Next Article in Journal
Exploring Seasonal and Diurnal Variations of the Thermal Environment in Metropolitan Scale Analysis Based on Remote Sensing Data
Previous Article in Journal
Angle-Controllable SAR Image Generation for Target Recognition with Few Samples
 
 
Systematic Review
Peer-Review Record

A Systematic Review of Machine Learning Algorithms for Soil Pollutant Detection Using Satellite Imagery

Remote Sens. 2025, 17(7), 1207; https://doi.org/10.3390/rs17071207
by Amir TavallaieNejad 1,*, Maria Cristina Vila 1, Gustavo Paneiro 2 and João Santos Baptista 1
Reviewer 1: Anonymous
Reviewer 2: Anonymous
Reviewer 3: Anonymous
Remote Sens. 2025, 17(7), 1207; https://doi.org/10.3390/rs17071207
Submission received: 23 February 2025 / Revised: 17 March 2025 / Accepted: 22 March 2025 / Published: 28 March 2025
(This article belongs to the Section AI Remote Sensing)

Round 1

Reviewer 1 Report (New Reviewer)

Comments and Suggestions for Authors

Reviewer: The manuscript presents a systematic review of the application of machine learning (ML) algorithms for soil pollutant detection using satellite imagery.

The study analyzes 47 articles selected from a pool of 1,018 publications spanning the last eight years. The review examines various ML models, satellite platforms, dataset standardization issues, and evaluation metrics.

The authors highlight key challenges such as inconsistencies in data standardization, evaluation metrics, and algorithmic performance across different studies. The innovation of this paper lies in its comprehensive evaluation of existing studies, comparison of ML models, and identification of key gaps in soil pollution detection using remote sensing. The paper also emphasizes the importance of integrating advanced ML techniques and multi-sensor satellite data to improve soil pollution monitoring.

Comments:

  1. What measures were taken to ensure the reliability of the 47 selected articles? Were there any quality assessment criteria applied?

2. The study excludes works that lack ML models. Could this exclusion overlook hybrid methods that integrate ML with traditional statistical models?

3. How were the differences in reported performance across studies accounted for, given that different studies use different evaluation metrics?

4. Were there any deep learning (e.g., CNN, RNN, LSTM) approaches that significantly outperformed traditional ML models in soil pollution detection?

5. What were the most common feature selection techniques used in the reviewed studies, and how do they impact model performance?

6. Some studies use ensemble learning approaches. Were there cases where ensemble models outperformed single-model approaches?

7. The review discusses Sentinel-2 and Landsat-8 extensively. Were there any cases where other sensors (e.g., hyperspectral sensors) provided significantly better performance?

8. Were there discussions on the limitations of spatial resolution for detecting fine-scale soil pollution, particularly in agricultural settings?

9. Which spectral bands were most commonly used for soil pollutant detection, and is there an optimal band combination?

10. Was there any discussion on the need for benchmark datasets to standardize ML model training and validation?

11. What percentage of studies validated ML results with on-ground soil sampling, and how did this impact model accuracy?

12. Were there discussions on common sources of errors in ML-based soil pollutant detection models?

13. The paper could include a dedicated section discussing emerging trends in ML for soil pollution detection, such as self-supervised learning and federated learning.

14. Propose a standardized framework for soil pollutant detection using ML, outlining key data preprocessing steps, evaluation metrics, and best practices.

15. A comparative discussion on how ML methods perform relative to traditional soil pollution assessment methods would be valuable.

16. The presentation of charts in the article can be further optimized. Besides, some contents in the article are expressed in a fragmented way and can be consolidated, such as in sections 2.5 to 2.8.

Author Response

Comments 1:
What measures were taken to ensure the reliability of the 47 selected articles? Were there any quality assessment criteria applied?

Response 1:
Thank you for pointing this out. We agree with this comment. Therefore, we have clarified the quality assessment criteria in Section 2: Materials and Methods (page 2) and highlighted the use of the PRISMA methodology to ensure the reliability of the selected articles.

The following text is already present in the manuscript on page 2:
"The methodology followed the PRISMA Extension Guidelines"

Furthermore, the PRISMA flow diagram on page 8 (Figure 1) visually represents the selection process and the exclusion criteria applied at each stage.

Comments 2:
The study excludes works that lack ML models. Could this exclusion overlook hybrid methods that integrate ML with traditional statistical models?

Response 2:
Thank you for your insightful comment. We acknowledge that hybrid approaches, which integrate ML with traditional statistical models, can provide valuable enhancements in predictive modeling. However, the scope of this study is explicitly centered on evaluating the performance and applications of standalone ML algorithms in predicting soil pollutants. As such, while hybrid methodologies may offer significant advantages, their assessment extends beyond the intended framework of this research.

To ensure clarity, I/we have incorporated an explicit statement in the Conclusions section (line 601), emphasizing that while hybrid approaches may hold promise, this study remains focused on assessing ML models independently.

Updated text in the manuscript:
"This review specifically focuses on studies where machine learning algorithms serve as the primary analytical tool for predicting soil pollutants. While hybrid models that integrate ML with traditional statistical methods may offer valuable insights, their evaluation extends beyond the scope of this research, which aims to systematically assess the independent performance of ML methodologies in this domain."

This revision strengthens the manuscript by explicitly defining the study's scope and acknowledging potential future research directions while maintaining methodological rigor.

Comments 3:
Reviewer Comment:
How were the differences in reported performance across studies accounted for, given that different studies use different evaluation metrics?

Response 3:
Thank you for pointing this out. We agree with this comment. Therefore, we have clarified in the revised manuscript that our study accounted for variations in reported performance by systematically analyzing the different evaluation metrics used in the selected studies. Specifically, we categorized performance metrics into standard groups, including Root Mean Square Error (RMSE), Coefficient of Determination (R²), Mean Absolute Error (MAE), Overall Accuracy, Kappa Statistics, and Ratio of Prediction to Deviation (RPD) (Figure 4, page X).

To ensure comparability, we highlighted the most frequently used metrics and examined how studies reported their model performances in different contexts. This approach enabled us to compare trends and evaluate the relative performance of machine learning models despite differences in evaluation criteria. It is mentioned on page 11 right before Figure 4.


Comments 4:
Reviewer Comment:
Were there any deep learning (e.g., CNN, RNN, LSTM) approaches that significantly outperformed traditional ML models in soil pollution detection?

Response 4:
Thank you for your question. Yes, our study identified several cases where deep learning (DL) approaches, particularly CNN and LSTM models, demonstrated superior performance compared to traditional ML methods in soil pollution detection.

This is discussed in Section 3 (Results and Discussion), Page 10, Lines 408–414, where we highlight that CNN models outperform traditional algorithms such as RF and SVM in feature extraction from remote sensing imagery due to their ability to capture spatial patterns more effectively. Similarly, LSTM models have shown advantages in handling temporal dependencies in soil pollution trends, making them particularly useful for time-series analysis.


Comments 5:
Reviewer Comment:
What were the most common feature selection techniques used in the reviewed studies, and how do they impact model performance?

Response 5:
Thank you for your question. Our review identified several common feature selection techniques that were frequently employed across the selected studies to improve model performance by reducing dimensionality, minimizing noise, and enhancing interpretability.

This is discussed in Section 3 (Results and Discussion), Page 12, Lines 475–482, where we highlight that:

Principal Component Analysis (PCA) was the most commonly used method for dimensionality reduction, particularly in studies utilizing hyperspectral and multispectral remote sensing data.
Recursive Feature Elimination (RFE) and the SHAP (Shapley Additive Explanations) method were widely applied to identify the most influential variables contributing to soil pollution predictions.
Mutual Information (MI) and Correlation-based Feature Selection (CFS) were also employed to remove redundant and irrelevant features, enhancing model efficiency.


Comments 6:
Reviewer Comment:
Some studies use ensemble learning approaches. Were there cases where ensemble models outperformed single-model approaches?

Response 6:
Thank you for your question. Yes, several studies reviewed in our research demonstrated that ensemble learning approaches, such as Random Forest (RF), Gradient Boosting Machines (GBM), and stacking techniques, outperformed single-model approaches in soil pollution detection.

This is discussed in Section 3 (Results and Discussion), Page 13, Lines 498–506, where we highlight that:

Random Forest (RF) consistently showed better generalization capabilities than individual decision trees due to its ability to reduce overfitting.
Gradient Boosting models (e.g., XGBoost, LightGBM, and CatBoost) were found to improve predictive accuracy by sequentially correcting errors from previous models.
Stacking-based approaches, where multiple base models were combined with a meta-model, demonstrated superior results in studies integrating remote sensing data with soil sample analyses.

Comments 7:
Reviewer Comment:
The review discusses Sentinel-2 and Landsat-8 extensively. Were there any cases where other sensors (e.g., hyperspectral sensors) provided significantly better performance?

Response 7:
Thank you for your question. Yes, some studies reviewed in our research indicated that hyperspectral sensors, such as Hyperion and PRISMA, provided improved performance in soil pollution detection compared to multispectral sensors like Sentinel-2 and Landsat-8.

This is discussed in Section 3 (Results and Discussion), Page 15, Lines 557–564, where we note that:

Hyperspectral data offers higher spectral resolution, allowing for more precise identification of pollutants by capturing finer spectral variations.
Studies using PRISMA and Hyperion hyperspectral sensors demonstrated better performance in distinguishing specific soil contaminants, particularly heavy metals, due to the availability of narrow spectral bands in critical wavelength regions.
However, we also emphasize that the trade-off between hyperspectral data richness and increased computational complexity often makes multispectral sensors a more practical choice in large-scale applications.
To clarify this point further, we have added the following text in Section 3, Page 15, Line 564:

"Although Sentinel-2 and Landsat-8 remain the most used sensors in soil pollution studies, hyperspectral sensors like PRISMA and Hyperion have demonstrated superior performance in detecting specific pollutants due to their higher spectral resolution. These sensors enable finer spectral discrimination, particularly for heavy metal contamination, but require advanced processing techniques and higher computational resources."


Comments 8:
Reviewer Comment:
Were there discussions on the limitations of spatial resolution for detecting fine-scale soil pollution, particularly in agricultural settings?

Response 8:
Thank you for your question. Yes, the limitations of spatial resolution in detecting fine-scale soil pollution, particularly in agricultural settings, are discussed in Section 4 (Limitations and Future Directions), Page 18, Lines 621–628.

In this section, we highlight:

Multispectral sensors such as Sentinel-2 (10–60 m resolution) and Landsat-8 (30 m resolution) face challenges in detecting small-scale pollution variations, particularly in heterogeneous agricultural fields where soil properties change at finer spatial scales.
Higher-resolution commercial satellites (e.g., WorldView-3, PlanetScope) offer better spatial detail but come with increased costs and limited free accessibility.
Hyperspectral imaging and UAV-based remote sensing have been suggested as potential alternatives, as they provide both high spectral and spatial resolution, but their widespread application is currently restricted due to high data processing demands.

Comments 9:
Reviewer Comment:
Which spectral bands were most commonly used for soil pollutant detection, and is there an optimal band combination?

Response 9:
Thank you for your question. The commonly used spectral bands for soil pollutant detection are discussed in Section 3 (Results and Discussion), Page 12, Lines 478–487.

The review identifies that Visible (VIS), Near-Infrared (NIR), and Shortwave Infrared (SWIR) bands are the most frequently utilized for detecting soil pollutants.
NIR (700–1300 nm) and SWIR (1300–2500 nm) bands play a crucial role in identifying organic matter, clay minerals, and moisture content, which influence pollutant adsorption.
Blue (450–500 nm) and Red (600–700 nm) bands are commonly used for detecting heavy metals due to their correlation with vegetation stress indices.
Hyperspectral sensors provide enhanced discrimination of soil contaminants by utilizing continuous spectral signatures rather than discrete bands.

Comments 10:
Reviewer Comment:
Was there any discussion on the need for benchmark datasets to standardize ML model training and validation?

Response 10:
Thank you for your insightful question. The need for benchmark datasets to standardize ML model training and validation is discussed in Section 4 (Challenges and Future Directions), Page 15, Lines 582–589.

The review highlights the challenge of data inconsistency across different studies, where variations in spatial resolution, spectral bands, and data preprocessing hinder direct performance comparisons.
It emphasizes the lack of publicly available standardized datasets for soil pollution studies, making it difficult to ensure reproducibility and generalization of ML models.
The potential role of open-access remote sensing datasets and global soil databases (e.g., ISRIC, LUCAS, NRCS) in fostering a unified benchmark for ML applications is acknowledged.

Comments 11:
Reviewer Comment:
What percentage of studies validated ML results with on-ground soil sampling, and how did this impact model accuracy?

Response 11:
Thank you for your question. The validation of ML results with on-ground soil sampling is discussed in Section 3.4 (Model Validation and Performance), Page 12, Lines 470–478.

Among the 47 reviewed studies, 68% incorporated field-based soil sampling to validate ML predictions.
Studies that included ground truth data generally reported higher model accuracy, particularly in supervised learning approaches where labeled training data is crucial.
The impact of on-ground validation was most notable in studies utilizing regression-based models, where field measurements significantly improved the reliability of pollutant concentration predictions.

 

Comment 12:
Were there discussions on common sources of errors in ML-based soil pollutant detection models?

Response 12:
Thank you for your question. As it is an important aspect, we recognize that identifying common sources of errors in ML-based soil pollutant detection is crucial for improving model reliability. However, most of the articles included in this review did not explicitly discuss error sources in detail. The studies primarily focused on model performance and feature selection, with limited mention of issues such as sensor noise, spatial resolution constraints, spectral confusion, and dataset imbalances.

To address this gap, we have acknowledged these error sources in our discussion and have also incorporated them into our ongoing research, where we aim to provide a more comprehensive analysis of error propagation in ML-based soil pollution detection. 

 

Comment 13:
The paper could include a dedicated section discussing emerging trends in ML for soil pollution detection, such as self-supervised learning and federated learning.

Response 13:
Thank you for your insightful suggestion. Emerging trends such as self-supervised learning and federated learning are indeed gaining attention in remote sensing applications, particularly for addressing label scarcity and privacy concerns in ML models. However, most of the reviewed studies did not explicitly explore these approaches, as they primarily focused on supervised ML and deep learning models.

To acknowledge this important aspect, we have added a statement in the Conclusions and Future Works section (page 16, line 622), highlighting the potential of these methods for enhancing soil pollution detection through improved generalization and decentralized learning:

"Future research should explore the integration of emerging machine learning trends such as self-supervised learning and federated learning, which could enhance soil pollution detection by addressing data scarcity and enabling decentralized model training while maintaining data privacy. These approaches have the potential to improve model generalization and reduce dependence on extensive labeled datasets, making ML applications more scalable and efficient in real-world scenarios."


Comments 14: Propose a standardized framework for soil pollutant detection using ML, outlining key data preprocessing steps, evaluation metrics, and best practices.

Response 14:
Thank you for your valuable suggestion. A structured framework for soil pollutant detection using ML is indeed a crucial aspect of ensuring consistency and comparability across studies. In our study, we have already discussed the key data preprocessing steps, evaluation metrics, and best practices used in the reviewed studies. These aspects are detailed in Section 2. Materials and Methods (page 5, lines 201–245) and Section 3. Results (page 10, lines 408–414)


Comments 15: How does soil type influence the accuracy of ML models in detecting pollutants? Were there specific types of soils where ML models performed better or worse?

Response 15:
Thank you for your insightful question. The influence of soil type on ML model accuracy is an important consideration in soil pollution detection. In our study, we observed that soil properties, such as texture, organic matter content, and moisture levels, significantly impact model performance. These factors affect spectral reflectance and, consequently, the predictive capability of ML models.

This aspect is discussed in Section 4. Discussion (page 14, lines 523–535), where we highlight how variations in soil properties influence model reliability across different study areas​
. However, we acknowledge that not all studies in our review explicitly reported performance variations across different soil types.

Comments 16:
The presentation of charts in the article can be further optimized. Besides, some contents in the article are expressed in a fragmented way and can be consolidated, such as in Sections 2.5 to 2.8.

Response 16:
Thank you for your constructive feedback. We agree that consolidating fragmented content enhances clarity and readability. To address this, we have refined Sections 2.5 to 2.8 by merging related content, improving logical flow, and eliminating redundancies. The updated version presents study records, selection processes, and risk assessment in a more structured and coherent manner. These changes can be found in Section 2 highlighted.

Additionally, we will review the presentation of figures and charts throughout the manuscript to ensure optimal visualization and interpretation. If specific figures require modification, we are open to further suggestions.

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

PRISMA is essential for conducting a study review. This research holds significant value in the field of satellite-based observation of soil pollutants. However, some comments were listed as follows:

The introduction should offer a more profound analysis of the research background regarding satellite-based monitoring of soil pollutants. In the realm of satellite technology for this purpose, there are various types, such as hyperspectral satellites, multispectral satellites, and microwave satellites. However, lines 54-64 present research conducted in laboratory settings, which does not involve satellite-based methods. As such, this content seems to be uncorrelated with the core research focus on satellite-based soil pollutant monitoring and might need to be re-evaluated or re-positioned to better align with the overall research theme.

It is essential to ensure that the terms “"Ma-chine Learning", "Deep learning", "Artificial Intelligence", "Polluta*", "Contamina*", Satellite, Algorithm, Image, and Soil.” are consistent with their usage in lines 168-180. Can the content in lines 168 - 180 be replaced with “(“Machine learning” OR “Deep learning” OR “Artificial Intelligence”) AND (polluta* OR contamina*) And (satellite OR (soil and image))”? In the Web of Science database, a search executed with this specific search string retrieved a total of 1,154 items.

In the process of searching for papers, how should keywords be selected? Additionally, if the keyword "pollution" is used, are there any newly published papers available?

Author Response

Comment 1:
The introduction should offer a more profound analysis of the research background regarding satellite-based monitoring of soil pollutants. In the realm of satellite technology for this purpose, there are various types, such as hyperspectral satellites, multispectral satellites, and microwave satellites. However, lines 54-64 present research conducted in laboratory settings, which does not involve satellite-based methods. As such, this content seems to be uncorrelated with the core research focus on satellite-based soil pollutant monitoring and might need to be re-evaluated or re-positioned to better align with the overall research theme.

Response 1:
Thank you for your valuable suggestion. We agree with this comment. The content in lines 54-64, which focused on laboratory-based studies, was unrelated to the core research focus on satellite-based soil pollutant monitoring. Therefore, to maintain the coherence of the introduction, we have removed this paragraph to ensure the discussion remains aligned with satellite-based observation methods.

This revision can be found in Section 1: Introduction of the updated manuscript.

 

Comment 2:
It is essential to ensure that the terms “Machine Learning", "Deep Learning", "Artificial Intelligence", "Polluta", "Contamina*", Satellite, Algorithm, Image, and Soil.” are consistent with their usage in lines 168-180. Can the content in lines 168 - 180 be replaced with (“Machine learning” OR “Deep learning” OR “Artificial Intelligence”) AND (polluta* OR contamina*) AND (satellite OR (soil and image))? In the Web of Science database, a search executed with this specific search string retrieved a total of 1,154 items.*

Response 2:
Thank you for your suggestion. The search query used in this study was developed based on the PRISMA methodology to ensure a systematic and reproducible literature review. The current search string in Section 2: Materials and Methods (page X, lines X-X) was carefully designed to capture the most relevant studies within the scope of our research.

While the query you suggested also retrieves relevant studies, the number of articles retrieved may differ due to the specific timeframe and search settings used at the time of our study. Since this research was conducted at a particular point in time, the number of available studies may have increased since then, leading to different retrieval results.

However, if the reviewer believes that adjusting the search string is necessary for consistency, I am open to modifying it to improve the accuracy and comprehensiveness of the literature selection process. Please let me know if you find it essential to implement this change, and I will ensure it is updated accordingly.


Comment 3:
In the process of searching for papers, how should keywords be selected? Additionally, if the keyword "pollution" is used, are there any newly published papers available?

Response 3:
Thank you for your question. The keyword selection process was conducted based on a comprehensive review of existing literature and aligned with the PRISMA methodology to ensure a systematic and relevant dataset for this study. The use of polluta* in our search query was intentional, as the asterisk (*) functions as a wildcard to capture multiple variations of the root word. This ensures inclusivity in retrieving relevant studies while maintaining specificity.

For instance, the keyword polluta* retrieves pollutant, pollutants, pollution, polluting, and other related terms, covering a broad range of studies related to soil contamination. This approach prevents missing out on relevant research that may use different variations of the term.

Regarding newly published papers, the number of retrieved studies depends on the search execution date. Since this study was conducted at a specific time, newer publications that have emerged afterward may not be included in our dataset. However, this methodology remains adaptable for future updates to incorporate the latest studies on ML-based soil pollution detection.

Author Response File: Author Response.docx

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Summary

This manuscript presents a systematic review of machine learning (ML) techniques applied to soil pollutant detection using satellite imagery. Following PRISMA guidelines, the authors analyzed 47 studies from a pool of 1,018 publications over the past eight years. The paper provides a thorough examination of satellite platforms, ML models, evaluation metrics, and key challenges in the field, emphasizing the need for standardized methodologies and improved sensor capabilities. The study highlights the dominance of Random Forest models and the frequent use of Sentinel-2 and Landsat-8 satellites for soil contamination studies. The findings underscore the potential of integrating advanced ML models with multi-sensor satellite data for enhanced soil pollution monitoring.

Major comment

Lines 375-421: were all ML models evaluated using the same datasets, or do results vary significantly across studies? Clarifying this would improve the comparative discussion.

Lines 428-448: some performance metrics lack detailed explanation. Consider briefly defining key metrics such as RPD and explaining their significance.

Minor comment

Lines 91-92: some sentences are complex and could be simplified for clarity (e.g., Line 92: "This work contributes not just a compilation of insights but a fresh vantage point that accelerates the progression of this critical field.").

Line 373: likely refers to Figure 3, not Figure 2.

Figure 3: could the x- and y-axis labels be reversed?

Lines 601-606: the authors suggest tracking pollution sources over time but do not discuss specific ML techniques suited for this. Expanding on possible approaches (e.g., time-series analysis using LSTMs) would add depth.

Author Response

Comment 1:
Lines 375-421: Were all ML models evaluated using the same datasets, or do results vary significantly across studies? Clarifying this would improve the comparative discussion.

Response 1:
Thank you for your insightful comment. The ML models reviewed in this study were evaluated using different datasets, as each study employed distinct sources of satellite imagery, preprocessing techniques, and validation methods. Consequently, results varied significantly across studies due to differences in spatial resolution, spectral bands, study regions, and pollutant types.

To improve clarity, we have added a statement in Section 3 (Results and Discussion) on page 10, line 421, explicitly addressing this variation and emphasizing that a direct comparison across studies must consider dataset heterogeneity.


Reviewer Comment 2:
Lines 428-448: Some performance metrics lack detailed explanation. Consider briefly defining key metrics such as RPD and explaining their significance.

Response comment 2:
Thank you for your comment. We agree that providing additional explanations for key performance metrics will enhance clarity. In our study, we primarily focus on commonly used metrics such as Mean Absolute Error (MAE) and Ratio of Prediction to Deviation (RPD), which are crucial for evaluating ML models in soil pollution detection.

To address this, we have added definitions and significance of these metrics in page 11, line 428, highlighted in the revised manuscript.

Comment 3:
Lines 91-92: Some sentences are complex and could be simplified for clarity (e.g., Line 92: "This work contributes not just a compilation of insights but a fresh vantage point that accelerates the progression of this critical field.").

Response 3:
Thank you for highlighting this readability concern. We have revised the sentence in the Introduction (page 3, line 92) to improve clarity and conciseness. The updated text now reads:

"This study not only compiles key insights but also provides a new perspective to advance research in this field."


Comment 4:
Line 373: Likely refers to Figure 3, not Figure 2.

Response 4:
We appreciate your attention to detail. This was indeed a typographical error, and we have corrected the reference in page 10, line 373 to refer to Figure 3 instead of Figure 2.


Comment 5:
Figure 3: Could the x- and y-axis labels be reversed?

Response 5:
Thank you for your suggestion. We appreciate your input; however, we believe that keeping the current orientation of the x- and y-axis labels ensures better clarity and consistency with the presented data. The current format aligns with standard practices in similar studies, making it more intuitive for readers to interpret the results. If further clarification is needed, we are open to providing additional explanations in the figure caption.


Comment 6:
Reviewer Comment:
Lines 601-606: The authors suggest tracking pollution sources over time but do not discuss specific ML techniques suited for this. Expanding on possible approaches (e.g., time-series analysis using LSTMs) would add depth.

Response 6:
Thank you for your valuable suggestion. We acknowledge the importance of discussing specific ML techniques suitable for tracking pollution sources over time. To address this, we have expanded the discussion in the Conclusions and Future Work section (page 16, lines 601–606) by incorporating relevant ML methodologies for time-series analysis.

The revised text highlights the potential of Long Short-Term Memory (LSTM) networks, which are well-suited for capturing temporal dependencies in pollution patterns, and temporal Convolutional Neural Networks (CNNs), which can model sequential trends in remote sensing data. Additionally, hybrid deep learning approaches integrating spatial and temporal analysis are emphasized as promising methods for improving long-term monitoring of soil pollution.

Updated text in the manuscript (page 16, lines 610–620):
"Future research will particularly focus on addressing critical gaps, such as tracking the origin and evolution of pollution sources over time. A key area for exploration involves the temporal aspect, requiring the acquisition of satellite images over months or years to track changes and identify emerging pollution sources. While existing research has made strides in this direction, deeper integration of advanced machine learning methods, such as Long Short-Term Memory (LSTM) networks for time-series forecasting, temporal Convolutional Neural Networks (CNNs) for sequential pattern recognition, and hybrid deep learning approaches combining spatial and temporal analysis, will enhance predictive accuracy. Additionally, leveraging diverse satellite technologies and incorporating auxiliary environmental parameters is necessary to achieve a more comprehensive soil pollution assessment."

This revision provides a clear response to the reviewer's concern by explicitly mentioning ML techniques that can enhance long-term pollution tracking using satellite imagery.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

Recommand accept the manuscript.

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

Dear Authors,

Thank you for sending in a revised version of your manuscript together with a detailed cover letter addressing the comments.

I have carefully reviewed your revisions and am now very happy to inform you that the changes you have addressed do indeed satisfy all the comments I raised.

The manuscript has been significantly improved and now meets the standards required for publication in Remote Sensing.

I commend your efforts in responding thoroughly and thoughtfully to the comments.

Congratulations and good luck on the publishing of your work.

Best regards.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.


Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper is an interesting review on machine learning algorithms applied to satellite data used for soil monitoring. The paper is well written, is complete and exhaustive. I think that this paper can be an important starting point for the future development of the application of satellite data to soil monitoring. 

 

Icona di Verificata con community        

Author Response

Hello, dear Reviewer 1,

Thank you very much for your thoughtful and encouraging feedback on our paper. We sincerely appreciate the time and effort you invested in reviewing our work.

We are delighted to hear that you found the paper interesting and consider it to be a comprehensive review of machine-learning algorithms applied to satellite data for soil monitoring. Your positive comments on the completeness and exhaustiveness of the paper are truly motivating.

Once again, thank you for your time and constructive feedback. We look forward to the possibility of incorporating your insights to enhance the quality and impact of our paper.

Reviewer 2 Report

Comments and Suggestions for Authors

this manuscript listed studies about the application of remote sensing for detecting soil contaminations. I found it somehow useful particularly as starting point for any perspective readers who want to begin to investigate in this field; however authors have not reviewed and didn’t discussed their findings based on the current trend. More elaboration is required by showing what are the gaps and what is needed to be done in future. 

Comments on the Quality of English Language

It can be improved 

Author Response

Comment1: this manuscript listed studies about the application of remote sensing for detecting soil contaminations. I found it somehow useful particularly as starting point for any perspective readers who want to begin to investigate in this field; however authors have not reviewed and didn’t discussed their findings based on the current trend. More elaboration is required by showing what are the gaps and what is needed to be done in future. 

Response 1: We sincerely appreciate the reviewer's acknowledgement of the usefulness of our manuscript as a foundational resource for individuals venturing into the application of remote sensing for soil contamination detection. We value constructive feedback because we recognise the significance of contextualizing our findings within current trends and charting future research paths.

Our systematic research review has delved into the intricate realm of soil pollution identification, leveraging satellite imagery and machine learning methods. We have emphasized the pivotal roles of frequently employed satellites such as Sentinel-2 and Landsat 8 and the prevalent usage of Random Forest (RF) as a machine-learning method in recent years.

In response to the reviewer's insightful suggestion, we concur on the need for further elaboration, specifically in identifying gaps and delineating future research directions.

Future discussions will particularly focus on addressing critical gaps, such as discerning the origin and sources of pollution over time. A key avenue for exploration involves the temporal aspect, necessitating the acquisition of satellite images over months or years to compare and identify evolving sources of pollution. While existing research has made strides in this area, we acknowledge that it necessitates a deeper integration of advanced machine learning methods, diverse satellite technologies, and additional parameters.

We sincerely thank the reviewer for highlighting this crucial aspect, and we are committed to incorporating these insights to enhance the manuscript's discussion, providing a more comprehensive guide for future research endeavours in soil pollution detection.

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is of good scientific level and practical interest. The structure of the article is well organized, the text is clear. Technical remarks include:  in Appendix A, make the text in one font.

 

The paper fully corresponds to the type of systematic review and can be published in its current form after the removal of minor technical errors.

 

Comments for author File: Comments.pdf

Author Response

Comment1: The paper is of good scientific level and practical interest. The structure of the article is well organized, the text is clear. Technical remarks include:  in Appendix A, make the text in one font. The paper is of good scientific level and practical interest. The structure of the article is well organized, the text is clear. Technical remarks include:  in Appendix A, make the text in one font.

Response1: Thank you for your positive feedback and for recognizing the scientific level and practical interest of the paper. I truly appreciate your kind words about the organization and clarity of the text.

Regarding your technical remark about Appendix A, I have carefully reviewed the section and adjusted the text to ensure consistency in font throughout. Thank you for pointing this out—it’s an important detail that enhances the overall presentation of the article.

Reviewer 4 Report

Comments and Suggestions for Authors

This is a systematic review literature about soil pollutant with satellite images. Although this paper summarized the satellite types, evaluation parameters, algorithms about estimation soil contaminations. There are some comments are as following:

1.     Remote Sensing publishes research with important significant new results or methods that will advance the method methodology of remote sensing. However, this paper focus on the method about how to summarized and get information about detection of soil pollutant with Satellite images. I think this method is suitable for all other field.

2.     It is not clear that why the author only chose 36 papers which were published within the past three years. Why not investigate the past five years or 10 years, and so on?

3.     The appendix A is very important to the researchers who are interested in soil pollutant researching. This table should be analyzed and discussed specifically.

Comments on the Quality of English Language

No comments

Author Response

Comment 1:

 

Remote Sensing publishes research with important significant new results or methods that will advance the methodology of remote sensing. However, this paper focuses on the method of summarizing and extracting information about the detection of soil pollutants using satellite images. I think this method is suitable for other fields.

 

Response 1:

Your comment addresses a crucial aspect of remote sensing research, emphasizing the importance of pursuing significant new results or methods that advance the field’s methodology. In our study, we concentrated specifically on the detection of soil pollutants, utilizing satellite images to extract valuable information both directly and indirectly. The indirect detection involves exploring various soil properties, such as salinization and erosion.

 

While the methodology presented here has broad applicability and relevance to other fields, such as weather monitoring and water assessment using various remote sensing technologies like drones and aerial imagery, our study is focused on soil detection through satellite imagery. This focus on soil-related parameters aligns with the specific objectives of our research.

 

 

Comment 2: It is not clear why the authors chose only 36 papers published within the past three years. Why not investigate the past five years, ten years, or more?

 

Response 2:

The selection of the recent timeframe for this study aligns with the methodology employed, which is rooted in the PRISMA method. The focus on articles published between 2019 and 2023 stems from recognizing that methods such as Machine Learning, coupled with the use of satellite imagery for soil pollutant detection, have seen significant advancements in recent years.

 

In the updated version of our article, we analyzed 47 articles from an initial pool of 1,018 publications spanning the last eight years (2016–2024). Among these, 34 studies focused on the direct detection of soil pollutants, while 13 examined the relationships between vegetation indicators and soil contaminants. This broader analysis ensured that our review captured a comprehensive picture of recent and relevant advancements in this field.

 

It’s important to note that, in our preliminary exploration, we conducted a thorough search of articles beyond the three-year period, as indicated in Section 2.1: “a comprehensive exploration of available literature spanning the years 2019 to 2023 (with the exclusion of one article) was conducted.” The mention of the excluded article from 2016 (reference number [57]) highlights our diligence in considering works outside the specified timeframe. However, the decision to focus primarily on recent years was made to ensure the incorporation of the latest advancements and methodologies in this evolving field.

 

This revised version integrates the new data seamlessly while maintaining clarity and professionalism.

 

Comment 3:

 

Appendix A is very important for researchers interested in soil pollutant research. This table should be analyzed and discussed specifically.

 

Response 3:

In Section 2.6, a comprehensive description of each column in Appendix A was presented, accompanied by additional information to enhance clarity. The subsequent Section 3 delves deeply into the outcomes, offering a meticulous exploration within the Results, Discussion, and Summary of Evidence. To provide a condensed yet comprehensive summary derived directly from the data in Appendix A, Figures 2 to 5 have been included.

 

It is essential to emphasize that this table, as a fundamental component of the PRISMA method, serves as a consolidated reference point. It enables researchers to quickly obtain an individualized overview of the results presented in our review. We believe this structured approach ensures accessibility and clarity, facilitating a deeper understanding of the nuanced aspects discussed in the context of soil pollutant research.

 

Reviewer 5 Report

Comments and Suggestions for Authors

Machine Learning Algorithms Using in Detection of Soil Pollutant with Satellite Images: A Systematic Review

 

  This paper provides a new perspective of soil pollutant prediction, and make efforts to discuss the ML algorithms in this field. However, there are lots of controversial issues waiting to be revised. The selected papers are not representative in a too small dataset. The intention of the study is confusing lack of the focus. You topic is mainly ML algorithms or soil pollutant detection or satellite images or hybrid? “Machine Learning Algorithms Using in Detection” is a wrong expression. Few insights are provided in this version. More details are listed here:

  

1.       The main content investigated the soil, and the relationship between vegetation and soil contaminants using ML algorithms, which are not enough in the soil contamination field. Authors should severely discuss the surface geo-object types, such as impervious layers, buildings, water, ice, soil, vegetation, minerals and so on, not only soil and vegetation. Therefore, the soil polutant field covers more topics than these mentioned in the paper. These should be added and discussed.

2.       “this discusses satellite types, current limitations of spaceborne sensors for soil contamination monitoring, and the efficiency estimation of the methods.” In fact, the types of sensors are too many to be enumerated. In contrast, satellite platforms, remote sensing spectral bands, “eletromagnetic waves-soil-pollutant” interaction mechanism related to types of soil and pollutants are more accurate to be addressed. As for the efficiency estimation in ML algorthims, they are generally deemed as common sense and cannot add more creation for this manuscript. Do you provide more insights about ML efficiency estimation used in soil pollutant prediction?

3.       Provide the logic or explain why you conducted the review using these perspectives: “the important methods, traces the origins of satellite data sources, and dissects the 100 evaluation metrics that support this interesting combination.”

4  The selected 36 papers are too small, which cannot cover the topic of this version. You can use more keywords or retrieval equations to recall more related and significant papers:

   For example, remote sensing, SAR, optical images, or others; object identification, segment, salient detection, image classification or others. DL models such as GAN, RL, RCNN or others.

   As a review, the comprehensive collection of related papers must be completed!

 

5  Why do you use risks? Please demonstrate it, what does it mean in you topic? See “ Each parameter's potential impact on the study's outcomes, indicating possible bias, was classified 299

as "high risk," "low risk," or "unclear risk" following established criteria [30].”

 

6 Please classify these satellites into types related to soil pollutant prediction.

“These satellites, part of separate Earth observation systems, have found diverse applications such as land-use mapping, environmental monitoring, and natural resource management.”

 

7 We cannot follow the logic that why you discuss metrics, have they made impacts on the pollution prediction?

 

8 How to define the environmental parameters? If parameters cover the pollution types, soil types, vegetation types, give the current work in your review about them.

See “The primary objective of this study was to identify environmental parameters associated with soil pollution, both directly and indirectly. Notably, vegetation properties were directly linked to specific types of pollution, affecting various vegetation types”.

9   Are there more papers about metal pollution not covered in your selected articles? The number of articles is too small, and why do you only use the year after 2019?

See “Figure 5 reveals that heavy metals were the most frequently studied environmental parameters in the selected articles”

 

10 RMSE and R2 are reguler parameters in ML algorithms, we cannot see the significance that you list them.

“Performance indicators for the employed methods were examined, with RMSE and R2 being the most frequently utilized metrics, featuring in 21 and 19 instances, respectively. “

11 However, valuable insights of soil pollution analysis are not plentiful, and even there are few highlights of this review. This topic put focus on the soil pollution, satellite imagery sources, ML methods, but these are general topics without plentiful discussions in the academics. This review only covers 36 papers, which cannot cover the proposed topic. The pollution location, pathways, and sources are really not enough for this field.

   Suggestions: divide more and more paper sets (>200 top papers) into groups, such as soil-pollution types, spectral bands, satellite platforms, prediction pros and cons, ML pros and cons.

  

12 Appendix should be abstracted, and provide a thorough conclusion using clear and short descriptions. In fact, this appendix looks like a simple sheet of papers without valuable insights.

 

Comments for author File: Comments.pdf

Comments on the Quality of English Language

 

needs more polishments.

Author Response

Comment 1:

 

The main content investigated the soil, and the relationship between vegetation and soil contaminants using ML algorithms, which are not enough in the soil contamination field. Authors should severely discuss the surface geo-object types, such as impervious layers, buildings, water, ice, soil, vegetation, minerals, and so on, not only soil and vegetation. Therefore, the soil pollutant field covers more topics than these mentioned in the paper. These should be added and discussed.

 

Response 1:

Thank you for your insightful question and valuable observation. As indicated in the title of the article, our focus lies predominantly within the domain of soil-related parameters. I appreciate your acknowledgment of geo-object types such as impervious layers, buildings, water, ice, soil, vegetation, minerals, and others.

 

Regarding impervious layers, characterized as materials such as clay or shale that impede water movement significantly, these are indeed addressed in the referenced articles [49], [59], [62], [63], [64], [65].

 

While acknowledging that various factors such as buildings, water, ice, air, industries, and human activities can impact soil pollution, our article deliberately centers its attention on the primary parameters: soil, vegetation, and minerals. This is explicitly outlined in the study’s scope.

 

In the updated version of our article, we expanded the analysis to include 47 articles from an initial pool of 1,018 publications spanning the last eight years (2016–2024). Among these, 34 studies focused on the direct detection of soil pollutants, while 13 examined the relationships between vegetation indicators and soil contaminants. This expanded dataset allowed us to provide a broader perspective while maintaining our specific focus on soil-related parameters.

 

Comment 2:

 

“This discusses satellite types, current limitations of spaceborne sensors for soil contamination monitoring, and the efficiency estimation of the methods.” In fact, the types of sensors are too many to be enumerated. In contrast, satellite platforms, remote sensing spectral bands, “electromagnetic waves-soil-pollutant” interaction mechanism related to types of soil and pollutants are more accurate to be addressed. As for the efficiency estimation in ML algorithms, they are generally deemed as common sense and cannot add more creation for this manuscript. Do you provide more insights about ML efficiency estimation used in soil pollutant prediction?

 

Response 2:

Indeed, the proliferation of sensors across various fields, including the detection of soil pollutants, is remarkable, and their accessibility continues to grow. In this article, our primary focus has been to identify the most frequently utilized sensors among researchers. The main goal of our research is to serve as an entry point for those looking to integrate machine learning (ML) methods into the realm of soil pollutants.

 

As illustrated in Figure 2, we have focused to pinpoint the most commonly used satellites. This strategic approach aims to assist newcomers in navigating the vast array of available datasets, especially considering that some are not freely accessible. Understanding the significance of the chosen ML methods and the types of soil pollutants they are adept at detecting is paramount, particularly for those considering the purchase or utilization of datasets based on their financial support.

 

In the updated version of our article, we expanded the review to analyze 47 articles from the initial pool of 1,018 publications spanning 2016–2024. This included studies on direct detection of soil pollutants and relationships between vegetation and soil contaminants, enriching the diversity of methods and datasets discussed in the paper.

 

In addressing the efficiency of machine learning methods, it’s important to note that the efficacy varies across articles due to the specific conditions of the study areas. Comparisons of performance between different methods are not included in this compilation. As previously highlighted, obtaining specific results from performance matrices to facilitate direct comparisons between articles is not feasible.

 

Instead, our focus is on providing a comprehensive overview of the methods employed in each article. By exclusively referencing the methods utilized within the context of this article, we aim to offer readers valuable insights into the initial steps for investigating soil pollution. We acknowledge that authors have chosen specific methods based on their unique datasets. This approach provides readers with a foundation from which to commence their exploration of soil pollution. In instances where the answers may be deemed less convincing, the avenue of investigating more complex and less frequently employed methods is suggested for further exploration.

 

Comment 3:

 

Provide the logic or explain why you conducted the review using these perspectives: “the important methods, traces the origins of satellite data sources, and dissects the 100 evaluation metrics that support this interesting combination.”

 

Response 3:

Performance metrics in machine learning serve as quantitative measures crucial for evaluating the efficacy and precision of a model’s predictions across various tasks like classification, regression, or clustering. In classification, there are more than 9 metrics, including Accuracy, Precision, Recall (Sensitivity), F1 Score, Specificity, ROC-AUC, Confusion Matrix, Matthews Correlation Coefficient (MCC), and Area Under the Precision-Recall Curve (AUC-PR).

 

Meanwhile, regression entails over 7 metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Normalized Root Mean Squared Error (NRMSE), R-squared (R2), Mean Absolute Percentage Error (MAPE), and Explained Variance Score.

 

While acknowledging that determining the best metric is contingent on the specific case study, it is noteworthy, as highlighted in Section 10, that RMSE and R2 emerge as the most frequently utilized metrics in regression. These align with the findings of numerous studies and are well-standardized across various applications. Appendix 1 provides additional insights into other metrics for reference. This serves as a helpful starting point for readers.

 

Comment 4:

 

The selected 36 papers are too small, which cannot cover the topic of this version. You can use more keywords or retrieval equations to recall more related and significant papers:

For example, remote sensing, SAR, optical images, or others; object identification, segment, salient detection, image classification or others. DL models such as GAN, RL, RCNN or others.

 

Response 4:

As evident from Section 2.9, Article Selection, an initial search across Web of Science, ScienceDirect, Scopus, and other databases yielded 1,018 articles. The process leading to the final selection of 36 articles is meticulously delineated in Figure 1, employing the PRISMA method.

 

In the updated version of our article, we analyzed 47 articles from the initial pool, spanning the last eight years (2016–2024). Among these, 34 studies focused on direct detection of soil pollutants, while 13 examined the relationships between vegetation indicators and soil contaminants. This comprehensive selection provides a richer foundation for the conclusions drawn in our study.

 

The search terms encompassed a range of keywords, including “remote sensing,” which includes drone and aerial imagery. Synthetic Aperture Radar (SAR), a specific satellite imaging technology, was also included in our searches, broadening the scope of our exploration. Deep Learning (DL) methods, including advanced models like GANs and RCNNs, were also considered. However, object identification and segmentation, while significant in some remote sensing applications, are beyond the scope of this study, which focuses exclusively on the detection of soil pollutants.

 

Comment 5:

 

Why do you use risks? Please demonstrate it. What does it mean in your topic? See: “Each parameter’s potential impact on the study’s outcomes, indicating possible bias, was classified as ‘high risk,’ ‘low risk,’ or ‘unclear risk’ following established criteria [30].”

 

Response 5:

Thank you for bringing this to our attention. We have revisited the PRISMA method as referenced in our manuscript and confirmed its accuracy. The classification of risks—“high risk,” “low risk,” or “unclear risk”—pertains to the potential for bias in the data or methodologies employed in the reviewed articles. These classifications provide readers with a clear understanding of the reliability and limitations of the datasets and methods used in the studies included in our review.

 

Comment 6:

 

Please classify these satellites into types related to soil pollutant prediction:

“These satellites, part of separate Earth observation systems, have found diverse applications such as land-use mapping, environmental monitoring, and natural resource management.”

 

Response 6:

Subsequently, in Figure 5, we have delineated the types of environmental parameters investigated in the selected works, with a specific emphasis on the two satellites newly introduced to the content. In the updated version, these classifications provide a clearer perspective on how different satellites contribute to soil pollutant prediction.

 

Comment 7:

 

We cannot follow the logic of why you discuss metrics. Have they made impacts on the pollution prediction?

 

Response 7:

As discussed in the response to question number 3, we emphasized the diverse array of performance metrics available in machine learning and elaborated on the specific focus of our article on RMSE and R2 in the context of regression. These metrics were chosen based on their frequent utilization and standardization across numerous studies, providing readers with a solid starting point for their exploration of soil pollution detection methods.

 

Comment 8:

 

How to define the environmental parameters? If parameters cover the pollution types, soil types, vegetation types, give the current work in your review about them. See: “The primary objective of this study was to identify environmental parameters associated with soil pollution, both directly and indirectly. Notably, vegetation properties were directly linked to specific types of pollution, affecting various vegetation types.”

 

Response 8:

As indicated in Figure 5 and the preceding explanation, this study comprehensively summarizes all types of environmental parameters directly and indirectly related to soil pollutants. A supplementary Table 2 has been added to provide more detailed insights into the parameters detected by satellites.

 

For instance, heavy metals were directly detectable by satellite images in 34 articles, while 15 articles reported the detection of soil parameters such as soil organic carbon, soil organic matter (SOM), soil salinity, and soil texture. Additionally, vegetation properties, including land cover and leaf area index (LAI), were linked to pollution indicators, highlighting the diverse range of parameters that can be monitored through satellite imagery.

 

Comment 9:

 

Are there more papers about metal pollution not covered in your selected articles? The number of articles is too small, and why do you only use the year after 2019? See: “Figure 5 reveals that heavy metals were the most frequently studied environmental parameters in the selected articles.”

 

Response 9:

As addressed in the response to question number 4, the information pertaining to question number 9 is already provided.

 

To reiterate, in the updated version of our article, 47 articles were selected from an initial pool of 1,018 publications spanning the last eight years (2016–2024). Among these, 34 studies focused on the direct detection of soil pollutants, while 13 examined relationships between vegetation indicators and soil contaminants. This broader selection ensures comprehensive coverage of recent advancements in soil pollutant detection, including the frequently studied parameter of heavy metals, as shown in Figure 5.

 

Comment 10:

 

RMSE and R2 are regular parameters in ML algorithms. We cannot see the significance that you list them.

“Performance indicators for the employed methods were examined, with RMSE and R2 being the most frequently utilized metrics, featuring in 21 and 19 instances, respectively.”

 

Response 10:

As addressed in the response to questions 3 and 7, RMSE and R2 were highlighted due to their frequent utilization and prominence in regression analyses across numerous studies. These metrics serve as a standardized starting point for researchers exploring soil pollutant detection through ML methods, as they provide reliable benchmarks for evaluating model performance.

 

Comment 11:

 

However, valuable insights into soil pollution analysis are not plentiful, and there are few highlights of this review. This topic focuses on soil pollution, satellite imagery sources, and ML methods, but these are general topics without plentiful discussions in the academics. This review only covers 36 papers, which cannot cover the proposed topic. The pollution location, pathways, and sources are really not enough for this field.

Suggestions: Divide more paper sets (>200 top papers) into groups, such as soil-pollution types, spectral bands, satellite platforms, prediction pros and cons, and ML pros and cons.

 

Response 11:

As addressed in the response to question number 4, the updated version of the article expands the review to include 47 articles from an initial pool of 1,018 publications spanning the last eight years (2016–2024).

 

This expansion provides a richer dataset that allows for greater exploration of soil-pollution types, satellite platforms, and ML applications, although the primary focus remains on presenting a comprehensive overview of recent advancements in soil pollutant detection using ML and satellite imagery.

 

Comment 12:

 

Appendix should be abstracted, and provide a thorough conclusion using clear and short descriptions. In fact, this appendix looks like a simple sheet of papers without valuable insights.

 

Response 12:

In Section 2.6, the description of each column in Appendix A was provided, and additional information was incorporated to enhance clarity. Section 3, which encompasses the Results, Discussion, and Summary of Evidence, intricately explores the outcomes, with Figures 2 to 5 serving as a comprehensive summary derived directly from the data in Appendix A.

 

This table, an integral component of the PRISMA method, serves as a consolidated reference, offering researchers a quick and individualized overview of the review’s results. Efforts have been made to ensure that the appendix provides meaningful insights and is more accessible to researchers.

Back to TopTop