Assessing Land-Cover Changes in the Natural Park ‘Fragas do Eume’ over the Last 25 Years: Insights from Remote Sensing and Machine Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

The msc fails to pass the core message on What is new for science? There is a lot of data and work done, but presented as a technical report. There is no hypothesis, nor a clear methodological proposal and the local relevance is the unique issue covered by the authors.

The core question is what are the global or macroregional implications on documenting land cover/ use changes in a protected area?

A few details to consider:

 

Landsat as source of information is very limited. Why not used Sentinell or Spot? The spatial and spectral resolutions are far better in the lasts two images than Landsat.

 

This section must be shortened since the information is already know and documented.

1.      Introduction

Biodiversity constitutes a fundamental element in the proper functioning of ecosystems 32 and the maintenance of services they provide, such as climate regulation, cultural, and 33 provisioning [1]….

 

Paragraphs one to three may be merged into one straight forward.

 

Paragraphs four to five may be merged into one straight forward. Here include the rational expressed in lines from 112-126.

 

Paragraphs six (objective) must be only limited to this:

 

The objective of this study is to quantify the changes in the main land uses and land 81 covers that have occurred in the ‘Fragas do Eume’ Natural Park since its declaration as a 82 protected area to the present day.

 

The rest is part of methods and not needed at all.

 

Proposal: The objective was to assess land cover / use changes in the ‘Fragas do Eume’ Natural Park, combining spatial, and field data sources.

 

 

2.      Methods

 

Define habitat (there is a detailed description of these forests but most readers do not know).

 

Table 3 and Figure 3 as additional data, these are not relevant.  

Do the authors verify the outcomes of the multispectral indexes? It muss be done or complemented with another source.

 

 

 

3.      Results

 

Define habitat (there is a detailed description of these forests but most readers do not know).

 

Figure 4 in percentage to see the proportional change, please.

 

Matrix of change?

 

 

References are lacking. Please review Land for the latest papers published.

 

English must be reviewed in the light of making it straight forward and avoid wording. 

Comments on the Quality of English Language

idem

Author Response

Response to Reviewer #1:

Comment: The manuscript fails to convey the core message about what is new for science. There is a lot of data and work done, but it is presented as a technical report. There is no hypothesis or a clear methodological proposal, and the local relevance is the unique issue covered by the authors. The core question is what are the global or macroregional implications of documenting land cover/use changes in a protected area?

#Response: We have revised the introduction and conclusions of the manuscript to clearly emphasize the core message and the main contribution of our study. In the introduction, we have added a section that explicitly states our working hypothesis, in addition to the specific objectives of the research. Our study not only documents land cover changes in the ‘Fragas do Eume’ Natural Park but also provides an innovative methodology that combines machine learning techniques with multispectral indices to improve the accuracy of satellite image classification. This methodology can be applied in similar studies at a global or macroregional level.

We have expanded the discussion section to address the global and macroregional implications of our findings. In this section, we discuss how the results obtained in ‘Fragas do Eume’ can be extrapolated to other similar contexts in different regions of the world, highlighting the importance of considering different learning machine techniques to deal with the uncertainty associated with the image classification procedures.

We have added references to previous and current studies that underscore the relevance of our findings in a broader, global context.

Comment: Landsat as a source of information is very limited. Why not use Sentinel or SPOT? The spatial and spectral resolutions are far better in the latter two images than Landsat.

Response: Thank you for your insightful comment regarding the choice of satellite imagery. We acknowledge that Sentinel and SPOT offer higher spatial and spectral resolutions compared to Landsat. We decided to use Landsat data due to the following considerations:

1. Landsat provides a continuous time series of imagery dating back to the 1970s, which is crucial for long-term land-cover change analysis. For our study period (1997-2022), Landsat offered a consistent dataset that allowed us to perform a comprehensive 25-year analysis.
2. Using Landsat data ensured consistency in our temporal analysis. While Sentinel and SPOT offer higher resolutions, they do not provide the same historical depth as Landsat for the entire study period. Mixing data sources with different resolutions and acquisition times could introduce inconsistencies and complicate the comparison over time.

However, we recognize the advantages of higher resolution imagery from Sentinel and SPOT and agree that future studies could benefit from these data sources. We have included a discussion on the potential improvements and additional insights that could be gained by incorporating Sentinel data in future research.

Comment: This section must be shortened since the information is already know and documented. 

1. Introduction 

Biodiversity constitutes a fundamental element in the proper functioning of ecosystems 32 and the maintenance of services they provide, such as climate regulation, cultural, and 33 provisioning [1]….Paragraphs one to three may be merged into one straight forward. Paragraphs four to five may be merged into one straight forward. Here include the rational expressed in lines from 112-126. Paragraphs six (objective) must be only limited to this: The objective of this study is to quantify the changes in the main land uses and land 81 covers that have occurred in the ‘Fragas do Eume’ Natural Park since its declaration as a 82 protected area to the present day. The rest is part of methods and not needed at all. Proposal: The objective was to assess land cover / use changes in the ‘Fragas do Eume’ Natural Park, combining spatial, and field data sources. 

Response: Thank you for your constructive feedback regarding the introduction section. We have revised the introduction to address your comments. Below is a summary of the changes made:

1. We have combined the first three paragraphs into one concise paragraph to streamline the information about biodiversity and its importance in ecosystem functioning and service provision.
2. Paragraphs four and five have been merged into a single, straightforward paragraph. This section now includes the rationale expressed in lines 112-126, ensuring a coherent flow of information.
3. We have revised the objective statement to be concise and focused as per your suggestion, while introducing a working hypothesis. The last paragraph now reads: "The objective of this study is to quantify the changes in the main land uses and land covers that have occurred in the ‘Fragas do Eume’ Natural Park since its declaration as a protected area to the present day. Our working hypothesis posits that the decline of Atlantic forest in the ‘Fragas do Eume’ Natural Park is primarily driven by the expansion of eucalyptus plantations. However, this expansion occurs at a lower rate than before the park's designation as a protected area due to its legal protection. If the legal protections are effective, we would expect a halt in the expansion of eucalyptus plantations and subsequent natural recovery of the native forest. Specifically, our study aims to document land cover changes within the park and introduce an innovative methodology that leverages machine learning techniques and multispectral indices to enhance the accuracy of satellite image classification. This methodology has the potential for application in similar studies at both global and macroregional levels."
4. The detailed methodological description has been moved to the methods section, where it is more appropriate.

Comment: 2.      Methods. Define habitat (there is a detailed description of these forests but most readers do not know).

Response#: We have now clearly defined what Evergreen and Deciduous forest types are in our region. Evergreen forests are dense areas where the majority of trees retain their leaves throughout the year, primarily composed of species that remain green and photosynthesize continuously. These forests are found in various climates, ranging from tropical to temperate regions, and are known for their biodiversity and ecological stability. Deciduous forests, on the other hand, are characterized by trees that shed their leaves annually, usually in the autumn. These forests experience distinct seasonal changes, with trees growing new leaves in the spring and shedding them in the fall.

Comment: Table 3 and Figure 3 as additional data, these are not relevant.  

Response#: Since we are not over the limit imposed by the journal on the number of figures and tables, we prefer to keep them in the main body of the manuscript. Table 3 provides crucial information about the training and validation areas, and Figure 3 shows the accuracy for each algorithm, which we believe is very important from a methodological viewpoint.

Comment: Do the authors verify the outcomes of the multispectral indexes? It muss be done or complemented with another source. 

Response#: We are not sure what the reviewer means by "verify the outcomes of the spectral indices." We are using standard functions in the ‘RStoolbox’ package to compute them (see Table 1). A visual assessment confirms that both vegetation and water content indices make sense from an observational standpoint.

Comment: 3. Results. Define habitat (there is a detailed description of these forests but most readers do not know).

Response#: We have now clearly defined what Evergreen and Deciduous forest types are in our region, in lines 519-526.

Comment: Figure 4 in percentage to see the proportional change, please. 

Response#: we have changed the Figure 4 according to then reviewer’s suggestion, and improved the style as suggested by another reviewer.

Comment: Matrix of change? 

Response#: We have replaced the transition matrix with a Sankey diagram, as suggested by Reviewer #2, to facilitate the visualization of the transitions (see new Figure). 

Comment: References are lacking. Please review Land for the latest papers published. 

Response#: we have added more updated references, some suggested by other reviewers.

Comment: English must be reviewed in the light of making it straight forward and avoid wording.

Response#: We have improved the clarity and conciseness of the manuscript by reviewing and refining the language to make it more straightforward and to avoid unnecessary wording.

Thanks for your time dedicated to revise the article and the suggestions that help us to improve the work.

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

Review for Land

Assessing land-cover changes in the Natural Park ‘Fragas do Eume’ over the last 25 years: Insights from remote sensing and machine learning

P. Diaz-Garcia and A. Regos

General Overall Comments:

This paper provides a well-intentioned attempt at exploring land cover change for a specific region involving a Natural Park in Spain. The description of the increasing land cover converted for use in Eucalyptus plantations is compelling and I think this topic could be quite interesting for Land readers but I had a few major concerns about the methodology. My biggest concern about the paper is the limited time aspect of the analysis which definitely limits the interpretation of the landscape changes that were identified. The authors base the change study on data from the satellite record of the Landsat series of satellites (this 30m record which is essentially global in coverage with fairly regular imaging—although the repeat imaging can be a bit more limited prior to 2000). While this global record is a very good sources for evaluating land cover change over time, the authors used two bookend years (1997 and 2022) and analyze March and August imagery during only those two years. The authors then describe change between these two book-end dates. The Landsat archive is freely available distributed by the U.S. Geological Survey (which was not mentioned in the paper). The authors need to justify why they only based their analysis those two dates? A more cohesive and complete study would have involved many more intervals of change than these two….even a study based on land cover change across  5 intervals (every five years) would be a vast improvement  to this book-end type exploration of change. The authors state that fire recovery would be quick in mixed forests (line 329) , but the lack of more intervals in this study means that speed of fire recovery is not explored very well explicitly in this study.

My other main comment concerns the methodology used for how map accuracy was performed and described. The practice of using the train/test feature of machine learning models to separate model training samples into separate sets of training and testing model fitness is not the same as assessing map accuracy with independent and randomly selected reference plots. Good practices for assessing map accuracy are very well described in the Olofsson et al. (2014) publication (https://doi.org/10.1016/j.res.2014.02.015 ).  The practices outlined should be followed in this analysis and if not possible, then explanations as to why they were not followed. I’m concerned that there were potentially large limitations introduced by using non-randomly sampling design for the plots used to estimate accuracy. Furthermore, the practice of (splitting training into train and test in machine learning does not provide a map accuracy assessment.  I would recommend that further analysis be performed to address this shortcoming before publication. 

I would also strongly advise against using the Kappa statistic for describing map accuracy. Please see Pontius and Millones (2011) https://doi.org/10.1080/01431161.2011.552923. This paper has been cited over 1300 times. Kappa is inherently a flawed metric of accuracy. Overall accuracy, producers’ and users’ accuracy, and F1 scores provide appropriate statistical measures of accuracy to support user understanding of map accuracy published in scientific papers.

 

Detailed comments:

There is no mention that the study area is in Spain in the title, abstract, or keywords. That limits the findability of the paper by users. I recommend adding Spain as a key word at a minimum.

Line 316 contains an error, please fix this.

Lines 307,308. The authors state that different sources of information as a possible source for the difference in the extent of change found in this study from other studies. Can the authors please explain this further.

Minor comments:

Tables:

Table 1. Could the authors provide the testing that they did in order to select the five indices shown. There should be evidence for these selections.

Figures:

Figure 4. What is mean by “Extension” as the label on the Y axis in the graph? By the caption, this should be pixels (since these data were the result of pixel counting).

Figure 5. This figure showing maps should have the map graticule descriptions be latitude and longitude. The map x and y (as part of some unknown map projection) is totally inadequate for readers to locate the map area. The authors should include a scale bar as well.

Author Response

Response to Reviewer #2:

Review for Land

General Overall Comments:

Comment: This paper provides a well-intentioned attempt at exploring land cover change for a specific region involving a Natural Park in Spain. The description of the increasing land cover converted for use in Eucalyptus plantations is compelling and I think this topic could be quite interesting for Land readers, but I had a few major concerns about the methodology.

Response#: Thank you for your feedback on our paper. We appreciate your positive comments regarding the significance of the topic.

Comment: My biggest concern about the paper is the limited time aspect of the analysis which definitely limits the interpretation of the landscape changes that were identified. The authors base the change study on data from the satellite record of the Landsat series of satellites (this 30m record which is essentially global in coverage with fairly regular imaging—although the repeat imaging can be a bit more limited prior to 2000). While this global record is a very good sources for evaluating land cover change over time, the authors used two bookend years (1997 and 2022) and analyze March and August imagery during only those two years. The authors then describe change between these two book-end dates. The Landsat archive is freely available distributed by the U.S. Geological Survey (which was not mentioned in the paper). The authors need to justify why they only based their analysis those two dates? A more cohesive and complete study would have involved many more intervals of change than these two….even a study based on land cover change across  5 intervals (every five years) would be a vast improvement  to this book-end type exploration of change.

Response#: Thank you for your insightful comment regarding the limited time aspect of our analysis and the suggestion for a more comprehensive study involving multiple intervals of change. We appreciate your emphasis on the potential benefits of utilizing Landsat data for a broader temporal analysis. We recognize the value of a multi-temporal approach for understanding temporal trends in land cover change. However, our decision to use only two bookend years was primarily driven by the specific objectives of our study, which aimed to quantify the changes since the designation of the park as a protected area in 1997. Additionally, we acknowledge the availability of Landsat data from the U.S. Geological Survey and regret the oversight in not mentioning it in the paper. We now include this information in the revised manuscript. In our case, a multi-temporal approach did not yield significant insights into the conversion between deciduous forest and other land covers over the long-term. While such an approach could potentially reveal temporal trends, it may not provide specific information about the conversion among different land covers. We have now added a Sankey diagram, as suggested by Reviewer #2, to facilitate the visualization of the land-cover conversions. Additionally, previous studies conducted in the natural park have already covered earlier time periods using aerial photography before the Landsat missions, thereby complementing our analysis. We have improved the last section on the introduction and discussion to clarify the main goal of the study.

Comment: The authors state that fire recovery would be quick in mixed forests (line 329) , but the lack of more intervals in this study means that speed of fire recovery is not explored very well explicitly in this study.

Response#: Thank you for bringing this to our attention. Upon reflection, we realize that the statement regarding the quick recovery of mixed forests after fire was not explicitly supported by our study data and may have been based more on general knowledge or assumptions about the region. We appreciate your feedback, and in light of this, we have removed this statement from the conclusions to ensure that our conclusions are firmly grounded in the findings of our study.

Comment: My other main comment concerns the methodology used for how map accuracy was performed and described. The practice of using the train/test feature of machine learning models to separate model training samples into separate sets of training and testing model fitness is not the same as assessing map accuracy with independent and randomly selected reference plots.

Response#: We fully agree with the reviewer. We have now clarified in the revised paper that a set of training areas and validation areas were defined independently to mitigate any potential issues associated with the use of cross-validation procedures with machine learning models. Thank you for highlighting this, and we hope that this clarification addresses your concern.

Comment: Good practices for assessing map accuracy are very well described in the Olofsson et al. (2014) publication (https://doi.org/10.1016/j.res.2014.02.015).  The practices outlined should be followed in this analysis and if not possible, then explanations as to why they were not followed. I’m concerned that there were potentially large limitations introduced by using non-randomly sampling design for the plots used to estimate accuracy. Furthermore, the practice of (splitting training into train and test in machine learning does not provide a map accuracy assessment.  I would recommend that further analysis be performed to address this shortcoming before publication.  

Response#: Thank you for this interesting publication. We acknowledge the importance of adhering to best practices in accuracy assessment and area estimation, as outlined in the recommendations you provided. Key recommendations include reporting error matrices, overall accuracy, user's accuracy, and producer's accuracy; estimating area based on reference classifications; quantifying uncertainty through confidence intervals; evaluating variability and potential errors in reference classifications; and documenting any deviations from best practices that could impact. We have computed error matrices, indicating the overall accuracy, user's accuracy, and producer's accuracy, and kappa index. Please also note that we have considered different algorithms and data sources. Now all this info is included as Supp. Material. Regarding the training and validation areas, we have ensured a minimum number of 10 areas/polygons per class, and have strived to make them homogeneous while still capturing environmental heterogeneity to account for spectral variability within each class. We have now added a more detailed explanation of this procedure.

Comment: I would also strongly advise against using the Kappa statistic for describing map accuracy. Please see Pontius and Millones (2011) https://doi.org/10.1080/01431161.2011.552923. This paper has been cited over 1300 times. Kappa is inherently a flawed metric of accuracy. Overall accuracy, producers’ and users’ accuracy, and F1 scores provide appropriate statistical measures of accuracy to support user understanding of map accuracy published in scientific papers.

Response#: We appreciate your reference, Pontius and Millones (2011), which provides a comprehensive critique of the Kappa statistic and its limitations. We understand that Kappa can be a flawed metric and that there are more appropriate statistical measures available. In response to your recommendation, we will refrain from using the Kappa statistic in our analysis. Instead, we will focus on reporting overall accuracy, producer's accuracy, user's accuracy, and F1 scores as our primary measures of map accuracy. These metrics provide a clearer and more reliable understanding of map accuracy and are widely accepted in the scientific community. We appreciate your guidance on this matter and will ensure that our revised manuscript reflects these changes.

Detailed comments: There is no mention that the study area is in Spain in the title, abstract, or keywords. That limits the findability of the paper by users. I recommend adding Spain as a key word at a minimum.

Response#: We added ‘NW Spain’ in the abstract, and name of the region ‘Galicia’ in the keyworks.

Comment: Line 316 contains an error, please fix this.

Response#: done.

Comment: Lines 307,308. The authors state that different sources of information as a possible source for the difference in the extent of change found in this study from other studies. Can the authors please explain this further.

Response#: The type and resolution of data sources used in landscape change analysis can significantly impact the results. Our study employed Landsat satellite imagery, while other studies have used different satellite datasets and aerial photography. Each data source has its own characteristics, including spatial and temporal resolution, which can lead to variations in the quantified extent of change.

Minor comments:

Tables: Table 1. Could the authors provide the testing that they did in order to select the five indices shown. There should be evidence for these selections.

Response#: We now add in Table 1 a justification for the selection of each index.

Comment. Figure 4. What is mean by “Extension” as the label on the Y axis in the graph? By the caption, this should be pixels (since these data were the result of pixel counting).

Response#: following the suggestion of the reviewer#1, Figure 4 now represents the percentage of coverage of each LC class.

Comment. Figure 5. This figure showing maps should have the map graticule descriptions be latitude and longitude. The map x and y (as part of some unknown map projection) is totally inadequate for readers to locate the map area. The authors should include a scale bar as well.

Response#: we have added UTM coordinates, and the scale bar. We also indicate the UTM coordinate system, WGS84 Zone 29N.

Reviewer 3 Report

Comments and Suggestions for Authors

This study has focused on the land use and vegetation classification within the 'Fragas do Eume' Natural Park, spanning the period from 1997 to 2022. The results clearly shown that deciduous forests in the northwest were replaced by evergreen forests, while there has also been a notable expansion in cropland. This study holds significant guiding implications for biodiversity conservation efforts. However, it is not without its limitations, which are as follows: 

1 The quality of the figures needs to be improved. 

(1) Table 5 & Fig. 4 may specify the units.

(2) The authors might consider using a Sankey diagram as an alternative to a table, which could potentially enhance the expressiveness and clarity of the data presentation. 

(3) Fig.5 needs further modification, including coordinate axis, scale, legends...

(4) Table 4, It doesn't seem necessary to put Pixels and Hectares together, because the pixels are the same size (30×30m). if the author want to use landscape parameter analysis.

2 The discussion encompasses a wide range of perspectives but two years of remote sensing images may not support the data that shown in Line 270-273, According to this study... or a negative trend. I believe the primary issue depicted in Figure 5 is that deciduous forests are being directly supplanted by evergreen forests, a phenomenon that warrants further exploration and discussion. 

3 Line 276-278 The authors emphasize the strong disturbances by forest fire in 2012, but native species are not threatened by invasive species. Let’s assume that the basic features of soil did not change much, however, human activities also play a critical role in “...conversion of native forest into Eucalyptus plantations... (Line 284)”. Because it's worth noting that cropland is also a major land use type in study area and is also expanding. 

4 The Conclusions section is overly extensive (some issues are not deeply analysis in this manuscript) and could benefit from a more concise presentation. Such as Line 327-331.

 

 

Comments on the Quality of English Language

Minor editing

Author Response

Reviewer#3

Comment. This study has focused on the land use and vegetation classification within the 'Fragas do Eume' Natural Park, spanning the period from 1997 to 2022. The results clearly shown that deciduous forests in the northwest were replaced by evergreen forests, while there has also been a notable expansion in cropland. This study holds significant guiding implications for biodiversity conservation efforts.

Response#: thanks for positive feedback and suggestions.

Comment. However, it is not without its limitations, which are as follows:

1 The quality of the figures needs to be improved.

Response#: we have improved different aspects in all figures, following reviewers’ suggestions.

• Table 5 & Fig. 4 may specify the units.

Response#: Table 5 is in hectares. Fig 4 now indicates the percentage of coverage. Captions now include the units.

Comment. (2) The authors might consider using a Sankey diagram as an alternative to a table, which could potentially enhance the expressiveness and clarity of the data presentation.

Response#: Thanks for this very helpful suggestion. We have added a Sankey diagram, which illustrates the transitions between different land use types over time. The categories include Croplands (yellow), Deciduous Forest (red), Evergreen Forest (green), Shrubland (light green), and Water (blue). The flow lines between the source (left) and target (right) nodes represent the changes in land use, measured in hectares. This visualization highlights the dynamic nature of land cover transitions within the study area and enhances the expressiveness and clarity of the data presentation.

Comment. (3) Fig.5 needs further modification, including coordinate axis, scale, legends...

Response#: we have added UTM coordinates, and the scale bar. We also indicate the UTM coordinate system, WGS84 Zone 29N.

Comment. (4) Table 4, It doesn't seem necessary to put Pixels and Hectares together, because the pixels are the same size (30×30m). if the author want to use landscape parameter analysis, I recommended to consult the following article for detailed calculation methodologies and best practices, Urban vegetation cooling capacity was enhanced under rapid urbanization in China.

Response#: we agree that using hectares and pixels is redundant, so we have removed the column ‘Pixels’. Thanks for the reference to this interesting study. If the referee refers to the geometric and radiometric procedures of the Landsat images needed, we have applied the standards methods recommended, as already mentioned in the subsection preprocessing analysis.

Comment. 2 The discussion encompasses a wide range of perspectives but two years of remote sensing images may not support the data that shown in Line 270-273, According to this study... or a negative trend. I believe the primary issue depicted in Figure 5 is that deciduous forests are being directly supplanted by evergreen forests, a phenomenon that warrants further exploration and discussion. It may be useful to select a few specific years to explore trends (Changes in spatio-temporal patterns of urban forest and its above-ground carbon storage: Implication for urban CO2 emissions mitigation under China's rapid urban expansion and greening. Environment international 129 (2019): 438-450.)

Response#: Yes, the primary objective of our study is to assess the long-term changes of deciduous forest and to understand the main conversion from this habitat, and to do so we used a diachronic approach by using the year just after the designation as natural park as reference level. We have clarified the objective of our study and also added a Sankey diagram, which helped us to show the transitions among the different land use types. We have changed the term ‘negative trend’ by ‘loss’ since, as the reviewer said, we focused on two snapshots rather than on a multitemporal analysis. Thanks again for the reference.

Comment. Line 276-278 The authors emphasize the strong disturbances by forest fire in 2012, but native species are not threatened by invasive species. Let’s assume that the basic features of soil did not change much, however, human activities also play a critical role in “...conversion of native forest into Eucalyptus plantations... (Line 284)”. Because it's worth noting that cropland is also a major land use type in study area and is also expanding.

Response#: Yes, right. We have now mentioned the impact of cropland activities, and the expansion found in this time period.

 

Comment. 4 The Conclusions section is overly extensive (some issues are not deeply analysis in this manuscript) and could benefit from a more concise presentation. Such as Line 327-331.

Response#: we have shortened the conclusions by focusing exclusively on the main findings and recommendations derived from our analysis.

Comment. Some other references may be also useful for revision.

(1) Rapid urbanization and meteorological changes are reshaping the urban vegetation pattern in urban core area: A national 315-city study in China-

(2) Understanding the cooling capacity and its potential drivers in urban forests at the single tree and cluster scales (About forest classification, this reference maybe helpful)

Response#: thanks for the references, they have been incorporated in the discussion to support some of our statements

Reviewer 4 Report

Comments and Suggestions for Authors

I have carefully reviewed the manuscript titled “Assessing land-cover changes in the Natural Park ‘Fragas do Eume’ over the last 25 years: insights from remote sensing and machine learning”. The study addresses an important topic for land use change using machine learning models. While the authors have presented a manuscript with a complete structure, I have several comments and suggestions for both the editors and authors to consider:

1. The introduction mentions the widespread use of machine learning methods in land use classification but lacks a detailed description of the current state-of-the-art methods and the challenges they address. Consider expanding this section to include a technical overview.

 

2.It is recommended that Figures include a legend, scale, compass, and coordinates to align with scientific standards.

3. I don't quite understand the meaning of the numbers in Table 2, an explanation could be added.

4.The font size in Figure 3 appears small, and it would be more logical to include horizontal and vertical axes for better data interpretation.

 

5. It is suggested that the figure consider using areas instead of pixels for a more accurate representation. Additionally, revising the color scheme of the bars could improve the aesthetic appeal.

6. Figures 2 and 5 lack clarity. If possible, replace them with clearer images. For Figure 5, consider using a grid based on latitude and longitude for accuracy. As with Figure 1, adding legends, scales, and compasses would aid in visualizing changes in different land use types.

7.  A detailed description of the machine learning methods used, including principles and hyperparameters, is suggested. Given that only five vegetation indices were selected using a method that omits spatial information, consider incorporating texture indices and geometric features. If there are too many features, feature extraction methods like PCA could be employed.

8. The title suggests a 25-year span, but the study primarily focuses on two years (1995 and 2022). Given the continuous and accessible nature of Landsat imagery during this period, consider expanding the analysis to include additional years.

9. While neural networks, SVM, and RF are known to produce good classification results, your study indicates the highest accuracy with PLS. Conducting cross-validation experiments could provide a more comprehensive analysis and support the accuracy claims.

Author Response

Reviewer#4

I have carefully reviewed the manuscript titled “Assessing land-cover changes in the Natural Park ‘Fragas do Eume’ over the last 25 years: insights from remote sensing and machine learning”. The study addresses an important topic for land use change using machine learning models. While the authors have presented a manuscript with a complete structure, I have several comments and suggestions for both the editors and authors to consider:

Comment. 1. The introduction mentions the widespread use of machine learning methods in land use classification but lacks a detailed description of the current state-of-the-art methods and the challenges they address. Consider expanding this section to include a technical overview.

Response#: We agree with the reviewer that these issues can enrich the introduction. We have added s few lines to highlight the uncertainty associated with the image classification methods and the benefits of using multiple algorithms to deal with and community such uncertainty.  

Comment  2.It is recommended that Figures include a legend, scale, compass, and coordinates to align with scientific standards.

Response#: as suggested also by other reviewers, we had included legend, and specify the coordinate system in the figures.

Comment 3. I don't quite understand the meaning of the numbers in Table 2, an explanation could be added.

Response#: we now indicate that these number are the habitat code on the habitat listed in the annexes of the EU habitat directive.

Comment 4.The font size in Figure 3 appears small, and it would be more logical to include horizontal and vertical axes for better data interpretation.

Response#: we have increased the font size and added labels to each axis.

 

Comment 6. Figures 2 and 5 lack clarity. If possible, replace them with clearer images. For Figure 5, consider using a grid based on latitude and longitude for accuracy. As with Figure 1, adding legends, scales, and compasses would aid in visualizing changes in different land use types.

Response#: We have improved both figures and indicated the coordinate system, UTM 29 North, adding legends, scales, and compasses.

Comment 7.  A detailed description of the machine learning methods used, including principles and hyperparameters, is suggested. Given that only five vegetation indices were selected using a method that omits spatial information, consider incorporating texture indices and geometric features. If there are too many features, feature extraction methods like PCA could be employed.

Response#: thanks for the suggestion, we had added a reference to the ‘caret’ package wherein by default parameters are indicated. We have not included texture indices or geometric figures in the classification process, since we don’t much ecological meaning on those indices and the accuracy of the final map is very high, with very litter for improvement in this regard.

Comment 8. The title suggests a 25-year span, but the study primarily focuses on two years (1995 and 2022). Given the continuous and accessible nature of Landsat imagery during this period, consider expanding the analysis to include additional years.

Response#: In our case, a multi-temporal approach would not yield significant insights into the conversion between deciduous forest and other land covers over the long-term. While such an approach could potentially reveal temporal trends, it may not provide specific information about the conversion among different land covers. We have now added a Sankey diagram, as suggested by Reviewer #2, to facilitate the visualization of the land-cover conversions. Additionally, previous studies conducted in the natural park have already covered earlier time periods using aerial photography before the release of the first Landsat satellite, thereby complementing our analysis.

Comment 9. While neural networks, SVM, and RF are known to produce good classification results, your study indicates the highest accuracy with PLS. Conducting cross-validation experiments could provide a more comprehensive analysis and support the accuracy claims.

Response#: We understand the reviewer‘s concern but cross-validation procedures (e.g. by partitioning the dataset into training and validation subsets) are usually carried out when independent validation datasets are not available. Artificial neural networks, SVM, and RF often yield high accuracy results in cross-validation validation procedures, but they might overestimate the accuracy due to the large number of iterative internal models. It is better to test the accuracy of those models using data independent of the calibration datasets. For this reason, we have defined validation and training areas independently. We think that doing internal cross-validation would not provide many insights in the accuracy of the final maps.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

Well revised.

Reviewer 4 Report

Comments and Suggestions for Authors

I think the revised manuscript can be accepted now.