Point Cloud Segmentation Based on the Uniclass Classification System with Random Forest Algorithm for Cultural Heritage Buildings in the UK
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript concerns the application of AI and specifically supervised machine learning to cultural heritage.
From an engineering and computer science point of view, and in particular with reference to machine learning and deep learning techniques, in order to make the manuscript more understandable even for those who are not specifically in the field, I suggest:
- Line 9 mentions hierarchical classification, which represents a specific sector of classification. Please explain more clearly what it consists of.
- Line 164 mentions ‘16% of the total dataset...’. What does this subset consist of? What parts does it consist of?
- Line 187 mentions customised Python scripts. Please explain how these scripts guarantee consistent partitioning.
- In addition, in order to improve the manuscript with useful ideas, I suggest reading the following articles:
10.1108/JEIM-02-2020-0059
10.1007/978-981-13-8331-1_23
10.3390/ijgi9050297
10.1007/s12518-021-00359-2
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper describes a workflow to segment point clouds of buildings according to a hierarchical framework and using classes defined according to Uniclass classification system.
Although I think the article proposes an interesting methodology and draws some interesting points in the discussion section and in the conlcusions section, I think it needs to be modified in many respects which I list below:
- It lacks a comprehensive state of the art that not only better details what previous studies do, but also describes what Uniclasses are and how they are constructed (the average reader of this journal may not be fully familiar with them, especially if reference is made to UK and not international standards)
-compared to the work currently cited, it is not well stated, or should be stated more emphatically, what the elements of innovation or improvement are compared to the work cited (e.g. compared to Teruggi et al);
-the choice of Uniclasses should be explained in more detail, why other codifications were not used (e.g. IFC ?)
-actually, Uniclasses were used as classes for point cloud segmentation. I think it is essential to include a table detailing well what all the segmentation classes are for the various cases L1, L2, L3, to make it more obvious how the work was done
-personally I had some difficulty in understanding the workflow used, I suggest inserting more diagrams or explanatory figures
-some tables and figures are not mentioned within the text and the numbering of both figures and tables should be checked
-the description of the case study needs to be implemented. What is the case study used? is it a historical building or a modern building? is it a church or an apartment block? what tools were used for the acquisition? I believe that this information is especially useful for a better understanding of the quality and type of data (a dataset from mobile mapping is different from one from a terrestrial laser scanner, and the classes in a church are different from those in an apartment building)
-Two methods are mentioned in section 2.2, of which only one is then used. It is not clear to me whether the two methods exist in the literature or whether they are two alternative implementation proposals proposed by the authors.
-point 3 mentions a pilot experiment, about which, however, there is no information. I suggest either explaining the pilot experiment in detail or not describing it and only describing the experiment with the full dataset.
-table 3: I suggest adding the mathematical formulae of the features mentioned (or a literature reference)
- the figure on page 9 needs further explanation in the caption, and I suggest giving an explanation within the text of both how it was made and what it represents and what information can be gained from it
-in the conclusion the bulleted list for future research could be revised, making it more discursive, some points are described, others are just terms, which should be more contextualised to the case study presented in the article.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors The paper addresses key challenges in point cloud classification by presenting an advanced hierarchical classification framework that employs the Random Forest algorithm, a classical supervised machine learning technique, to segment and classify large-scale point clouds of heritage buildings. Furthermore, the study contributes to the field of Heritage Building Information Modelling (HBIM) by proposing optimized hierarchical structures and scalable machine learning techniques, particularly relevant for advancing AI applications within HBIM. The manuscript deals with several important aspects of the current and highly relevant topic of point cloud segmentation involving large datasets of complex geometries characterized by structural irregularities and notable class imbalance. Building on prior research, appropriately cited, it distinguishes itself by explicitly aligning the methodology with the UK-based Uniclass taxonomy, aiming to develop a robust hierarchical classification framework tailored specifically to Heritage point clouds across multiple spatial resolutions. The paper is structured into seven sections and includes a case study from the Royal Greenwich Museum in the UK, which demonstrates that taxonomy refinement and multi-scale feature extraction significantly enhance classification accuracy. The paper is interesting and proposes an innovative workflow that critically evaluates the limitations, practical constraints, and potential advantages of integrating Uniclass within the heritage BIM context. However, the state of the art is limited to foundational works and would benefit from a broader discussion to more effectively situate the research within its academic context. A more comprehensive review of the background literature is recommended. In addition, the number of references is very limited. Please consider expanding the reference list to include more recent and diverse contributions in the field, particularly in relation to the background literature. Figures are generally clear and informative, although the font size in some of them is quite small and could be increased for better readability. The English language used in the manuscript is accurate and understandableAuthor Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsAuthors have coherently arranged the article following my previous comments