Review Reports - Individual Tree-Level Biomass Mapping in Chinese Coniferous Plantation Forests Using Multimodal UAV Remote Sensing Approach Integrating Deep Learning and Machine Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript focuses on China’s typical coniferous forests and presents a hybrid framework that couples the NB-YOLOv8 algorithm with Random Forest (RF) to achieve high-accuracy individual-tree detection and biomass mapping. The results improve tree-level AGB prediction and highlight the potential of UAV remote sensing for large-scale, high-resolution forest-carbon monitoring and management. The findings provide actionable ecological evidence for hotspot identification of carbon sinks, stand-structure optimization, and differentiated thinning strategies. The topic is novel, the logic is clear, the structure is sound, the data are reliable, and the methods are feasible; the figures and tables are properly formatted. Nevertheless, the following issues still need to be addressed:

The authors should elaborate on the specifics of the field survey, e.g., sampling design, size and shape of sample plots, data-collection procedures, and survey duration.
The measured data should be reported by species and diameter class, and detailed information on the allometric growth models selected for different species should be added.
In addition to R² and RMSE—similar to comparable studies—other accuracy metrics such as MAE and MSE should be included, and the uncertainty should be better analyzed in the Discussion.
The optimal implementation process and parameter-tuning procedures of the machine-/deep-learning algorithms need to be described in greater detail. Explain how the optimal OOB, mtry, and ntree values in the RF model were determined.
The description of the Boruta feature-selection method is insufficient; the number of iterations or the significance threshold was not given. Moreover, why was Boruta chosen, and was the modelling accuracy compared between variables selected by Boruta and by other methods?
More details on image selection and preprocessing are required, e.g., the temporal range of the selected images and the total number of processed images.
All abbreviations used in the text should be listed in a unified glossary at the end of the manuscript to improve readability.
The accuracies of individual-tree segmentation and species classification from the imagery should be explicitly reported.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors of 'Individual tree-level biomass mapping in Chinese coniferous plantation forests using multimodal UAV remote sensing approach integrating deep learning and machine learning' present a novel YOLOv8 model (NB-YOLOv8) used to detect individual trees and predict AGB estimates across numerous plots based on UAV optical and lidar data. The results, formed from a validation on trees of several species and size classes, demonstrate an improvement over both previous YOLOv8 methods and watershed segmentation (a common practice in the field today). While this new methodology is effective, the paper is quite limited in 3 regards.

First, there is an severe lack of detail throughout the methodology, making this research not reproducible.

In Section 2.2 (Ground data), it is clear that there are a lot of trees being measured in the field. How were these trees selected and how are they spatially arranged? Are they distributed across plots? If so, how many?
In Section 2.2, scientific names are necessary for each species.
In Section 2.3, the manufacturers information is needed for each piece of UAV hardware.
for Section 2.3.1, was this optical image processing conducted using a software or programming package? Depending on the process, specific settings/parameters must be documented as they have a notable impact on modeling quality. A far greater level of detail is needed here.
In Section 2.3.2, similarly a programming package or software application and its relevant settings must be documented.
In Section 2.5.1, how was the YOLOv8 tree detection algorithm applied?
In Section 2.5.3, which band(s) was the watershed segementation algorithm performed on?
- How was this method applied to the UAV data?
For table 2, what are the sources / support for these optical bands? They are not common, and are likely highly correlated with original opitcal imagery bands. Please support why a color space transformation (IHS) or Principal Component Analysis (PCA) is not more reliable.
Figure 5, please explain why these coefficients differ from Table 4. There is not enough detail in this section.

Second, the paper could be improved by reviewing the terminology and formatting throughout. For example, there are several instances where new acronyms are introduced without being defined. Advanced topics, acronyms, and field-specific terminology should be clarified.

Line 20, 'improves' is used here while only a single accuracy statistics is presented in this first highlight. 'Improves' should be used for a comparative statement.
Line 26, SHAP is not defined.
Line 68, the journal-specific formatting of citations is needed throughout the paper.
Captions throughout the paper need to clearly present all of the values (units) which are contained within the Table / Figure. For example, it is unclear what values are presented in Table 1. In Figure 4, what the bar- and line- graphs refer to.
In Sections 2.6.2. and 2.7.2., citations are needed for prominent methods such as GLCM (Harralick) and RF (Breiman).
Line 461, 'WA' is a new acronym which is not a common usage in remote sensing.
Table 4 caption, 'accuracy' seems misused here. The RMSE (error) and coefficient of determination are given here; not accuracy.
Line 545, '_ME' is not defined before this.
Figure 7, it is unclear, based on the caption, why there are three charts given here.
- The should each be defined either in the text or within the caption.

Lastly, there are several instances in which the methods presented here are described as practical or operational, however, this lacks clear evidence; based on both point #1 above and the slow adoption of advanced technologies and methods by most practitioners.

Line 28, a 'hybrid deep-learning + machine-learning workflow' is far from 'operational' for many foresters or land managers. Effective may be a better word here. The discussion section would need to clearly describe how collecting UAV data using both sensors and then using this processing framework is practically available before it could be called operational.
Lines 74-77, conventional plot-level methods using remote sensing may fail to meet these standards but not when using field-based measurements. This sentence should be reconsidered.
Lines 672 and 686, while it is important that these characteristics can be measured using such technologies and methods, it would again be a benefit if this paper spoke more to what might be 'practical' or 'operational' about the analysis that has been conducted here. It may be true that this will be easier to understand after the level of detail and transparency (points #1 and #2 above) are improved.

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 done a tremendous job improving the quality and clarity of this manuscript.

I would now recommend this work for publications.

Author Response

Thank you for your assistance in the publication of this thesis.