Lane-Level Road Extraction from High-Resolution Optical Satellite Images
Round 1
Reviewer 1 Report
Overview
The authors utilize high-resolution (50-80 cm) optical imagery to test road extraction methods on both single- and double-lane roads. The models are tested on difficult roadways: those with bends, blurred boundaries, shadows, and structures (e.g., trees and buildings) blocking parts of the road.
General Comments
This is a well written paper. I only have a few small line-by-line comments below.
Line-by-Line Comments
Line 63: Please define DSM and OSM.
Line 87: Please define MLSOH. I see it is defined in Line 100, but typically acronyms are defined the first time they are used.
Lines 137-138: The sentence "In this paper, the L0 smoothing method Xu et al. [36] is used to filter the original image" sounds awkward. Is it supposed to say "...smoothing method from Xu et al..."?
Line 186: Add an open parenthesis "(" to "formula 1)."
Line 196: Please disregard this comment if I missed it earlier in the paper (I double-checked and cannot seem to find it), but what is μ?
Line 216: Is "Lie et al." supposed to be in parentheses here?
Lines 263-264: Reword to say "...the road is sufficient to ensure that..."
Line 275: What is α, a user-defined threshold?
Author Response
Ln 55 Do you have a citation / source for “other methods”?
Response: Thanks to the experts for their comments. The paper has added references to this section (Ln 55).
Ln 344 - You used a conditional statement to determine how to proceed with the single-lane tracking mode process: “If the length of the line segment is greater than or equal to 5 times the road width, set a tracking point (blue point in Figure 10b) at every other road width along the tracking direction based on the reference point until the distance between the last tracking point and the reference point is greater than the length of the line segment, and then proceed to step (7). If the length of the line segment is less than 5 times the road width, or if ② in (3) occurs, proceed to step (5).”
Response: Thanks to the experts for their comments. The paper has been re-edited for this section as follows “Depending on the length of the line segment, we get the tracking points in two ways: For the line segments with a length greater than or equal to 5 times the width of the road, based on the reference point, set a tracking point (blue point in Figure 10b) every other road width along the tracking direction, and stop when the distance between the last tracking point and the reference point is greater than the length of the line segment, and then proceed to step (7). Otherwise (the length of the line segment is less than 5 times the road width, or if ② in (3) occurs), proceed to step (5).” (line 351-357)
Was there a method used to define the line segment length to road width threshold? Or was this threshold determined through trial and error or discovery? I don’t see a methodological issue with this, but it may be prudent to state why this threshold was chosen. I am curious if there was an underlying explanation behind choosing that value as a determinant on which step to proceed with the process.
Response: Thank you for your comments.
First question：There is no clear way which is used to define the line segment length threshold for road width.
Second question：This threshold is the empirical threshold obtained through repeated experiments.
Third question：An underlying explanation behind choosing that value is，the line segments obtained by noise are usually shorter, and the line segment information exceeding the threshold can only be obtained through the road or adjacent object edges. The direction of the adjacent object boundary，such as building or farmland edge is mostly the same as or different from the direction of the road (or even vertical), the former can provide auxiliary information for the determination of the tracking direction, and the latter is usually directly regarded as the interference term, which can be determined by angle threshold α.
In Section 3.2 ln 405-414 – It would have been interesting to see a comparison between your method and DL/ML methods, but I think you provided a good enough explanation for why you did not evaluate your model against such methods.
Response: Thanks for the expert's suggestions. However, it is very difficult to use the deep learning algorithm as the comparative method for this paper, because:
Firstly, this is due to inconsistencies in the research ideas between the two methods. The deep learning method is a migration learning method, which requires a large number of training data sets to be established in advance. The proposed method belongs to the traditional manual design method and does not require a large amount of prior data.
Secondly, And deep learning algorithms are highly automated, but less complete (usually less than 85%) (Wei et al., 2017, Remote Sens. Lett; Xu et al., 2018, Remote Sens; Gao et al., 2018, Remote Sens); the proposed algorithm needs manual participation, but the accuracy and completeness are high (both above 95%), so they can't make a contrast.
Finally, the experimental area of this paper includes urban area and rural area. Because deep learning needs to establish a large number of similar data sets in advance, for this paper, the comparison is not significant, and I am sorry that it is difficult to complete in one week.
This section has been added to lines 410-416.
Ln 423 – Could you elaborate on your usage of the word “better” to describe resolution. Do you mean better as in higher spatial resolution?
Response: Thank you for your comments. This is our mistake. We have corrected the expression of this part to "image with resolution less than 1 meter." (Ln 406)
Could you also provide some additional temporal and phenological information about the Pleiades image? When was this image taken? Is the image leaf-off/on? Same with the GF2 panchromatic images.
Response: Thanks for the expert's suggestions. It has been added in the paper: “Figure 13shows the Pleiades satellite panchromatic image covering the rural area of Chrysanthemum Island (Huludao, China), which taken at January 2016 (leaf-off). The image size is 5000*5000 pixels with a resolution of 0.5 m and WGS-84 coordinate system. (LN464-466)”, “Figure 14 shows the Pleiades satellite panchromatic image covering the urban area of Huludao, China, which taken at July 2015 (leaf-on). The image size is 4000*4000 pixels with a resolution of 1m and WGS-84 coordinate system. (LN513-515)”, “Figure 15 shows the Pleiades satellite panchromatic image covering the urban area of Shenyang, China, which taken at September 2016 (leaf-on). The image size is 3000*3000 pixels with a resolution of 0.8m and WGS-84 coordinate system. (LN531-533)”.
Figure 11 is very well done, it is very interesting to see the differences between each method, especially the contrast in accuracy between the eCognition extraction and the proposed method, and how susceptible the OBIA approach was to road breakage. However, would it be possible to add coordinate graticules to one image to help identify the study area within Huludao, China?
Response: Thanks to the experts' suggestions, the first picture in each experiment has been modified to add information such as the coordinate system, see Figure 13, Figure 14 and Figure 15 for details.
Figure 13 – Perhaps add a coordinate graticule around one of the larger images as well?
Response: Thanks to the expert's suggestions, the first picture in each experiment has been modified.
Ln 571 – “With the amount of manual participation being 1/10, 1/18, and 1/6” – Is this being measured by time or the number seed pixels, or something else? How were these fractions determined?
Response: Thank you for your comments.
First, the amount of human involvement mentioned in the paper is measured by the number of seed points.
Second, the author separately calculates the sum of the seed points needed for the road algorithm and the other three algorithms to extract the three images，the ratios are close to 1 / 10, 1 / 18 and 1 / 6, respectively. (1 represents the seed points needed for the method of the paper)(Lines 586-580)
Ln 685 – Source #35. Zhang, J. appears to be the first author on this paper, not Lin, X. You state Zhang et al. on ln 76, but it’s backwards in the references.
Response: Thanks for the expert's comments. This is our mistake. We have corrected the references (line 712).
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors investigate lane-level road extraction through a novel method which combines an adaptive correction model and several descriptors (MLSOH, sector, multiangle beamlet) as an alternative method to traditional approaches for road network extractions such as object-based image segmentation, deep learning, and template matching. The goal of this work is to provide highly accurate and automated road extraction from high resolution imagery.
The manuscript is concise and offers a good explanation of a novel method. The structure of the work is logical and easy to follow. The outlined methods, despite their complexity, are well explained. It was very interesting that the authors used road interior features, which would be normally seen as a hindrance to road extraction in OBIA segmentation approaches to assist with road extraction.
Comments:
ln 55 - Do you have a citation / source for “other methods”?
Ln 344 - You used a conditional statement to determine how to proceed with the single-lane tracking mode process: “If the length of the line segment is greater than or equal to 5 times the road width, set a tracking point (blue point in Figure 10b) at every other road width along the tracking direction based on the reference point until the distance between the last tracking point and the reference point is greater than the length of the line segment, and then proceed to step (7). If the length of the line segment is less than 5 times the road width, or if ② in (3) occurs, proceed to step (5).”
Was there a method used to define the line segment length to road width threshold? Or was this threshold determined through trial and error or discovery? I don’t see a methodological issue with this, but it may be prudent to state why this threshold was chosen. I am curious if there was an underlying explanation behind choosing that value as a determinant on which step to proceed with the process.
In Section 3.2 ln 405-414 – It would have been interesting to see a comparison between your method and DL/ML methods, but I think you provided a good enough explanation for why you did not evaluate your model against such methods.
Ln 423 – Could you elaborate on your usage of the word “better” to describe resolution. Do you mean better as in higher spatial resolution?
Ln 432 – Could you also provide some additional temporal and phenological information about the Pleiades image? When was this image taken? Is the image leaf-off/on? Same with the GF2 panchromatic images.
Figure 11 is very well done, it is very interesting to see the differences between each method, especially the contrast in accuracy between the eCognition extraction and the proposed method, and how susceptible the OBIA approach was to road breakage. However, would it be possible to add coordinate graticules to one image to help identify the study area within Huludao, China?
Figure 13 – Perhaps add a coordinate graticule around one of the larger images as well?
Ln 571 – “With the amount of manual participation being 1/10, 1/18, and 1/6” – Is this being measured by time or the number seed pixels, or something else? How were these fractions determined?
Ln 685 – Source #35. Zhang, J. appears to be the first author on this paper, not Lin, X. You state Zhang et al. on ln 76, but it’s backwards in the references.
Author Response
Line 63: Please define DSM and OSM.
Response: Thanks for the expert's comments. DSM refers to Digital surface model，OSM refers to Open Street Map，and they were added in line 63 of the paper。
Line 87: Please define MLSOH. I see it is defined in Line 100, but typically acronyms are defined the first time they are used.
Response: Thanks to the experts for their suggestions. This is our problem and has been adjusted in the 88th line of the paper.
Lines 137-138: The sentence "In this paper, the L0 smoothing method Xu et al. [36] is used to filter the original image" sounds awkward. Is it supposed to say "...smoothing method from Xu et al..."?
Response: Thanks for the expert's suggestions. Adjustments have been made in lines 139-140 of the paper.
Line 186: Add an open parenthesis "(" to "formula 1)."
Response: Thanks for the expert's suggestions. Adjustments have been made in line 193 of the paper.
Line 196: Please disregard this comment if I missed it earlier in the paper (I double-checked and cannot seem to find it), but what is μ?
Response: Thanks for the expert's suggestions. The value here should be the gradient threshold μ, defined in line 203, and which is based on line 432-438.
Line 216: Is "Lie et al." supposed to be in parentheses here?
Response: Thanks for the expert's suggestion. Adjustments have been made in line 223 of the paper.
Lines 263-264: Reword to say "...the road is sufficient to ensure that..."
Response: Thanks for the expert's suggestions. Adjustments have been made in lines 269-270 of the paper.
Line 275: What is α, a user-defined threshold?
Response: Thank you for your comments. In this paper, α is angle threshold, which is difined by user, and used to measure the difference between two angles. This value is defined in line 281, and based on lines 444 - 450.
Author Response File: Author Response.docx
