is necessary to explain how to ensure that the augmented samples conform to operational logic. The authors may provide consistency checks on the distributions before and after augmentation to reduce the risk of label contamination.

6) Table 4 shows that on the ±180s metric, CatBoost+DDA+SIM (89.77%) outperforms the complete model CatBoost+DDA+SIM+ATT (89.12%). The performance actually decreases in this interval after adding ATT. The possible reasons for this need analysis and explain.

7) The best model in Table 5 includes the full set of enhancements of DDA+SIM+ATT, while other models do not use any corresponding enhancement strategies. This comparison cannot answer whether the improvement comes from CatBoost itself or from the enhancement framework. The parameters such as the number of training rounds and early stopping strategy of the model also need to be explained.

8) There are many format errors that need to be revised. For instance, in line 413 of the original text, Figure 7 is referenced, but the text is describing the process of the attention mechanism. The actual corresponding figure should be Figure 8. Additionally, some figure names use lowercase “figure 5”, “figure 6”, and “figure 7”. however, line 368 use uppercase “Figure 6”, and line 413 uses "Figure 7". It is recommended to use consistent capitalization.

9) The discussion on the application value of the project is overly general.

10) Since this study uses data augmentation approaches, several studies may be helpful as they also emphasize data augmentation in proposing prediction model for aviation operation/event. The authors may consider these papers and cite them in the current manuscript.

https://doi.org/10.1016/j.ress.2025.111524

https://doi.org/10.1016/j.ress.2025.111921

https://link.springer.com/article/10.1007/s10115-024-02168-6?error=server_error

Author Response

Reviewer1:

Comments 1: [There is an obvious error in the chapter numbering. The next main chapter title after "1.Introduction" is "1. Influencing Factors of Departure Flight Taxi-Out Time and Their Correlation Analysis", please revise.]

Response 1: [Thank you for identifying the chapter numbering error. We have revised the title from "1. Influencing Factors of Departure Flight Taxi-Out Time and Their Correlation Analysis" to "2. Influencing Factors of Departure Flight Taxi-Out Time and Their Correlation Analysis" to maintain correct sequential numbering after the Introduction. The changes are marked in red in the manuscript.]

 

Comments 3: [In section 2.2, the definitions of departure queue x₃ and arrival queue x₄ rely on the takeoff and landing information of other flights within the taxiing interval [AOBT(i), ATOT(i)]. In real-world prediction scenarios, the future ATOT may not be not known at the AOBT moment, so the future information within this window is not available at the prediction. The author needs to clarify the prediction horizon and demonstrate that all input features are available at that time point.]

Response 3: [Thank you for your insightful comment regarding the definitions of departure queue x3 and arrival queue x4 in Section 2.2. You correctly point out that in a real-world prediction setting, the actual take-off time (ATOT) is unknown at the actual off-block time (AOBT); consequently, information within the taxiing interval [AOBT(i),ATOT(i)] would not be available at the prediction moment.We acknowledge this limitation and would like to provide a detailed clarification:

（1）Research scope: This study is a proof-of-concept aimed at validating the feasibility and effectiveness of the proposed DDA-SIM-ATT-CatBoost model using historical data. In this retrospective analysis, we assume that all features—including those derived from the taxiing interval—are known at the prediction time. This assumption is acceptable for initial methodological research to demonstrate the model's potential under ideal conditions.

（2）Practical deployment: In real-world applications, once the model is trained on historical relationships between features and taxi-out time, predictions can be made using only the information available at the prediction moment (e.g., AOBT, current queue lengths, weather conditions, etc.). The features defined in Section 2.2 are used during the training phase to capture the underlying patterns; after training, the model does not require future ATOT or any information from the taxiing interval. Therefore, the proposed method can be directly applied in operational settings, provided that the necessary real-time data (which are typically available at AOBT) can be obtained.]

 

Comments 4: [It is necessary to clarify whether data augmentation occurs before or after the data training/testing split, and to ensure that there is no cross-set mixing of augmented samples with original samples.]

Response 4: [Thank you for your valuable comment. We agree that it is essential to clarify whether data augmentation is performed before or after the data splitting, and to ensure that augmented samples are not mixed with original samples across the training and test sets.

In the revised manuscript, we have added a clear explanation in the methodology section (see Section 3.1 highlighted in red). Specifically, we now state that data augmentation is conducted after data preprocessing but before the training/testing split, as illustrated in Figure 5, which presents the overall flowchart of the departing flight taxi-out time prediction model. This ensures that augmented samples are generated exclusively from the original training data and are not incorporated into the test set. The augmentation process introduces reasonable perturbations based on various influencing factors, thereby expanding the training set while preserving the integrity and independence of the test set for unbiased evaluation. This setup prevents any potential data leakage and ensures that the test set remains representative of real-world conditions.]

 

Comments 5: [It is necessary to explain how to ensure that the augmented samples conform to operational logic. The authors may provide consistency checks on the distributions before and after augmentation to reduce the risk of label contamination.]

Response 5: [Thank you for your insightful comment. We fully agree that it is important to explain how the augmented samples are ensured to conform to operational logic. In the revised manuscript, we have clarified this point in the methodology section. Specifically, as described in Section 3.1 (second paragraph) and Section 4.1(2), the proposed Dynamic Data Augmentation (DDA) method operates within well-defined time window constraints. Features are categorized into numerical, categorical, and temporal types, and reasonable perturbations are applied to each type while respecting the underlying physical constraints and temporal dependencies inherent in taxi-out operations. This ensures that the augmented samples remain realistic and adhere to the same operational logic as the original data.]

 

Comments 6: [Table 4 shows that on the ±180s metric, CatBoost+DDA+SIM (89.77%) outperforms the complete model CatBoost+DDA+SIM+ATT (89.12%). The performance actually decreases in this interval after adding ATT. The possible reasons for this need analysis and explain.]

Response 6: [Thank you for your valuable comment. In response, we have added a detailed analysis and explanation in Section 4.3 (line 577) to address the possible reasons why CatBoost+DDA+SIM (89.77%) outperforms CatBoost+DDA+SIM+ATT (89.12%) on the ±180s metric. The revisions are highlighted in red in the manuscript. We appreciate your guidance.]

 

Comments 7: [The best model in Table 5 includes the full set of enhancements of DDA+SIM+ATT, while other models do not use any corresponding enhancement strategies. This comparison cannot answer whether the improvement comes from CatBoost itself or from the enhancement framework. The parameters such as the number of training rounds and early stopping strategy of the model also need to be explained.]

Response 7: [Thank you for your valuable comment. Table 6 is a comparative experiment designed to identify the most advantageous base model using preprocessed original data without data augmentation. The first six rows of Table 6 present results obtained from the original preprocessed data, from which the superiority of CatBoost can be clearly observed. In contrast, Table 5 presents an ablation study that progressively adds or removes modules (DDA, SIM, ATT) to evaluate the contribution of each component to the overall prediction performance. These two tables serve different purposes and together provide a comprehensive evaluation of both the base model and the enhancement framework.

Regarding model parameters, the detailed settings—including the number of training rounds and early stopping strategy—are described in Section 4.1 (4) and summarized in Table 3. The relevant parts have been highlighted in red in the revised manuscript.]

 

Comments 8: [There are many format errors that need to be revised. For instance, in line 413 of the original text, Figure 7 is referenced, but the text is describing the process of the attention mechanism. The actual corresponding figure should be Figure 8. Additionally, some figure names use lowercase “figure 5”, “figure 6”, and “figure 7”. however, line 368 use uppercase “Figure 6”, and line 413 uses "Figure 7". It is recommended to use consistent capitalization.]

Response 8: [Thank you for pointing out the citation error. We have corrected the reference in line 413: Figure 7 has been replaced with Figure 8, as it correctly corresponds to the description of the attention mechanism. The correction is highlighted in red in the revised manuscript. Additionally, we have standardized the capitalization of all figure and table references (e.g., "Figure" and "Table") throughout the manuscript, and the changes are also marked in red. We appreciate your careful review.]

 

Comments 9: [The discussion on the application value of the project is overly general.]

Response 9: [Thank you for your valuable comment regarding the discussion on application value. We agree that the original description was overly general. In response, we have substantially strengthened the Conclusion section (Section 5.2) to provide a more concrete and practice-oriented discussion of the model's application value. Specifically, we now explicitly highlight that the proposed model can be integrated into airport collaborative decision-making (A-CDM) systems to support ground controllers in taxi-out time prediction, thereby enabling optimized taxi routing, push-back scheduling, and conflict prediction under dynamic operational conditions. We also elaborate on its potential to enhance runway sequencing and surface traffic management in real-world airport environments. These revisions are highlighted in red in Section 5.2 of the manuscript. We appreciate your guidance in improving the practical relevance of our work.]

 

Comments 10: [Since this study uses data augmentation approaches, several studies may be helpful as they also emphasize data augmentation in proposing prediction model for aviation operation/event. The authors may consider these papers and cite them in the current manuscript.]

Response 10: [Thank you for your insightful suggestion regarding relevant literature on data augmentation in aviation research. We agree that incorporating appropriate references helps to better contextualize our work within the field. In response, we have reviewed and added several recent studies that also employ data augmentation techniques in aviation operation and event prediction. These references have been incorporated in 1.Introduction and References of the revised manuscript and are highlighted in red for clarity. We appreciate your guidance in enriching the literature foundation of our study.]

 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

1. The abstract is well written. All the abbreviations need to be mentioned on their first appearance.
2. The research gap is not clearly articulated in the manuscript.
3. Section numbers need to be properly assigned.
4. Figure quality needs to be enhanced for better visibility.
5. The entire research methodology framework needs to be illustrated as a single unit for better understanding and ease.
6. The features are selected on what basis? What is the criterion?
7. How was the optimal number of features fixed?
8. Table 5 reveals the enhanced performance of the proposed methodology. What is the reason behind the enhanced performance of the proposed approach? Literature support is needed.
9. Results of model hyperparameter tuning need to be reported.
10. A statistical validation of th results is needed.
11. Ensure the non-occurrence of overfitting of the model.
12. Enahnce the conclusion with key findings.

Author Response

Reviewer 2:

Comments 1: [The abstract is well written. All the abbreviations need to be mentioned on their first appearance.]

Response 1: [Thank you for your positive feedback. We have revised the manuscript accordingly: all abbreviations are now defined at their first appearance, and the changes have been highlighted in red for your reference.]

 

Comments 2: [The research gap is not clearly articulated in the manuscript.]

Response 2: [Thank you for your insightful comment. We have revised the Introduction section to more clearly articulate the research gap. The changes are highlighted in red in the revised manuscript for your reference.]

 

Comments 3: [Section numbers need to be properly assigned.]

Response 3: [Thank you for your comment. We agree that the section numbers need to be properly assigned. Following your suggestion, we have carefully reviewed and corrected the chapter numbering throughout the manuscript to ensure logical consistency and proper sequence. The revisions have been highlighted in red in the revised manuscript for your easy reference.]

 

Comments 4: [Figure quality needs to be enhanced for better visibility.]

Response 4: [Thank you for your suggestion. We have improved the quality of all figures in the manuscript to ensure better visibility and clarity. All figures have been updated with higher resolution, improved labeling, and sharper details. The revised figures are highlighted in red in the manuscript for your reference.]

 

Comments 5: [The entire research methodology framework needs to be illustrated as a single unit for better understanding and ease.]

Response 5: [Thank you for your valuable suggestion. In the manuscript, Section 3: Construction of the Departure Flight Taxi-Out Time Prediction Model serves as the overall framework of the research methodology. Specifically, Figure 5 illustrates the complete process and main steps. Subsequently, Subsections 3.1–3.4 provide detailed explanations of each module—including data preprocessing, feature engineering, model construction, and evaluation methods—along with their corresponding flowcharts in Figures 6–8. The specific parameter settings for each module are presented in Section 4: Experimental Results and Analysis. The revised text is highlighted in red in the manuscript.]

 

Comments 6: [The features are selected on what basis? What is the criterion?]

Comments 7: [How was the optimal number of features fixed?]

Response 6 & 7: [Thank you for raising these important questions regarding the feature selection process and the determination of the optimal number of features. We fully agree that clarifying the basis and criteria for feature selection is essential for ensuring the transparency and reproducibility of our study. In response, we provide a detailed explanation below.

In this study, feature selection was conducted through a two-stage approach that combines statistical filtering with empirical evaluation using the proposed DDA-SIM-ATT-CatBoost model.Stage 1: Correlation Analysis：After data preprocessing, we first performed a Pearson correlation analysis to examine the linear relationship between each candidate feature and the target variable (taxi-out time). The correlation coefficients, as visualized in Figure 4, provided an initial indication of feature relevance. Features with extremely low correlation were considered potentially redundant or noisy. However, correlation analysis alone is insufficient for feature selection in complex nonlinear prediction tasks, as it does not capture multivariate interactions or nonlinear dependencies. Therefore, we proceeded to a model-based evaluation.Stage 2: Model-Based Feature Selection Experiment：Using the proposed DDA-SIM-ATT-CatBoost model as the evaluation platform, we conducted a systematic feature selection experiment. Starting from the full set of 13 candidate features (denoted as x₁——x₁₃, sorted in descending order of their correlation with taxi-out time), we sequentially removed the least correlated features and evaluated model performance at each step. The performance was assessed using multiple metrics to ensure a comprehensive evaluation: prediction accuracy within ±120 s, ±180 s, and ±300 s, along with MAPE (%), MAE (s), and RMSE (s). The detailed results of this selection procedure are presented in Table 4.Determining the Optimal Feature Set：Our experimental results revealed that removing the two features with the weakest correlation led to the best overall predictive performance across all evaluation metrics. This finding indicates that eliminating low-relevance features reduces noise and mitigates overfitting, thereby enhancing the model's generalization capability. Further removal of additional features resulted in performance degradation, suggesting that the remaining features collectively capture the essential information required for accurate taxi-out time prediction. Consequently, the optimal feature set was determined as the one that excludes only the two least relevant features.

It is worth noting that this feature selection approach is consistent with best practices in predictive modeling, where a combination of statistical screening and empirical validation is employed to balance model simplicity and predictive power. The selected feature set not only improves model performance but also enhances interpretability by retaining only the most informative variables.The relevant descriptions and results have been highlighted in red in the revised manuscript for your convenience.]

 

Comments 8: [Table 5 reveals the enhanced performance of the proposed methodology. What is the reason behind the enhanced performance of the proposed approach? Literature support is needed.]

Response 8: [Thank you for your insightful comment regarding the enhanced performance shown in Table 6. The improvement achieved by our proposed approach can be attributed to the synergistic integration of the three core modules—DDA, SIM, and ATT. Detailed definitions, purposes, and underlying principles of these modules are provided in Sections 3.1–3.4 of the revised manuscript, with the relevant descriptions highlighted in red for your convenience.In addition, literature support for the design of these modules has been incorporated in the 1.Introduction, where key references related to data augmentation, similarity modeling, and attention mechanisms are cited. These references are also highlighted in red. We believe this combination of methodological innovation and established literature collectively underpins the enhanced performance observed.]

 

Comments 9: [Results of model hyperparameter tuning need to be reported.]

Response 9: [Thank you for your comment. The detailed settings for model hyperparameters, including the number of training rounds and the early stopping strategy, have been described in Section 4.1(4) and summarized in Table 3. The relevant parts have been highlighted in red in the revised manuscript for your reference.]

 

Comments 10: [A statistical validation of th results is needed.]

Response 10: [Thank you for your valuable comment regarding the statistical validation of our results. In response, we have conducted comprehensive statistical analyses to verify the significance of the performance improvements achieved by our proposed DDA-SIM-ATT-CatBoost model. Specifically, we employed independent-sample Mann–Whitney U tests to compare our model against each of the six baseline models (Random Forest, GRU, LSTM, XGBoost, CatBoost, and BP). The results show that all p-values are less than 0.001, indicating statistically significant improvements. Table 6 has been revised accordingly, with the modifications highlighted in red in the manuscript.]

 

Comments 11: [Ensure the non-occurrence of overfitting of the model.]

Response 11: [Thank you for your valuable comment regarding the need to ensure that our model does not overfit. In the proposed DDA-SIM-ATT-CatBoost model, we have implemented multiple strategies to prevent overfitting and enhance generalization. These measures are now explicitly described in the revised manuscript (highlighted in red in Section 4, Table 3). The key anti-overfitting techniques include:(1) Early Stopping: The CatBoost component employs early stopping with early_stopping_rounds = 50.

(2) Regularization: Gradient clipping with max_norm = 1.0 is applied to prevent gradient explosion.(3) Model Complexity Control: The complexity of both the attention network and the CatBoost model is dynamically adjusted based on the sample size of each cluster. Specifically, when a cluster contains fewer than 100 samples, we use iterations = 500 and depth = 4; for clusters with 100–500 samples, iterations = 800 and depth = 6; for clusters with ≥500 samples, iterations = 1000 and depth = 8. The learning rate is set to 0.05 across all settings. This adaptive strategy prevents overly complex models from being trained on small data subsets.]

 

Comments 12: [Enahnce the conclusion with key findings.]

Response 12: [Thank you for your suggestion. We have strengthened the conclusion section to highlight the key findings more clearly and concisely. The revised conclusion is presented in Section 6 and highlighted in red in the manuscript.]

 

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

MLLM–human comparison precludes definitive conclusions; findings should be regarded as hypothesis-generating for future confirmatory studies. The authors need to illustrate that their proposed methodcan improve interactivity which the authors claimed that this issue is better than previous work that they reviewed in Introduction. The proposed DDA-SIM-ATT-CatBoost is not explained for combind modules, theory, and mechanism. An additional discussion section should be included in this paper.

The contributions seem to be limited by DDA-SIM-ATT-CatBoost. The authors need to clearly specify the contributions in the background when discussing relevant work in wound care assessment and specify how their method is different from existing method. 

If you have the developing the synergistic multi-module fusion model, author need to explain the contents of model design and implementation. The authors need to recheck and reorganize the paper. Also, there is short literature reviews related to what authors want to describe.

Author Response

Reviewer 3:

Comments 1: [MLLM–human comparison precludes definitive conclusions; findings should be regarded as hypothesis-generating for future confirmatory studies. The authors need to illustrate that their proposed methodcan improve interactivity which the authors claimed that this issue is better than previous work that they reviewed in Introduction. The proposed DDA-SIM-ATT-CatBoost is not explained for combind modules, theory, and mechanism. An additional discussion section should be included in this paper.]

Response 1: [Thank you for your insightful and detailed comments. We have carefully considered each point and made corresponding revisions to the manuscript, as outlined below:

1. Regarding the MLLM–human comparison and the hypothesis-generating nature of the findings:We acknowledge that the current comparison between the model and human performance is exploratory in nature and does not support definitive conclusions. Accordingly, we have revised the Discussion section to explicitly state that these findings should be regarded as hypothesis-generating, providing a foundation for future confirmatory studies. This clarification has been added in Section 5.3 (Limitations and Future Directions) and is highlighted in red.

2.Regarding the improvement in interactivity:We agree that interactivity is a key advantage of our proposed method over prior work. In response, we have expanded the Introduction and Section 3 to articulate how the integration of dynamic data augmentation (DDA), similarity theory (SIM), and attention mechanism (ATT) enhances model interactivity—particularly in terms of adaptive feature weighting and dynamic pattern matching. These improvements enable the model to better capture real-time operational dynamics, a capability that was limited in earlier studies. Relevant revisions have been made in the Section 3.1-3.4, and are marked in red.

3.Regarding the explanation of module integration:We have strengthened the description of how the three core modules (DDA, SIM, ATT) are combined and how they interact within the overall framework. Section 3 now provides a more explicit explanation of the theoretical basis and functional roles of each module, as well as their synergistic effects. Figures 5–8 have been updated to better illustrate these relationships. The revised text is highlighted in red in Sections 3.1–3.4.]

 

Comments 2: [The contributions seem to be limited by DDA-SIM-ATT-CatBoost. The authors need to clearly specify the contributions in the background when discussing relevant work in wound care assessment and specify how their method is different from existing method.]

Response 2: [Thank you for your valuable comment regarding the contributions of our work. We agree that the contributions should be clearly articulated in relation to the problems addressed and distinguished from existing methods.As demonstrated in the manuscript, the contributions of our proposed DDA-SIM-ATT-CatBoost model are twofold. First, the ablation study (Table 5) systematically quantifies the positive contribution of each module. By progressively adding DDA, SIM, and ATT to the baseline CatBoost model, the results show that each component improves prediction accuracy, and their synergistic combination yields the best performance. This directly validates the necessity and effectiveness of each proposed module.

Second, the problem-driven model construction (Section 3) explicitly links each module to the inherent challenges in taxi-out time prediction—such as data sparsity, class imbalance, missing values, and complex feature interactions. Specifically, DDA mitigates data imbalance, SIM handles missing data via similarity theory, and ATT captures nonlinear feature dependencies. Thus, the proposed framework is not merely a new algorithm but a systematic solution tailored to the practical difficulties of the problem domain.

To enhance clarity, we have revised the manuscript to explicitly connect the ablation results (Table 5) with the problem analysis in Section 3. The relevant sentences are highlighted in red in the revised version. We believe these revisions adequately clarify that our contribution lies in both identifying key operational challenges and developing a multi-module framework that jointly addresses them.]

 

Comments 3: [If you have the developing the synergistic multi-module fusion model, author need to explain the contents of model design and implementation. The authors need to recheck and reorganize the paper. Also, there is short literature reviews related to what authors want to describe.]

Response 3: [Thank you for your comment. The design and implementation of the proposed synergistic multi-module fusion model are elaborated in detail in Section 4 (Experimental Results and Analysis), including the parameter settings and implementation methods for each module. Through feature selection, ablation experiments, and comparative experiments, we comprehensively demonstrate the rationale behind feature selection, module configuration, and baseline model comparison, fully validating the scientific significance and practical value of the proposed model for departure flight taxi-out time prediction.

Regarding the concern about insufficient references, we have supplemented the Introduction section with relevant literature for all modules used in this study. The revised parts have been highlighted in red in the manuscript.]

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The reviewer comments have been addressed, and the manuscript is recommended to be considered for acceptance.

Reviewer 2 Report

Comments and Suggestions for Authors

Congrats to the authors.