Sardegna grassland mapping for livestock management: a practical Intra-Annual NDVI contrasts approach

  1. Marco Mura

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Mapping grassland locations and area is of paramount importance for the management of these areas, particularly when the primary use is for traditional livestock production over common lands. Often the institutions responsible for drafting land use policies lack the tools to support their decisions. This is the case in the mountains of Sardegna. The availability of satellite imagery with fine temporal resolution and the variety of classification techniques, offer several methods to obtain an efficient and accurate land cover map. The objective of this study is to offer an accurate, easy-to-use method for mapping grasslands. Using Sentinel-2 data, an analysis of NDVI is conducted through the whole year. For known grasslands, these analyses provide contrasting evidence for different months. Using the difference between the month with the NDVImax and NDVImin the Intra-Annual delta NDVI layer is constructed. To this layer three classification algorithms are applied to discriminate grasslands from other land covers: random forests (RF), clustering and thresholding. The accuracies of these maps are evaluated in terms of OA, PA, UA and area estimates. All three methods produced similar results. Clustering showed the greatest OA, 89% with area estimate of 613.67 ha; RF produced OA of 87% with area estimate of 546.56 ha and tresholding produced OA of 88% with area estimate of 579.69 ha. Because the results are quite similar, the operator has the opportunity to use a classification method of choice. The results are auspicious and replicable each year. Further work is recommended to determine whether these results apply in other regions.

Article activity feed

  1. PREreview

    This Zenodo record is a permanently preserved version of a PREreview. You can view the complete PREreview at https://prereview.org/reviews/10115353.

    This review reflects comments and contributions from Femi Arogundade.

    The study employs high-temporal-resolution Sentinel-2 data and open-source tools to propose a practical and replicable method for mapping grassland areas in the Mediterranean mountains of Sardegna. By analyzing the Intra-Annual delta NDVI, the study applies three classification methods, including Random Forest, Thresholding, and Clustering, demonstrating their effectiveness in identifying grasslands for improved livestock management.

    General Comments:

    The study is well-structured, presenting a comprehensive approach to grassland mapping with valuable insights into methodological choices and robust evaluation metrics. It effectively addresses the practical needs of land managers, providing a reliable and easily replicable method. Minor formatting issues in the text and figures should be addressed for enhanced clarity. Overall, the study makes a significant contribution to the field of grassland mapping in complex landscapes.

    Major Comments:

    The use of Sentinel-2 data with high temporal resolution is commendable for capturing seasonal changes in vegetation, essential for grassland mapping.

    The selection of three classification methods (Random Forest, Thresholding, and Clustering) adds robustness to the study, allowing for a comprehensive evaluation of the proposed approach.

    The incorporation of ground truth data from various sources, including field surveys and fine-resolution images, enhances the accuracy and reliability of the results. The approach to collecting the ground truth dataset is commendable.

    The study demonstrates thoroughness in experimental design, with a well-defined study area, clear selection criteria, and detailed data acquisition methods.

    The validation process, including the use of control points and comparison with existing maps, contributes to the reliability of the results.

    The analysis of the Intra-Annual delta NDVI is a novel and valuable contribution, providing a seasonal perspective that aligns with the phenological aspects of grasslands.

    The comparison of three classification methods adds depth to the analysis, allowing for a nuanced understanding of their performance.

    The conclusions drawn from the study are well-supported by the presented results and analyses. The discussion on the applicability of the proposed method for land managers is insightful.

    The study acknowledges the limitations and suggests avenues for further research, indicating a balanced interpretation of the findings. The paper transparently discusses limitations and challenges in mapping grassland areas, such as the difficulty in distinguishing between classes at ecotones. Acknowledging these challenges adds credibility to the study.

    The methodology is generally well-explained, but there are instances where the description could be more explicit. For example, a clearer explanation of the thresholding process and the rationale behind the selection of certain parameters would enhance the reader's understanding.

    The study relies on a combination of fine-resolution images and various data sources for ground truth. It would be beneficial to elaborate on the challenges or limitations encountered during the collection and integration of this diverse ground truth dataset.

    The temporal analysis of NDVI is a strong point of the study. However, a more in-depth discussion on the specific ecological reasons for the observed NDVI patterns in different months could add depth to the interpretation.

    While the study emphasizes the use of open-source tools, providing additional details on code availability or a step-by-step guide for reproducibility would enhance the transparency and applicability of the proposed methodology.

    Minor Comments:

    The abstract could benefit from a more concise and direct statement about the specific contribution of the study and the key findings.

    When discussing the differences between classification methods in the Results section, consider adding information on the statistical significance of these differences to strengthen the conclusions.

    In the Materials and Methods section, when discussing the Sentinel-2 bands, consider providing a brief explanation of what each band represents or detects to aid readers who may not be familiar with remote sensing.

    When discussing the analysis of the probability density function (PDF) in the Results section, provide a bit more context on why this analysis was conducted and how it contributes to the overall interpretation.

    The study uses Sentinel-2 data with a temporal resolution of 10 days (5 days with the twin satellite). How was this temporal frequency determined to be suitable for capturing the intra-annual dynamics of grasslands? Were there considerations for specific phenological events?

    Provide more details on the composition of the ground truth dataset. How many samples were collected, and how were they distributed across different land cover classes? Were there any challenges or biases in collecting ground truth data?

    Explain the rationale behind selecting NDVI and the chosen months (May and August) for analysis. How sensitive is the method to variations in these months? Did the study consider other vegetation indices, and if not, why?

    Discuss the decision-making process for choosing Random Forest, Thresholding, and Clustering as the classification algorithms. Were there specific considerations for selecting these methods, and how do they complement each other in achieving the study objectives?

    Discuss the rationale behind choosing the specific accuracy metrics (OA, PA, UA) and their significance in evaluating the classification results. Were there considerations for potential biases in accuracy assessment?

    Address the potential challenges or limitations in replicating the proposed method in different regions or over multiple years. Are there factors that might influence the method's performance in diverse settings?

    The paper mentions resampling bands at 20m to 10m using the 'terra' package. Could you provide more details on why this resampling was necessary and the impact on the analysis? Offer insights into the necessity of resampling and its potential effects on the results.

    The sample size is determined by Equation 1, considering the proportion of each class, overall accuracy standard error (S(Ô)), and user accuracy. Could you elaborate on the choice of S(Ô) as 0.025 and how it impacts the reliability of the accuracy assessment?

    The total sample size is mentioned as 308 points. How does this number relate to the size of the study area and the diversity of land cover classes? Additionally, how does the allocation balance between classes?

    How does the choice between proportional and optimal allocation impact the overall accuracy and reliability of the land cover classification? Were sensitivity analyses conducted to assess the robustness of the chosen allocation strategy?

    Given the low pixel count for water bodies, how does the allocation (both proportional and optimal) for this class impact the reliability of the accuracy assessment, especially considering the allocation proportions are 0 for proportional allocation?

    Table 5 provides a structured approach to allocating testing points for accuracy assessment, with the choice between proportional and optimal allocation methods influencing the balance between fairness and overall accuracy in the land cover classification. The following should be considered for improvement:

    • Consider the relative importance of different land cover classes. Classes with higher ecological or management significance may require adjustments in the allocation to enhance their accuracy.

    • Explore dynamic allocation strategies that adapt to the characteristics of each land cover class. This could involve a more granular approach to the allocation process based on class-specific considerations.

    Statistical Analysis:

    The statistical analyses, especially those related to accuracy metrics, are well-presented. However, a discussion on the statistical significance of the differences observed in accuracy metrics between the classification methods would strengthen the conclusions.

    Suggestions for future studies:

    The paper concludes with well-defined suggestions for future research, indicating a commitment to continuous improvement and a broader understanding of the proposed methodology's strengths and limitations.

    Explore the inclusion of data from other satellites, such as Landsat, with different spectral characteristics to evaluate its impact on classification accuracy and the ability to discriminate between land cover classes.

    Extend the study over multiple years to observe how grassland dynamics change over time and to capture potential cyclical patterns or trends that may not be evident in a single-year study.

    Compare the performance of NDVI with other vegetation indices (e.g., Normalized Difference Water Index, Enhanced Vegetation Index) to evaluate if alternative indices provide better discrimination for grassland mapping in Mediterranean mountain regions.

    Investigate the influence of additional climatic variables, such as temperature and precipitation, on grassland dynamics to enhance the understanding of the environmental factors affecting land cover changes.

    Explore the use of machine learning algorithms beyond Random Forest, considering advanced techniques and fine-tuning parameters to optimize classification accuracy and handle the complexity of the landscape.

    Conduct a cross-validation of the results by comparing the proposed methodology with ground truth data collected in different years to assess the consistency and reliability of the approach over time.

    Integrate more extensive field-based data, possibly using advanced techniques like drone-based remote sensing, to validate and refine the accuracy of the classification results.

    Investigate the impact of scale on the accuracy of grassland mapping, considering variations in spatial resolution and exploring how different scales may affect the classification results.

    Explore how the proposed methodology performs in areas experiencing transitions from grassland to shrubland (shrubland encroachment) and identify potential challenges in mapping such dynamic landscapes.

    Investigate the applicability of the proposed methodology at different altitudes and in diverse geographical regions to assess its robustness and generalizability.

    Competing Interests:

    The author declares no competing interests

    Competing interests

    The author declares that they have no competing interests.

  2. Version published to 10.31219/osf.io/p83ru on OSF Preprints