Urban Green Space Mapping from Sentinel-2 and OpenStreetMap via Weighted-Sample SVM Classification
, Liangqing Wu 1, Anhao Zhang 1, Xianfang Yang 1, Xiujuan Li 1 and Chaoyun Chen 2
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsManuscript ID: remotesensing-4064718
Title: Urban Green Space Mapping from Sentinel-2 and OpenStreetMap via Weighted-Sample SVM Classification
Authors: Bin Yuan , Zhiwei Wan , Liangqing Wu , Anhao Zhang , Xianfang Yang , Xiujuan Li , Chaoyun Chen
This study presents a clear and technically robust framework for mapping large-scale urban green spaces in the Guangdong Hong Kong Macao Greater Bay Area. It makes a significant contribution by integrating data from multiple sources and employing weighted machine-learning techniques to yield consistent, transferable results across diverse urban environments. This work is highly relevant for both remote sensing research and urban ecological planning. Hereunder are some comments, hoping to enhance the paper:
First: Major Comments
- Explain Novelty More Explicitly: Although the study presents a fully automated framework for UGS maps at the multi-city level, the novelty remains implicit. It is recommended to clarify:
- What specifically distinguishes this framework from previous Sentinel-2 + OSM studies?
- Is the novelty in the sample weighting strategy, the cross-city balanced training, or both?
- It is recommended to add an explicit paragraph at the end of the introduction or the beginning of the discussion that clearly explains the scientific contribution.
- Independent Validation: The current evaluation relies on samples taken from the same OSM system used for this. This may raise questions about circular validation bias.
- Recommended:
- Clarify whether the validation samples are spatially and graphically independent of the training data. Or explicitly discuss this limitation as a methodological boundary in the Discussion section.
- Justify the choice of SVM compared to newer algorithms. The choice of SVM is theoretically justified, but given the prevalence of Random Forest, XGBoost, and Deep Learning, the reader might ask:
- Why wasn't the performance compared to at least one alternative algorithm?
- Even a limited benchmark on a single city would strengthen the methodology.
- Provide more precise details about the weighting scheme. The description of the multidimensional weighting scheme is good, but it remains more descriptive than analytical.
Second: Minor Comments
- Improving the wording in some long sentences: There are relatively long sentences in the Methods and Discussion sections that could be shortened to improve clarity without compromising the scientific content.
- Standardising terminology: Sometimes used: UGS maps, green space maps.
It is preferable to use consistent terminology throughout the manuscript to avoid ambiguity.
- Maps (Figures 3–5)
It is recommended to add: A standardised scale bar or a more precise explanation of city boundaries in the detailed maps
- Minor typographical errors
There are some minor errors (spacing/hyphenation), such as: ShenZhen → Shenzhen
The Greater Bay Area is sometimes written inconsistently.
- The manuscript language is clear, but some sentences need further proofreading and review, particularly for clarity.
Overall Assessment of the Manuscript
The study is methodologically sound, well-organised, and of high practical value, making it suitable for the journal Remote Sensing. With major comments addressed, particularly regarding novelty and verification, the manuscript is ready for acceptance after revisions.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis study focuses on the GBA and proposes a UGS mapping framework that integrates Sentinel-2 annual time-series imagery with OSM data. The primary innovation of this research lies in the introduction of a multi-dimensional sample quality assessment and weighting strategy, utilizing a Weighted SVM classification to effectively overcome challenges such as inconsistent OSM tagging quality and sample noise. The research findings hold significant academic reference value. The specific comments and suggestions are as follows:
1. The manuscript states that the comprehensive quality score is classified into three levels—high, medium, and low—with each level assigned a numerical reliability weight. Please clarify the specific numerical values assigned to these weights(0.9?0.6?0.3?) and the rationale for their determination. Furthermore, have these manually assigned weights undergone sensitivity analysis?
2. The study utilizes a Weighted Support Vector Machine (W-SVM) with an RBF kernel. When processing large-scale remote sensing data, Random Forest (RF) often demonstrates superior robustness and computational efficiency compared to SVM. Please provide the technical considerations for choosing SVM instead of algorithms such as RF or XGBoost, or include a brief performance comparison.
3. The authors mention using a "city-level quality factor" derived from OpenStreetMap (OSM) completeness metrics to adjust sample weights. For cities where OSM data is extremely sparse, such as some of the smaller cities, will this weighting strategy lead to a model dominated by data-rich cities like Guangzhou or Shenzhen? Please explain how the "balanced multi-city training dataset" ensures that the unique characteristics of cities with sparse data are not overshadowed in the global model.
4. While Table 2 provides Accuracy, Precision, Recall, and F1 scores, a comprehensive global confusion matrix is missing. Please add a summary confusion matrix to analyze the primary land-cover types where "non-UGS" is misclassified as "UGS"—for example, distinguishing between errors caused by cropland, artificial turf, or building shadows.
5. Please conduct a comparison between the proposed results and existing global or regional products (such as ESA WorldCover 10m or FROM-GLC10) within the Guangdong-Hong Kong-Macao Greater Bay Area (GBA).
6. The study utilizes 10m resolution Sentinel-2 data. While 10m resolution is superior to previous medium-resolution products , it still faces challenges in highly urbanized regions—such as the Macao Peninsula—where street trees, roadside plantings, pocket parks, and courtyard vegetation are often smaller than a single pixel. It is recommended to provide a more in-depth discussion regarding the underestimation of UGS extent caused by the 10m spatial resolution in the "Discussion" section.
7. The manuscript acknowledges that shadows cast by high-rise buildings frequently depress reflectance values. The authors currently rely on a median composite approach to reduce contamination from shadows. Please clarify the extent to which this method effectively mitigates permanent or deep urban shadows cast by high-rise buildings.
8. The title and highlights refer to "Annual GBA UGS maps" , yet the research results and analysis primarily present data from the year 2020. Please clarify！
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis manuscript proposes an automated UGS mapping framework based on weighted SVM and OSM-Sentinel-2 fusion, which is validated in the highly heterogeneous urban agglomeration of the Greater Bay Area. It demonstrates strong technical feasibility and practical application value. The overall structure is clear, the methodology is systematic, and the conclusions are reasonable, but there are several areas for improvement before accepted:
- The abstract mentions 16 cities, while the study area description and figures indicate only 11 cities. This discrepancy requires further verification.
- P4 L126-128，what is the basis for constructing the external classification table? Has it been validated?
- Fig6, the highly clustered sample distribution may lead to overestimated accuracy. Was spatial cross-validation considered?
- Table 2,the F1 scores vary significantly. Further analysis should be conducted to explain these differences, such as urban structure or OSM completeness.
- The policy recommendations are somewhat generic. Could more specific spatial priorities be proposed based on the MSPA results? Additionally, the applicability of the method to other climatic zones and urban forms should be discussed.
- The units for area measurements in the text should be unified. Avoid using both square kilometers and hectares interchangeably.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe topic of the manuscript is relevant to the scope of Remote Sensing. While the overall methodology is sound, several minor issues should be addressed before acceptance.
The methodological components adopted in the study (spectral indices, SVM classifier, and MSPA) are well established in the existing literature. The main contribution of the paper therefore lies in the integration of these elements into a unified and automated multi-city workflow, rather than in methodological novelty.
The proposed sample weighting strategy is clearly described and logically motivated. Additional discussion on its quantitative impact on classification performance would further strengthen the manuscript.
In addition, some statements in the manuscript tend to overemphasize the capability of Sentinel-2 imagery at 10 m spatial resolution, whereas the Discussion section appropriately acknowledges important limitations related to shadow effects and the detection of small-scale vegetation. These claims should be more consistently balanced throughout the text.
Details on minor issues:
– Missing spacing or formatting issues are present at the beginning of several subsections (e.g., lines 184, 207, 209, 243, and 279) and should be corrected.
– In 2.3.3 Sentinel-2 data Feature Extraction (lines 184 to 206)
Clarification of the NDRE definition is required, particularly in relation to the subsequent use and definition of NDVIre, to avoid ambiguity.
– Some figures use color schemes that reduce interpretability. In particular, green is used to represent non-UGS in Figure 3, while green is used to represent UGS in Figure 4, which may be confusing for the reader. A more consistent color convention is recommended. (line 302)
Comments for author File: Comments.pdf
The English language is generally clear and correct, although some long sentences could be simplified to improve readability, particularly in the Discussion.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThanks for authors for enhancing the manuscript.
Reviewer 2 Report
Comments and Suggestions for AuthorsI have carefully read the revised manuscript and the authors’ point-by-point response. Overall, the revision has greatly improved the readability and interpretability of the paper, and the methodology and results are now presented in a much clearer and more convincing way. The proposed UGS mapping framework based on annual Sentinel-2 time-series imagery and standardized OSM data is meaningful and practical, with clear contributions in cross-city generalization, handling noisy samples, and potential for re-use in real applications. In general, the revised manuscript has reached a publishable level.
The authors have responded to the previous comments well, and the key methodological details have been strengthened. In particular, the rationale and usage of the reliability weights are explained more clearly than before, and the added sensitivity analysis shows that the overall performance does not change much within a reasonable range of weight settings. This indicates the method is relatively robust and reduces the concern about subjective parameter choices.
In terms of data and training strategy, the revision further clarifies how cross-city sample balancing and weight normalization are conducted, which helps avoid possible bias caused by large cities dominating the training data and by differences in city sizes. Meanwhile, the model selection is better supported by comparative experiments and error analysis, showing that the proposed method can effectively reduce important false positives (Non-UGS → UGS), which is more suitable for application-oriented mapping tasks.
In addition, the discussion section now gives a more careful and well-bounded explanation of limitations related to Sentinel-2 10 m resolution, shadow effects, and spatial autocorrelation, and the authors also add comparisons with existing land-cover products. These improvements provide useful external reference and further enhance the reliability of the results.
