the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 05 May 2026
Quantifying Mangrove Extent and Uncertainty: An Embedding-Driven Approach for Southeast Asia and Papua New Guinea
Abstract. Mangroves are an important ecosystem across Southeast Asia and Papua New Guinea, and accurate estimates of their occurrence and extent are critical. However, area estimates vary widely across studies, often because they rely on direct pixel counts without stratified sampling or design-based inference, making them sensitive to image and model biases that vary across space and time. This study evaluates whether satellite foundation-model embeddings can provide a more temporally consistent representation for mangrove monitoring while examining spatial transferability. We analyze the spatiotemporal structure of 64-dimensional AlphaEarth Foundations embeddings and apply a gradient-boosted tree classifier to generate annual mangrove probability maps for 2017–2024. These maps are combined with stratified probability sampling and design-based accuracy assessment to derive accuracies and unbiased area estimates with quantified uncertainty. The embeddings show clear mangrove–non-mangrove separability and relative temporal consistency in the embedding space, supporting generally stable regional classification performance (overall accuracy ≈0.80–0.82; AUC ≈0.92–0.93). While performance declines in heterogeneous coastal systems indicate remaining transferability limits, the results demonstrate that foundation-model embeddings provide a temporally consistent and operational basis for large-scale mangrove monitoring with robust, design-based area estimation. More broadly, embeddings offer a compact representation that supports feature-space analysis and more consistent interpretation of changes over time.
Status: closed
-
RC1: 'Comment on egusphere-2026-2205', Akkarapon Chaiyana, 29 Jun 2026
-
AC1: 'Reply on RC1', Vanna Teck, 28 Aug 2026
Dear Referee 1,
We sincerely appreciate your thorough review and the constructive, insightful feedback you provided on our manuscript. Thank you for the significant time and effort you invested in evaluating our research. Your valuable suggestions have been instrumental in improving both the clarity and overall quality of our work.
Each of your points has been carefully addressed, and we have provided comprehensive, point-by-point responses in the attached response document. For clarity, all revisions and corresponding updates within the modified manuscript have been explicitly highlighted.
Kind regards,
Vanna Teck, on behalf of all authors
-
AC1: 'Reply on RC1', Vanna Teck, 28 Aug 2026
-
RC2: 'Comment on egusphere-2026-2205', Anonymous Referee #2, 20 Jul 2026
This manuscript showcases an embedding-driven approach for mangrove extent mapping with uncertainty estimates. While this study clearly uses Alpha Earth embeddings for mangrove mapping, the manuscript requires several improvements to clarify the main message, presentation, and methodology.
General comment
The manuscript presents AE embeddings with 64-dimensional features combined with gradient boosting to generate annual mangrove maps. While the authors confirm that this approach achieves the greatest classification stability, they do not provide evidence that it improves conventional S1/S2-based methods or existing global maps such as WorldCover in terms of mangrove delineation. Before introducing a complex, large dataset (64 layers) as a go-to method for mapping, its efficiency relative to simpler approaches needs to be demonstrated.
The manuscript mentions integrating an embedding-based map with design-based area estimation (line 99). It also mentions correcting the map-based area estimation with design-based area estimation. It is not clear where the integration is performed or what it entails. The manuscript seems to suggest that the design-based area estimation corrects the map-based pixel counting and thereby integrates the two. In reality, these two are different approaches to estimating areas. The map-based approach is subject to mapping error and therefore biased. The design-based (sample-based) approach is unbiased as it uses statistical probability sampling. The Olofsson formulas for area estimation, in fact, only use the reference (sampled data) equation 6. Hence, this is a sample-based area estimation under design-based inference. So there is no integration of design-based area estimation and AE-embedding-based maps, other than that the generated maps are used for stratification to draw the reference sample. If one wants to integrate two estimates (map-based and sample-based), model-assisted area estimation is a way to increase the precision of the area estimation. These differences should be clearly presented in this manuscript and the “integration” needs to be clarified.
Spatial similarity is assessed using a distance-based similarity metric and Montecarlo. It is not clear why more common moran’s I and kriging variances are not used. In particular, the latter can clearly show the spatial dependency of the variable with clear spatial structure and range. In addition, the motivation of spatial similarity assessment and its link with the AE variables are not clear. Spatial similarity is highly dependent on underlying landscape homogeneity and will vary naturally from place to place. What value does this bring besides an area being spatially homogeneous or heterogeneous? In particular, this assessment is not compared against Sentinel 2 or other satellite-based similarity assessments; therefore, the value of this analysis or the advantage of using AE is not clear. Also, how is the temporal variability calculated? Please clarify.
The sample size calculation and allocation need further explanation; what was the sample size per region? What was the expected accuracy or standard error used for estimating and allocating sample size? Equation 7, what is ph? What value did you use? A multi-year sample was collected; are they different (in terms of spatial location) for each year, or are they collocated? In other words, only once sample locations were selected, and the interpretation done for each year? Please clarify. If for the latter, change in mangrove area can be estimated, it would be informative to see the change area with uncertainty estimates.
Specific comments
52-55: please provide evidence or reference for this statement.
Figure 2: Are the forest samples the training sample or a stratified random sample used for accuracy and area estimation?
Table 2: Please show this table as a figure with simple error ranges on a bar; this way, the temporal fluctuations can be clearly seen. In fact this can be directly compared with the map based pixel counting.
Table 3: Please provide the confidence ranges for the accuracy estimates. There is a systematic tendency of high omission errors in the classification. This needs to be clearly communicated and highlighted in the manuscript.
Samples: multiple times, the word “samples” was used when referring to the sample size or sample locations. A sample means a set of sample units (see the Tyukavina et al 2025 glossary and terminology explanation). I recommend using the correct term sample units or sample sites rather than samples, as I believe this study does not use multiple sets of samples.
Citation: https://doi.org/10.5194/egusphere-2026-2205-RC2 -
AC2: 'Reply on RC2', Vanna Teck, 28 Aug 2026
Dear Referee 2,
We sincerely appreciate your careful evaluation of our manuscript and your constructive and technically insightful feedback. Thank you for the time and effort you invested in reviewing our work. Your comments and suggestions have helped us substantially improve the clarity and methodological presentation of the manuscript, particularly regarding the contribution of AlphaEarth Foundations (AEF) embeddings, the distinction between map-based area estimates (MBAE) and sample-based area estimates (SBAE), the motivation for the spatial similarity analysis, the sampling design, and the terminology used for reference sample units.
We have carefully considered and addressed each of your comments. Our detailed, point-by-point responses are provided in the attached response document, and all corresponding changes in the revised manuscript have been highlighted.
Kind regards,
Vanna Teck, on behalf of all authors
-
AC2: 'Reply on RC2', Vanna Teck, 28 Aug 2026
Status: closed
-
RC1: 'Comment on egusphere-2026-2205', Akkarapon Chaiyana, 29 Jun 2026
This study proposes a framework for quantifying mangrove extent across Southeast Asia and Papua New Guinea using a Gradient Boosting Tree (GBT) model trained on the AlphaEarth Embedding dataset. The manuscript is generally well organized, and the overall study design is presented according to scientific standards. However, the manuscript requires substantial improvement in clearly articulating its key scientific contributions and novelty. At present, the manuscript reads more like a project description than a research study, with insufficient emphasis on the methodological innovations, scientific significance, and broader implications of the proposed approach. The authors should better clarify how their work advances the current state of the art and distinguish it from existing mangrove mapping studies. Therefore, I recommend major revision before the manuscript can be considered for publication.
- Line 85, citation
- Why Gradient Boost Tree?
- Why 95% training and 5% testing split, do this have logic behind?
- How many samples of each country?
- Figure 1, suggest revising as professional style if necessary.
- Table 2. Checking entire table, Malaysia (2017 and 2018), why this duplicated number?
- Do you have any discussion about Table 2, why some year of a country are suddenly drop such as Cambodia (2019 and 2020)?
- Table 1, p stands for? Relevant S1-S5 considers individual country? Please simplify this.
- Section 2.5, how GBT works? Please simplify 1-2 sentences.
- Why UMAP is better to others, please emphasize how is this suited as 1 sentence.
- Adding research diagram will be helpful for readers, this highly suggest.
- In result section, the comparison when prior available mangrove products are highly recommended to emphasize EM dataset how are better to be consider for future study. Please also discuss
- Figure 3, color RGB contains value if yes, this fig is to have legend.
- In discussion, this lacks such a claim as mentioned in abstract (“However, area estimates vary widely across studies, often because they rely on direct pixel counts without stratified sampling or design-based inference, making them sensitive to image and model biases that vary across space and time.” Please discuss how this better to others?
- In addition, Why EM dataset is better, please discuss in Discussion section.
- Figure 5, This is not Country-level mangrove which is an example of each mangrove extent? Scale bars suggest adding. Why probability gradient bar is only 2024, why this year is matter? Why only 8 countries?
- Unbiased area estimates with ±95% confidence intervals (CI), in section 3.3.2 is not clear. Please elaborate.
- This study can provide key message as recommendation to improve study. What limitation, and future improvement are?
- In conclusion, “Compared with traditional approaches based on spectral indices…”, did this study compare?
- Suggest showing overview of mangrove extent at all countries.
Citation: https://doi.org/10.5194/egusphere-2026-2205-RC1 -
AC1: 'Reply on RC1', Vanna Teck, 28 Aug 2026
Dear Referee 1,
We sincerely appreciate your thorough review and the constructive, insightful feedback you provided on our manuscript. Thank you for the significant time and effort you invested in evaluating our research. Your valuable suggestions have been instrumental in improving both the clarity and overall quality of our work.
Each of your points has been carefully addressed, and we have provided comprehensive, point-by-point responses in the attached response document. For clarity, all revisions and corresponding updates within the modified manuscript have been explicitly highlighted.
Kind regards,
Vanna Teck, on behalf of all authors
-
RC2: 'Comment on egusphere-2026-2205', Anonymous Referee #2, 20 Jul 2026
This manuscript showcases an embedding-driven approach for mangrove extent mapping with uncertainty estimates. While this study clearly uses Alpha Earth embeddings for mangrove mapping, the manuscript requires several improvements to clarify the main message, presentation, and methodology.
General comment
The manuscript presents AE embeddings with 64-dimensional features combined with gradient boosting to generate annual mangrove maps. While the authors confirm that this approach achieves the greatest classification stability, they do not provide evidence that it improves conventional S1/S2-based methods or existing global maps such as WorldCover in terms of mangrove delineation. Before introducing a complex, large dataset (64 layers) as a go-to method for mapping, its efficiency relative to simpler approaches needs to be demonstrated.
The manuscript mentions integrating an embedding-based map with design-based area estimation (line 99). It also mentions correcting the map-based area estimation with design-based area estimation. It is not clear where the integration is performed or what it entails. The manuscript seems to suggest that the design-based area estimation corrects the map-based pixel counting and thereby integrates the two. In reality, these two are different approaches to estimating areas. The map-based approach is subject to mapping error and therefore biased. The design-based (sample-based) approach is unbiased as it uses statistical probability sampling. The Olofsson formulas for area estimation, in fact, only use the reference (sampled data) equation 6. Hence, this is a sample-based area estimation under design-based inference. So there is no integration of design-based area estimation and AE-embedding-based maps, other than that the generated maps are used for stratification to draw the reference sample. If one wants to integrate two estimates (map-based and sample-based), model-assisted area estimation is a way to increase the precision of the area estimation. These differences should be clearly presented in this manuscript and the “integration” needs to be clarified.
Spatial similarity is assessed using a distance-based similarity metric and Montecarlo. It is not clear why more common moran’s I and kriging variances are not used. In particular, the latter can clearly show the spatial dependency of the variable with clear spatial structure and range. In addition, the motivation of spatial similarity assessment and its link with the AE variables are not clear. Spatial similarity is highly dependent on underlying landscape homogeneity and will vary naturally from place to place. What value does this bring besides an area being spatially homogeneous or heterogeneous? In particular, this assessment is not compared against Sentinel 2 or other satellite-based similarity assessments; therefore, the value of this analysis or the advantage of using AE is not clear. Also, how is the temporal variability calculated? Please clarify.
The sample size calculation and allocation need further explanation; what was the sample size per region? What was the expected accuracy or standard error used for estimating and allocating sample size? Equation 7, what is ph? What value did you use? A multi-year sample was collected; are they different (in terms of spatial location) for each year, or are they collocated? In other words, only once sample locations were selected, and the interpretation done for each year? Please clarify. If for the latter, change in mangrove area can be estimated, it would be informative to see the change area with uncertainty estimates.
Specific comments
52-55: please provide evidence or reference for this statement.
Figure 2: Are the forest samples the training sample or a stratified random sample used for accuracy and area estimation?
Table 2: Please show this table as a figure with simple error ranges on a bar; this way, the temporal fluctuations can be clearly seen. In fact this can be directly compared with the map based pixel counting.
Table 3: Please provide the confidence ranges for the accuracy estimates. There is a systematic tendency of high omission errors in the classification. This needs to be clearly communicated and highlighted in the manuscript.
Samples: multiple times, the word “samples” was used when referring to the sample size or sample locations. A sample means a set of sample units (see the Tyukavina et al 2025 glossary and terminology explanation). I recommend using the correct term sample units or sample sites rather than samples, as I believe this study does not use multiple sets of samples.
Citation: https://doi.org/10.5194/egusphere-2026-2205-RC2 -
AC2: 'Reply on RC2', Vanna Teck, 28 Aug 2026
Dear Referee 2,
We sincerely appreciate your careful evaluation of our manuscript and your constructive and technically insightful feedback. Thank you for the time and effort you invested in reviewing our work. Your comments and suggestions have helped us substantially improve the clarity and methodological presentation of the manuscript, particularly regarding the contribution of AlphaEarth Foundations (AEF) embeddings, the distinction between map-based area estimates (MBAE) and sample-based area estimates (SBAE), the motivation for the spatial similarity analysis, the sampling design, and the terminology used for reference sample units.
We have carefully considered and addressed each of your comments. Our detailed, point-by-point responses are provided in the attached response document, and all corresponding changes in the revised manuscript have been highlighted.
Kind regards,
Vanna Teck, on behalf of all authors
-
AC2: 'Reply on RC2', Vanna Teck, 28 Aug 2026
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 322 | 0 | 4 | 326 | 0 | 0 |
- HTML: 322
- PDF: 0
- XML: 4
- Total: 326
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This study proposes a framework for quantifying mangrove extent across Southeast Asia and Papua New Guinea using a Gradient Boosting Tree (GBT) model trained on the AlphaEarth Embedding dataset. The manuscript is generally well organized, and the overall study design is presented according to scientific standards. However, the manuscript requires substantial improvement in clearly articulating its key scientific contributions and novelty. At present, the manuscript reads more like a project description than a research study, with insufficient emphasis on the methodological innovations, scientific significance, and broader implications of the proposed approach. The authors should better clarify how their work advances the current state of the art and distinguish it from existing mangrove mapping studies. Therefore, I recommend major revision before the manuscript can be considered for publication.