Application of AHP to Road Selection

Round 1

Reviewer 1 Report

SUMMARY OF RECOMMENDATION:

The authors propose an extension of stroke-based road network selection algorithms, adding additional attributes to describe properties of strokes (apart from topological measures), AHP to cope with ranking of multiple descriptive attributes, and a connectivity maintenance algorithm. The work potentially has merits, but there are a number of things that need to be improved before the manuscript can be accepted:

A) Language editing (see details below)

B) Many stroke-based road network selection algorithms have been published over the past few years. All the extensions presented in this manuscript -- except AHP-based attribute ranking -- have been proposed before in the literature in a similar form. Hence, the authors must more clearly identify in the Introduction what research gaps still exist and how they respond to these with their proposed methods.

C) Most importantly, the evaluation of the proposed methodology is insufficient and must be improved. In your evaluation, you compare against a manual generalization, taking this as a benchmark. That’s a good start, but firstly more could have been learned from the comparison with manual generalization, secondly you must compare also with other algorithms, and thirdly more varied test data sets must be used. Details follow below.

D) It’s not clear what scale range this algorithm is designed for. Also, you work only with numerical reduction factors (i.e. proportion of strokes retained), but the relationship to map scales is not revealed.

 

DETAILED COMMENTS

1) Language editing: Generally, the level of the English is quite good, but some language editing is required to clean up grammatical mistakes, awkward style (L31: “wildly researched”) and typos (“indictor”, several times). This should be done by a professional editing service and requires somewhat more than simple proof-reading.

2) Clearly identify and explain the research gaps existing in the literature and the contributions made in this paper. All the elements of the proposed methodology, except the use of AHP, have already been introduced in the literature one way or the other. Also, the results of some of the existing algorithms are similarly good as the ones presented in this paper, and some of these algorithms have been evaluated extensively. So, it is important that in the review of the literature, the deficiencies of existing algorithms are clearly identified and that it is clearly explained how the proposed approach responds to these weaknesses. That is, the reader should be able to tell what exactly is new in this work (contributions) and why it is necessary (research gaps).

3) Improve the evaluation by better exploitation of comparison with manual generalization results. The presentation in Figure 9 is confusing. It would be easier if you would use only a single map and use different colors to highlight (a) the strokes that were removed by your algorithm but had been retained in the manual solution, and (b) the strokes that were retained by your algorithm but removed in the manual solution. That is, basically highlight the false negatives and the false positives differently. Furthermore, it would be important to delve more into the reasons for the deviations. A first step would be to provide not only a small overview map as in Figure 9, but zoom in on details that can be used to explain what exactly happened. Also, if you can get your “experienced cartographers” to comment on the differences, that would help even further. Note, for instance, the work by [25] where this was done.

4) Improve the evaluation by using more diverse test data sets. Testing an algorithm on a single data set is not enough to establish how the algorithm can deal with different situations. The current data set only contains city roads/streets, with an emphasis on higher road classes. As a result, all characteristics of the data set are very homogeneous. First, the road density is more or less equal across the entire study area. That is, the capacities of the proposed approach to maintain density variations in the input data are not tested at all. However, we know from previous work (Weiss & Weibel, 2014, see References) that e.g. centrality-based algorithms have difficulty reducing the density of the road network in urban areas when urban and rural areas occur together in the same data set. Second, the junction angles are similar and typically near 90 degrees, making stroke formation almost trivial. Again, as Weiss & Weibel (2014) have shown, if roads in rural areas as well as roads and paths of very different classes are included, the standard threshold for the deflection angle of 60% will often not lead to optimal results, and variable angles have to be used. Since all stroke-related measures, however, depend on the quality of stroke formation, different data sets with differing density and junction angles would provide a more honest test of the proposed methodology. Typical features of test data sets should include: urban and rural areas covered, variable density, irregular shapes of strokes. It is clear that this cannot be met in a single data set; so, use several data sets.

5) Improve the evaluation by using better benchmarks. So far, only the manual generalization is used as a benchmark. However, since so many similar algorithms already exist, the results should also be compared to these. In particular, the authors should attempt to isolate the improvement in quality that is due to their main contributions, the AHP-based ranking and the inclusion of POIs. Many different baselines would be possible, but I suggest the authors use at least the following algorithm as a baseline: stroke formation as in the proposed approach; edge betweenness centrality as the only attribute used for ranking of strokes; and connectivity maintenance as in the proposed approach. This baseline would implement what can be typically expected from a state-of-the-art stroke-based road selection method, and it would allow to isolate the effect of adding (a) AHP-based stroke ranking and (b) POI-based stroke ranking.

6) No generalization algorithm is valid for all scale ranges, including road network selection. Typically, in road network selection, there is a break at around 1:200k or 1:250k. Combined line-mesh algorithms [24, 25] work best for scales between 1:10k and approximately 1:200k, while stroke-based algorithms such as the one presented here seem to work best for scales smaller than 1:200k or 1:250k. It is not clear what scale range the proposed approach was designed for. Also, the relationship of the numerical reduction factors (i.e. proportion of strokes retained) used in the experiments to map scales is not revealed.

7) I suggest adding the paper by Weiss & Weibel (2014, see References below). That paper proposes a stroke-based method with the following capacities: stroke formation with variable deflection angle thresholds; adaptation of selection thresholds in response to network density; an alternative connectivity maintenance method; and quantitative and qualitative evaluation on multiple data sets.

8) P1L35: Add Reference [24] by Li & Zhou (2012) at the end of this sentence. They cover exactly this.

9) P2L46-48: “Because the semantic information of roads tends to be incomplete and inaccessible, studies in road selection is phasing out of semantic characteristics in recent years” -- This sentence is not clear. Language?

10) P2L68 and following: Explain why AHP is required, and why this is better than, for instance, using machine learning (trained on manually generalized data) to find adequate parameter weights and thresholds. (This is also related to Comment 2.)

11) P2L76-77: “However, it is unable to handle inherent uncertainty due to the neglect of quantitative information” -- This sentence is not clear. Please clarify.

12) P4L135-136: “The deflection angle threshold within the range of 40º ~ 60º is optimal” – Some authors have used variable thresholds (e.g. Weiss & Weibel 2014).

13) P4L143-144: What do you mean by “high practicability”?

14) I recommend removing Table 2 and restricting the coverage of the stroke measures to references to the literature (e.g. Brandes 2001, see References below). These network measures are well known.

15) P5L154: What are “build-ups”? Language?

16) P5L171: “According to existing studies and standard specifications” -- Would that not differ depending on the type and size of the POIs? 30 m may be fine for a corner store but too small for a shopping center.

17) P6L185-203: I recommend to shorten the description of the Delphi method, which is rather well-known.

18) P7L222: “In the mode of AHP” -- What is meant by “mode” here? Do you mean “model”? Language?

19) P7L224: “but also it is in line with human cognition” -- Add a reference that supports this claim.

20) P7L225-226: “L_s, k, C_B, C_c, P” -- I know you have introduced these notations above, but this is the first time they are mentioned together, and hence I recommend to briefly repeat what the mean (i.e. mention again the names).

21) P8L242: “scope of [0,1]” --> “interval of [0,1]”

22) P8L271-274: This is where references to work investing into stroke formation would be useful, such as [35] and Weiss & Weibel (2014).

23) P10L306: “regarded as a separate tree” -- Do you mean “as the root of a separate tree”?

24) P10L312: “The step ends until there are no white leaf nodes in ?? _(Figure 4(c)).” -- I still see a white leaf node in Figure 4(c).

25) Figure 5 is too small to see anything clearly. This hold also for Figures 6, 7 and 9.

26) In Figure 6, you select a proportion of 0.01, 0.1, 0.2 and 0.4. Does that translate to 1%, 10%, 20% and 40%? Also, what is the relationship of these proportions to approximate map scales? For instance, if I select 1’%, what map scale would that approximately represent?

27) P11L333: “The relative density distribution of the generalized networks in different scales can be reflected” -- How can you say that? See also my comments regarding evaluation, in particular Comment 5.

28) P11L336-337: “Generally speaking, our method can meet the basic requirements of automatic generation of road network.” -- But can it meet these requirements better than other methods that also meet these requirements? Again, see my Comment 5 above.

29) P12L346-347: You mention street names. Please use detail maps (zoomed in) and highlight these streets, as well as other interesting cases that support your discussion.

30) P12L350-354: Again, you are making claims that cannot be conclusive unless you have evaluated your methods more broadly (Comments 3 to 5).

31) In Figure 7, what is the unit that is used in the legend?

32) Table 5: Who defined the categories for the Degree of Similarity, and how? Who chose the labels? (“Extreme similarity” sounds a bit awkward, for instance.)

33) Figure 8: Using degree centrality as the basic measure for similarity assessment, even a simple stroke-based algorithm using edge betweenness (as e.g. in [4]) might have resulted in high similarity. You MUST compare your results to other algorithms. See Comment 5 above.

34) Figure 9: Add detail maps to better present and discuss these results (see Comment 3).

35) P14L417-419: The proposed future work is restricted to fine-tuning only. Work of this nature would not warrant further publication(s).

 

REFERENCES

Weiss, R. & Weibel, R. (2014): Road Network Selection for Small-Scale Maps Using an Improved Centrality-based Algorithm. Journal of Spatial Information Science, 9: 71-99. https://doi.org/10.5311/JOSIS.2014.9.166

Brandes, U. (2001): A faster algorithm for betweenness centrality. The Journal of Mathematical Sociology 25, 2, 163–177. https://doi.org/10.1080/0022250X.2001.9990249

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Congratulations on very interesting and applicable for the multiple scenarios research results.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

This is much better now! I still have a few minor points that will need to be improved, however:

1) Some effort has been spent on language editing, but I’m afraid the manuscript still needs further improvements, particularly in the text that has been amended (red font). A few examples:

- L47: “in respond to” --> “in response to”

- L63: “is the point-like representations” (singular or plural?)

- L90: “build” seems to be some copy/paste left-over

- L123: “formula (2) and (3)” --> either use the plural “formulae” or use “equations”

- L150: “is a China’s provider” ??

2) L29-30: The first sentence is not really needed (but it would not hurt if it stayed).

3) L30: Change this sentence to “Road selection is one of the main operations in topographic map generalization.” -- The reason being that depending on the scale, building generalization might actually be more important. Also, road selection is important for topographic not thematic maps.

4) L125-126: The numbering and formatting of Equations (2) and (3) is somehow mixed up.

5) L176: What you call the “dual graph” is actually called a “line graph”. (It is a dual graph to the primary road network, but the technical term is “line graph”.)

6) Figure 2: It’s hard to see for the reader how the dual line graph relates to the primary graph of the road network. In fact, it seems almost as if the line graph was flipped (i.e. mirrored) horizontally. I suggest to superimpose the line graph on the primary graph, so it becomes clear how the edges are turned into nodes, and vice-versa.

7) P10: You use the density of Voronoi cells as a proxy for network density. But from which objects is the Voronoi diagram formed? Are you using the nodes of the road network intersections as seed points? Or the midpoints of strokes? Or are you even forming a line Voronoi diagram from the strokes? You need to say that.

8) Figure 7: Use solid lines for the pink ellipses. The pink color is so subtle that there remains sufficient contrast to the road network. If you use a dashed/dotted line style, the ellipses are hardly visible.

9) Figure 8: Label the three subfigures as (a), (b) and (c) and mention in the figure caption what these stand for. In fact, it took me quite a while to figure out what was displayed in the three subfigures.

10) Figure 11: What is meant by “disparate strokes with POIs” and “without POIs”, respectively? Disparate in what sense? What do they differ from? Please be more specific about this in the accompanying text.

11) Figure 12: Same thing here. What does “disparate” mean in this case? Please provide a clear explanation in the accompanying text.

12) Figure 12: Same as in Figure 7, don’t use dashed or dotted line styles for circles/ellipses. Use solid line styles. Particularly the orange ellipses are currently hardly visible.

13) L510-511: “the extra computing time … can be ignored”. What do you mean by that? And why can it be ignored? You don’t even tell us in Section 3.2.2. how much extra computational effort is incurred.

14) Please check all references. Reference [12] now contains two references (Thomson/Richardson and Jiang/Harrie). Same for reference [25], which currently contains three references (Tomko/Winter/Claramunt; Li/Zhou; and Benz/Weibel). Possibly, there are other references that got mixed up. As a result, the numbering will probably be wrong throughout the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.docx