Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The document entitled “High-Resolution Forest Structure Mapping with Deep Learning  to Evaluate Restoration Outcomes” represents a relevant advance in forest inventorying and monitoring for management purposes. The proposed approach offers an alternative with adequate spatial and temporal resolution, addressing challenges commonly associated with the assessment of fine-scale forest structural attributes, which are often constrained by low revisit frequency and the high costs of airborne LiDAR data. In this context, the use of moderate-resolution satellite imagery combined with convolutional neural networks (CNNs) constitutes a valuable contribution to the field.

Regarding the references, it is recommended that the authors carefully review the “Reference List and Citations Style Guide” provided by MDPI to ensure full compliance with the required formatting standards.

The results obtained at different fine spatial scales are generally acceptable and are supported by the reference data used for model training, which appear to be spatially extensive and consistent. In this case, the California Forest Observatory 3 m canopy cover product was employed, and it can be considered a high-quality proxy for ground truth, as it is primarily derived from LiDAR datasets complemented by other satellite imagery. The authors also acknowledge a limitation related to pixels containing sparse vegetation, which are classified as gaps. This binary classification approach may introduce uncertainties, as remaining vegetated material could still act as fuel and contribute to forest fire ignition.

The use of SRTM elevation data could be reconsidered, as it may be replaced by more accurate datasets derived from LiDAR or other moderate-resolution sensors, such as ALOS PALSAR. Such alternatives would better match the spatial resolution of the other inputs and could potentially improve the overall results.

Finally, the paragraph beginning at line 387 does not correspond to Section 2.8 and should be revised for consistency. In addition, the conclusions could be expanded to better highlight the broader implications and potential applications of the study.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper developed a forest structure monitoring workflow based on PlanetScope imagery and a U-Net model, validated its potential using extensive independent LiDAR data, filled the gap in low-cost, high-resolution monitoring, and holds value for engineering applications.

1. page9, line357-367.The model relies on spectral discrimination at approximately 5-meter resolution, whereas the HRV (High Resolution Vegetation) metric is structurally defined based on tree height and a 40-meter radius neighborhood. The coarse resolution definition of HRV tends to smooth over small gaps, while the high-resolution model in this study may detect more fragmented gaps. Therefore, directly comparing their area distributions (Figure 4) is statistically unsound.
   -
2. page15, line568-581
   The model classifies the land surface into only two categories: "forest" and "non-forest." For applications like fire behavior analysis and ecological recovery, this classification is too coarse. It fails to distinguish between bare ground, shrubs, herbaceous plants, or non-closed canopy forests of different densities. In fire-prone areas, the fuel load differences between shrubs and bare ground are significant, but this model cannot capture these critical distinctions.
3. The model was not trained on ground truth data generated from field surveys or raw LiDAR, but instead used an existing product from the "California Forest Observatory (CFO)" as training samples. Using the output of one model to train another introduces the risk of potential error propagation. Consequently, the model may learn the algorithmic characteristics of the CFO product rather than the true characteristics of the ground surface.
4. The model was trained using only imagery from the summer (third quarter). This limits its applicability to other seasons. Particularly in spring or early summer, when deciduous vegetation or understory herbaceous plants exhibit high spectral reflectance, the model might misclassify these as forest canopy, thereby underestimating the gap rate.
5. page10, line409-418.The validation results indicate that the main errors are concentrated at the boundaries (edges) between forest and non-forest areas. Considering that the 4.77-meter resolution of PlanetScope is comparable to the size of an individual tree crown (approximately 6 meters), the mixed pixel problem at edges is severe. Using a 50% coverage threshold at these boundaries is prone to introducing random errors.

Author Response

Please see the attachment uploaded to Reviewer 1 response (attached again here).

Author Response File: Author Response.docx