High-Resolution Land Use Land Cover Dataset for Meteorological Modelling—Part 2: ECOCLIMAP-SG-ML an Ensemble Land Cover Map
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsHigh-resolution land use land cover dataset for meteorological modelling – Part 2: ECOCLIMAP-SG-ML an ensemble land cover map
Introduction
Although the paper discusses the importance of increasing the resolution of land cover maps and the use of AI/ML, the central argument as to why the n-to-n approach is superior needs further reinforcement. Perhaps including more details on the practical benefits of this approach compared to other AI techniques mentioned (e.g., random forests, multi-layer perceptrons) would help strengthen the reasoning.
Lines 48-50: References [17] and [18] are mentioned without any explanation of the content or methodology used in these studies. The reader would have a better understanding if you provided a brief description of these approaches. Example: "In [17], the authors explore a novel technique for n-to-n map translation, which..."
Methodology
Adequate.
Results
Lines 319-330: The text mentions that the merging described in Section 3.3 was not applied, but the justification that this "would lead to unrealistically good scores" requires more technical support. Additionally, the explanation of why the inferences are compared directly with ECOSG+ is unclear. Provide a more detailed explanation of the effects of the merging process on accuracy and why it would introduce bias. Moreover, clarify the choice of ECOSG+ as a reference, highlighting its advantages over the original ECOSG. At least a few lines of explanation should be included if deemed pertinent.
Lines 343-351: While the improvements are discussed, the potential impact of significant misclassifications, especially in sensitive areas (e.g., coastal and urban regions), is not addressed in depth. Provide a more detailed discussion of the practical implications of these misclassifications, perhaps offering clearer examples of the negative impact in the context of weather forecasting.
Lines 409-417: The description of the geographical areas is quite factual but does not explain why these areas were selected beyond their diversity of landscapes and latitudes. Include a more detailed explanation of the importance of these areas for the evaluation of land cover maps. For instance, what makes the Snaefell glacier, Nanterre, or El Menia particularly relevant for the validation of resolution and accuracy of the maps?
Discussion
Lines 485-488: The explanation of obviously wrong classifications is too vague. It is not clear why these inconsistent classifications occur or what impact they have on the practical use of the map. Provide a more detailed explanation of the causes of these misclassifications, such as limitations in input data or specific challenges related to the neural network.
Lines 513-516: The suggestion to improve the loss function is mentioned but lacks technical details on how this could be implemented or what has been attempted so far. Provide a more precise description of which types of weighted loss functions or class adjustment methods could be experimented with. For instance, you could explore using class-weighted cross-entropy to address class imbalance or employ focal loss to focus more on hard-to-classify examples. Additionally, consider discussing the potential of similarity-aware loss functions that could prioritize minimizing errors within related class groups (e.g., within the same primary label).
Lines 528-530: The conclusion that improving input data might be more effective than enhancing ML methods is an important claim, but the evidence provided is insufficient. Expand the discussion by detailing the marginal improvements achieved with changes to ML architectures and how these compare to the improvements obtained by enhancing input data. For example, explain the specific alterations made to the ML models (e.g., adding attention layers or adjusting latent space size) and how these led to only slight gains in accuracy. Contrast this with the more significant improvements observed when refining input data, such as incorporating additional variables like elevation or bioclimatic factors. Providing concrete examples or quantitative comparisons would help strengthen this argument.
Conclusion
Adequate.
Author Response
We would like to thank the reviewer for its helpful and constructive review. Please find our responses in the PDF document joined to this message.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper titled "Highresolution land use land cover dataset for meteorological modelling – Part 2: ECOCLIMAPSGML an ensemble land cover map" presents a significant advancement in the field of meteorological modeling through the development of an ensemble land cover dataset. The methodology is innovative, particularly in its use of autoencoders for map translation and the integration of multiple land cover maps. However, there are several areas that require clarification and enhancement to strengthen the manuscript.
Data Section:
1. Sample Partitioning Strategy:
Stratified Sampling: The authors should consider implementing stratified sampling to ensure all land cover types are adequately represented in the training, validation, and testing datasets. This will enhance the model's robustness and generalizability.
2. Clarity on Sample Distribution:
Geographic and Class Balance: It would be beneficial to provide a clear explanation of how the larger patches are divided and how this division maintains a balanced distribution of land cover types across the subsets. This information is crucial for understanding the training dataset's representativeness.
Method Section:
3. Discussion Section:
Parameter Justification: In Figure 3, adding context on the choice of 600 × 600 pixels and 50 channels would provide readers with insights into how these parameters impact model performance and the quality of the output maps.
4. Training Process Details:
Specifics on Training: The manuscript should include more details about the training process, such as the specific optimization algorithms used (e.g., Adam or SGD), learning rates, and the number of epochs. This information will help readers better understand the training dynamics and effectiveness.
5. Loss Function Explanation:
Role of Embedding Loss: If Lemb is relevant, the authors should clarify its role within the overall loss function and discuss its expected contributions to the model's performance and training process.
6. Graphical Details:
Label Key Components: In Figure 3, labeling the critical components (such as the input and output of the autoencoder) would enhance the figure's clarity and aid in understanding the methodology.
Discussion Section:
7. Integration of Temporal and Spatial Scales: Lines 397-403 discussing temporal and spatial scales should be moved to the discussion section. A more thorough analysis of how these scales affect data interpretation and model outcomes is essential for contextualizing the results.
8. Utilization of Additional Datasets: The authors might benefit from validating their model using additional datasets from various sources or time periods. This would provide a more comprehensive evaluation of the model's generalizability and robustness.
Author Response
We would like to thank the reviewer for its helpful and constructive review. Please find our responses in the PDF document joined to this message.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Editor,
After review, we confirm that the manuscript titled "High-Resolution Land Use Land Cover Dataset for Meteorological Modelling–Part 2: ECOCLIMAP-SG-ML an Ensemble Land Cover Map" has fully addressed the previous comments. The revised text is now more fluid and dynamic, enhancing readability and understanding of the content.
We congratulate the authors on their excellent work and dedication to improving the manuscript.
We recommend, therefore, the acceptance of this article.
Sincerely,
