An improved global land cover mapping in 2015 with 30&thinsp;m resolution (GLC-2015) based on a multisource product-fusion approach

Li, Bingjie; Xu, Xiaocong; Liu, Xiaoping; Shi, Qian; Zhuang, Haoming; Cai, Yaotong; He, Da

doi:https://doi.org/10.5194/essd-15-2347-2023

Articles | Volume 15, issue 6

https://doi.org/10.5194/essd-15-2347-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/essd-15-2347-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 15, issue 6

Data description paper

|

07 Jun 2023

Data description paper |

| 07 Jun 2023

An improved global land cover mapping in 2015 with 30 m resolution (GLC-2015) based on a multisource product-fusion approach

Bingjie Li, Xiaocong Xu, Xiaoping Liu, Qian Shi, Haoming Zhuang, Yaotong Cai, and Da He

Download

Final revised paper (published on 07 Jun 2023)
Supplement to the final revised paper
Preprint (discussion started on 03 Aug 2022)
Supplement to the preprint

Interactive discussion

Status: closed

RC1:
'Comment on essd-2022-142', Anonymous Referee #1, 02 Sep 2022

Please find my comments in the supplement.

Citation: https://doi.org/10.5194/essd-2022-142-RC1
- AC1: 'Reply on RC1', Bingjie Li, 07 Nov 2022
  
  We are particularly grateful for your constructive comments. Please find the attached file of our point-by-point responses.
  
  Citation: https://doi.org/10.5194/essd-2022-142-AC1
RC2:
'Comment on essd-2022-142', Anonymous Referee #2, 04 Sep 2022
This paper developed an improved global land cover map at 30m resolution in 2015 by fusing multi-source products of land covers and other thematic mappers. Two sets of global samples with points and patches have been developed and used to evaluate the performance of derived GLC-2015. This work is high-intensive in terms of the labor involved, and the evaluation is sound with clear logic. Before recommending it for publication, I raised several concerns below, which might be helpful to improve this paper.

(1) Although the authors adopted the DSET approach to generate the GLC-2015 product and compared it with similar products such as FROM-GLC and GLC_FCS30, the improvements gained from the DSET approach should be highlighted in those common approaches such as major voting and other common approaches. Otherwise, the highlights of the DSET in the manuscript should be reconsidered.

(2) How about the mapping performance if using these samples (80%) do the classification directly? Because these samples have been manually visualized and are qualified for the classification task. Please add some test results or discuss this issue in the manuscript.

(3) The proposed work can also be applied in regions with adequate and high-quality data, such as NLCD in the US and China. This can be improved or discussed in the revised manuscript.

Minor Comments:

Page 108: BPA function. This term should be fully spelled when it first appears in the main text.

Page 172: The selection of 4°4° should be discussed.

Page 178-179: I wonder why the initial samples generated from the FROM_GLC were used in this study, not other land cover products. Explanations about this topic should be discussed in the manuscript.

Page 198: Why select these 1507 samples randomly? Can they be determined according to their ecoregions or cover types? It is better to explain it here clearly.

Fig. 3 and Fig.4 can be combined. Also, the main text uses “Fig” whereas the figure caption uses “Figure”. Please make them consistent.

Page 281: how to determine these two thresholds: 25% and 75% in Eq. (4). Please explain.
Citation: https://doi.org/10.5194/essd-2022-142-RC2
- AC2: 'Reply on RC2', Bingjie Li, 07 Nov 2022
  
  We are grateful for your comments and suggestions. Please find the attached file of our point-by-point responses.
  
  Citation: https://doi.org/10.5194/essd-2022-142-AC2

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

AR by Bingjie Li on behalf of the Authors (09 Nov 2022) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (11 Nov 2022) by Hanqin Tian

RR by Anonymous Referee #2 (26 Nov 2022)

RR by Anonymous Referee #1 (09 Dec 2022)

Suggestions for revision or reasons for rejection

The manuscript has been improved in terms of additional patch-based validation and asset name prefixes. However, due to the following concerns, I am still not convinced that this dataset and manuscript could be candidates for ESSD.

The method is defective when input GLC maps contain errors (see my comments below) and their reliability could not be fully evaluated by the samples. In such a case, the accuracy of some LC maps was overestimated, leading to the wrong contribution to the fusion. This can be confirmed by the misclassification of bare land from water (check my previous comments), where FROM-GLC and GLC_FCS30 identified a volcano as water. This also indicates that input samples are not representative.

The 4° grid is large in the case of global mapping, as many studies have employed smaller grid (such as the 1° grid in GAUD). Generally, smaller grids will help assess each input LC map's accuracy (e.g., geographical variations). Therefore, it's suggested to investigate the relationship between grid size and mapping performance (e.g., global OA as a function of grid size), and find a proper size.

I understand the high quality of GlobeLand30. But I don't think the LC changes caused by the 5-year gap of GlobeLand30 can be largely avoided through the input samples. Due to the simple stratified sampling, these samples could rarely capture the LC changes during the interval. Thus, the accuracy of GlobeLand30 (against the actual LC in 2015) will be overestimated. In such a way, the contribution of Globeland30 (when LC changes) to fusion would not be as small as the authors suggested.

It's not sound to determine the location of input samples by FROM_GLC. The results show that FROM_GLC has the lowest UA for shrubs, grasslands, and wetlands, suggesting a large number of omissions. In this way, the area ratio of these LCs will be underestimated, as well as their number within the input samples. Consequently, the mapping performance of the final map will be biased. This may be one of the possible reasons for the poor performance of these LCs in the GLC-2015. Perhaps the GlobeLand30 is a better choice, given its high quality.

The incorporation of GLC_FCS30 is problematic, as it adopted a detailed classification system (level-2) in some places. I think this will lead to geographical accuracy biases, which however have not been tested in the current assessment.

Finally, I'm not sold by the story that some problems (e.g., some LCs always possess lower accuracies and poor performance in areas with more disagreements) could be solved when more reliable maps are available in the future. It would be more useful for the community to tackle these issues here, wouldn't it?

Hide

RR by Anonymous Referee #3 (16 Dec 2022)

RR by Anonymous Referee #4 (19 Dec 2022)

Suggestions for revision or reasons for rejection

The other two reviewers have provided very professional suggestions for data and manuscript improvements. I fully agree with their opinions. Generally, I think this data has some value but I have two major concerns. The first is about the classification system used. This study assimilated various information from other land use data products. However, the definitions of different land types greatly differ in each classification system. For example, in your data, 'cropland' is defined as 'Land areas used for food production and animal feed.', which means pasture was classified as cropland. However, pasture is classified as grassland in GLC_FCS30. When all these signals were combined, the approach used could cause problems depending on the LC definitions adopted. Moreover, during the validation process, how this definition differences were treated? I think, for example for the cropland, it does not indicate that GLC_FCS30 was wrong (as claimed in Line565), but because of the differed definitions. Besides, it is unclear how you deal with fallow land, which is often been mixed with grassland in classifications (need to be separated by examine the temporal information).
Another example, in FROM_GLC, the forest was defined as tree cover >=10%, and GlobeLand30 defined forest as land with tree cover above 30% and also include sparse woodland with tree cover between 10%-30%. These are different from the fixed threshold of 30% adopted in this study. In other words, I do not agree that other datasets are 'wrong' (as claimed in Line565 and other places).

My second major concern is the current manuscript reads really arrogant. The authors' attitudes against other data products are not polite. I strongly suggest the authors tone down their remarks against other data products. As I said, the differences are partially from LC various definitions in each product. Please be moderate and modest. You actually built your data relying on these products. The words such as "wrongly", "worst performance" should be avoided, they are too harsh (see a few examples below).

Other suggestions:
1. Are the validation samples checked and used separately in 2015 and 2010? For example, the GlobeLand30 in 2010 was used, in which the validation samples of 2015 might not be appropriate to be used (as claimed to have been visually checked using google earth).
2. Line282, why 75% and 25% been chosen? what if these weights changed?
3. Line315, error matrix?
4. Lines453 and 462, 'worst' should be replaced by, for example, lowest.
5. Line537, 'overwhelming superiority' is too far. As I mentioned, it might because of the definition differences.
6. Line601, 'show great superiority over others', same as above.
7. Line585, 'capture most human activity'. I don't agree, 'capture the footprint of human activities' might be better.

Hide

ED: Reconsider after major revisions (16 Jan 2023) by Hanqin Tian

AR by Bingjie Li on behalf of the Authors (30 Mar 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (05 Apr 2023) by Hanqin Tian

RR by Anonymous Referee #1 (07 Apr 2023)

Suggestions for revision or reasons for rejection

This study developed a 30m resolution GLC product by integrating multiple products using the DSET method. Accuracy assessment with two validation datasets demonstrates the high quality of the GLC-2015. The comparison between GLC-2015 and other GLC products (Globeland30, FROM_GLC, GLC_FCS30) is comprehensive and the analysis is reasonable. This data is valuable and can provide accurate information for many applications. Some improvements are needed before publication.

1. The authors used three national-scale products to improve the quality of the GLC-2015 in China and America. However, it is unknown that whether the GLC-2015 performed better than these national-scale products. It is advised that the authors quantitatively compared the GLC-2015 with NLCD, CLCD and CLUD. Also, area difference for various LC classes in GLC-2015 and national-scale data can be analyzed.
2. National-scale land cover products, such as CLUD, NLCD 2016, have a two-level classification system, how did you transform these classification systems into the adopted classification system? This should be explained in the study.
3. The national-scale land cover datasets, such as CLCD for 2015, were used in the fusion, why these products were not listed in the framework (Figure 4)?
4. Figures S8-S11 exhibit visual comparisons for various land cover classes at local scale. However, the detailed locations of examples were not clearly showed. It would be better to tell readers the specific locations with graticules or central point as well as area size.
5. In abstract, “LC” should be fully spelled.
6. Line 116: do you mean “the final fused results based on the DSET method”?
7. Line 128: “are” should be “were”. And line 211: “resulted” should be “result”.
8. Some relevant papers may be reviewed in the references:
Mapping 10 m global impervious surface area (GISA-10m) using multi-source geospatial data. Earth System Science Data, 2022, 14: 3649–3672.
30-m global impervious surface area dynamics and urban expansion pattern observed by Landsat satellites: from 1972 to 2019. SCIENCE CHINA Earth Sciences, 2021, 64(11): 1922-1933.

Hide

RR by Anonymous Referee #4 (12 Apr 2023)

ED: Publish subject to minor revisions (review by editor) (12 Apr 2023) by Hanqin Tian

AR by Bingjie Li on behalf of the Authors (19 Apr 2023) Author's response Author's tracked changes Manuscript

ED: Publish as is (23 Apr 2023) by Hanqin Tian

AR by Bingjie Li on behalf of the Authors (28 Apr 2023) Manuscript

Post-review adjustments

AA: Author's adjustment | EA: Editor approval

AA by Bingjie Li on behalf of the Authors (30 May 2023) Author's adjustment Manuscript

EA: Adjustments approved (30 May 2023) by Hanqin Tian

Download

Article (13697 KB)
Full-text XML

Short summary

A global land cover map with fine spatial resolution is important for climate and environmental studies, food security, or biodiversity conservation. In this study, we developed an improved global land cover map in 2015 with 30 m resolution (GLC-2015) by fusing the existing land cover products based on the Dempster–Shafer theory of evidence on the Google Earth Engine platform. The GLC-2015 performed well, with an OA of 79.5 % (83.6 %) assessed with the global point-based (patch-based) samples.