Application of Deep Learning Approach for the Classification of Buildings’ Degradation State in a BIM Methodology

Reviewer #1

Reviewers Comments to Author            

Authors Response to Reviewers Comments

Presented article with Title ”Application of Deep Learning Approach for the classification of buildings’ degradation state in a BIM methodology” is writing on 13 pages with 9 figures, 1 tables and 34 references. The paper is processed to the required extentand and I think deserves pubblication.

We appreciate your valuable and constructive feedback. The authors hope that the revised version addresses your concerns being now suitable for publication.

Sugestion:

The article is written well but is too short.

Introduction is too brief, requires a more detailed art of state.

The authors considered the suggestion. The introduction was improved (was given to the previous text a more structured form, divided in two sections), and the state of art was extent as well as the case study.

In materials and methods are missing input parameters for experiment. Authors describe just Workflow of the methodology.

Thank you for the suggestion. The methodology is explained in the text before the workflow that was improved introducing the dataset and the dictionary of images. Also, in the text was introduced this part that was missing in the first submission. All the methodology is deeply explained in the case study section, along its application. In section 4  at 4.1 a table presenting the input images was added as the inputs parameters that supported all the development of the images’ recognition model (the input images used in the DL model). The development of the maintenance actions database is referred as future works that will be published in a next paper (please see in 4.4).

Experiment is interesting but the application of deep learning is unclear.

A CNN chapter (in section 2) was added, where the role of Deep Learning is explored and explained.

Revit software is a full - featured application used deep learning, what is the added value  in article.

The authors appreciate your comment.

Subchapter 2.4 requires a more detailed description.

Thank you for the suggestion. Table 2 was added and it presents the organization of the dataset and the input images used in the DL model.

In the article is missing a presentation of CNN.

The authors considered the suggestion. A chapter about CNN was added in section 2 (at 2.2)

Results require redrafting. They are too short and inconsistent.

The authors considered the suggestion. The results were rewritten and the special contribute of this paper was highlighted.

Conclusion is missing.

Conclusions, please describe your future directions.

The authors considered the suggestion. The conclusions chapter was added and the future developments were pointed out.

Final recommendations: After studying the article, I can recommend publish publication after Major revision.

Authors thanks your positive comment about our work.

Please see the paper in which with track changes the changes are highlighted.

Reviewer #2

Reviewers Comments to Author            

Authors Response to Reviewers Comments

General Comment: In this manuscript, the authors develop a methodology which include building digitalization techniques and 3D building modelling, in order to provide an automated tool to recognize building anomalies inside of the BIM environment. Although my feeling is that the novelty is limited, this paper is well organized and show extensive case. I’m willing to recommend it for publication if the authors could address the following concerns.

We appreciate your valuable and constructive feedback. The authors hope that the revised version addresses your concerns and the paper is suitable for publication.

Comment 1: One case study is not enough to evaluate the effectiveness of the proposed method. I suggest the authors adding more building datasets in their experiments.

Thank you for your comment.

Several images of different buildings were considered in the dataset used to train the DL model.

The quality of the desired outputs from the DL model was evaluate by the F1 Score, which assesses the quality of the data obtained.

The application of a case study was considered to be the pilot to validate the linkup between DL and BIM models.

Comment 2: The Introduction section is too long and needs to be reorganized. In addition, the author needs to highlight the innovation and contribution of the manuscript. My suggestion is to describe it separately at the last two ends of the Introduction.

The suggestion was considered by the authors.

The introduction was reorganized (structured in 2 sections) and the innovation and the contribution of the paper was highlighted in the last paragraph of the introduction.

Comment 3: Page 8, the experiment part. Why does the classification accurate get lower as the number of network layers increases? In my intuition, the effect of the network will be more excellent as the depth increases.

Thanks for pointing this issue out.

More layers in a learning model not always means a better learning model, with less rate error.

After Table 2, the follow text was added:

“In this case, more layers in the learning model does not always means a better learning model with less rate error. When the dataset is not so big, adding more layers to the deep neural networks, makes the model become stagnant or starts to degraded. When deeper networks are able to start converging, with the network depth increas-ing, accuracy gets saturated and then degrades rapidly [16].

According to [16] that the 18-layer plain/residual nets are comparably accurate, but the 18-layer ResNet converges faster. According to this author, the deeper ResNet has smaller magnitudes of responses, and when there are more layers, an individual layer of ResNets tends to modify the signal less.”

Please see the attached paper in which the changes are highlighted.