Tunnel Rapid AI Classification (TRaiC): An Open-Source Code for 360° Tunnel Face Mapping, Discontinuity Analysis, and RAG-LLM-Powered Geo-Engineering Reporting

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents an innovative and timely open-source platform, TRaiC, that integrates panoramic tunnel imaging, AI-powered discontinuity detection, and RAG-LLM-based automated geotechnical reporting. The work is well-structured and offers a promising tool for enhancing safety, efficiency, and reproducibility in tunnel face mapping and digital rock mass characterization. The user-friendly GUI (which makes it suitable for a broad range of users) and openly shared code via GitHub will help the work attract both active researchers and practitioners to read and apply the current work.

After reviewing the manuscript and testing the software using the provided GitHub repository and sample data, I suggest the following minor revisions to improve usability, clarity, and practical value:

- While testing the code, I noticed that users are required to re-enter the same parameter settings multiple times for different tunnel facings (e.g., left wall, right wall, roof, invert). This repetitive input can be time-consuming and may hinder the user experience. Hence, I recommend considering implementing a feature that allows parameter settings to be stored in memory after the first input, and then automatically apply the same configuration to all other facings unless the user explicitly opts to change them. This would streamline the workflow and reduce redundancy, especially for large-scale tunnel applications.

- The manuscript describes automated detection of discontinuities and the derivation of parameters such as dip, dip direction, and discontinuity length. However, in practice, some known geological features (e.g., faults or joints identified in previous site investigations) may not be automatically captured but are critical for geotechnical analysis. Please clarify whether the current platform supports the manual addition or correction of discontinuity features not identified in the automatic detection phase. If such a feature exists, a brief description of the procedure would greatly enhance practical utility and flexibility.

- The current RAG-LLM component appears to rely on access to the OpenAI API, which may present limitations for some users (e.g., due to internet connectivity, licensing, or privacy constraints). Could the authors comment on the potential adaptability of TRaiC to use locally hosted Large Language Models (e.g., LLaMA, Mistral, GPT4All)? If such compatibility exists or is planned, a short note or configuration guidance would be useful for developers or users aiming for offline deployment or enhanced data privacy.

Author Response

The manuscript presents an innovative and timely open-source platform, TRaiC, that integrates panoramic tunnel imaging, AI-powered discontinuity detection, and RAG-LLM-based automated geotechnical reporting. The work is well-structured and offers a promising tool for enhancing safety, efficiency, and reproducibility in tunnel face mapping and digital rock mass characterization. The user-friendly GUI (which makes it suitable for a broad range of users) and openly shared code via GitHub will help the work attract both active researchers and practitioners to read and apply the current work.

We sincerely appreciate your thorough review and evaluation of our article. Your positive and encouraging feedback is of great value to us.

After reviewing the manuscript and testing the software using the provided GitHub repository and sample data, I suggest the following minor revisions to improve usability, clarity, and practical value:

- While testing the code, I noticed that users are required to re-enter the same parameter settings multiple times for different tunnel facings (e.g., left wall, right wall, roof, invert). This repetitive input can be time-consuming and may hinder the user experience. Hence, I recommend considering implementing a feature that allows parameter settings to be stored in memory after the first input, and then automatically apply the same configuration to all other facings unless the user explicitly opts to change them. This would streamline the workflow and reduce redundancy, especially for large-scale tunnel applications.

We greatly appreciate this valuable suggestion, as it will contribute significantly to the further development of our software. While achieving maximum automation is indeed one of the central objectives of this research, the first version of the platform must also support flexible parameter adjustment. This flexibility is essential for systematically investigating how variations in these parameters influence the results. Consequently, it is preferable that the initial release of the software not operate as a complete “black box,” but instead allow researchers full control to independently adjust all parameters for each tunnel face. Once the parameter sensitivity study has been completed, integration and full automation can be implemented with relative ease.

- The manuscript describes automated detection of discontinuities and the derivation of parameters such as dip, dip direction, and discontinuity length. However, in practice, some known geological features (e.g., faults or joints identified in previous site investigations) may not be automatically captured but are critical for geotechnical analysis. Please clarify whether the current platform supports the manual addition or correction of discontinuity features not identified in the automatic detection phase. If such a feature exists, a brief description of the procedure would greatly enhance practical utility and flexibility.

This is an excellent and insightful question. In the AI-based discontinuity trace detection phase, we have incorporated a supervising tool that allows users to manually delineate structural features on tunnel faces. This ensures that deterministic features within the rock mass are explicitly included in the analysis. Furthermore, the open-source code of the TNA software is accessible on this platform, enabling users to integrate the 3D geometry of these structural features into the calculations through simple MATLAB coding. Your suggestion is indeed valuable, and in future versions of the platform, we plan to implement a dedicated graphical interface to facilitate the manual entry of 3D deterministic geo-structural features in a more intuitive and user-friendly manner.

- The current RAG-LLM component appears to rely on access to the OpenAI API, which may present limitations for some users (e.g., due to internet connectivity, licensing, or privacy constraints). Could the authors comment on the potential adaptability of TRaiC to use locally hosted Large Language Models (e.g., LLaMA, Mistral, GPT4All)? If such compatibility exists or is planned, a short note or configuration guidance would be useful for developers or users aiming for offline deployment or enhanced data privacy.

We sincerely appreciate your thoughtful and valuable comment. As noted in the article (section 3.12), the platform already supports integration with local LLMs. However, as also discussed, the use of local models entails certain limitations. The necessary scripts for connecting the platform to local models have been previously published and are freely accessible to all researchers via our GitHub repository. For ease of access, the repository link has been provided in the article (subsection 3.12.3), allowing readers to readily locate and utilize these resources when needed.

 

We sincerely thank the reviewer for their careful evaluation and constructive feedback, which have been invaluable in improving our work.

Best Regards

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors submitted an interesting and well written manuscript dealing with tunnel face mapping using AI-based technology. Specifically, the study introduces Tunnel Rapid AI Classification platform which is an open-source solution designed to enhance the efficiency, accuracy, and safety of tunnel face mapping and rock mass characterisation. The conceptual framework is well described and the conclusions are supported by the experimental results. Given the current trend into infrastructure constructions and the need for the more robust and efficient technology, this study could heavily contribute to the geological, geotechnical and geo-engineering activities in tunnel face mapping. However, the manuscript should be improved before it could be considered for publication. The authors should revise the way the references are cited in the main manuscript, they should refer to the instructions for authors provided by Remote Sensing journal. Additionally, they could provide quantitative results of accuracy assessment.

Best wishes

Author Response

The authors submitted an interesting and well written manuscript dealing with tunnel face mapping using AI-based technology. Specifically, the study introduces Tunnel Rapid AI Classification platform which is an open-source solution designed to enhance the efficiency, accuracy, and safety of tunnel face mapping and rock mass characterisation. The conceptual framework is well described and the conclusions are supported by the experimental results. Given the current trend into infrastructure constructions and the need for the more robust and efficient technology, this study could heavily contribute to the geological, geotechnical and geo-engineering activities in tunnel face mapping.

We are deeply grateful for your careful review and evaluation of our article, and we greatly value your positive and encouraging feedback.

However, the manuscript should be improved before it could be considered for publication. The authors should revise the way the references are cited in the main manuscript, they should refer to the instructions for authors provided by Remote Sensing journal.

Yes, this has been addressed. Thanks.

Additionally, they could provide quantitative results of accuracy assessment.

This platform is the result of several years of research in the field of digital rock mass characterization, during which multiple studies have been conducted and their results published. Therefore, to evaluate the accuracy of the algorithms used in this software, we should refer to the related articles cited in this research study. The open-source software presented here is the outcome of integrating previous findings, methods and algorithms with new techniques such as panoramic photography and the use of large language models. The purpose of publishing this article is to introduce the architecture and functionality of this open-source software, as well as to outline its limitations and the opportunities for further research. As mentioned in the manuscript, assessing the accuracy of the results, particularly the outputs of large language models, requires additional research and development, and this platform provides the necessary foundation and tools for research development and application purpose.

Best wishes

We sincerely thank the reviewer for their careful evaluation and constructive feedback, which have been invaluable in improving our work.

Best Regards

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In the manuscript, the authors present a highly technical and novel approach to tunnel face mapping through an integrated system, TRaiC, which combines 360° photography, AI-based discontinuity detection, 3D modeling, and LLM-powered reporting. The work is multidisciplinary and extensive, embracing computer vision, geotechnics, AI, and software engineering. It has the potential to impact the field of digital geotechnical engineering. The manuscript demonstrates a high level of technical competence in image processing, artificial intelligence integration, and knowledge in geotechnical fields. The algorithms are also well explained, especially for trace detection, segment connection, and stereo mesh analysis.

However, the reviewer believes that it would benefit from a clearer structure, greater scientific precision, and more critical validation of all claims. Below are some comments and recommendations:

• Regarding the structure and length of the manuscript, it is evident that it is incredibly long. It could be divided into two parts: the first dedicated to the design and architecture of the system, and the second to performance evaluation, validation, and applications.
• A simplified introduction will improve readability at the beginning of the text. The literature review should be condensed, and the research gaps and objectives should be stated clearly.
• The abstract should be more concise and factual, stating briefly the purpose of the research, the principal results, and the major conclusions. Emphasize performance gains with specific metrics, if available.
• The manuscript lacks experimental validation, and no quantitative comparative analysis is provided comparing TRaiC results to manual mapping, photogrammetric, and lidar methods, as well as to existing commercial software solutions (e.g., 3GSM). How can the authors prove the claims of improved accuracy, speed, and safety without control comparisons? The authors can include a dedicated results section and compare TRaiC outputs with baseline methods using precision, recall, accuracy, or expert ratings.
• It is indicated that future work will focus on complete automation capabilities, integration with GSI/Q-class, and adaptive thresholding. The presented work can be considered more as a prototype than a ready-for-production solution.
• Regarding integration and security risks of LLM (based on GPT-4) there are some privacy and security concerns due to cloud-based dependencies. No local implementation of the proposed mapping and assessment approach has been demonstrated or tested. How accurately does LLM generate the geological descriptions compared to a human expert? Some examples of generated reports can be provided.
• The text is technically comprehensive, but in places it is too promotional, with expressions that sound exaggerated, such as “revolutionary” or “expert-level automation”. Some of the individual sections, for example, cubemap rendering and LLM prompting, could be presented more concisely and summarized.
• Please indicate in the manuscript text in appropriate places the figures that are shown but not mentioned.
• Limitations of the proposed approach should be pointed out.
• It is recommended to add case studies or examples of implementation of real tunnel projects, as well as to obtain feedback from users or experts in the field.

Comments on the Quality of English Language

It is recommended to use more neutral and scientific wording, avoiding an advertising tone. The text should be checked for minor grammatical errors and repetitions.

 

Author Response

In the manuscript, the authors present a highly technical and novel approach to tunnel face mapping through an integrated system, TRaiC, which combines 360° photography, AI-based discontinuity detection, 3D modeling, and LLM-powered reporting. The work is multidisciplinary and extensive, embracing computer vision, geotechnics, AI, and software engineering. It has the potential to impact the field of digital geotechnical engineering. The manuscript demonstrates a high level of technical competence in image processing, artificial intelligence integration, and knowledge in geotechnical fields. The algorithms are also well explained, especially for trace detection, segment connection, and stereo mesh analysis.

We truly appreciate the time and care you’ve taken to review and evaluate our article, and we greatly value your positive and encouraging feedback.

However, the reviewer believes that it would benefit from a clearer structure, greater scientific precision, and more critical validation of all claims. Below are some comments and recommendations:

• Regarding the structure and length of the manuscript, it is evident that it is incredibly long. It could be divided into two parts: the first dedicated to the design and architecture of the system, and the second to performance evaluation, validation, and applications.

We sincerely appreciate the reviewer’s thoughtful observation and we fully acknowledge that the current article is indeed extensive, and that including detailed discussions on performance, validation, and applicability would further lengthen the manuscript and risk negatively affecting its focus. The primary aim of the present work is to introduce the architecture and framework of the open-source platform. For this reason, validation of the algorithms and verification of the methods have not been addressed in depth.

 

There are two main considerations behind this decision. First, the algorithms and core methodologies implemented in the platform have been described and validated in detail in our prior publications, to which we have referred readers for a thorough examination of these aspects. Second, certain topics, such as the application of language models in geological analysis and the preparation of engineering reports, are complex and merit independent, dedicated studies, falling beyond the intended scope of this manuscript.

 

This work is therefore focused on presenting the platform, its components, and its operational workflow, while providing the tools necessary to enable future research in these areas. We respectfully request the reviewer’s consideration in supporting the publication of the current study in its present structure, as it forms a foundational step toward more specialized follow-up investigations.

 

• A simplified introduction will improve readability at the beginning of the text. The literature review should be condensed, and the research gaps and objectives should be stated clearly.

 

We sincerely thank the reviewer for this constructive suggestion to improve the clarity and readability of the manuscript. While the introduction was originally concise and focused on the article’s topic, we have carefully revised it in line with the reviewer’s recommendation. Specifically, the section has been condensed from 1,270 words to 1,026 words, with greater emphasis placed on clearly stating the research gaps and objectives.

 

• The abstract should be more concise and factual, stating briefly the purpose of the research, the principal results, and the major conclusions. Emphasize performance gains with specific metrics, if available.

We would like to thank the esteemed reviewer for their constructive suggestion to improve the clarity and focus of the abstract. In line with this recommendation, the abstract has been thoroughly revised to be more concise and factual, emphasizing the purpose, key results, and main conclusions of the study. As a result, its length has been reduced from 274 words to 182 words.

 

• The manuscript lacks experimental validation, and no quantitative comparative analysis is provided comparing TRaiC results to manual mapping, photogrammetric, and lidar methods, as well as to existing commercial software solutions (e.g., 3GSM). How can the authors prove the claims of improved accuracy, speed, and safety without control comparisons? The authors can include a dedicated results section and compare TRaiC outputs with baseline methods using precision, recall, accuracy, or expert ratings.

 

We appreciate the reviewer's important observation regarding experimental validation and quantitative comparative analysis. The TRaiC platform represents the culmination of several years of research in digital rock mass characterization, where the core algorithms implemented in this software have been rigorously described, verified, and validated in our previous publications over multiple studies. These foundational works provide the necessary scientific basis for the computational methods employed in TRaiC. Therefore, to evaluate the accuracy of the algorithms used in this software, researchers and practitioners should refer to the related articles cited in this research study, where detailed validation procedures and accuracy assessments have been conducted.

The present manuscript focuses on introducing the architecture and functionality of this open-source platform, which integrates our previously validated methods with innovative techniques such as panoramic photography and large language models. The primary objective is to present the software's capabilities, outline its limitations, and identify opportunities for further research and development. Regarding comparison with commercial software such as 3GSM, it is important to note several fundamental considerations. 3GSM is proprietary commercial software that is not publicly available for evaluation or benchmarking purposes. Moreover, fundamental methodological differences exist between the two approaches, as 3GSM employs photogrammetry-based methods while TRaiC utilizes image processing-based techniques. As explained and emphasized in the introduction, TRaiC is designed to fill gaps and address limitations of existing 3D digital models, positioning it as a complementary solution rather than a direct replacement. This fundamental difference in approach and purpose makes direct quantitative comparison methodologically inappropriate and potentially misleading.

We fully recognize that further verification is essential, particularly for results derived from large language model analysis. The potential directions for comprehensive validation studies are discussed in Section 4 of the manuscript, and we remain committed to conducting rigorous validation studies, publishing outcomes of follow-up investigations that include quantitative comparisons with manual mapping methods, developing appropriate metrics such as precision, recall, and accuracy for image processing-based discontinuity detection, and collaborating with industry partners to establish validation datasets. This open-source platform provides the necessary foundation and tools for research development and practical applications. By making TRaiC publicly available, we enable the research community to conduct independent validation studies, contribute improvements, and develop standardized benchmarking procedures for image processing-based rock mass characterization methods.

In conclusion, while we acknowledge the importance of comprehensive validation, the current manuscript serves its intended purpose of introducing an innovative open-source platform and establishing a foundation for future validation research. We respectfully submit that the manuscript's contribution lies in its methodological innovation and open-source accessibility rather than in comparative performance metrics, which will be the subject of dedicated future studies. The platform provides the necessary foundation and tools for research development and application purposes, enabling the broader scientific community to advance the field of digital rock mass characterization through collaborative validation efforts.

 

• It is indicated that future work will focus on complete automation capabilities, integration with GSI/Q-class, and adaptive thresholding. The presented work can be considered more as a prototype than a ready-for-production solution.

We fully agree with the reviewer’s observation. As an open-source platform, TRaiC represents a foundational prototype that requires further development before it can be considered a ready-for-production solution. We have not made any claims to the contrary in the manuscript; however, to improve clarity, we will explicitly state in the conclusion section that the current version is a prototype and highlights the need for ongoing development toward complete automation:

“Overall, TRaiC represents a significant advancement toward a fully automated, intelligent, and secure digital ecosystem for underground construction by combining AI-based methodologies with practical engineering requirements. While it lays a strong foundation for future developments in digital rock mass characterization and geotechnical decision-making systems, the current work serves primarily as a prototype rather than a ready-for-production solution.”

 

• Regarding integration and security risks of LLM (based on GPT-4) there are some privacy and security concerns due to cloud-based dependencies. No local implementation of the proposed mapping and assessment approach has been demonstrated or tested. How accurately does LLM generate the geological descriptions compared to a human expert? Some examples of generated reports can be provided.

 

We appreciate your valuable comment. As noted in the article (Section 3.12), the platform already supports integration with local LLMs, addressing the core security concerns raised. The necessary scripts for connecting the platform to local models have been previously published and are freely accessible to all researchers via our GitHub repository. For ease of access, the repository link has been provided in the article (Section 3.12.3), allowing readers to readily locate and utilize these resources when needed.

However, as also discussed in the manuscript, the use of local models has certain limitations that need careful consideration. Our preliminary testing of open-source LLMs was constrained by available hardware capabilities, limiting us to distilled large language models with lower parameters rather than full-scale implementations. This limitation prevented comprehensive evaluation of original-size LLMs and their potential performance in geological description generation. If we set aside hardware requirements, the transition from cloud-based to locally deployed LLMs introduces key challenges related to accuracy and performance that require systematic investigation.

While open-source models demonstrate strong potential, their capabilities can vary significantly compared to advanced commercial models, particularly in specialized technical domains such as geological assessment. Addressing this discrepancy necessitates extensive research and development for evaluation and validation to ensure locally hosted models deliver acceptable accuracy. The complexity of geological terminology, the nuanced nature of rock mass characterization, and the critical importance of accurate technical documentation in tunneling applications all contribute to the challenging requirements that local models must meet.

Therefore, we emphasize the need for systematic research development to measure performance differences between cloud and local deployments, assessing factors such as technical accuracy, compliance with engineering standards, response consistency, and overall reliability in tunnel face reporting. This comparative analysis must encompass not only the linguistic quality of generated descriptions but also their technical precision, adherence to established geological classification systems, and consistency with expert human assessments. Such a comprehensive evaluation requires installation of a full-size local LLM, controlled testing environments, and standardized evaluation metrics to establish acceptable performance thresholds.

Consequently, we plan to undertake comprehensive comparison research work in this regard in the future, which will systematically evaluate the trade-offs between security benefits of local deployment and potential accuracy limitations compared to cloud-based solutions. Until this systematic evaluation is completed, suitable comparison results are not available to provide definitive conclusions about the relative performance of local versus cloud-based LLM implementations in geological description generation. We acknowledge that this represents an important area for future development and commit to publishing these findings to benefit the broader research community once this comparative study is conducted.

 

 

 

• The text is technically comprehensive, but in places it is too promotional, with expressions that sound exaggerated, such as “revolutionary” or “expert-level automation”. Some of the individual sections, for example, cubemap rendering and LLM prompting, could be presented more concisely and summarized.

 

Thank you for your constructive feedback regarding the promotional language and section length in our manuscript.

Regarding the promotional language: I carefully searched the manuscript for the specific expressions mentioned (such as "revolutionary" and "expert-level automation"), but I was unable to locate these exact phrases in our text. However, I understand the concern about overly promotional tone and have systematically reviewed the manuscript to identify words and phrases with similar promotional tendencies. I have modified language throughout the document to adopt a more measured and objective academic tone, replacing any overstated claims with more balanced expressions supported by evidence.

Regarding the section length: I have worked to make the individual sections more concise, particularly the cubemap rendering section (3.3) and the LLM prompting discussion (3.12), while ensuring that essential technical details necessary for reproducibility are retained. The sections now provide more focused summaries of the methodologies while maintaining scientific rigor.

 

• Please indicate in the manuscript text in appropriate places the figures that are shown but not mentioned.

Thank you for your comment regarding figure references. I have carefully reviewed the entire manuscript and confirmed that all figures (Figures 1-12) are properly cited in the text at appropriate locations where they are discussed. If there are specific figures that appear to be unreferenced, could you please indicate which ones? This would help us address any potential formatting issues or provide clarification.

 

• Limitations of the proposed approach should be pointed out.

We thank the reviewer for this consideration, though we respectfully note that the limitations have been addressed in detail throughout the manuscript. We have made efforts to maintain transparency about the system's current constraints, including explicit discussions of parameter sensitivity issues in trace detection and segment linking algorithms (Sections 3.6.3 and 3.7), acknowledged dependencies on LLM capabilities for geological interpretations (Section 3.10), identified data security concerns with cloud-based APIs (Section 3.12.3), and provided a comprehensive future work roadmap that directly addresses current limitations (Section 4). Additionally, our conclusions (Section 5) summarize these key limitations to ensure readers understand both the system's capabilities and areas requiring further development.

 

• It is recommended to add case studies or examples of implementation of real tunnel projects, as well as to obtain feedback from users or experts in the field.

regarding the inclusion of real tunnel project case studies, as this would undoubtedly strengthen the practical validation of our proposed platform, we are pleased to inform the reviewer that engineering data from an actual tunneling site has been collected and analyzed using our platform, and preliminary feedback from field experts has been gathered during the development process. However, we regret to inform that project stakeholders do not permit us to publish the information, as it involves ongoing commercial tunnel construction projects with confidentiality requirements. We therefore kindly request the reviewer's understanding and consideration in evaluating the current manuscript based on the methodological contributions and technical framework presented.

Comments on the Quality of English Language

 

It is recommended to use more neutral and scientific wording, avoiding an advertising tone. The text should be checked for minor grammatical errors and repetitions.

We have conducted a thorough proofreading to eliminate grammatical errors,  ensure all language maintains a scientific tone, appropriate for academic publication.

 

Finally, we thank the reviewer for this valuable feedback and appreciate their attention to these important details.

Best Regards

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This research has made significant contributions in the field of digitalization of tunnel engineering, but there are still some deficiencies that need to be improved.

(1) Section 2.1's review of existing technologies does not cover the latest applications of deep learning.

(2) Although the paper describes in detail the two line detection algorithms, Hough Transform and LSD, it lacks systematic performance comparison and quantitative analysis. The visual comparison in Figure 5(c-d) and the simple statistics of the number of line segments alone are not sufficient to prove the superiority of the LSD algorithm.

(3) The paper repeatedly mentions that the system is sensitive to parameter thresholds, but fails to provide a scientific basis for threshold selection.

(4) The LLM-assisted geological interpretation system described in Section 3.1 does not specify the specific LLM model and training data used.

(5) The paper lacks system verification data from real tunnel engineering.

Author Response

This research has made significant contributions in the field of digitalization of tunnel engineering, but there are still some deficiencies that need to be improved.

We sincerely appreciate your thorough review and thoughtful evaluation of our article, and highly value your positive feedback.

(1) Section 2.1's review of existing technologies does not cover the latest applications of deep learning.

Thank you for your insightful observation. You are absolutely correct that recent deep learning applications are not extensively covered in Section 2.1. However, the scope of this review was deliberately limited to methods specifically applied to digital tunnel face mapping to maintain focus and conciseness. Given the overall length of the manuscript, including broader or less directly relevant studies could detract from its clarity and engagement. Nonetheless, if you are aware of any recent research or publications pertinent to digital tunnel face mapping that we have not included, we would greatly appreciate your recommendations for further consideration.

(2) Although the paper describes in detail the two line detection algorithms, Hough Transform and LSD, it lacks systematic performance comparison and quantitative analysis. The visual comparison in Figure 5(c-d) and the simple statistics of the number of line segments alone are not sufficient to prove the superiority of the LSD algorithm.

Thank you for your valuable observation. We fully acknowledge that this manuscript does not provide a systematic performance comparison or quantitative analysis between the Hough Transform and LSD algorithms. The primary reason is that our focus here is to introduce the architecture and framework of the open-source platform rather than delve deeply into methodological details and algorithmic results. In this research, our intention is to demonstrate that various line segment detection methods can be integrated within this platform, thereby enabling broader comparative research in the future. Based on our previous work, the Hough Transform tends to perform optimally for detecting line segments in three-dimensional space; however, since our current approach identifies traces initially in 2D space before projecting them into 3D, the LSD method was also incorporated to illustrate alternative viable options. We therefore fully agree with the reviewer’s suggestion regarding the importance of systematic comparison and respectfully propose that such detailed analyses be reserved for future investigations, while this article remains focused on presenting the platform’s architecture and capabilities.

(3) The paper repeatedly mentions that the system is sensitive to parameter thresholds, but fails to provide a scientific basis for threshold selection.

Thank you for highlighting this important point. In this article, we deliberately emphasize the sensitivity of the system to parameter thresholds to draw readers’ attention to the critical need for further research and development in this area. We fully agree with the reviewer that establishing a scientific basis for selecting threshold values is essential to achieving robust and reliable results. However, developing such a scientific foundation, particularly for parameters relevant to rock engineering, requires dedicated and independent studies beyond the scope of this work. More detailed discussions on threshold determination for the AI-aided trace detection method and TNA analysis have already been provided in prior related publications, which is why this article does not address these aspects in depth. Our intention here is to highlight this challenge and encourage future investigations to establish rigorous, scientifically grounded threshold criteria.

(4) The LLM-assisted geological interpretation system described in Section 3.1 does not specify the specific LLM model and training data used.

We are very grateful to the reviewer for highlighting this important point. Upon revisiting the manuscript, we acknowledge that while Section 3.12.1 mentions the use of the OpenAI language model and Section 3.12.2 specifically refers to the ChatGPT-4 model employed in the software, some related details were inadvertently omitted in other parts of the text. To clarify, in the modified version, Section 3.10 also explicitly states the use of the ChatGPT OpenAI language model. We appreciate the reviewer’s careful reading and will ensure these references are made more consistent and clearly highlighted in the revised manuscript. Thank you again for your valuable feedback.

(5) The paper lacks system verification data from real tunnel engineering.

We sincerely appreciate the reviewer’s observation on this important matter. It is worth noting that the algorithms implemented in this software have been described in detail in prior publications of the authors, where the necessary verification and validation were performed. Nonetheless, we fully agree that further verification is essential, particularly regarding results derived from the analysis of language models. In this study, engineering data from an actual tunneling site is indeed available for validation purposes; however, until present, we have not been granted permission to publish this information. We therefore kindly request the reviewer’s understanding and consideration in evaluating the current manuscript in its present form. The potential directions for further development, as well as the scope for future validation studies, are discussed in Section 4, and we remain committed to publishing the outcomes of these follow-up investigations in the near future.

 

We sincerely thank the reviewer for their careful evaluation and constructive feedback, which have been invaluable in improving our work.

Best Regards

 

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have responded in detail to the reviewer's comments and recommendations, reflecting all of them in the edited version of the manuscript.