Building Damage Detection from Post-Event Aerial Imagery Using Single Shot Multibox Detector

Round  1

Reviewer 1 Report

The goal of the paper is very inetresting but it needs some improvements especially as regards paragraph 2.2.2. SSD Model, 3.2.3 The proposed method. The model should be describe better because the process is not completely clear. Some details should be specified and improved (see the attached file)

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 1 Comments

Point 1: The goal of the paper is very interesting but it needs some improvements especially as regards paragraph 2.2.2. SSD Model, 3.2.3 The proposed method. The model should be described better because the process is not completely clear. Some details should be specified and improved (see the attached file)

Response 1: Thank you for your suggestion. We describe the SSD model in more detail in 3.2.2. This includes the role of additional convolutional layers. In 3.2.3 we use a table to express the hyperparameters that need to be deployed before training. In addition, in order to make the experimental results more convincing, the Hurricane IRMA data set was added to verify the experiment. The details you think need to be improved in the article are answered in the table below.

Reviewer 2 Report

Thank you for an interesting paper with a very interesting approach. I have several questions and remarks regarding this paper:

Regarding data source, was it obtained by UAV? "UAV" was mentioned in the introduction but was never specified when describing the dataset used. Are the photos geo-referenced/photogrammetrically processed or at least geo-tagged? A geo-referenced dataset can have the advantage of not only classifying the image, but giving directly the coordinates of said damages. I think this would be a good implementation of your algorithm.

I have doubts with the data augmentation process. The authors did mention in the text that subjectively their approach may seem to generate model overfitting, but they claim that the machine does not think so. As I see it, although the analysis parameters seem to improve by using the proposed approach, the improvement is still not very big and only based on one dataset observation. I would suggest to implement this method to other datasets in order to give a more objective point of view and less bias. This is also linked to the discussions part of the paper, which I found rather underwhelming. 

In terms of English language, I find that it was good in the beginning (introduction and related work) but deteriorates as the paper progresses. Especially during the description of the method, paragraphs sound more like training manuals than a scientific paper. Many sentences are not even sentences at all. This is very annoying to read, more so for the most important part of the paper which describes the actual algorithm. I suggest to rewrite these paragraphs, and also to perform proofreadings (many typos).

Author Response

Response to Reviewer 2 Comments

Point 1: Regarding data source, was it obtained by UAV? "UAV" was mentioned in the introduction but was never specified when describing the dataset used. Are the photos geo-referenced/photogrammetrically processed or at least geo-tagged? A geo-referenced dataset can have the advantage of not only classifying the image, but giving directly the coordinates of said damages. I think this would be a good implementation of your algorithm.


Response 1: Our original intent was to obtain data from the air including UAV and so on. We have already made improvements in the introduction. The image of the raw data we obtained contains only basic information such as basic axes and does not contain damage coordinates.

Point 2: I have doubts with the data augmentation process. The authors did mention in the text that subjectively their approach may seem to generate model overfitting, but they claim that the machine does not think so. As I see it, although the analysis parameters seem to improve by using the proposed approach, the improvement is still not very big and only based on one dataset observation. I would suggest to implement this method to other datasets in order to give a more objective point of view and less bias. This is also linked to the discussions part of the paper, which I found rather underwhelming.

Response 2: Thank you for your suggestion. We realized that we needed to add another dataset to validate our experiment. So, we used the scene image of Hurricane IRMA to verify our experiment again. We have added the dataset and experimental results to 3.1 Study Areas and 4.1.2 Discussion

Point 3: In terms of English language, I find that it was good in the beginning (introduction and related work) but deteriorates as the paper progresses. Especially during the description of the method, paragraphs sound more like training manuals than a scientific paper. Many sentences are not even sentences at all. This is very annoying to read, more so for the most important part of the paper which describes the actual algorithm. I suggest to rewrite these paragraphs, and also to perform proofreadings (many typos).

Response 3: We have modified the language in the article and added a description of the method.

Round  2

Reviewer 2 Report

Thank you for submitting a revised manuscript. I think that the authors have addressed all of my concerns. The addition of another dataset is very much appreciated, as I think it reinforces the conclusions. Some minor errors in English are still present, but otherwise the paper is OK.