Mapping Public Urban Green Spaces Based on OpenStreetMap and Sentinel-2 Imagery Using Belief Functions

Round 1

Reviewer 1 Report

Well written paper, dealing with the fusion of available open access remotely sensed data (Sentinel-2 images) and OSM data. The authors present with all relevant tables and equations their proposed methodology for mapping urban green spaces. They implement the Dempster-Shafer theory for this purpose at the Dresden city. 

Overall the paper, contributes to the direction of further elaboration of open access data, based on existing and known algorithms (i.e. Dempster-Shafer theory), and for this purpose i recommend its publication. 

I have only three comments that i would like to authors to reply:

1) The level of processing of the Sentinel-2 images (i.e. were the images radiometrically and atmospherically corrected?)

2) please replace the NDVI equations with reflectance at the NIR and RED bands

3) Please discuss the novelty of this paper in comparison with this one

Ludwig, C.; Hecht, R.; Lautenbach, S.; Schorcht, M.; Zipf, A. Mapping of Public Urban Green Spaces based on OpenStreetMap and Sentinel-2 imagery using Belief Functions: Data and Source Code, 2020. 756 oi:10.11588/data/UYSAA5.

The Python source code is missing in this link.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Review Summary

The author proposes a technique to map public urban green spaces by fusing Sentinel and OSM data using Dempster Shafer theory and a Bayesian model. The author claims 95% accuracy of their proposed technique. The study area is in Dresden City Center. A 2- 2-stage Dempster-Shafer fusion is applied initially at OSM and Sentinel Imagery and later fuses publicness and greenness. The authors used a Bayesian hierarchical logistic regression model to identify the publicness from OSM, while greenness is calculated using NDVI.

Proofread errors

• Page 2, line 29 extra '-'
• Page 2, line 35-40... multiple 'but' usage makes the sentence monotonous. It can be replaced with a better connector.
• Page 5, line 172 population -> 563,011 instead of 563.011

Strength

• Interesting and timely problem in OSM space.
• Quality presentation and language is easy to understand.
• Good motivational example discussed regarding issues with 'public green spaces', Green Urban Areas, and datasets.
• The author claims Dempster-Shafer's theory is used for the first time to fuse Sentinel and OSM data to map green spaces.
• Comprehensive literature review and background information.
• The code is publicly accessible for review.
• The approach is validated using Aerial Imagery of 40 cm resolution.
• Good approach to create land use polygons by identifying city blocks using traffic networks and creating sub-blocks based on land use information.
• The greenness approach is robust, which is evaluated by identifying peak NDVI across the 2019 year and later clustered using fuzzy c-means.
• The results demonstrate the efficacy of the proposed technique.

Suggestions for improvements

1. Page 2, Line 72-77: The Research Question is more thesis centric style. I would suggest aligning the RQ's with challenges/problem requirements (such as handling uncertainties during fusion).
2. Before related work, it would be good to add explicit contributions (techniques, models, etc. ) that authors have proposed in the paper for the above challenges/RQ's.
3. What is 'B' in equation (2). Is it a set of all other classes other than A (looks like that after reading eq. 3). Similarly, what is 'C' in eq. 4,5. It would be good to clarify the same.
4. Page 8 Line 254 -255. How common green space polygons are merged. Is this is an automated data-driven approach or is it done manually after identifying the common spaces? Clarification is required.
5. Figure 3 (c) Legends for Land Use Information are missing while the same is shown in 3(a).
6. It would be good to give an example to explain how the probability masses for greenness is calculated in Figure 5. I found it a little cumbersome while analyzing it from Fig. 4(a), shift values.
7. Table 2 needs more clarification behind the rationale of selecting 's' values. What do you mean by manually adjusted there?
8. What aerial imagery is used for validating greenness?
9. What resampling strategy is used for creating 1 m resolution. Is superpixel resolution or any other technique is used for this? Clarification required?
10. Why the grey area is more in the extreme left middle part in Fig 8(c), which is not the case in The Sentinel 8(a)and OSM 8(b) greenness classification.

The author solves an interesting problem in OSM and mapped public green spaces by fusing Sentinel data using the Dempster-Shafer technique and strongly validated the results. The overall presentation and organization of the paper is brilliant, and  I would recommend the paper to be accepted after minor suggested revisions.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Christina Ludwig, Robert Hecht, Sven Lautenbach, Martin Schorcht and Alexander Zipf,

Thank you for your submission to the ISPRS International Journal of Geo-Information and your associated effort. Your paper presents in clear, understandable terms a method to fuse information derived from Sentinel-2 and OSM data for the mapping of public green spaces. The topic is motivated convincingly and the objective to compensate the limitation of both data sources by fusing information is explained comprehensibly. The introduction to the underlying Dempster-Shafer theory is appropriately summarized. Furthermore, the proposed approach is presented in a comprehensible manner. The validation of the estimates of ‘greenness’ and ‘accessibility’ as well as the evaluation of the final map using an aerial imagery and manually annotated land use polygons, demonstrate the added value of the approach for a test area in Dresden.

However, I have several major and some minor points that I would like to see addressed in a revised version of the paper.

1. Transferability of the approach to other geographic areas

As already indicated in your paper in section 6 (Discussion), the success of the proposed approach depends to a large extent on the level of completeness of the OSM data as well as the selection of suitable contextual indicators and their coverage in the data. In particular, the contextual indicators 'benches' and 'playground' could affect the generalizability of the model, as they may be only sparsely represented in OSM data in other areas. In order to assess the transferability of the model, and since the evaluation on an independent different data set will not be feasible, I propose the following approaches:

1. 1. An independent validation of the estimations about the accessibility for polygons that belong to the class ‘public’ but do not contain a ‘bench’ or ‘playground’ tag. Which accuracy can still be achieved to identify such polygons as ‘public’ using the proposed model?

1. 1. A recalculation of a Bayesian model, which does not include the two contextual indicators ‘benches’ and ‘playground’. It should be considered to what extent this modification of the model affects the final mapping and in which cases a successful mapping is no longer possible

2. Use of the access-tag

My second major criticism is the erroneous behavior of the model with regard to the estimation of accessibility for polygons, which have the unambiguous access=* tag. This tag already explicitly specifies the membership of the polygon to the class public or private. I strongly suggest, to check this tag as part of a preprocessing step and use it as a unique feature. Since the access-tag is cosequently omitted as contextual indicator in the following processing, the Bayesian model is affected in a direct way, so that a recalculation is necessary.

In addition to these points, I noticed a few more issues which, however, only require minor changes:

3. Confusion between the terms set and class

In Section 3, the Dempster-Shafer theory is summarized precisely and understandably. However, in some parts (line 127/128 and line 138) there occurs some confusion between the expressions set and class. In line 125, A is introduced as one of the mutually exclusive classes. Later in line 138 and 139, A is referred to as set, denoting a union of classes.

4. Description of Equation (3)

Equation (3) is described incorrectly. It is stated: "The plausibility of a set A is defined as the sum of all probability masses of sets which contain the set A". However, the correct description of the equation would be: "The plausibility of a set A is defined as sum of all the masses of the sets B that intersect the set A".

5. Overlapping OSM polygons

In Section 4.3.1, the method for generating homogeneous land use polygons is outlined. You explain in which manner overlaps are handled, if a polygon is located completely within another polygon. But what happens if polygons overlap only partially? In that case, would the overlapping area be transformed into a new polygon? What properties would be assigned to this new polygon? Outside the scope of the paper, I am personally interested what happens for areas where no tagged OSM polygon is present. Will such areas be assigned a probability mass m_osm({green, grey}) = 1.0?

6. Grouping OSM tags into land use classes

Further, in Section 4.3.1, for the description of the rule based approach to group OSM tags into land use classes, reference is given to Figure 5. However, the table shown there is not consistent with the previously mentioned example, which states that both tags landuse=forest and natural=wood are assigned to the class forest. In Figure 5, forest and wood form separate classes with different basic probability assignments.

7. Motivation for the choice of fuzzy c-means clustering

To determine appropriate thresholds for NDVI values based on which the basic probability assignment for ‘greenness’ is performed, you propose the use of fuzzy C-means clustering. At this point, I am missing a short motivation (one sentence would suffice here) why this particular method is applied. If I understand the approach correctly, the clustering is only used to determine the tresholds h-high, h-mixed and h-grey, but not to assign a probability of class membership to each NDVI value directly. Therefore, I would like to see a short explanation why a soft clustering method is chosen over a hard clustering method, such as k-means.

8. Priors of the Bayesian model

One last point is the seemingly arbitrary choice of the priors of the Bayesian model. The origin of the assumptions about the distribution of the coefficients of the contextual parameters and the intercept has not become clear to me. Are the corresponding distributions obtainable from the data at hand? I would like to see a short reasoning for the chosen parameters/distributions.

I hope my notes and criticisms are comprehensibly formulated and a helpful input for a revision. Since I consider your contribution very interesting, both in terms of the idea of fusing OSM and remote sensing data, and the methodological approach, I look forward to reading a revised version of your submission.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Dear Christina Ludwig, Robert Hecht, Sven Lautenbach, Martin Schorcht and Alexander Zipf,

Thank you for this revised version of your paper. In my opinion, the paper has improved considerably and several points have become more comprehensive. However, I would like to revisit some of my comments from the last round and how they have been answered and implemented. Since you already have all of the data needed to address my concerns, I would consider this a minor revision, which is what I suggested to the editors.

To: 1. Transferability of the approach to other geographic areas

I am glad to see my suggestions towards the analysis of transferability of your approach tested based on a recalculated Bayesian model. However, instead of just mentioning overall accuracy in the discussion, I would like to see a more in-depth consideration within the result section. In particular, it could be addressed where the decrease in accuracy from 95% to 90% originated. To this end, I suggest the following:

• Select a sample region that differs in the overall result of identified public green spaces between the two models (with and without ‘benches’ and ‘playgrounds’)
• For the selected region: present the belief in public green (similar to Figure 11) for both models, so that the reader can assess the impact of the benches and playgrounds on the overall accuracy directly and possibly find substitutes in his/her region of interest.
• To the same end, I would like to see such a case discussed in detail (i.e. how much does the belief in public green decrease and why is this so… e.g. because the other indicators for public green such as footpath intersections are not sufficiently present?)

To: 3. Confusion between the terms set and class

There is still confusion in the use of the terms sets and classes. Expressions in the Dempster-Shafer theory (mass, belief, plausibility) refer to sets, not to classes. I would like to explain the issue on one example:

In line 134ff you write: ‘The amount of evidence which speaks for an object belonging to a certain class is called a belief. However, in the Dempster-Shafer theory, the belief function does not measure membership to a particular class. Instead, the belief function quantifies the amount of evidence given to a set (which consists only in exceptional cases of only one class). This misconception continues throughout the section (line 146, line 150 and line 164).

Furthermore, the same issue arises in Section 4. The probability mass functions should always be assigned to sets, e.g. m({green}), m({grey}), m({green, grey}) and not to classes, e.g. m(green) (see line 300, 306, 308 etc.).

To: 8. Priors of the Bayesian model

You explain the choice of the priors by: ‘The sensitivity of the model towards different priors was evaluated to ensure model convergence and reasonable model results.’ However, this still does not make clear to me how you arrive at the assumption of normal distributions and a Half Cauchy distribution of the model parameters. And, since this choice of distribution is crucial to the Bayesian model, I think this should be made clear to the reader.

Beyond these points, I noticed some further minor issues that should be improved:

1. Correction of Equation 1

Correction:  0 ≤ m(A) ≤ 1;...

2. Unassigned references

In line 112ff and 237 ‘[?]’ occurs instead of the reference number.

3. Inconsistent Referencing of Figures, Tables and Sections

Both abbreviations and full names are used (e.g. fig/figure), as well as upper and lower case (fig/Fig).

4. Inconsistency in the use of italic font

Sometimes class names (e.g. green, grey, public, private) are written in italics and sometimes normally. The same applies to units. It would be helpful if you could find a consistent usage here.

5. Axis label in Figure 12

The y-axis is labelled by accuracy in %, however the value range is [0.93, 1.0].

Overall, the paper is already very appealing and the presented approach should gain much attention in the scientific community. In order to improve the paper, especially with regard to the validation of the transferability, even further, I would ask you to consider my remaining issues in a second revision of the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.docx