Cascaded Residual Attention Enhanced Road Extraction from Remote Sensing Images

Round 1

Reviewer 1 Report

This paper proposed a Cascaded Residual Attention Enhanced Road Extraction method for Remote Sensing Images. The method is effective with two kennel modules. The method is of importance and the performance is good enough. There are still some items that need improvements. 

1. This paper proposed a multi-scale network for spatial information extraction and a coarse-to-fine segment module for smooth boundary recognition. Some SOTA researches need to be cited and analyzed, e.g., (1) multi-scale/multi-branch networks: [1] Multi-scale and multi-task deep learning framework for automatic road extraction [2] Fractional Gabor Convolutional Network for Multisource Remote Sensing Data Classification. (2) coarse-to-fine module for boundary region: [3] Corse-to-fine road extraction based on local Dirichlet mixture models and multiscale-high-order deep learning [4] Joint classification of hyperspectral and LiDAR data using hierarchical random walk and deep CNN architecture [5]Feature Extraction for Classification of Hyperspectral and LiDAR Data Using Patch-to-Patch CNN. 
2. The proposed method includes several modules, e.g., attention enhancement block. The efficiency is concerned. Thus a theoretical computational analysis section is suggested. 
3. For the proposed method, the motivation of using specific module needs further explanations, e.g., the selected attention module, ASPP, etc. 
4. The mathematical expression of proposed method is missing. Specific network designing is also needed.
5. An overall pseudo-code algorithm is required for a better understanding of the proposed method.

Author Response

First of all, we would like to express our sincere appreciation for the detailed comments and suggestions from the reviewer. We have made a serious response to each comment, please refer to the attachment and the revised manuscript for details.

Author Response File: Author Response.docx

Reviewer 2 Report

2 applications related to navigation and Geographic Information System information updating.

Two times information is double:  updating the GIS or updating the GIS database or similar will do

data-driven methods based on semantic segmentation recognize roads from images pixel4

wisely,

Intelligent use of data driven ?? or clever use of data driven systems ???  wisely is maybe not correct in this context.

pixel4

wisely, which generally uses only local spatial information and

The “snakes” method is quite famous  and uses much more than only local spatial information  here any arbitrary reference:

https://link.springer.com/chapter/10.1007/BFb0055700

Do not confuse the traditions in remote sensing ( moving windows based or kernel based) and those in computer vision

22 research on automatic road extraction in urban areas,

If urban areas are the main focus, use this info in the introduction too !!

118 results from images through adversarial learning between generative and discriminative

119 models.

adversarial machine learning???

especially at the edges of roads so that

196 the context characteristics of the road boundary are unsmooth, making the boundaries of

197 the roads are difficult to be recognized as smoothly as realistic.

Making the boundary of the roads difficult to recognize (or similar.) Making and are in this sentence is overdone.

“222 Road extraction from remote sensing images is a typical binary segmentation

223 task that seeks to classify every pixel as road or background.”

Or classify image objects and not single pixels as potential roads

In this section, we describe the experimental datasets, including the Massachusetts

237 Road Dataset [55], the DeepGlobe Road Extraction Challenge dataset [52], and the

238 Huawei Cloud competition dataset [56]. Then, we describe extensive comparative exper239

iments and analyze the results of our method and some classical semantic segmentation240

based methods on the datasets.

Normally, all sections are described directly after the introduction or in chapter 2.

In Table 1  “OURS” is simple  to be referred to by other researchers, please put your name for your approach so others can more easily discuss your model.

Figure 9

As the dataset is well known and used by many, is there not a final result for comparison for the Massachusetts dataset with ALLL road vectors  in a separate GIS vector file ??????

Author Response

First of all, we would like to express our sincere appreciation for the detailed comments and suggestions from the reviewer. We have made a serious response to each comment, please refer to the attachment and the revised manuscript for details.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors proposed a cascaded attention architecture to extract boundary-refined roads from remote sensing images.  According to the results, they obtained a better trade-off between precision and recall (i.e., F1 score) than other state-of-the-art work. 

Author Response

Thanks for the reviewer's recognition of our work.

Round 2

Reviewer 1 Report

I have no more further comment.