Building a Vision Transformer-Based Damage Severity Classifier with Ground-Level Imagery of Homes Affected by California Wildfires

Round 1

Reviewer 1 Report

Great job! Continue working on improving the data to make it more robust.

Author Response

We thank the reviewer for comments. 

Reviewer 2 Report

The publication addresses the important issue of identifying the severity of fire damage to buildings.  

Among the strengths of the publication is the fact that there is information on open source software provided and a training dataset. This is an important usability aspect but not scientific.  

The structure of the publication is logical but should be improved: 

In the abstract very briefly but in the introduction the research problem, hypothesis and methods used could be described more extensively. 

In Materials and Methods one should: 

- describe the principle of operation and the main differences between the systems analysed,

- justify what guided and why the classification systems described in the publication were tested,

Most importantly, in the Discussion section:

- describe why only software testing was applied using the black box model, why the internal structure of the software was not considered,

- complete the discussion by comparing the results of the analyses to the results of other authors in other publications,

- describe the results obtained in the context of the research problem posed.

Author Response

Reviewer 2

Comments and Suggestions for Authors

• The publication addresses the important issue of identifying the severity of fire damage to buildings.
• Reply: We thank the reviewer for comments
• Among the strengths of the publication is the fact that there is information on open source software provided and a training dataset. This is an important usability aspect but not scientific.
• Reply: We thank the reviewer for comments
• The structure of the publication is logical but should be improved:
• In the abstract very briefly but in the introduction the research problem, hypothesis and methods used could be described more extensively.

Reply: The abstract and addressed was updated accordingly in line 147-158

• In Materials and Methods one should:
  • describe the principle of operation and the main differences between the systems analysed,
  • justify what guided and why the classification systems described in the publication were tested,

Reply: It was added and addressed accordingly in line 225-258

• Most importantly, in the Discussion section:
  • describe why only software testing was applied using the black box model, why the internal structure of the software was not considered,

Reply: It was addressed accordingly in Line 519-543

• complete the discussion by comparing the results of the analyses to the results of other authors in other publications,

Reply: It was addressed accordingly in Line 428-517

• describe the results obtained in the context of the research problem posed.

Reply: It was addressed accordingly in Line 684-696

Author Response File: Author Response.pdf

Reviewer 3 Report

The study proposed a damage severity classification model using ground-level imagery, focusing on residential structures damaged by wildfires. This classifier, a Vision Transformer (ViT) model trained on over 18,000 professionally labeled images of damaged homes from the 2020-2022 California wildfires, has achieved an accuracy score of over 95%. In total, it is a interesting paper and helpful for the forest-city connection fire. However, there are still some comments to improve this paper. 

1. Various damages are proposed, but how to define the different damage?

2. The detailed compution process of model, such as training time, should be introduced.

3. More traning images for every damage should be provided.

4. How about the accuracy degree of  the present model, and how to judge it? Further comparison of other common model and present model should be still added.

Good

Author Response

Reviewer 3

• The study proposed a damage severity classification model using ground-level imagery, focusing on residential structures damaged by wildfires. This classifier, a Vision Transformer (ViT) model trained on over 18,000 professionally labeled images of damaged homes from the 2020-2022 California wildfires, has achieved an accuracy score of over 95%. In total, it is a interesting paper and helpful for the forest-city connection fire. However, there are still some comments to improve this paper.
• Various damages are proposed, but how to define the different damage?

Reply: It was addressed accordingly in Line 202-223

• The detailed compution process of model, such as training time, should be introduced.

Reply: It was addressed accordingly in Line 341

• More traning images for every damage should be provided.

Reply: It was addressed accordingly in Line 202-223

• How about the accuracy degree of the present model, and how to judge it? Further comparison of other common model and present model should be still added.

Reply: It was addressed accordingly in Line 428-517

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Accept

Good