Predictive Modelling of Land Cover Changes in the Greater Amanzule Peatlands Using Multi-Source Remote Sensing and Machine Learning Techniques

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents a robust analysis of land cover changes within the GAP. The authors are to be commended for their innovative approach, which leverages a fusion of multi-spectral and radar remote sensing data, complemented by machine learning algorithms. This methodology not only provides a detailed snapshot of the current land cover but also offers predictive insights into future scenarios, which is highly valuable for conservation and land use planning. And there are several issues the authors need to consider before final publication.

1. Abstract: It would be beneficial to include a brief mention of the specific machine learning techniques used, as this will immediately inform readers about the methodology's sophistication. Additionally, highlighting the key findings or the most significant changes in land cover (such as the 85% increase in rubber plantations) will make the abstract more informative.

2. Introduction: The introduction could be strengthened by providing a more detailed background on the GAP, including its ecological significance and why it is a critical area for study. This will help readers appreciate the importance of the research and its relevance to broader environmental conservation efforts.

3. The research is noted for its strategic selection of time points (2010, 2015, and 2020) for analysis. However, the choice to use only three data points might limit the capture of the full spectrum of land cover dynamics, particularly within the complex ecosystem of tropical peatlands where more frequent changes might occur. It is recommended that the authors discuss how this limitation could potentially affect the fidelity of their predictive models. Additionally, the inclusion of an interim time point or consideration of annual data, if available, could provide a more nuanced understanding of the transitional processes at play.

4. The study adeptly lays the groundwork for future investigative paths, particularly in enhancing predictive models and integrating variables such as climate change impacts and socio-economic dynamics. It would be beneficial for the authors to expand on how future research might advance the predictive capabilities of their model. This could include discussions on the potential for incorporating real-time data streams, the impact of global environmental changes on model outcomes, and the scalability of their methodology to other geographical contexts. Furthermore, exploring the model's sensitivity to various parameters and the robustness of predictions under different environmental scenarios would be a valuable contribution to the field.

 

Author Response

Comment 1: Abstract: It would be beneficial to include a brief mention of the specific machine learning techniques used, as this will immediately inform readers about the methodology's sophistication. Additionally, highlighting the key findings or the most significant changes in land cover (such as the 85%increase in rubber plantations) will make the abstract more informative.

Response 1: Thank you for your insightful feedback on our abstract. We have revised the abstract to mention the specific machine learning methods employed in our study. We now explicitly state the use of Random Forest (RF) classification for land cover analysis and Cellular Automata Artificial Neural Networks (CA-ANN) for predictive modelling. We have also included the 85% increase in rubber plantations, underscoring the substantial anthropogenic impact on the Greater Amanzule Peatland ecosystem.

Comment 2: Introduction: The introduction could be strengthened by providing a more detailed background on the GAP, including its ecological significance and why it is a critical area for study. This will help readers appreciate the importance of the research and its relevance to broader environmental conservation efforts.

Response 2: We have revised the introduction to provide background information on the Greater Amanzule Peatlands (GAP), emphasising its ecological significance and why it is a critical area for study. We note the essential ecosystem services provided and emphasise the threats posed by agricultural plantation expansion, urbanisation, and the burgeoning oil and gas industry since 2011. We also note the limited studies conducted on African peatlands compared to those in Southeast Asia, highlighting why the GAP is a critical area for research.

Comment 3: The research is noted for its strategic selection of time points (2010, 2015, and 2020) for analysis. However, the choice to use only three data points might limit the capture of the full spectrum of land cover dynamics, particularly within the complex ecosystem of tropical peatlands where more frequent changes might occur. It is recommended that the authors discuss how this limitation could potentially affect the fidelity of their predictive models. Additionally, the inclusion of an interim time point or consideration of annual data, if available, could provide a more nuanced understanding of the transitional processes at play.

Response 3: We acknowledge that using only three time points (2010, 2015, and 2020) may limit the ability to capture the full range of land cover dynamics, particularly in the complex tropical peatland ecosystem of the GAP. The choice of these three time points was however driven primarily by the availability of satellite data and the occurrence of significant anthropogenic events, such as the onset of oil and gas activities and small-scale gold mining, which we assumed could have substantially influence on land use during the period. We agree that the lack of interim data may reduce the model's ability to capture short-term fluctuations or minor transitional changes which could impact the precision of long-term predictions. The limitation has been highlighted in the “Data/Satellite Imagery” section. A recommendation has also been added to the section “Projected Land Cover Changes” for future research to incorporate annual or more frequent data points, where available, to improve the model’s ability to capture these dynamics. Despite this limitation, the selected time points still provide robust insight into the dominant trends and drivers of land cover change in the GAP, which remain crucial for developing conservation and management strategies.

Comment 4: The study adeptly lays the groundwork for future investigative paths, particularly in enhancing predictive models and integrating variables such as climate change impacts and socio-economic dynamics. It would be beneficial for the authors to expand on how future research might advance the predictive capabilities of their model. This could include discussions on the potential for incorporating real-time data streams, the impact of global environmental changes on model outcomes, and the scalability of their methodology to other geographical contexts. Furthermore, exploring the model's sensitivity to various parameters and the robustness of predictions under different environmental scenarios would be a valuable contribution to the field.

Response 4: We have carefully considered your suggestions regarding the enhancement of the predictive model and have expanded the \subsection {Projected Land Cover Changes} to incorporate discussions on how future research can advance the predictive capabilities by incorporating real-time data streams, assessing the impact of climate change and socio-economic dynamics, enhancing temporal resolution, and evaluating model sensitivity and robustness.

Reviewer 2 Report

Comments and Suggestions for Authors

General:

The manuscript provides a comprehensive analysis of land cover changes within the Greater Amanzule Peatlands (GAP) in Ghana. The integration of multi-source remote sensing data and machine learning techniques for predictive modeling offers a robust framework for understanding and forecasting land cover dynamics in tropical peatlands. The authors have effectively utilized Landsat serious data, SAR, and SRTM data to achieve their objectives. Additionally, the socio-economic context provided helps readers appreciate the significance of the study's findings for local communities and policy-makers.

However continuous monitoring and adaptive management recommendations should be strengthened to ensure the resilience and sustainability of the GAP ecosystem.

 

Specific Comments

  

1. The land cover transition analysis reveals interesting trends, how they might impact the local environment and communities in results part.

2. The manuscript could discuss how the predictive models can assist in planning sustainable development practices that balance ecological preservation with socio-economic benefits in discrssion.

3. "The Greater Amanzule Peatland(GAP) in Ghana represent an important ecosystem facing dynamic land cover changes as a result of both natural and anthropogenic factors." Should be "The Greater Amanzule Peatland (GAP) in Ghana represents an important ecosystem facing dynamic land cover changes as a result of both natural and anthropogenic factors."

4. "This study integrates multispectral and radar remote sensing data from Landsat-7 and-8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these ch" The sentence is incomplete and should continue with "changes."

5. Land Cover Changes: "GAP underwent significant land cover changes — either through reduction or expansion of land cover types — from 2010 to 2020." To avoid redundancy, consider, "Between 2010 and 2020, the GAP experienced significant land cover changes involving the reduction or expansion of various land cover types."

Comments on the Quality of English Language

1. "The Greater Amanzule Peatland(GAP) in Ghana represent an important ecosystem facing dynamic land cover changes as a result of both natural and anthropogenic factors." Should be "The Greater Amanzule Peatland (GAP) in Ghana represents an important ecosystem facing dynamic land cover changes as a result of both natural and anthropogenic factors."

2. "This study integrates multispectral and radar remote sensing data from Landsat-7 and-8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these ch" The sentence is incomplete and should continue with "changes."

3. Land Cover Changes: "GAP underwent significant land cover changes — either through reduction or expansion of land cover types — from 2010 to 2020." To avoid redundancy, consider, "Between 2010 and 2020, the GAP experienced significant land cover changes involving the reduction or expansion of various land cover types."

Author Response

Comments 1:

1. The land cover transition analysis reveals interesting trends, how they might impact the local environment and communities in results part.
2. The manuscript could discuss how the predictive models can assist in planning sustainable development practices that balance ecological preservation with socio-economic benefits in discussion.

Response 1: Thank you for this insightful suggestion. We have expanded the discussion (section: Projected Land cover changes) to address how the predictive models can assist in planning sustainable development. We now emphasise how the model provides critical insights into balancing ecological preservation with socio-economic benefits, particularly by identifying areas of stability and vulnerability within the GAP. Additionally, we discuss how incorporating factors such as climate change, large-scale projects, and socio-economic dynamics into future models would allow for more nuanced and actionable strategies, ensuring long-term sustainability and resilience in the region.

Comment 2: "The Greater Amanzule Peatland (GAP) in Ghana represent an important ecosystem..." should be "The Greater Amanzule Peatland (GAP) in Ghana represents an important ecosystem..."

Response 2: The suggested correction has been implemented, the sentence now reads, "The Greater Amanzule Peatland (GAP) in Ghana represents an important ecosystem facing dynamic land cover changes as a result of both natural and anthropogenic factors."

Comment 3: The sentence "This study integrates multispectral and radar remote sensing data from Landsat-7 and -8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these ch" is incomplete and should end with "changes."

Response 3: The sentence has been revised to, "This study integrates multispectral and radar remote sensing data from Landsat-7 and -8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these changes."

Comment 4: The sentence "GAP underwent significant land cover changes — either through reduction or expansion of land cover types — from 2010 to 2020." could be rephrased to avoid redundancy. Consider "Between 2010 and 2020, the GAP experienced significant land cover changes involving the reduction or expansion of various land cover types."

Response 4: We agree with the suggestion and have revised the sentence to, "Between 2010 and 2020, the GAP experienced significant land cover changes involving the reduction or expansion of various land cover types."

Reviewer 3 Report

Comments and Suggestions for Authors

Introduction:

Line 50. “Previous research has extensively covered land cover mapping and change detection in tropical peatlands. Various remote sensing techniques have been employed to monitor and analyze land cover changes over time, contributing to a better understanding of these ecosystems' dynamics [14–17]”. It should be made clear which bibliographical references belong to the first of the themes, and if would not have cited those “Previous research”, and which ones belong to the second.

In this section it is considered that the main and secondary objectives of the research should be cited and categorized much more clearly.

 

Methodology:

It is recommended to include at the beginning of this section a figure (similar to Figure 3) summarizing the methodological scheme of the process followed for the development of the research carried out.

 

Results – Discussion of results:

The discrimination between the methodological description of the study and the results obtained is extremely correct, but in the discussion of results it would be advisable not to include information that could be redundant (for example that which appears in the first paragraph of the Landcover changes subsection), which already It was presented in the results section.

 

Conclusion:

The conclusions are clear, but it would be advisable to establish a clear and direct link with the brilliant fulfillment of the research objectives.

Author Response

Comment 1: Introduction

Line 50. “Previous research has extensively covered land cover mapping and change detection in tropical peatlands. Various remote sensing techniques have been employed to monitor and analyse land cover changes over time, contributing to a better understanding of these ecosystems' dynamics [14–17]”. It should be made clear which bibliographical references belong to the first of the themes, and if would not have cited those “Previous research”, and which ones belong to the second. In this section it is considered that the main and secondary objectives of the research should be cited and categorized much more clearly.

Response 1: We have revised the introduction to distinctly categorise the bibliographical references, differentiating between studies focused on land cover mapping (e.g., [14, 16, 17]) and those addressing change detection (e.g., [18, 20, 21]). The introduction has also been modified to explicitly highlight the aim of the study, which is to fill the gap in predictive modelling of future land cover changes in tropical peatlands, particularly in the Greater Amanzule Peatlands (GAP). We have also clearly stated the specific objectives: (1) to utilise combined optical, Synthetic Aperture Radar (SAR), and digital elevation data for tropical peatland mapping and change detection in the GAP, and (2) to predict future land cover scenarios up to 2040 using Random Forest classification and Cellular Automata Artificial Neural Networks (CA-ANN).

 

Comment 2: Methodology

"It is recommended to include at the beginning of this section a figure (similar to Figure 3) summarising the methodological scheme of the process followed for the development of the research carried out."

Response 2:
We have moved Figure 3 up to the "Satellite Data" section, as it was originally designed to serve as a summary of the methodological scheme. By repositioning the figure, we ensure it provides a clear overview of the methodological process at the appropriate point in the manuscript.

Comment 3: "The discrimination between the methodological description of the study and the results obtained is extremely correct, but in the discussion of results, it would be advisable not to include information that could be redundant (for example, that which appears in the first paragraph of the Landcover changes subsection), which was already presented in the results section."

Response 3: We have revised the discussion to avoid repeating information from the results section, particularly in the first paragraph of the Landcover changes subsection.

Comment 4: Conclusion:

The conclusions are clear, but it would be advisable to establish a clear and direct link with the brilliant fulfilment of the research objectives.

Response 4:
Thank you for your constructive feedback. In response, we have revised the conclusion to establish a clearer and more direct link with the research objectives. Specifically, the conclusion now emphasises how the study fulfilled its primary objectives by successfully utilising multi-source remote sensing data and machine learning techniques to (1) analyse land cover changes in the GAP from 2010 to 2020 and (2) project future land cover scenarios up to 2040. It highlights the significant findings related to land cover changes, their broader implications for conservation and development strategies, and the need for continuous monitoring and adaptive management.

Reviewer 4 Report

Comments and Suggestions for Authors

This paper focuses on mapping the land cover changes in the Greater Amanzule Peatlands from 2010 to 2020, using multisource remote sensing data and machine learning techniques. It also explores predictive modelling to look into land cover changes in the region. While the idea is interesting and relevant with recent discussions around the sustainable management of peatlands, there are several significant concerns regarding the methodology that need to be addressed to strengthen the study.

 

Below are the major concerns regarding the methods, followed by some specific points.

 1.)   Could the authors clarify the rationale for using TOA reflectance products instead of Bottom of Atmosphere (BOA)/surface reflectance (SR), particularly given the importance of atmospheric correction for accurate spectral signatures? SR products, which account for atmospheric factors such as aerosols, water vapour, and varying solar angles, seem especially crucial for temporal mapping, where consistency across different time periods is essential for reliable change detection. Given that land cover mapping relies on precise spectral information and median filtering is not adequate to deal with this issue, it would be helpful to either justify the use of TOA reflectance or consider switching to SR products, which are readily available in GEE.

 2.)   Landsat-7 data is known to have issues due to the Scan Line Corrector (SLC) failure, which results in gaps and missing data. This is a significant limitation for spatial analysis and land cover mapping. The methods section does not explain how these gaps were addressed or accounted for in the study. Did the authors consider using Landsat-5, which does not have this issue and can serve as a better alternative? While the methodology mentions using a median filter, this doesn’t resolve the missing data problem. Additionally, there seems to be a stripping effect visible in Figure 7(a) towards the east end possibly introduced by the SLC issue. Could the authors clarify why Landsat-7 was chosen despite its limitations and explain how these data gaps were managed?

 3.)   There is a mismatch in the spatial resolutions of the radar data (25 meters) and the optical/DEM data (30 meters). The methods section does not address how this was considered/handled. It is important to clarify whether resampling or other techniques were used to align these different resolutions. This is important for the consistency and accuracy of the datasets that were integrated. Please explain how this resolution mismatch was managed and discuss any potential impact on the results of this study. Without these details and improvements, the conclusions drawn could be significantly affected.

 The methodology should provide a clear and step-by-step description of the methods used in this study without being clouded by extensive background information such as model comparisons or other literature reviews. This can confuse the reader and detract from understanding the actual methodological approach. If there is a need for it, it would be more appropriate to move it to the introduction section.

  

Following are some specific comments

 Abstract

 The abstract is well written, however, the structure needs some work and some important elements are missing.  The background needs to be expanded a little and some key aspects of the study such as the uncertainty and the novelty should be highlighted. The acronyms should be spelt out. Below are some specific points.

  

The background is very brief and does not provide sufficient context for the readers unfamiliar with the topic e.g., expand a bit on the  “natural and/or anthropogenic factors” What are some of these factors?

 

This should be followed up by the main aim of the paper. It is written that “This study 2

integrates multispectral and radar remote sensing data from Landsat-7 and -8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these changes from 2010 to 2020 and to predict future scenarios up to 2040”. This is a long sentence that should be broken down into two, separating the methods from the aim for more clarity. It should outline the primary research question or objective to help readers understand the focus of the study.

 

Line 3: multispectral is a subtype of optical remote sensing as correctly mentioned in the manuscript later, please revise this accordingly. Furthermore, instead of referring to the specific product (i.e., L-band SAR), the platform should be mentioned: as ALOS/PALSAR to maintain consistency with how Landsat is referenced. Lastly, SRTM is the mission name, not the product or sensor itself. It should be clarified that it refers to an SRTM-derived DEM or topographic data for better clarity, and this should be corrected throughout the manuscript.

  

The abstract does not sufficiently discuss the uncertainty associated with the results. Including information on the overall accuracy, F1 score, or other relevant uncertainty metrics of the machine learning models. It is important to discuss these to highlight the reliability and robustness of the methods used and the results obtained.

 

The concluding sentence of the abstract mentions “methodological framework” but does not specify what it entails. It should also clearly highlight what makes this work novel. Is it the integration of a multisource dataset? the focus on a previously understudied region like the GAP? The novelty should be highlighted.

 Introduction

 The introduction is well written however it is brief and light on citations. Furthermore, the first two paragraphs rely heavily on reports (e.g., Reference 1 and 6) rather than peer-reviewed studies and lack recent research references. Citations appear multiple times for different statements (e.g., [6]), suggesting a lack of diverse sources. Multiple related studies have been published in the past 2 years alone and these should be incorporated. Below are some specific points.

Line 22-23: Cite the original sources.

Line 33: Cite the original source i.e., Joosten H, Clarke D (2002) Wise Use of Mires and Peatlands. International Mire Conservation Group and International Peat Society, Devon

 

Line 36-38: Please cite the original sources and bring in references from recent peer-reviewed research articles as well e.g.,

Fluet-Chouinard, E., Stocker, B.D., Zhang, Z. et al. Extensive global wetland loss over the past three centuries. Nature 614, 281–286 (2023). https://doi.org/10.1038/s41586-022-05572-6

 Minasny, B., Adetsu, D.V., Aitkenhead, M. et al. Mapping and monitoring peatland conditions from global to field scale. Biogeochemistry 167, 383–425 (2024). https://doi.org/10.1007/s10533-023-01084-1

 

Paragraphs 2-3 transition from global peatlands to GAP. To enhance the context it would be beneficial to mention the tropical peatlands in the region and then move on to GAP.

 

The last paragraph discusses the methodology. It provides a very broad overview of methods and gaps. It should explicitly identify the gaps in existing studies and how this study strives to fill these gaps. It would be worthwhile looking into the following recent publications.

  

de Waard, Farina, et al. "Remote sensing of peatland degradation in temperate and boreal climate zones–A review of the potentials, gaps, and challenges." Ecological Indicators 166 (2024): 112437.

 

Habib, W., Connolly, J. A national-scale assessment of land use change in peatlands between 1989 and 2020 using Landsat data and Google Earth Engine—a case study of Ireland. Reg Environ Change 23, 124 (2023). https://doi.org/10.1007/s10113-023-02116-0

 Methods

 Study area

The subsection contains superfluous details. The section should be more concise, focusing specifically on the aspects of the study area that directly impact the research, such as the geography, climate and hydrology of the GAP. Some parts of the section contain information that reads more like an introduction, such as discussions on regional biodiversity and socio-economic contexts, which should be relocated to the introduction. The aim should be to briefly describe the physical and environmental characteristics of the GAP that are directly relevant to the study. Below are some specific points.

 

Line 65-67: It is written that “WR is bordered to the west by Cote D’Ivoire, to the east by the Central Region, to the north by Ashanti and Western North regions, and to the south by the Gulf of Guinea”. I think this information is not necessary, it is already evident from the map (Figure 1), consider condensing this to a brief mention of the GAP's location within Ghana and its general geographic setting relevant to the study.

 

Line 86-93: The detailed explanation of IUCN Red List categories reads more like an introduction. And I don’t see how this is relevant to the study.  This could be placed better in the introduction to maybe establish the broader significance of biodiversity conservation in peatlands. But please explain the relevance. Similar comment about the undocumented species and socio-economic context.

 

Figure 1: The map is too small, it should at least be double the size of what it is currently to improve readability. The caption mentions “Base map imagery sourced from Google Satellite”. This is not accurate, the Google base map imagery is sourced from various third-party sources there is no “Google satellite”. Please correct.

 

Figure 2: Is this necessary? I think it could be incorporated into Figure 1 as a base map whereas I don’t think a satellite image base map adds much to that. If it is necessary it should be highlighted in the caption i.e., elevation difference within the context of the study.

 

Table 1: The class description for “Mangrove” just mentions the geographical coverage. Could improve with the details of species like other class descriptions. Additionally, the source of the classification schema should also be cited in the caption.

 Data

Satellite Imagery

 Line 150-158: The rationale for choosing the years for the data acquisition and analysis should be moved to the introduction as it provides important context.

 Line 160: Please change to very high-resolution imagery “available” in Google Earth Pro. As mentioned before the software is just the platform.

 Preparation of image feature

 It is written that “Annual composite images for the years 2010, 2015, and 2020 were generated from the Landsat series using pixel-based compositing in Google Earth Engine (GEE). This method employed median statistics for composite generation, a robust technique for maintaining the representativeness of surface conditions by mitigating anomalies due to cloudiness and shadows..”  

 However, there is no mention of the total number of scenes used or how cloud cover was managed. Since cloud cover can significantly impact pixel values, it is essential to mask out clouds before applying a median filter, especially in tropical regions where cloud cover is frequent. Without this information, it is unclear how effectively the median filter would reflect true land cover conditions. This is also emphasised in one of the sources cited i.e., [47] in this study. Please also have a look at the methods used in below similar studies:

 Amani, Meisam, et al. "Canadian wetland inventory using Google Earth Engine: The first map and preliminary results." Remote Sensing 11.7 (2019): 842.

 Habib, W., Connolly, J. A national-scale assessment of land use change in peatlands between 1989 and 2020 using Landsat data and Google Earth Engine—a case study of Ireland. Reg Environ Change 23, 124 (2023). https://doi.org/10.1007/s10113-023-02116-0

  Line 242-246: This reads more like an introduction. I also don’t think this discussion is needed as it is well established the RF is the best-performing classifier.

 Equation (4): The formula is incorrect. It is missing a factor of 2. Please correct and explain if it is just a typo.

 Line 290-312: This sounds more like a literature review/background and should be moved to discussion if necessary. At present, the method section is clouded by extensive background information that at best belongs to the introduction.

 

Table 4: The precision, recall, and F-scores reported for several land cover classes are unacceptably low, with many values falling below 0.5 indicating an inconsistency in classification performance. OA is high but it is not a good indicator and can be misleading when there is an imbalance of classes. I would like to see what sort of training sample strategy was used. This brings me back to the comment regarding the use of TOA instead of SR and no cloud masking applied.  Could that also be the reason that the optical bands have the lowest contribution? Please address this.

 Results

 Figure 6: It is really hard to visualise the change here. Could the author do a simple overlay analysis with three layers and add a graduated colour scheme to reflect the areas of change over different years?

 Figure 7: I recommend making these maps bigger.

 Discussion:

 Line 535: It is written, “This may result from the spectral similarity between these two classes..” Did the author try spectral unmixing? I would suggest moving Figure A1 to the methods section as it reflects on the effectiveness of sampling as well. Perhaps move the variable importance figure to the supplementary material.

 

 Line 536: There is a discussion regarding the “spatial proximity”. In terms of sample design did the author consider introducing a minimum distance between the samples to cater for spatial autocorrelation?

 

Author Response

Comment 1: Could the authors clarify the rationale for using TOA reflectance products instead of Bottom of Atmosphere (BOA)/surface reflectance (SR), particularly given the importance of atmospheric correction for accurate spectral signatures? SR products, which account for atmospheric factors such as aerosols, water vapour, and varying solar angles, seem especially crucial for temporal mapping, where consistency across different time periods is essential for reliable change detection. Given that land cover mapping relies on precise spectral information and median filtering is not adequate to deal with this issue, it would be helpful to either justify the use of TOA reflectance or consider switching to SR products, which are readily available in GEE.

Response 1: Thank you for your valuable feedback regarding the use of TOA versus SR reflectance products. We agree that Surface Reflectance (SR) products, which account for atmospheric correction, are important for ensuring consistency in spectral signatures across different time periods, particularly in temporal mapping and change detection analyses. However, for this study, we chose to use TOA reflectance products due to the limited availability of cloud-free SR imagery in the GAP. The high cloud cover in this area significantly reduced the number of usable SR images, making it challenging to perform a consistent multi-temporal analysis. TOA data, on the other hand, provided a more complete and temporally consistent dataset, ensuring broader coverage across the study period. 

We would also like to note that this study involves land cover classification, where the primary focus is on creating land cover maps rather than directly comparing spectral pixel values over time. Each image is trained and classified independently, meaning that while atmospheric correction may have some influence, it is less critical when spectral data are converted into land cover properties. Land cover maps generated from the TOA data was independently assessed for accuracy before performing change detection analysis, ensuring the reliability of the classification outputs. This minimises the concern that global atmospheric correction measures would have a significant impact on the classification outcomes. The accuracy of each input land cover map (2010, 2015, and 2020) is reported and acknowledged prior to change detection. This clarity has been added to the methodology.

 

Comment 2: Landsat-7 data is known to have issues due to the Scan Line Corrector (SLC) failure, which results in gaps and missing data. This is a significant limitation for spatial analysis and landcover mapping. The methods section does not explain how these gaps were addressed or accounted for in the study. Did the authors consider using Landsat-5, which does not have this issue and can serve as a better alternative? While the methodology mentions using a median filter, this doesn’t resolve the missing data problem. Additionally, there seems to be a stripping effect visible in Figure 7(a) towards the east end possibly introduced by the SLC issue. Could the authors clarify why Landsat-7 was chosen despite its limitations and explain how these data gaps were managed?

Response 2: Thank you for your insightful comment. While we acknowledge the known issues with the Landsat-7 Scan Line Corrector (SLC) failure, Landsat-7 was chosen due to its superior cloud-free coverage in the Greater Amanzule Peatlands (GAP) region, particularly during the years of interest. Landsat-7 provided more usable data than Landsat-5, making it preferable for multi-temporal analysis. To address the gaps caused by the SLC failure in Landsat-7, as well as gaps in other datasets due to cloud cover and missing data, we employed a median compositing technique to fill the gaps and ensure spatial consistency. Persisting gaps were filled using helper images from adjacent periods (±1 year). Temporal changes between consecutive years were assumed to be minimal. For the 2010 data, Landsat-5 imagery from 2009 to 2011 was used to fill missing data. This approach helped compensate for both SLC-related gaps and regions with high cloud cover, ensuring data completeness across the study period. Additionally, the CFMask algorithm was applied to mask clouds and shadows before compositing. Again, the conversion from spectral values to thematic land cover classes in this study reduces the impact of missing or inconsistent spectral data. Each classified land cover map undergoes accuracy assessment, ensuring the reliability of the classification outputs before proceeding to change detection. The methodology section has been expanded to include this explanation.

 

Comment 3: There is a mismatch in the spatial resolutions of the radar data (25 meters) and the optical/DEM data (30 meters). The methods section does not address how this was considered/handled. It is important to clarify whether resampling or other techniques were used to align these different resolutions. This is important for the consistency and accuracy of the datasets that were integrated. Please explain how this resolution mismatch was managed and discuss any potential impact on the results of this study. Without these details and improvements, the conclusions drawn could be significantly affected.

The methodology should provide a clear and step-by-step description of the methods used in this study without being clouded by extensive background information such as model comparisons or other literature reviews. This can confuse the reader and detract from understanding the actual methodological approach. If there is a need for it, it would be more appropriate to move it to the introduction section.

Response 3: Thank you for raising this important point. We acknowledge the spatial resolution mismatch between the radar data (25 meters) and the optical/DEM data (30 meters). To ensure consistency across all datasets, the radar data were resampled to 30 meters using bilinear interpolation, aligning it with the optical and DEM data. Bilinear interpolation was chosen as it maintains the spatial characteristics of the radar data while providing a smooth transition to the 30-meter resolution, thereby reducing potential errors or biases that could arise from differing spatial resolutions. This clarification has been added to the methodology. 

 

Comment 4: The abstract is well written, however, the structure needs some work and some important elements are missing. The background needs to be expanded a little and some key aspects of the study such as the uncertainty and the novelty should be highlighted. The acronyms should be spelt out. Below are some specific points.

• The background is very brief and does not provide sufficient context for the readers unfamiliar with the topic e.g., expand a bit on the “natural and/or anthropogenic factors” What are some of these factors? This should be followed up by the main aim of the paper.
• It is written that “This study integrates multispectral and radar remote sensing data fromLandsat-7 and -8, L-band SAR, and SRTM to conduct a machine learning-based analysis of these changes from 2010 to 2020 and to predict future scenarios up to 2040”. This is a long sentence that should be broken down into two, separating the methods from the aim for more clarity. It should outline the primary research question or objective to help readers understand the focus of the study.
• Line 3: multispectral is a subtype of optical remote sensing as correctly mentioned in the manuscript later, please revise this accordingly. Furthermore, instead of referring to the specific product (i.e., L-band SAR), the platform should be mentioned: as ALOS/PALSAR to maintain consistency with how Landsat is referenced.
• Lastly, SRTM is the mission name, not the product or sensor itself. It should be clarified that it refers to an SRTM-derived DEM or topographic data for better clarity, and this should be corrected throughout the manuscript.
• The abstract does not sufficiently discuss the uncertainty associated with the results. Including information on the overall accuracy, F1 score, or other relevant uncertainty metrics of the machine learning models. It is important to discuss these to highlight the reliability and robustness of the methods used and the results obtained.
• The concluding sentence of the abstract mentions “methodological framework” but does not specify what it entails. It should also clearly highlight what makes this work novel. Is it the integration of a multisource dataset? the focus on a previously understudied region like the GAP? The novelty should be highlighted.

Response 4: Thank you for your constructive feedback. We have revised the abstract to address the issues you raised. Specifically:

• Background: Expanded to include anthropogenic factors (e.g., agricultural expansion and oil and gas activities) influencing the GAP.
• Aim and Methods: The long sentence has been split for clarity, and the primary research aim is clearly stated, followed by the methodology.
• Terminology: The term “multispectral” was revised to “optical remote sensing,” and “L-band SAR” was replaced with “ALOS/PALSAR.” We clarified that SRTM refers to the Shuttle Radar Topography Mission-derived DEM.
• Uncertainty Metrics: We added key performance metrics, including overall accuracy (93–94%), F1 scores (0.43 to 0.99), Kappa (0.70), and predictive accuracy (80%) to highlight the robustness of the models.
• Novelty: The novelty of integrating multi-source data and machine learning in the understudied GAP region is now highlighted.

Comment 5: The introduction is well written however it is brief and light on citations. Furthermore, the first two paragraphs rely heavily on reports (e.g., Reference 1 and 6) rather than peer-reviewed studies and lack recent research references. Citations appear multiple times for different statements (e.g., [6]), suggesting a lack of diverse sources. Multiple related studies have been published in the past 2 years alone and these should be incorporated. Below are some specific points.

• Line 22-23: Cite the original sources.
• Line 33: Cite the original source i.e., Joosten H, Clarke D (2002) Wise Use of Mires and Peatlands. International Mire Conservation Group and International Peat Society, Devon
• Line 36-38: Please cite the original sources and bring in references from recent peer-reviewed research articles as welle.g., Fluet-Chouinard, E., Stocker, B.D., Zhang, Z. et al. Extensive global wetland loss over the past three centuries. Nature,281–286 (2023). https://doi.org/10.1038/s41586-022-05572-6
• Minasny, B., Adetsu, D.V., Aitkenhead, M. et al. Mapping and monitoring peatland conditions from global to field scale. Biogeochemistry, 383–425(2024). https://doi.org/10.1007/s10533-023-01084-1
• Paragraphs 2-3 transition from global peatlands to GAP. To enhance the context it would be beneficial to mention the tropical peatlands in the region and then move on to GAP.
• The last paragraph discusses the methodology. It provides a very broad overview of methods and gaps. It should explicitly identify the gaps in existing studies and how this study strives to fill these gaps. It would be worthwhile looking into the following recent publications.
• de Waard, Farina, et al. "Remote sensing of peatland degradation in temperate and boreal climate zones–A review of the potentials, gaps, and challenges." Ecological Indicators 166(2024): 112437.
• Habib, W., Connolly, J. A national-scale assessment of land use change in peatlands between 1989 and 2020 using Landsat data and Google Earth Engine—a case study of Ireland. Reg Environ Change 23, 124 (2023). https://doi.org/10.1007/s10113-023-02116-0

Response 5: Thank you for your thoughtful comments regarding the introduction. We have made several revisions based on your feedback:

• We have incorporated a wider range of references, including some of the recommended peer-reviewed articles.
• We have improved the flow of the introduction by providing a clearer transition from global peatland issues to tropical peatlands, before focusing specifically on the GAP region.
• The final paragraph has been revised to emphasise the limited predictive modelling available for African tropical peatlands. It now clearly outlines how this study contributes to addressing this gap by integrating multi-source remote sensing data with advanced machine learning models for the analysis and future land cover predictions in the GAP.

Comment 6: Study area

The subsection contains superfluous details. The section should be more concise, focusing specifically on the aspects of the study area that directly impact the research, such as the geography, climate and hydrology of the GAP. Some parts of the section contain information that reads more like an introduction, such as discussions on regional biodiversity and socio-economic contexts, which should be relocated to the introduction. The aim should be to briefly describe the physical and environmental characteristics of the GAP that are directly relevant to the study. Below are some specific points.

• Line 65-67: It is written that “WR is bordered to the west by Cote D’Ivoire, to the east by the Central Region, to the north by Ashanti and Western North regions, and to the south by the Gulf of Guinea”. I think this information is not necessary, it is already evident from the map (Figure 1), consider condensing this to a brief mention of the GAP's location within Ghana and its general geographic setting relevant to the study.
• Line 86-93: The detailed explanation of IUCN Red List categories reads more like an introduction. And I don’t see how this is relevant to the study. This could be placed better in the introduction to maybe establish the broader significance of biodiversity conservation in peatlands. But please explain the relevance. Similar comment about the undocumented species and socio-economic context.
• Figure 1: The map is too small; it should at least be double the size of what it is currently to improve readability. The caption mentions “Base map imagery sourced from Google Satellite”. This is not accurate, the Google base map imagery is sourced from various third-party sources there is no “Google satellite”. Please correct.
• Figure 2: Is this necessary? I think it could be incorporated into Figure 1 as a base map whereas I don’t think a satellite image base map adds much to that. If it is necessary it should be highlighted in the caption i.e., elevation difference within the context of the study.
• Table 1: The class description for “Mangrove” just mentions the geographical coverage. Could improve with the details of species like other class descriptions. Additionally, the source of the classification schema should also be cited in the caption.

Response 6: Thank you for your valuable feedback regarding the study area section. We have made the following revisions:

• Biodiversity and Livelihood Aspect: We have relocated some of the detailed discussions on biodiversity and socio-economic contexts to the introduction.
• Geography and Redundancy: Information on the bordering regions of Ghana have been removed
• IUCN Red List and Species Information: The detailed explanation of the IUCN Red List categories and undocumented species has been removed from the study area section. A brief mention of the importance of biodiversity conservation in peatlands is now included in the introduction, ensuring its relevance to the broader context.
• Figures 1 & 2 have been resized
• Table 1: The class description for “Mangrove” has been enhanced to include details on species (i.e., Rhizophora and Avicennia). Additionally, the source of the classification schema has been cited in the caption for clarity and accuracy.

Comment 7: Data, Satellite Imagery
Line 150-158: The rationale for choosing the years for the data acquisition and analysis should be moved to the introduction as it provides important context.
Line 160: Please change to very high-resolution imagery “available” in Google Earth Pro. As mentioned before the software is just the platform.

Response 7: The rationale for choosing the years for data acquisition and analysis has been moved to the introduction to provide better context for readers. Line 160 has been changed to available in Google Earth Pro section

Comment 8: Preparation of image feature

It is written that “Annual composite images for the years 2010, 2015, and 2020 were generated from the Landsat series using pixel-based compositing in Google Earth Engine (GEE). This method employed median statistics for composite generation, a robust technique for maintaining the representativeness of surface conditions by mitigating anomalies due to cloudiness and shadows..”  

However, there is no mention of the total number of scenes used or how cloud cover was managed. Since cloud cover can significantly impact pixel values, it is essential to mask out clouds before applying a median filter, especially in tropical regions where cloud cover is frequent. Without this information, it is unclear how effectively the median filter would reflect true land cover conditions. This is also emphasised in one of the sources cited i.e., [47] in this study. Please also have a look at the methods used in below similar studies:

 Amani, Meisam, et al. "Canadian wetland inventory using Google Earth Engine: The first map and preliminary results." Remote Sensing 11.7 (2019): 842.

 Habib, W., Connolly, J. A national-scale assessment of land use change in peatlands between 1989 and 2020 using Landsat data and Google Earth Engine—a case study of Ireland. Reg Environ Change 23, 124 (2023). https://doi.org/10.1007/s10113-023-02116-0

  Line 242-246: This reads more like an introduction. I also don’t think this discussion is needed as it is well established the RF is the best-performing classifier.

 Equation (4): The formula is incorrect. It is missing a factor of 2. Please correct and explain if it is just a typo.

 Line 290-312: This sounds more like a literature review/background and should be moved to discussion if necessary. At present, the method section is clouded by extensive background information that at best belongs to the introduction.

 

Table 4: The precision, recall, and F-scores reported for several land cover classes are unacceptably low, with many values falling below 0.5 indicating an inconsistency in classification performance. OA is high but it is not a good indicator and can be misleading when there is an imbalance of classes. I would like to see what sort of training sample strategy was used. This brings me back to the comment regarding the use of TOA instead of SR and no cloud masking applied.  Could that also be the reason that the optical bands have the lowest contribution? Please address this.

Response 8: Thank you for your insightful comments. 

• We have clarified the total number of scenes used for each year (between 53 and 65) and expanded on how cloud cover was managed by applying the CFMask algorithm to mask clouds before applying the median filter. 
• The unnecessary introductory and literature review elements in the methods section have been either removed or relocated to the appropriate sections to streamline the methodology.
• The missing factor of 2 in Equation (4) has been corrected.

Comment 9: Results

 Figure 6: It is really hard to visualise the change here. Could the author do a simple overlay analysis with three layers and add a graduated colour scheme to reflect the areas of change over different years?

 Figure 7: I recommend making these maps bigger.

 

Response 9: We have implemented an overlay analysis with the three layers and a graduated colour scheme to better visualise the areas of change across different years added. We have also increased the size of Figure 7 for better clarity and readability.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This paper could be accepted in present form.

Author Response

Thank you 

Reviewer 4 Report

Comments and Suggestions for Authors

There seems to be a misunderstanding regarding the cloud cover issue and the use of TOA reflectance in this context. While it is true that high cloud cover can limit the availability of cloud-free SR images, this issue affects TOA products just as much. TOA reflectance does not solve the problem of cloud cover, it merely provides the raw reflectance without correcting for atmospheric distortions like aerosols, water vapour, and scattering effects. These distortions can lead to inconsistent spectral signatures, especially when conducting temporal analyses or change detection, which relies heavily on accurate SR data. Furthermore, it is written in the revised manuscript that "TOA data provided a more complete and consistent dataset across the time points of interest, ensuring broader temporal and spatial coverage. Additionally, TOA products offer consistent calibration and pre-processing [41,42]" The cited reference [42] is a good example where rigorous pre-processing was applied to obtain SR products, including gap-filling using the F-mask cloud cover as well as SLC-off error. This raises concerns that the methods discussed in the manuscript may not have been thoroughly reviewed, particularly since the cited work addresses the issues that I have raised.

Regarding my comment about the SLC-off error, the revised version mentions that Landsat-5 data was used as "helper" data to fill these gaps. However, this approach may introduce further inconsistencies in the analysis. If Landsat-5 data is available why not just use Landsat-5 data? It would be more logical to directly use it rather than blending it with TOA data from other sources, which could create further discrepancies in the dataset. Either do the gap-filling for SLC-off error or just use Landsat-5.

There are also missing responses to my comments e.g., in comment 8, the author does not address the following comment "Table 4: The precision, recall, and F-scores reported for several land cover classes are unacceptably low, with many values falling below 0.5 indicating an inconsistency in classification performance. OA is high but it is not a good indicator and can be misleading when there is an imbalance of classes. ". I have put a lot of time and effort into the review could you please ensure that a point-by-point response is provided?

Author Response

Comment 1: There seems to be a misunderstanding regarding the cloud cover issue and the use of TOA reflectance in this context. While it is true that high cloud cover can limit the availability of cloud-free SR images, this issue affects TOA products just as much. TOA reflectance does not solve the problem of cloud cover, it merely provides the raw reflectance without correcting for atmospheric distortions like aerosols, water vapour, and scattering effects. These distortions can lead to inconsistent spectral signatures, especially when conducting temporal analyses or change detection, which relies heavily on accurate SR data. Furthermore, it is written in the revised manuscript that "TOA data provided a more complete and consistent dataset across the time points of interest, ensuring broader temporal and spatial coverage. Additionally, TOA products offer consistent calibration and pre-processing [41,42]" The cited reference [42] is a good example where rigorous pre-processing was applied to obtain SR products, including gap-filling using the F-mask cloud cover as well as SLC-off error. This raises concerns that the methods discussed in the manuscript may not have been thoroughly reviewed, particularly since the cited work addresses the issues that I have raised.

Response 1: 

We appreciate your detailed feedback. However, we would like to clarify that this study focuses on land cover classification rather than direct spectral comparison over time. Each image is trained and classified independently, which reduces the significance of atmospheric correction in this context, as the spectral data is converted into land cover categories rather than compared pixel-by-pixel across time points.

Land cover maps generated from TOA data were independently assessed for accuracy before performing the change detection analysis. This ensures the reliability of classification outputs, minimizing the concern that atmospheric correction would significantly impact the final outcomes. Additionally, the accuracy of each input land cover map (2010, 2015, and 2020) was reported and acknowledged prior to conducting the change detection, as clarified in the revised methodology. Reference 42 has been removed as part of the revision.

We hope this clarifies the rationale behind our methodological choices, and we have ensured that this process is transparently described in the manuscript.

Comment 2: Regarding my comment about the SLC-off error, the revised version mentions that Landsat-5 data was used as "helper" data to fill these gaps. However, this approach may introduce further inconsistencies in the analysis. If Landsat-5 data is available why not just use Landsat-5 data? It would be more logical to directly use it rather than blending it with TOA data from other sources, which could create further discrepancies in the dataset. Either do the gap-filling for SLC-off error or just use Landsat-5.

Response 2:

Thank you for your valuable comment regarding the use of Landsat-5 as "helper" data to address the gaps introduced by the SLC-off error in Landsat-7. We understand the concern that blending data from multiple sources may introduce inconsistencies. However, our decision to blend Landsat-7 with Landsat-5 was driven by the need to mitigate significant data gaps caused not only by the SLC-off error but also by cloud-masking, which removed large portions of the data. While Landsat-5 data was available, using it exclusively for 2010 would have left us with fewer cloud-free observations and thus reduced the overall temporal consistency with later years, where Landsat-7 and Landsat-8 were the primary sources. To mitigate potential inconsistencies, we ensured that Landsat-5 and Landsat-7 data were calibrated to maintain radiometric consistency, minimising discrepancies between the two datasets. The calibration process was designed to ensure that the combined dataset remained suitable for multi-temporal analysis. The paragraph has been revised to read 

“Given the persistent cloud cover in tropical regions, helper images from adjacent periods (±1 year) were utilised when insufficient cloud-free scenes resulted in data gaps for the target year [55 ,56]. In the case of 2010, Landsat 5 data from 2009 to 2011 were incorporated as helper data to address significant gaps caused by both cloud masking and the scan line corrector (SLC) failure in Landsat 7 imagery. To ensure the final image was consistent, we calibrated the reflectance values between Landsat 5 and Landsat 7 by applying a scaling process, ensuring that the combined data was radiometrically aligned and suitable for analysis.”

Comment 3: There are also missing responses to my comments e.g., in comment 8, the author does not address the following comment "Table 4: The precision, recall, and F-scores reported for several land cover classes are unacceptably low, with many values falling below 0.5 indicating an inconsistency in classification performance. OA is high but it is not a good indicator and can be misleading when there is an imbalance of classes. ". I have put a lot of time and effort into the review could you please ensure that a point-by-point response is provided?

Response 3: 

We appreciate the concerns regarding the low precision, recall, and F-scores for certain land cover classes. As highlighted in the Discussion {classification accuracy} section, the observed misclassification patterns, particularly among peatland vegetation types, are attributed to the spectral similarities between certain land cover classes (Figure A1). In the case of sparse vegetation, which consistently showed lower metrics, this can be explained by its broad definition, encompassing diverse vegetation types such as young plantation trees, rainfed croplands, small-scale agriculture and others. Consequently, their misclassification with other land cover types was expected. We agree that Overall Accuracy (OA) may not fully capture these nuances, particularly in the presence of class imbalances. To address this, we have included the weighted F1-score, which considers both precision and recall while adjusting for class prevalence, offering a more representative metric of classification performance across all classes. Nevertheless, given the detailed discussion of misclassification patterns and their possible causes, we believe that the current metrics, while imperfect, reflect the inherent complexity of classifying diverse land covers in tropical peatland ecosystems.