Factory Extraction from Satellite Images: Benchmark and Baseline

Round 1

Reviewer 1 Report (Previous Reviewer 3)

The author has answered all my concerns. I have no other questions.

Author Response

Thank you for your constructive comments!

Reviewer 2 Report (Previous Reviewer 2)

The authors have done a nice job of revising this manuscript and addressing the reviewer comments and queries.

Author Response

Thank you for your constructive comments!

Reviewer 3 Report (Previous Reviewer 1)

the paper title has to be improved such as:

 

Factory Extraction from Satellite Images: Benchmark and Baseline 

Author Response

Thank you for your constructive comments!

We have replaced the title with “Factory Extraction from Satellite Images: Benchmark and Baseline”.  

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Titel

“Remote Sensing Image” is very general, please be more specific.

 Abstract

You name the images “FE4395”, and in the title “Remote Sensing Image”, the two descriptions are vague. You must use a clear description in the title or in the abstract.

Moreover, when you focus on the factory buildings, that means you focus on the large area buildings either they represent factories or another type. I believe before reading the paper that the suggested algorithm has not any importance because it extracts only the easy buildings.

I am sorry to say that the abstract is not well written, you spend ten lines introducing the paper topic and the employed data. The abstract must focus on the suggested approach and result in quality. Please rewrite the abstract.

Please excuse me, I feel sorry when the paper is of low-level quality.

References

There are 53 references cited in the paper, 2 are 2022 despite the machine learning topic representing a hot research spot. In 2022, hundred of papers were published about this topic.

Introduction

You said “Although the above methods achieve good results in building extraction from remote sensing images, it can not be directly applied to factory extraction task. On the one hand, since some non-factory buildings are similar to factory buildings at pixel level, directly using building extraction can easily lead to incorrect classification results” Why? Please prove it.

You said: “On the other hand, the current building extraction usually treats all buildings as one class without distinguishing detailed classes for each building, which makes it impossible to obtain the desired factory information directly from the conventional building extraction results.” Why? Please prove it.

I believe that these reasons are not realistic and need to be proved, else the selection of a factory building as the target for your algorithm is not acceptable.

The related work analysis is so weak regarding the wide application of machine learning techniques.

The main contributions

You said: “A comprehensive remote sensing image benchmark dataset is presented for factory extraction, which includes 4395 remote sensing factory images collected from Google Earth and corresponding accurate annotations. In addition, we also classify all the images according to different challenging attributes. It will facilitate community research on specific building extractions.” I believe that you need one architect in your team to realise this study. It will be very important and could be a topic of an independent paper.

You said: “We propose a general segmentation network based on encoder-decoder architecture, in which a novel gated dense aggregation module is designed to adaptively fuse multilayer features, as well as proposes a simple and effective information compensation module.” As you develop this algorithm, you must test it on several building typologies, one of them could be the factories. In this case, the paper becomes valuable.

At this stage, I advise you to rewrite all the paper after considering two or more datasets one of them representing factories.

 

You said: “Extensive experiments and visualization results analysis on our FE4395 benchmark dataset show the effectiveness of the proposed method, which is robust for complex scenarios and various types of factories. It achieves excellent performance in terms F-measure and IoU metrics compared with existing segmentation methods.” This is not a contribution.

Materials and Methods

2.1. Dataset

Please separate the section titles by transition paragraph.

2.1.1. Image Acquisition

 

Question: does Google Earth provide the original images, or does it provide manipulated images? Which satellite is ‎used? what are the characterises of these images? I believe that dataset selection is not an easy task.‎ 

Reviewer 2 Report

The paper presents a novel encoder-decoder network for factory extraction. Some major comments are listed as follows:

1) The originality of the method is not clear. For example, an ASPP module is adopted in the feature encoding stage to obtain multi-receptive field features, which is not new. Also, a simple post-processing step called hole filling (HF) is used to refine the results, which does not add too much to the novelty of work. The only scientific contribution, which can be derived from this paper is presenting a comprehensive remote sensing image benchmark dataset.

 

2) More state-of-the-art deep learning models should be compared and then discuss as the UNet, SegNet, FCN etc are old methods.

3) I think it would be interesting to check the superiority of the proposed model in factory extraction from various examples of challenges, which are Dense Building Disturbance (DBD),

Haze Occlusion (HO), Appearance Contamination (AC), and Complex Background and discuss the results to show how the model can deal with these issues to accurately map the buildings.

4) Discussion section is shallow. The Discussion needs rewriting with a full in-depth discussion of the comparison results where the choice of the various methods tested has been justified.

5) If possible, the authors are suggested to verify the presented method on more open-source datasets.

6) For further research on this topic, it would be good to depict the limitations of the proposed method.

7) Moreover, the following manuscripts can be reviewed to improve the Introduction/Literature.

[1] https://doi.org/10.3390/rs10111768

 

[2] https://doi.org/10.1016/j.mlwa.2021.100194

Reviewer 3 Report

Comments

This paper presents a new method for extracting factory scenes from remote sensing scene images. In general, the paper is very good, but there are still several problems need to be modified.

1. I am confused about the partitioning of the FE4395 dataset, and the basis for partitioning the dataset should be explained in detail in the manuscript.

2. Are there any specific scenarios for FE4395 datasets other than the four "DBD", "CB", "HO", "AC" scenarios?

3. In the Encoder module, X1 and X2 are used as information compensation modules, although good results have been achieved, but whether X3 and x4, or X1 and X3, or x2 and X4 as information compensation modules have better results, I suggest you should do an ablation experiment to verify.

4. In the information compensation module, “X_ Low” and “X_high” is defined should be explained in the manuscript.

5. The full name of GradCAM should be given in the manuscript, not just an abbreviation.