Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors classified salt marsh vegetation communities by leveraging UAV-acquired imagery through the construction of a multidimensional feature space and the implementation of a Random Forest (RF) algorithm. The experimental details are described in a clear and thorough manner, data processing was conducted appropriately, and literature references are comprehensive. Notably, the authors have shared their datasets and source code, which significantly enhances the transparency and reproducibility of their findings. The discussion section is robust and effectively addresses the key queries arising from the methodology. Overall, this is a meaningful study with considerable practical value for the community. Therefore, I recommend acceptance after minor revisions to address the specific points listed below.

A few minor points:

1. In the Keywords, what does "UAS" refer to? Should it be "UAVs"?
2. In the abstract, six sites are mentioned, whereas Section 2.1 refers to three different locations. It is suggested to unify the description.
3. For figures of the type shown in Figures 3 and 4, where values that reflect differences between categories are typically above 0.6, it may be better to set the starting value of the axes to a number above 0.5. This would help magnify differences between groups more clearly.
4. The terminology describing newly created and established salt marshes is inconsistent throughout the manuscript, e.g., "restored and established", "new and established", "newly created and established". It is recommended to standardize the terminology throughout the text and in figure labels.
5. Regarding the balance of sample numbers among different classes in the training data, previous studies have proposed an alternative to equal allocation. Specifically, based on field surveys and existing data, the number of samples can be proportionally assigned according to the relative area of each class within the study region. This is particularly relevant for studies testing models across multiple sites, where class types and spatial distributions differ. Using the same number of training samples for each site may affect classification accuracy. Therefore, I suggest including a discussion on the strategy for determining sample numbers per class in the Discussion section.
6. In the first point of the "Main Findings", I suggest the authors explicitly emphasize the specific number of salt marsh communities successfully identified. As noted in the discussion, a lower diversity of vegetation types often simplifies the classification task and yields higher accuracy. Therefore, highlighting the complexity of classifying multiple distinct communities will better demonstrate the robustness of your multi-dimensional feature space and significantly enhance the impact and attractiveness of your research to a broader audience.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

 

This paper investigates whether UAV multispectral imagery fused with photogrammetric DSMs and texture features, classified using Random Forests, can accurately map National Vegetation Classification (NVC) plant communities in heterogeneous, species-rich salt marshes of northwestern Europe. Across six marshes (established and restored) at three UK sites and multiple years, the authors report high test accuracies (mean overall accuracy ≈94%) and show DSM-derived elevation/structural features are the most important predictors, with no clear dependence of performance on restoration age. The work positions UAV-ML community mapping as a practical tool for monitoring restoration trajectories and informing marsh condition and ecosystem services assessments.

Below are my questions:

1. Can you provide spatial cross-validation results (e.g., block CV or leave-one-subarea/site-out) to assess generalization beyond the specific quadrats and dates used for training?

2. How sensitive are the results to the choice of resolution alignment? Did you test resampling the DSM to the 8 cm multispectral resolution instead of downsampling the imagery, and what was the impact on accuracy and speckling?

3. Can you include ablation studies quantifying the contribution of feature groups (bands+VIs only; +GLCM; +DSM+geomorphometry) to overall and class-wise performance?

4. What are the vertical and horizontal accuracies of the DSMs and orthomosaics (from Appendix A), and how might DSM vertical uncertainty in vegetated areas affect feature importance and classification stability?

5. Did you record and control for tide state and illumination during flights? Given the sensitivity of spectral and structural signals to inundation and sun angle, how reproducible are results across different tidal/lighting conditions?

6. Have you explored post-processing (e.g., majority filters, CRF/MRF, or OBIA segmentation) to reduce speckling at ecotones, and if so, how did it affect class-wise F1 and boundary accuracy?

7. How representative are the “purest” quadrats (selected via NMDS centroid proximity) of the full site heterogeneity, and could this selection bias inflate performance relative to boundary/mixed areas commonly encountered in operational mapping?

 

Thanks for considering my questions in advance.

Reviewer RS

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed a majority of my concerns. However, I still have some minor comments that I would like to draw the attention of the authors to:

1. "Limmonium vulgare" or "Limonium vulgare"?

2. "Zenmuse X3" earlier but later as "Zenmuse 3" in the Figure 6 caption.

3. Several software/packages are mentioned without version numbers at first mention (RMAVIS; R), lacking explicit version details.

4. As the main text refers to subfigures (e.g., "Figure 6e", "Figure 6a"), better to add the subfigures' titles also in the Figure 6 caption.

5. Figure 1 misses lat/lon.

6. The layout of the manuscript (Too many blanks)

Comments on the Quality of English Language

I recommend that the authors make the final check for grammar before publication.

Author Response

Please see the attachment

Author Response File: Author Response.docx