the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Refined mapping of tree cover at fine-scale using time-series Planet-NICFI and Sentinel-1 imagery for Southeast Asia (2016–2021)
Feng Yang
Zhenzhong Zeng
Abstract. High-resolution mapping of tree cover is indispensable for effectively addressing tropical forest carbon loss, climate warming, biodiversity conservation, and sustainable development. However, the availability of precise high-resolution tree cover map products remains inadequate due to the inherent limitations of mapping techniques utilizing medium-to-coarse resolution satellite imagery, such as Landsat and Sentinel-2 imagery. In this study, we have generated an annual tree cover map product at a resolution of 4.77 m for Southeast Asia (SEA) for the years 2016–2021 by integrating Planet-Norway’s International Climate & Forests Initiative (NICFI) imagery and Sentinel-1 Synthetic Aperture Radar data. we have also collected annual samples to assess the accuracy of our Planet-NICFI tree cover map products. The results show that our Planet-NICFI tree cover map products during 2016–2021 achieve high accuracy, with an overall accuracy of ≥ 0.867 ± 0.017 and a mean F1 score of 0.921, respectively. Furthermore, our tree cover map products exhibit high temporal consistency from 2016 to 2021. Compared to existing map products (FROM-GLC10, ESA WorldCover 2020 and 2021), our tree cover map products exhibit better performance, both statistically and visually. Yet, the imagery obtained from Planet-NICFI performs less in mapping tree cover in areas with diverse vegetation or complex landscapes due to insufficient spectral information. Nevertheless, we highlight the capability of Planet-NICFI datasets in providing quick and fine-scale tree cover mapping to a large extent. The consistent characterization of tree cover dynamics in SEA's tropical forests can be further applied in various disciplines. The annual Planet-NICFI V1.0 tree cover map products from 2016 to 2021 at 4.77 m resolution are publicly available at https://cstr.cn/31253.11.sciencedb.07173 (Yang and Zeng, 2023).
Feng Yang and Zhenzhong Zeng
Status: open (until 22 Jun 2023)
-
CC1: 'Comment on essd-2023-143', Clement Atzberger, 08 May 2023
reply
In my opinion, this is a weak paper and it should not be published. I have also strong doubts with respect to the findings. Both, methodology and results remain obscure. The work is neither well presented nor inspiring confidence.
Major shortcomings
The authos claim having produced six annual tree cover maps from 2016 to 2021 based on a random set of 1515 reference samples (visually interpreted) that remain fixed over the 6-year period. However, no example maps are shown that would allow the reader to assess the stability of the derived tree cover over time at the pixel level. The stability of the derived forets cover (at 5m resolution) is also not summarized into statistical numbers - we have to assume that such statistics would show a very large number of implausible forest/non-forest trajectories.
The quality of the reference data itself remains unclear and doubtful. In particular, the fixed set of (1515) reference samples shows inter-annual variations that are far from plausible (Tab.1). For example (Tab.1), the samples indicate a 13% tree cover loss between the two consecutive years 2017-2018 followed by a 12% gain the next year (2018 to 2019). The authors do not even mention/discuss this issue - they also fail to indicate possible spill-over effects on the maps produced with this reference data of questionable quality (see above comment).
The authors claim having combined Planet multi-spectral imagery and S1 (two polarisation) to produce the annual tree cover maps. However, we learn nothing about the respective contribution of the two sensor modalities.
Minor comments
What is labeled as "validation" data in Fig.2 is indeed "training" data.
The authors elaborate on the fact that different forest definitions exist (e.g., FAO) but fail to tell the reader which definition was finally adopted. We also learn only in the "Discussion" section, that plantations were excluded from the class "forest" during manual labeling.
The authors propose a "stability index" (year-to-year change in overall accuracies) "to evaluate tree cover accuracy". Unfortunately, tracking year to year changes in statistical measures will not tell us much about the tree cover accuracy. A good/better plausibility check would have been to compare (pixel-by-pixel) the forest/non-forest trajectories between 2016 and 2021 ... and to analyse if they are at least plausible.
Not clear how accuracies are assessed - I guess the authors use the OOB error provided by the RF algorithm?
Best regards
C Atzberger
Citation: https://doi.org/10.5194/essd-2023-143-CC1 -
RC1: 'Comment on essd-2023-143', Anonymous Referee #1, 16 May 2023
reply
This manuscript presents a new method to map distribution of trees between 2016 and 2021 using Planet and Sentinel-1 and RF model, offering high-resolution mapping capabilities. Although this dataset has the potential to make a significant contribution, there are several areas that require improvement, as identified below:
1. The manuscript claims to characterize tree cover and its changes, but it only provides a map distinguishing between tree and non-tree, without indicating the percentage of tree coverage. It is essential to reconsider the definition of tree cover in the revised manuscript.
2. The resolution of Sentinel-1 images is 10 meters, so it is unclear how the authors obtained tree distribution data at a resolution of 4.77 meters. Further clarification is needed regarding the methodology employed to achieve this higher resolution.
3. While the aim of this dataset is to provide information about changes in tree distribution, the validations conducted thus far seem to focus solely on the spatial pattern of trees. To strengthen the manuscript, it is necessary to include statistical validation regarding changes in tree distribution, ensuring a comprehensive assessment.
4. It remains unclear how the model performed over complex landscapes and regions with isolated trees. Additional information regarding the model's performance in such settings would greatly enhance the manuscript's credibility and applicability to various environments.
5. To differentiate this manuscript from the unpublished work of Yang et al. (2023), it is important to highlight the distinguishing features. Providing a clear outline of the unique contributions and methodologies employed in this manuscript compared to Yang et al.'s work will help readers understand the specific advancements and insights presented in each study.
Citation: https://doi.org/10.5194/essd-2023-143-RC1 -
CC2: 'Comment on essd-2023-143', lipeng jiao, 23 May 2023
reply
This manuscript generated an annual tree cover map product at a resolution of 4.77m using Planet and Sentinel-1 data in the period 2016-2021. In general, the research is significant and related-works are well investigated. However, there are still several issues that need to be clarified.
1. The author mentions that 1515 validation data are used, but the training data of the RF model is not clearly defined in this manuscript. The manuscript requires explicit definitions of training and validation samples.
2. In section 2.3, the author can complement some clear descriptions of how to make comprehensive use of Planet and Sentinel-1 data.
3. The manuscript has made an evaluation of the forest cover products produced in terms of quantitative assessment and detailed comparison, and the analysis of the results is convincing. However, it lacks a complete display and description of the annual forest cover results in Southeast Asia.Citation: https://doi.org/10.5194/essd-2023-143-CC2 -
RC2: 'Comment on essd-2023-143', Anonymous Referee #2, 26 May 2023
reply
This paper describes the production of a set of high-resolution forest cover dataset covering the Southeast Asia. Their method is solid, and the data are very useful. The authors also extensively analyzed the uncertainty of their data products. This paper provides enough details of understanding and using their data products.
Since I am not familiar with the technical details, my questions are mainly from the points of view of a potential user.
Major questions:
- Data validation
The text says, “We collected other validation datasets to assess the tree cover products during 2016-2021” (Line 111). Does it mean the authors went to the validation regions to check the forest states in person? Or, you are using other datasets or interpretation methods to do validation?
In the following paragraph, the authors says, “we randomly generated 1,515 points to ensure the representativeness of collected visual samples”. It seems there were field checking. However, in the following sentence “these points were labeled these points as forests or non-forests by four human interpreters using Planet Explorer of QGIS.”. Please clarify it.
- In the section “Statistical accuracy assessment”, the authors mentioned “the user’s accuracy, producer’s accuracy, and overall accuracy”. Please explain what they are. These terms may be well known in the authors’ discipline. However, as a reader of this paper, I don’t know what they are, and many readers (and potential users) may have the same issue.
- Section “6 Conclusions”
For a data paper, we have read the abstract, and understand how the data were generated/collected, the scope and uncertainty of the data, and know how to get them. Do we really need a “conclusion” section?
Minor questions and edits:
- Line 28: Please clarify what “annual samples” are.
- Lines 29~30 “with an overall accuracy of 0.867±0.017 and a mean F1 score of 0.921, respectively.” Please explain what “overall accuracy and F1 score” are. My opinion, either explain it clearly or don’t mention it. In the previous sentence, authors have said “high accuracy”. Since I don’t know “overall accuracy and F1 score”, it is still “high accuracy” to me.
- Lines 31~34 “Compared to existing maps …”. These sentences can be removed. Add more details about your data published with this paper.
- Lines 37 “The annual Planet-NICFI V1.0 tree cover map products from 2016 to 2021 at 4.77 m resolution”. I would replace it with “Our data”. It is not necessary to repeat the same information in the abstract.
- Lines 112~115. Please rephrase this section so we can understand why “except 2019”.
- Line 118: remove “these points”. repeated.
- Line 119 “four human interpreters”: Do you mean you asked four collogues (Homo sapiens) to do a test of identification? Are they acknowledged?
- Line 139: remove “For example,”.
- Line 176. Please explain what are “user’s accuracy, producer’s accuracy, and overall accuracy”.
- Lines 186~191: I cannot get what the first approach is.
- Line 188 “based on a study by Tsendbazar”: Please explain what it is.
12. Line 195 “The results for 2019 were provided by ..” can be move to the method section
- Line 261 “minimum tree height …”: is this about the definition of forest?
- Lines 268~269. I think this sentence is about algorithm “random forest”. Please rephrase this sentence.
- Line 270; “U-net”?
- Line 301 Section “Acknowledgements” who are those four “human interpreters” mentioned in the main text?
Citation: https://doi.org/10.5194/essd-2023-143-RC2
Feng Yang and Zhenzhong Zeng
Data sets
Fine-scale maps of tree cover generated using time-series Planet-NICFI imagery for Southeast Asia (2016-2021) Feng Yang and Zhenzhong Zeng https://cstr.cn/31253.11.sciencedb.07173
Feng Yang and Zhenzhong Zeng
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 455 | 128 | 17 | 600 | 5 | 5 |
- HTML: 455
- PDF: 128
- XML: 17
- Total: 600
- BibTeX: 5
- EndNote: 5
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1