PRTNet: Combustion State Recognition Model of Municipal Solid Waste Incineration Process Based on Enhanced Res-Transformer and Multi-Scale Feature Guided Aggregation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author,

I have some comments on the research paper that I seriously require to be revised to meet the publication guidelines.

                                              Thank you 

Comments for author File: Comments.pdf

Author Response

Comments 1: The title needs to be modified and should not contain abbreviations; it should be written in full without any abbreviations, whether in MSWI or Res.

 

Response 1: We sincerely appreciate the reviewers’ valuable comment and acknowledge that the inclusion of an abbreviation in the title was an oversight; this has now been corrected. We would like to clarify that “Res-Transformer” is the complete module name and does not contain any abbreviation.

 

 

Comments 2: The main problem and objectives are unclear, so it is necessary to focus on what the problem is with the burning of solid waste, what the objectives are clear and organized so that the overall picture is clear to the reader. I recommend that the author rewriting the research methods in a clearer order and adding the most important results and conclusions in a more organized manner for greater clarity.

 

Response 2: We sincerely appreciate the reviewers’ constructive remarks. Acknowledging the necessity of articulating the principal research problems and objectives with greater precision, we have substantially revised the abstract to enable readers to apprehend the overarching scope expeditiously. We further clarify that the present investigation is confined to the deployment of deep-learning architectures for the identification of flame combustion regimes, thereby furnishing decision-support data for subsequent operational stages; no aspects of the municipal solid-waste incineration chain beyond this scope are examined. The Methodology and Conclusions sections have been comprehensively reformulated in strict accordance with the reviewers’ recommendations.

 

 

Comments 3: It should address the different types of combustion, their impact on human health and the environment, how to maintain a clean environment despite polluting emissions, and how to control emissions through appropriate treatment.

 

Response 3: We fully agree with the reviewers’ suggestions and have examined how normal combustion and the three abnormal modes affect human health and the environment. This paper focuses on flame combustion states and does not address emission treatment; however, achieving accurate identification and real-time control of different combustion states is the key to curbing pollutant formation at the source and ensuring an environmentally clean incineration process.

 

 

Comments 4: How can artificial intelligence be applied to achieve optimal flame quality, enhance combustion efficiency, and reduce carbon and dioxin emissions? Furthermore, how can it be identified as the appropriate mechanism for treating emissions effectively, and can it be directly applied in pre-combustion segregation?

 

Response 4: Artificial-intelligence technologies do not exert a direct influence on combustion-efficiency augmentation; rather, their function is to deliver high-fidelity identification of flame combustion regimes, thereby furnishing technicians with evidence-based guidance for sustaining the normal combustion state. Continuous maintenance of this state within the furnace maximally attenuates the emission of hazardous substances. Pre-combustion preprocessing stages may likewise be subjected to exploratory design of dedicated artificial-intelligence methodologies.

 

 

Comments 5: The diagram does not clearly illustrate the nature of the processing, nor does it show the application of artificial intelligence. The diagram must be modified to align with the nature of the research.

 

Response 5: We sincerely appreciate the reviewers’ invaluable comments. We wish to clarify that this study is strictly focused on flame combustion states and does not address any other stages of the treatment process. In response to the reviewers’ suggestions, we have annotated the figures to explicitly indicate the application objective of the artificial-intelligence system: to identify the current flame regime and thereby guide operators in making decisions that maintain normal combustion.

 

Comments 6: The author needs to clarify this paragraph, including the advantages of applying artificial intelligence software, explaining the pros and cons of the mechanism used. Does it actually improve incineration efficiency and reduce emissions, or does it depend on the nature and type of waste and the treatment used? Please clarify.

 

Response 6: We fully endorse the reviewers’ recommendation. The principal advantage of employing artificial intelligence lies in its capacity to intelligently identify the current flame combustion regime without reliance on experiential knowledge accumulated by experts over extended periods. This approach primarily furnishes technicians with diagnostic references, enabling them to formulate optimization decisions that indirectly enhance incineration efficiency and curtail pollutant emissions. We have explicitly articulated this point in the revised manuscript as requested.

 

Comments 7: The author mentioned that the incineration of waste relied on three main stages: drying, ignition, and coal burning. However, he did not mention the key factors affecting the burning.

 

Response 7: We sincerely appreciate the reviewers’ invaluable suggestions. In the revised manuscript, we have itemized the pivotal factors governing the efficiency of each of the three sequential regimes of municipal solid-waste incineration.

 

 

Comments 8: The author did not specify the nature and type of fuel used in the combustion, how the combustion is fueled, whether for a specific period, what the usage standards are, how long each combustion cycle takes to complete the process and convert the fuel to ash, whether the ash is tested after combustion, whether it is sieved, and how it is utilized. Do the images represent four stages of the combustion process to reach ash? Why does oxygen decrease during combustion? Is oxygen supplied automatically, or does it rely on spontaneous combustion?

 

Response 8: We are sincerely grateful to the reviewers for their inquiry. In clarification, the present study is exclusively devoted to the identification of flame regimes during the incineration of solid waste, thereby furnishing decision-support information for subsequent operational stages; it does not encompass investigations pertaining to waste feeding, ash handling, oxygen supply, or any other ancillary segments of the overall treatment process.

 

 

Comments 9: I recommend that the author to prepare a table of images depicting waste burning be prepared and described according to the nature and type of solid waste, with the program being designed to display images of the flame from beginning to end.

 

Response 9: We sincerely appreciate the reviewers’ valuable recommendation and fully acknowledge that tabulating the categories of solid waste together with their corresponding combustion images would be highly effective. Regrettably, project-contractual obligations prohibit public dissemination of the data, precluding the presentation of specific details; we therefore express our profound apologies for this constraint.

 

 

Comments 10: What does the mean of 1x1 convolution and 7 x 7 according to the figure.

 

Response 10: We thank the reviewers for their inquiry and wish to clarify that the residual blocks employed in our architecture utilize 1 × 1 and 3 × 3 convolutions exclusively; no 7 × 7 convolution is implemented. Within the core residual unit, the 1 × 1 and 3 × 3 convolutions assume distinct yet complementary responsibilities. The 1 × 1 convolution performs channel-wise dimensionality reduction and expansion together with information fusion, regulating the number of feature-map channels to constrain computational complexity and introduce additional non-linearity. The 3 × 3 convolution executes spatial feature extraction, concentrating on the capture of local details and patterns within the image. Arranged in a “compress–learn–restore” sequence, these two operations collaborate to preserve strong representational capacity while markedly enhancing network efficiency and gradient flow.

 

 

Comments 11: How are categories distinguished during and after the incineration process? Does the temperature vary between different types of solid waste? And does the flame intensity vary according to the product?

 

Response 11: We thank the reviewers for their inquiry and wish to clarify that post-incineration residue handling lies outside the scope of the present study. According to our investigation, heterogeneous waste categories exhibit pronounced disparities in combustion temperature and flame intensity attributable to differences in calorific value, moisture content, and chemical composition. Upon complete oxidation, all organic constituents are converted into carbon dioxide, water vapour, and thermal energy, whereas non-combustibles fuse into a homogeneous mixture of bottom ash and fly ash. Consequently, the incineration process itself cannot discriminate among waste categories in the final products; efficient and clean operation hinges upon rigorous pre-combustion segregation and pretreatment, as well as intelligent real-time adjustment of feedstock and operating conditions informed by flame and temperature feedback during the process.

 

 

Comments 12: How does self-attention occur in overall modeling, and how does it improve the representation of key flame features to maintain computational efficiency and low data requirements, as the author mentioned? Please clarify, and are there any resulting implications that need to be explained?

 

Response 12: Thank you for the reviewer's question. In the article, we analyzed the limitations of using self-attention and introduced deformable convolution to improve the modeling process. By employing learnable sampling offsets, the receptive field can adaptively resize and reshape according to the geometric variations of flames, enabling more accurate extraction of key features. This adaptive sparse attention mechanism not only reduces dependency on global computations but also, owing to its powerful geometric deformation modeling capability, allows the model to efficiently learn the intrinsic characteristics of flames from limited data.

 

Comments 13: What does the author mean by the different layers focusing on different aspects of flame representation: surface layers focusing on local texture variations, middle layers focusing on transitional shapes, and deep layers focusing on the overall state of combustion? Does this mean that combustion begins in the outer layers, progresses to the middle layers, and then to the inner layers? Are there factors that shape the flame and directly affect the duration of complete combustion?

 

Response 13: We thank the reviewer for this pertinent question. The phrase “different layers focusing on different aspects of flame representation” refers to distinct levels of visual abstraction learned by the neural network during feature extraction, not to any physical stratification of the flame into outer, middle, and inner zones. Specifically, early convolutional layers (surface layers) encode fine-grained local details such as flame edges and texture motifs; intermediate layers integrate transitional shape information including regional segments and elongation tendencies; deepest layers capture global semantic descriptors such as overall combustion stability and intensity. These hierarchical descriptors are extracted in parallel and subsequently fused; they do not imply that combustion proceeds physically from the flame’s exterior toward its interior. With respect to the second point, flame morphology and the temporal duration required for complete oxidation are indeed governed by multiple factors, foremost among them fuel properties, air-supply regime, and intrafurnace temperature distribution.

 

 

Comments 14: clarify what is meant by accuracy: the percentage of expected positive samples that actually turn out to be positive; recall the percentage of actual positive samples that were correctly identified; does this refer to confirmation of specific samples and how they were burned completely, or does it refer to the accuracy of flame year measurement? Please clarify.

 

Response 14: We thank the reviewer for this inquiry. The central objective of the present study is to classify municipal solid-waste incineration flame images into four discrete combustion regimes, constituting a multi-class classification task. The metrics reported—Accuracy, Precision, Recall, and F1-score—are performance indicators for this categorical assignment, not measures of the precision with which any physical attribute of the flame is quantified. Accuracy specifically denotes the proportion of test images correctly classified by the model, thereby reflecting its overall discriminative capability.

 

 

Comments 15: How is the performance of the proposed PRTNet model evaluated in this study? The three-dimensional confusion matrices in Figure 10 illustrate the results of the model test.

 

Response 15: In this study, the performance of the PRTNet model is comprehensively evaluated through a combination of quantitative metrics—accuracy, precision, recall, and F1-score—and visual analytics. The three-dimensional confusion matrix presented in Figure 10 is a visual representation employed to intuitively illustrate the model’s per-class classification outcomes across the four combustion regimes; fundamentally, it serves as a detailed graphical exposition of the model’s categorical performance.

 

 

Comments 16: I recommend that the author clarifying how to verify the performance and final evaluation of modules added to a traditional ResNet architecture to assess their performance. This includes outlining their specifications and the factors that contribute to and influence performance efficiency.

 

Response 16: We sincerely appreciate the reviewers’ valuable suggestion. Within the conventional ResNet backbone, we have embedded the LSEA module, which is cascaded by an ELA submodule followed by an SGE submodule; their architectural details have been thoroughly elaborated in the Modeling Strategy section of the manuscript. Furthermore, an ablation study has been conducted to evaluate the LSEA module in isolation, corroborating the individual efficacy of each constituent component.

 

 

Comments 17: what you mean network from 1 to 8, Is it the ablating material and please insert the key words of the table such as LESA to Rec. Is √ mean the ablation or x is ablation process.

 

Response 17: The designations “Network1” through “Network8” are merely nominal labels assigned to the model configurations examined during the ablation study; they are adopted solely for descriptive convenience and carry no intrinsic significance. A check-mark (√) under any module in the attribute column indicates that the corresponding module was included in the respective ablation configuration.

 

 

Comments 18: Please explain the matching between accuracy comparison and F1 score in Figure 12. I believe there is agreement between the two models.

 

Response 18: In the ablation experiments, both Accuracy and F1-score are reported to appraise model performance from two complementary perspectives. Accuracy provides a global, intuitive summary by indicating the overall proportion of correctly classified samples, whereas F1-score integrates per-class precision and recall, thereby sensitively revealing inter-class performance balance and the model’s ability to recognize minority categories. Although the two metrics yield similar values on our balanced dataset, their joint presentation reinforces the credibility and rigor of the results.

 

 

Comments 19: This is conclusion only not discussion. I recommend that the author to add discussion to this paper.

 

Response 19: We sincerely thank the reviewers for their valuable comments; accordingly, we have substantially expanded the Discussion section.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a novel and effective hybrid architecture (PRTNet) for Municipal Solid Waste Incineration (MSWI) combustion state recognition, achieving high-level performance that significantly outperforms several state-of-the-art models. The methodology is clearly structured, and the ablation studies successfully validate the contribution of the three core modules (LSEA, FAFT, CFGA).

The conclusion rightly points out that the high model complexity from the multi-module cascades and computationally intensive operations (like deformable convolution) limits its real-time deployment on resource-constrained edge devices. To make the discussion more compelling, it is recommended to quantify this trade-off in the results section. Please consider adding a comparative table that shows the number of parameters (or FLOPs) and inference time for PRTNet compared to the baseline models (DenseNet, EfficientNet, ConvNeXt V2, ViT, PVT, FastViT). This data would strengthen the discussion on the model's current limitations and fully justify the proposed future work on model lightweighting.

Comments for author File: Comments.pdf

Author Response

Comments 1: The high model complexity (due to multi-module cascades and computationally intensive operations like deformable convolution) currently limits its real-time deployment on resource-constrained edge devices.

 

Response 1: We sincerely appreciate the reviewers’ valuable feedback. We acknowledge that the current model’s substantial parameter count precludes its deployment on resource-constrained edge devices, and addressing this limitation constitutes a primary focus of our forthcoming research.

 

 

Comments 2: The arguments and discussion ofthe findings, while coherent, could be further improved to be more compelling and balanced.

 

Response 1: We sincerely thank the reviewers for their valuable comments; accordingly, we have substantially expanded the Discussion section.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is interesting, but it should be noted that the data will not be made publicly available. This way, the authors can evaluate if the data used can be inserted in the paper, if the paper can be approved. 

The model is composed of specific modules, each justified by a technical challenge; The analyses were performed to the highest technical standards; The methods, tools, and software are described in sufficient detail to allow other researchers to reproduce the results.
The limiting factor for complete reproducibility is the unavailability of the raw dataset. The article explicitly states that, "Due to project constraints, the data will not be made publicly available." This is the big problem analysed in the paper.

Although the technical details of the architecture (LSEA, FAFT, CFGA) are of limited interest to deep learning specialists, the application and practical results of PRTNet—which provides a robust method for ensuring high efficiency and low emissions in an essential environmental technology—guarantee broad interest in various sectors related to sustainability, energy, and industrial automation.

There is a clear overall benefit to publishing this work, as it addresses a critical and long-standing industrial problem with an innovative and high-performance solution. But it should be pointed out that industrial applicability is low due to the high complexity of the model.

 

Author Response

Comments 1: There is a clear overall benefit to publishing this work, as it addresses a critical and long-standing industrial problem with an innovative and high-performance solution. But it should be pointed out that industrial applicability is low due to the high complexity of the model.

 

Response 1: We sincerely appreciate the reviewers’ valuable feedback. We acknowledge that the current model’s substantial parameter count precludes its deployment on resource-constrained edge devices, and addressing this limitation constitutes a primary focus of our forthcoming research.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author

Thank you so much for the correction, but I think the manuscript needs more editing and correction to be suitable for publication because some gaps are not clarified.

                                                 Thank you 

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Dear Sir 

I believe the manuscript should be reviewed because it is necessary, as it is a fundamental requirement for publication.

                            Thank you 

                                      

              

Author Response

Comments 1: If these items are considered researchobjectives, then they must necessarilyrepresent the main problem, as theauthor has not clarified what the mainproblem is, so it cannot be directly reliedupon.

 

Response 1: Thank you for the valuable suggestions from the reviewers. In the previous revision, we have already clarified the main problem addressed in the paper according to your requirements; therefore, these can serve as the primary research content of this article.

 

Comments 2: The author has mentioned that to overcome the challenges of extracting features resulting from complex shapes and strong noise, examples of these challenges and conditions must be given for example, the maximum.

 

Response 2: Thank you for the reviewers’ valuable comments. Regarding the issues of complex flame shapes and noise interference in our images, we have already addressed them when describing the four combustion states of the flame and have provided example images.

 

Comments 3: I recommend that the author to prepare a table of images depicting waste burning be prepared and described according to the nature and type of solid waste, with the program being designed to display images of the flame from beginning to end.

 

Response 3: We sincerely appreciate the reviewers’ valuable recommendation and fully acknowledge that tabulating the categories of solid waste together with their corresponding combustion images would be highly effective. Regrettably, project-contractual obligations prohibit public dissemination of the data, precluding the presentation of specific details; we therefore express our profound apologies for this constraint.

 

Comments 4: What is background noise, and how can noise levels be reduced clearly and effectively? Is the software used to determine the high temperature and reaction rate between oxygen and waste to ensure complete combustion?

 

Response 4: We sincerely thank the reviewers for their questions. Background noise primarily denotes non-flame disturbances such as lens contamination and smoke occlusion within flame images. The LSEA module leverages ELA for spatial flame localization and SGE for semantic purification, establishing a “localization–purification” two-stage cooperative mechanism that enhances fine-scale flame structures while suppressing noise. Although the proposed model concentrates on vision-feature-based combustion-state recognition and does not directly measure temperature or reaction rate, its recognition results can be supplied to combustion-control systems as inputs, thereby indirectly optimizing combustion conditions and promoting complete combustion.

 

Comments 5: What does the mean of 1x 1 convolution and 7 x 7 according to the figure.

 

Response 5: We thank the reviewers for their inquiry and wish to clarify that the residual blocks employed in our architecture utilize 1 × 1 and 3 × 3 convolutions exclusively; no 7 × 7 convolution is implemented. Within the core residual unit, the 1 × 1 and 3 × 3 convolutions assume distinct yet complementary responsibilities. The 1 × 1 convolution performs channel-wise dimensionality reduction and expansion together with information fusion, regulating the number of feature-map channels to constrain computational complexity and introduce additional non-linearity. The 3 × 3 convolution executes spatial feature extraction, concentrating on the capture of local details and patterns within the image. Arranged in a “compress–learn–restore” sequence, these two operations collaborate to preserve strong representational capacity while markedly enhancing network efficiency and gradient flow.

 

Comments 6: How are categories distinguished duringand after the incineration process? Does the temperature vary between different types of solid waste? And does the flame intensity vary according to the product?

 

Response 6: We thank the reviewers for their inquiry and wish to clarify that post-incineration residue handling lies outside the scope of the present study. According to our investigation, heterogeneous waste categories exhibit pronounced disparities in combustion temperature and flame intensity attributable to differences in calorific value, moisture content, and chemical composition. Upon complete oxidation, all organic constituents are converted into carbon dioxide, water vapour, and thermal energy, whereas non-combustibles fuse into a homogeneous mixture of bottom ash and fly ash. Consequently, the incineration process itself cannot discriminate among waste categories in the final products; efficient and clean operation hinges upon rigorous pre-combustion segregation and pretreatment, as well as intelligent real-time adjustment of feedstock and operating conditions informed by flame and temperature feedback during the process.

 

Comments 7: What is meant by a strong relationship between fine-grained textures and stereoscopic combustion patterns is that it greatly improves the accuracy and robustness of flame zone identification against background noise, and howcombustion occurs.

 

Response 7: We appreciate the reviewer's comment. Your understanding of the "strong relationship between fine-grained textures and stereoscopic combustion patterns" is entirely accurate. The LSEA module precisely establishes this relationship through the synergistic design of ELA and SGE: ELA captures the macroscopic spatial distribution of the flame, while SGE enhances the corresponding microscopic texture features. This correlation mechanism effectively distinguishes the true flame structure from background noise, thereby significantly improving recognition accuracy and robustness against noise.

 

Comments 8: How does self-attention occur in overall modeling, and how does it improve there presentation of key flame features to maintain computational efficiency and low data requirements, as the author mentioned?

 

Response 8: Thank you for your valuable feedback. As previously explained, to overcome the limitations of standard self-attention in modeling the non-rigid geometric deformations of flames, we introduce deformable convolution and construct a Deformable Multi-head Attention module. By learning adaptive sampling offsets, the module enables the receptive field to scale and deform in line with flame morphology, achieving precise focus on key contours. This design reduces reliance on global computation and, thanks to its strong geometric modeling capacity, improves the model’s ability to learn intrinsic features from limited data.

 

Comments 9: This paragraph is unclear and requiresclarification. What does the author meanby the different layers focusing ondifferent aspects of flamerepresentation: surface layers focusingon local texture variations, middle layersfocusing on transitional shapes, anddeep layers focusing on the overall stateof combustion? Does this mean thatcombustion begins in the outer layers,progresses to the middle layers, andthen to the inner layers? Are therefactors that shape the flame and directlyaffect the duration of completecombustion?

 

Response 9: Thank you for the reviewer’s question. We have addressed this point in our previous response. The phrase “different layers focusing on different aspects of flame representation” refers to the hierarchical visual abstractions learned by the neural network during feature extraction, rather than any physical stratification of the flame. Specifically: shallow layers capture local details such as edges and textures; intermediate layers integrate regional shapes and transitional forms; deep layers focus on global semantics such as the overall combustion state and intensity. These hierarchical features are extracted and fused in parallel and do not correspond to a physical combustion process from the exterior to the interior of the flame. Regarding flame morphology and the time required for complete oxidation, these are indeed influenced by multiple factors, including fuel properties, air supply methods, and the temperature distribution within the furnace.

 

Comments 10: clarify what is meant by accuracy: the percentage of expected positive samples that actually turn out to be positive; recall the percentage of actualpositive samples that were correctly identified; does this refer to confirmation of specific samples and how they were burned completely, or does it refer to the accuracy of flame year measurement? Please clarify.

 

Response 10: We thank the reviewer for this inquiry. The central objective of the present study is to classify municipal solid-waste incineration flame images into four discrete combustion regimes, constituting a multi-class classification task. The metrics reported—Accuracy, Precision, Recall, and F1-score—are performance indicators for this categorical assignment, not measures of the precision with which any physical attribute of the flame is quantified. Accuracy specifically denotes the proportion of test images correctly classified by the model, thereby reflecting its overall discriminative capability.

 

Comments 11: Does the author intend to divide the combustion process into four stages: normal combustion, partial combustion, flash combustion, and incandescent combustion? The combustion process proceeds according to a rate, starting slowly, then progressing to intermediate or partial combustion, and finally incandescent combustion, where the flame's activity increases until complete combustion is achieved.

 

Response 11: Thank you for the reviewer's question. We are not dividing the combustion process into four consecutive stages. The four combustion states mentioned are classified based on real-time visual characteristics and do not imply a time-sequential progression of the combustion process. These states are identified through flame image analysis, and their occurrence depends on specific operational conditions, without following a fixed sequence from slow to complete combustion.

 

Comments 12: How is the performance of the proposed PRTNet model evaluated in this study? The three-dimensional confusionmatrices in Figure 10 illustrate theresults of the model test.

 

Response 12: Thank you for the reviewer's question. We have already addressed this point in our previous response. This study comprehensively evaluates the performance of the PRTNet model through a combination of quantitative metrics—accuracy, precision, recall, and F1-score—and visual analysis. The three-dimensional confusion matrix shown in Figure 10 provides an intuitive illustration of the model’s classification results across the four combustion states, serving essentially as a detailed graphical representation of the model’s categorical performance.

 

Comments 13: I recommend that the author clarifying how to verify the performance and final evaluation of modules added to a traditional ResNet architecture to assess their performance. This includes outlining their specifications and the factors that contribute to and influence performance efficiency.

 

Response 13: Thank you for the reviewers’ valuable feedback. As previously addressed, we have embedded the LSEA module—constructed by cascading the ELA and SGE submodules—into the conventional ResNet backbone. The architectural details of this module have been thoroughly described in the Modeling Strategy section of the manuscript, and its individual efficacy, along with that of each submodule, has been validated through dedicated ablation studies.

 

Comments 14: What you mean network from 1 to 8, Is it the ablating material and please insert the key words of the table such as LESA to Rec. Is √ mean the ablation or x is ablation process.

 

Response 14: The designations “Network1” through “Network8” are merely nominal labels assigned to the model configurations examined during the ablation study; they are adopted solely for descriptive convenience and carry no intrinsic significance. A check-mark (√) under any module in the attribute column indicates that the corresponding module was included in the respective ablation configuration.

 

Comments 15: Please explain the comparison between accuracy and F1 score in Figure 12. I believe there is agreement between the two models.

 

Response 15: In the ablation experiments, both Accuracy and F1-score are reported to appraise model performance from two complementary perspectives. Accuracy provides a global, intuitive summary by indicating the overall proportion of correctly classified samples, whereas F1-score integrates per-class precision and recall, thereby sensitively revealing inter-class performance balance and the model’s ability to recognize minority categories. Although the two metrics yield similar values on our balanced dataset, their joint presentation reinforces the credibility and rigor of the results.

Author Response File: Author Response.pdf