Global-PCG-10: a 10-m global map of plastic-covered greenhouses derived from Sentinel-2 in 2020

Manuscript number: essd-2024-538

General comments.

This very interesting paper use machine learning and deep learning on Sentinel-2 10 m GSD images to obtain a global map of plastic-covered greenhouses (PCGs). Really, it is not the first global PCG maps since Tong et al. (2024) already published other global PCG map derived from PlanetScope images, a commercial satellite with 3 m GSD, but using also in the first steps Sentinel-2. Although both works have a similar objective (i.e., to attain a global PCG map), the strategies used were quite different.   

Thanks to the attained global PCG map, Niu et al. (2025) give interesting data about the area of PCG around the world, the major concentrations, spatial distribution, etc.

The manuscript is well written and it is worth being published. However, a few specifics comments should be taken into account.

Specific comments.

- Some cites in the manuscript appears with an extra comma. For example, in Page 2, Line 64, the cites “Aguilar et al., (2016) and Yang et al., (2017) independently developed…” should be “Aguilar et al. (2016) and Yang et al. (2017) independently developed…” Similarly, “Zhang et al., (2022a)” in Page 2, Line 66, should be “Zhang et al. (2022a)”. Please, correct this issue throughout the manuscript.

- Page 3, Line 71. The cite (Zhang et al., 2024) should be (Zhang et al., 2024a).

- Page 3, Line 76. Zhang et al., 2023 should be 2023a or 2023b. Please, review it.

- Page 5, Line 139-140. “Actually, Sentinel-2 is a constellation consisting of two satellites, i.e., Sentinel-2A and Sentinel-2B, which are in the same sun-synchronous orbit while phased at 180° to each other”. In fact, there is a new Sentinel-2C. You should speak a little about it.

- Page 6, Line 163. In Figure 2 (Stage 2) the train/validation ratio is 7:3, and in the manuscript you wrote 8:2. Is it a mistake in the Figure?

- Page 7, Line 175. In the caption of Figure 3, you should clarify that the size of the reference samples (512×512) are pixels and not meters.

- Page 9, Line 200. In the caption of Figure 4, it is written Multiple-temporal NDVI. Is not more appropriated multi-temporal NDVI?

- Page 9, Line 205. “1> Spectral features”. Strange login method.

- Page 10, Line 212. “2> Textural features”. Strange login method.

- Page 11, Line 237. You should clarify also in the manuscript that the size of the reference samples (512×512) are pixels and not meters.

- Page 12, Line 273. In Figure 2 (Stage 2) the train/validation ratio is 7:3, and in the manuscript you wrote 8:2. Please review it.

- Page 13, Line 286. Fu et al. (2021) is not in reference section.

- Page 15, Line 332-335. There are some numbers without thousands separation (e.g., 9874.51 km2, 2530.56 km2, 8224.90 km2).

- Page 16, Line 344. Figure 8a is not cited in the manuscript, and it should be.

- Page 18, Line 375. Why 20500 points for GH and 20500 for Non-GH. Justify this figure.

- Page 19, Line 381. Table 1 shows the confusion matrix where OA, User Accuracy (UA) and Producer Accuracy (PA) are depicted. Really, UA=Recall and PA=Precision, so, Table 2 is not necessary. The only data useful in Table 2 is F1 Score. I think that you should rewrite the methods and results about the accuracy assessment. Furthermore, Why is the classification so biased? For example, UA is 99.99% and PA is 86.30% for Non-GH and, UA is 84.18% and PA is 99.99% for GH.

- Page 22, Line 421-422. “… and in May 2024, the University of Copenhagen published a global 3-m PCGs dataset also in 2019”. Please, you should cite Tong et al. (2024) here.

- Page 22, Line 430. You should cite Tong et al. (2024) properly in the caption of Figure 12.

- Page 23, Line 434. “Tong et al., (2024) acquired from 3-m …”. Again, this cite appears with an extra comma.

- Page 30, Line 667. “Zhang, X., Liu, L., and Chen, X.: Global annual wetland dataset Data Descriptor at 30 m with a fine classification system from 2000 to 2022, Sci. Data, https://doi.org/10.1038/s41597-024-03143-0, 2024c”. This reference do not appear in the manuscript.

Final Comments:

It is very important that the global PCG map and the code are accessible to researchers. I have tested that the code for generating the initial labels of PCGs is publicly available via the following link on Google Earth Engine: https://github.com/MrSuperNiu/Greenhouse_Classification_GEE. It consists of feature extraction, RF classification, etc. Additionally, the code of APC-Net is accessible through the following link: https://github.com/MrSuperNiu/APCNet. The Global-PCG-10 dataset is stored on figshare, and can be downloaded here: https://doi.org/10.6084/m9.figshare.27731148.v2 (Niu et al., 2024).

The manuscript presents the development of a global plastic greenhouse distribution map (Global-PCG-10) based on Sentinel-2 10 m imagery. Weak labels were generated via a random forest classifier, and these were refined through a deep learning framework integrated with an active learning strategy. The spatial patterns, extent, and proportional coverage of global PCGs were analyzed, yielding results of clear scientific merit and practical utility. However, several issues and deficiencies must be addressed prior to publication:

1. Some in-text citations do not conform to the journal’s formatting requirements. For example, in line 64 the citation “Aguilar et al., (2016)” incorrectly includes a comma. Please review and standardize all reference notations.

2. In Figure 2 (Phase 2) the train/validation split is stated as 7:3, whereas in the main text it is described as 8:2. Please verify and ensure consistency in both places.

3. In lines 162-163 authors state that the initial labels were generated via GEE and RF classification. Was any quality control applied to these labels (e.g., manual verification rate, error‐rate assessment)? Please clarify.

4. In lines 165-166 (Stage 3) mention “post-processing was applied to the PCGs classification results to eliminate isolated noises” without specifying the algorithm or parameter settings. Please provide details of the post-processing method so that the results are reproducible.

5. The manuscript alternates between “PCG” and “PCGs” (e.g., line 166 “PCGs classification results” vs. Figure 2 “PCG classification results”). Please unify the terminology throughout.

6. In lines 239-243 outline the overall network architecture but omit implementation details of the MDCN and non-local modules. Schematic diagrams or detailed descriptions of these submodules are required.

7. Section 3.2.2 provides an overview of the active‐learning strategy but lacks specifics on the number of iterations, per-iteration sample‐selection criteria, and stopping conditions. Please supply quantitative stopping rules rather than the current qualitative phrase “until performance stabilizes or results are satisfactory.”

8. In lines 253-254, authors state that “if results are unsatisfactory, significantly altered initial labels are selected for further training,” but “expected standards” and “significant alteration” are not quantified. Please define the thresholds and the human‐involvement workflow in detail.

9. In lines 270–275, authors note the use of the Adam optimizer and data partitioning but omit critical hyperparameters such as total epochs, batch size, and learning‐rate decay strategy. Please provide these training details.

10. In lines 272-273, the dataset is described as comprising 14 825 training images and 3 707 validation images. Please explain how the sample was divided (randomly or by region? Is global diversity guaranteed?).

11. In line 292, the term “false positives (FP)” is incorrectly written as “false negatives.” This typo may confuse the metric definitions. Please verify that all terminology and formulas in the text and tables are consistent.

12. Section 4.2 presents only examples with no significant false alarms, which seems inconsistent with the overall recall rate of 84%. Please include error‐case analysis and discuss the causes of missed detections.

13. In lines 374-375, authors apply balanced sampling between PCG and non‐PCG in the test set, which does not reflect their real-world scarcity and may lead to overestimated performance. Please discuss the impact of balanced sampling on evaluation or provide results under the true class distribution.

14. In lines 381-382 the confusion matrix shows that the PCG class has a UA of 84.18% (FN = 3 243), yet no discussion of the false-negative causes is provided. Please analyze the reasons for these missed detections.

15. In lines 381-382, non‐PCG exhibits UA = 99.99% and PA = 86.30%, while PCG shows UA = 84.18% and PA = 99.99%, indicating a strong asymmetry in false‐positive and false‐negative distributions. Please discuss the origin of this bias.

16. In section 4.3.2, authors compare visually against Tong et al. (2024) using 3 m data (Figure 12) but lack objective quantitative metrics. Please add quantitative evaluation and report specific comparative values.

17. Recommend including a discussion of potential applications and the limitations of the produced dataset.