Applications of Snow-Covered Areas from Unoccupied Aerial Systems (UAS) Visible Imagery: A Demonstration in Southeastern New Hampshire

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study presents a high-resolution snow-covered area (SCA) mapping approach using visible imagery (RGB) acquired from Unoccupied Aerial Systems (UAS) and LiDAR-derived terrain data. 1. The reported accuracy improvement  by the random forest model is not supported by statistical significance testing. This makes it difficult to rule out random variation. I recommend conducting statistical tests on key results  and reporting p-values or confidence intervals. 2. The proportion of pixels misclassified as bare ground due to shadows or ice is not clearly quantified, which weakens the discussion of classifier limitations. It is advisable to include detailed misclassification rates for shadow/ice, either in the main text or as a supplementary table, along with an analysis of how illumination conditions affected overall accuracy. 3. Figures such as Fig. 6 suffer from low resolution and unclear boundaries between shadow and ice. Figure 11 lacks a scale bar. 4. The physical meaning of the reported fractal dimension is insufficiently discussed. Moreover, comparisons with related studies (e.g., Shook et al., 1993) are missing. A deeper exploration of the implications of fractal dimension for snow energy balance models would enhance the theoretical contribution of this work.

Author Response

Please see the attached PDF file, (Reviewer 1)

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

• The keywords section is a bit too long and should be limited to five keywords.

• The Introduction currently focuses only on the types of remote sensing data; it lacks information on the methodology and recent applications of computer technologies, such as machine learning and deep learning models.

• It would be better to include a scale bar in Figure 3.

• Random Forest is mentioned in the Methods section but is not introduced in the Introduction.

• The Conclusions should mention the models used, such as Random Forest, and provide suggestions for future studies.

Author Response

Please see the attached PDF file

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This study demonstrates the potential of high-resolution snow-covered area  mapping by integrating unoccupied aerial system visible imagery with satellite data. The proposed random forest based downscaling model significantly improves classification accuracy under mixed snow cover conditions, particularly in shadowed and topographically complex environments. 

Lines 14-15: the expression "during a snowmelt period" is rather vague and can be specific to a certain time period, such as "during the snowmelt period from [specific start time] to [specific end time]".

Line 21: "satellite observations"，Please specify what kind of satellite data it is.

Line 120: The reasons for the selection of this research area and its representativeness can be supplemented.

Lines 176-177: The specific application of this flight data in the research can be supplemented.

Line 200: The UAS RGB and LiDAR data collection process is well-documented. However, details on flight path planning or quality control measures (e.g., illumination consistency during orthomosaic generation) are lacking. Clarify how lighting variations were minimized to reduce classification errors.

Line 204: Please explain the role of satellite remote sensing data in this study.

Lines 278-279: The application steps and advantages of the K-means clustering method in snow mapping can be further explained.

Line 280: The implementation process and parameter Settings of the algorithm in MATLAB can be further explained.

Line 282: While four classes (snow, ice, shadow, bare earth) are used, the rationale for selecting this number (e.g., elbow method, silhouette coefficient) is not discussed. Manual merging of sub-classes may introduce subjectivity; address potential biases in this step.

Line 325：The data sources and processing methods of DTM and DSM can be further explained.

Line 398: The parameter optimization (e.g., NumTrees=100) requires further justification. Although the training dataset (~300,000 features) was balanced via subsampling, cross-validation or independent test sets were not mentioned, raising concerns about generalization performance.

Line 452: The formation causes and influences of this peak snow depth can be further analyzed.

Line 467: Table 1 highlights significantly lower accuracy in the northern test area (33% on March 7), attributed to shadow-ice confusion. Quantify the relationship between shadow coverage and misclassification rates (e.g., confusion matrices for shadowed regions).

Line 763: Compared with other scaling methods (such as neural networks), the unique advantages of choosing random forests need to be explained.

Figure 9 shows strong correlation between Sentinel-2 and UAS fSCA but overlooks mixed-pixel effects due to Sentinel-2’s 20 m resolution, especially near forest-field edges. Discuss limitations in resolving subpixel heterogeneity.

 

Author Response

Please see the attached PDF file, section 3 (Reviewer 3)

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

I agree with the revised manuscript