Internal Tree Trunk Decay Detection Using Close-Range Remote Sensing Data and the PointNet Deep Learning Method

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this study, a new method based on close-range remote sensing data was tested to automatically detect the probability of decay in the trunks of trees found in city parks or along roads. This method specifically utilized close-range photogrammetry technique and iPhone LiDAR. The PointNet deep learning algorithm was employed for 3D data classification. While the study highlights the versatile use of mobile devices in contemporary life, it also provides valuable data suggesting that certain parameters crucial for forestry can be obtained.

However, the dataset used in the study is a crucial factor in the success of the obtained results. Therefore, further explanation regarding the generalizability of the dataset in the study is needed.

 

Although the accuracy values obtained are quite remarkable, it is recommended to conduct additional verification with an adequate number of repetitions and various datasets.

 

More information should be provided on the hyperparameter selection using Grid Search. The values selected and their impact on the results should be clarified.

 

The article could emphasize more on how the obtained results can be practically applied. The discussion on how this technique can be used in field studies and its suitability for real-world applications can be expanded.

 

The language usage in some parts of the text is complex. The study could be expressed more effectively by using a clearer language.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Comments to the Author
Manuscript “Internal Tree Trunk Decay Detection Using Close Range Remote Sensing Data and the PointNet Deep Learning Method”introduced one new method based on close-range remote sensed data, specifically close-range photogrammetry and iPhone LiDAR, which was tested to detect decayed standing tree trunks automatically. And the mean achieved validation accuracies of the models were good for three different datasets consisting of pure coniferous trees, pure deciduous trees or mixed data.

 Major revision comments:

he conclusions and result of the new methods to detect decayed standing tree trunks are simple, please explain the relationship between the parameters and tree decay, and why different tree species have different results.

Please explain the physiological principles of the tree decay and the parameters used in the paper.

 Minor  comments:
1. Page 11, Line 357.  The table 4 “the best-performing training hyperparameters for this case were number of points = 2048, batch size = 32, learning rate = 0,001 and number of epochs = 70. ” why is it?  Such as Table 2 and Table 3.

 2.Page 7, Line 255 “PointNet trains a classifier based on labelled 3D meshes and predefined hyperparameters: the number of points sampled on the 3D mesh, batch size, number of epochs and 256 learning rate.”Please explain why choose the four parameters to train a classifier?

 3.Page 8, Line 293. “a matrix of potential hyperparameters was made” , please creat a table to show the name of potential hyperparameters and explain the each parameter, or show the calculation formula for each parameter if possible.

4.Page 14, Line 425. There are no two empty spaces at the beginning of each paragraph.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript used PointNet to classify decayed or healthy stems. The topic is exciting, but reviewer is not sure about the soundness of the article, due to the lack of some contents. Maybe authors are letting the PointNet building a relation that may not actually exist. Please see the detailed comments below.

 

From the introduction, it seems impossible to directly judge the state of decay by eye, as it is inside the stem. But Lidar and image data can only reflect 3D surface information, it is hard to believe PointNet can build the mechanism between 3D information and the inside decay. Authors should give some explain about the mechanism (e.g., relation between shape change and inside decay, you mentioned in line432). I have a similar example: AI can hardly judge the tree species only using 3D information (spectral information or others are required) and only coniferous forest or broad-leaved forest can be recognized as it is a representation of geometric form. So you force the AI to building a relation (that may not actually exist) between appearance data and inside decay

 

Both lidar-based and image-based point clouds are captured for the same tree. What is the relation between them and what is the role of the different two types of data? Have you registered them?

 

lines 277-279, modifying the tree point clouds are not reliable here. It is quite different from the case in city, e.g., we all know the shape of buildings, so we can create analog data for buildings. But it seems we are not sure what the geometry characteristic of a stem having inside decay is. You may add noise or cause classifier overfitting with so many similar trees.

 

line 375, suggest authors list the study recommending point cloud images, as it is now widely recognized that using directly point cloud for deep learning is better than point cloud image.

 

Section results. Authors should introduce the test data detailedly, e.g., how many trees are checked and how many healthy and decay trees are respectively, how many are used for training and testing, how many modified trees are used.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

Reviewer is grateful for the added contents, making many details clearer. Most importantly, authors show some evidence for the relation between shape change and decay. But it is not mentioned the wood rotting in [25] and [26] refer to heart rot or sap rot. From the experiment, reviewer thinks authors try to detect heart rot with DL. So maybe the relation between heart rot and shape change should be explained more clearly or directly. Besides this, reviewer has no other comments to add.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf