Semi-Automatic Extraction of Rural Roads under the Constraint of Combined Geometric and Texture Features
Round 1
Reviewer 1 Report
The language quality is in most parts ok. A small number of mistakes may be corrected in the publishing process, such as: “… Then, the manually input seed points are modified The figures and tables are well set. …”; here the adverb “manually” is wrong, an adjective is needed “manual”, or “manually set seed points” is also possible. “…The fuzziness of the resulting selections local road direction pre-diction methods such as that of Dai et al. [31] can eliminated with the method proposed in this paper; …” here a “be” is necessary “… can be eliminated …”; In the caption of Fig.4 “The matching model mainly solves the problems. (a) The convergence of …” should be “The matching model mainly solves the following two problems: (a) The similarity of …”; etc. …
A bit of proofreading will improve the manuscript.
Often the wording “convergence of road and surrounding surface textures” is used. In my mind “convergence” is a technical term from mathematics referring to infinite sums or sequences. The problem addressed here is rather “similarity of road and surrounding surface textures”. And I am also not so sure about the term “texture”. Isn’t it rather “color” or even just “intensity”?
Literature reference: There is a well written extensive overview on the state-of-the-art, distinguishing between model-driven and data-driven methods. The latter include the latest development in deep-learning convolutional perceptrons.
It is not clear to me, what a “gradient-mutation threshold” should be? Is that borrowed from [27]? Is it gradient change? And what is a “morphological gradient”? In my understanding of morphological image processing no gradients are used.
Topic: The paper is on road-extraction vom satellite imagery. Small country roads can be problematic due to the similarity in appearance with the environment around them. This definitely suits the journal.
Method: A semi-automatic approach is presented. The outline is given with Fig. 1 – I would call this a "flow-process-chart" of the method, instead of “Technological roadmap”. The presented method does not appear revolutionary new to me. It is a diligent refinement of existing template matching methods.
Empirical validation: The claims are high: “… While improving the accuracy and quality of the initial road centre point, the problems associated with convergent road and neighbourhood textures are largely solved. The experimental results show that, compared to other methods, the proposed method can not only guarantee the road extraction accuracy but can also greatly improve the degree of automation …” However, the empirical validation is based on three satellite images only. Advantages of the proposed method on these images are obvious, and they are quite divers in scenery and in part really challenging. Still it remains hard to extrapolate the behavior on all imagery, and giving guarantees is fairly bolt – based on such a small validation. I recommend being a bit more modest in the claims.
Comparison is with a handful of other template matching methods – resulting in a clear advantage for the proposed variant.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The paper proposes a new road extraction approach from satellite images. Although the paper is understandable, there are several issues I found with it. Following are my key comments:
- The presented related work in wide, however, not focused enough. I would suggest to the authors to shorten the description of unrelated methods by explaining only their shortcomings in regard to the template matching, while digging deeper in template matching concepts themselves.
- Section 2.2 should be rewritten in its entirety, as it is a mix of methodology description, its comparison to the related work (i.e. a discussion) as well as the motivation. Similarly, Section 3.1 is a mix of related work (that should be explained in Section 1), while results are also mixed with discussion. Accordingly, the paper is extremely confusing and should be significantly restructured, e.g. by following traditional IMRaD format.
- The paper does not explain how the actual ground-truth data was obtained. This is a mayor drawback of the paper that makes it unfit for publishing. The authors should convincingly address this issue in order to justify that the reported results have infect real implications.
Here are also some minor issuer, related mainly with the style of writing:
- “without distinguishing the data”: It is not clear to me what do authors mean by that and why would this allow a higher level of automation?
- “In section 2.2.2, after determining…” and also “In section 2.2.3, first,…”: Please rephrase as there is nothing done in this sections, it is just explained.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The work is well written, and presents a work of relevance to the research area. The extraction proposed show good scientific results that contribute to the literature. However, some points can be improved. However, some points must be revised.
Page 4: "The existing direction prediction model is improved" - Which one?
Table 1 presents data for bands Red, Green and Blue, but these bands are not used in the work.
Figure 1: Is initial point a question/test?
Figure 1: Are the results obtaind after a "NO Successful Match"?
Page 5: Wouldn't it be interesting to describe a little about the mentioned methods (L0, Dai et al and MLSOH)?
Page 12: Extracted as non-roads or just not extracted?
I didn't find Figures 5(a) and 6(a) cited in the text.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
See the uploaded file for the detailed report.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 4 Report
This version of the manuscript is better. However, the two operation blocks of "Adaptive road width extraction" and "Road direction prediction" should be combined into a single block of "Adaptive road width extraction and road tracking direction prediction model" in Figure 1 to ensure consistency with the method description.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
