Intelligent Estimating the Tree Height in Urban Forests Based on Deep Learning Combined with a Smartphone and a Comparison with UAV-LiDAR

Round 1

Reviewer 1 Report

The manuscript is of high quality - with well-processed data and an introduction supported by a sufficient number of up-to-date references. I only have a few minor comments.

L103-L104: Please provide species names in italics and in full scientific name at first appearance. I also think it would be more appropriate to place the species names outside the parentheses and instead put the abbreviations in the parentheses.

L146-L147: A chapter with the same name, which immediately introduces a subchapter with the same name. I might leave out the chapter name.

Results – mainly L327-L328: Please enter the same type of values rounded to the same number of decimal places - here, for example, Phone_HeightALL is 5.976 and 12.58. It also seems pointless to me to report a value rounded to three decimal places, if the estimate of its absolute error is rounded to two decimal places.

L381: You mention weeds around the root/person as one of the sources of possible measurement error. At the same time, in the picture in the manuscript, the height of the weed is not very high. In many European cities, in recent years, the strategy of leaving higher vegetation and flowery strips for the benefit of insects, especially pollinators, has been adopted. Could you calculate more precisely, or at least estimate, from what height of the weeds it is already a problem, and for example also estimate what percentage of other images within your original dataset (i.e. including training data) would be affected by this error?

L386-L388: You list the main sources of error: improper shooting distance and not accurate marking of the person and tree by model. Can you describe more practically and explicitly what "improper shooting distance" means and what the user should avoid, or on the contrary, what he should do, to do it properly?

 

L395-L405: obstructions, photographers in the middle of the traffic, background trees – I'm really glad you listed the main obstacles to getting a good picture of a person and a tree for a model. So you can describe more explicitly under which conditions your model can work very well, and you can estimate (really at least in a personal opinion) what percentage of trees in the whole city can meet the criteria for error-free use of your model (not only along the street, where the conditions for taking pictures can be only better than, for example, in some clump of trees between buildings).

Author Response

Thanks for your comments and suggestion. Meanwhile, we have made corresponding revisions and explanations in the manuscript. Detailed responses are provided in specific comments.

Author Response File: Author Response.pdf

Reviewer 2 Report

The topic of the article is very interesting and still relevant. The list of references contains the most important bibliographic sources from the subject area and also includes articles that appeared only recently in 2020 and 2022. Articles that are dedicated to measuring the height of trees and the diameter of tree crowns use means of artificial intelligence such as e.g. convolutional neural networks (eg Ultralytics LLC YOLOv5) and modern communication devices such as smartphones, UAV equiped by Lidar.

 

The content of the article follows modern trends in the subject area. The achieved results contribute to the development of new effective methods of measuring the height of trees. In agreement with the authors, I consider the following to be significant contributions of the article:

1.  Smartphone was used to obtain person-tree images, LabelImg was used to label the images, and a dataset was  constructed

 2.  Based on a deep learning method called YOLOv5  and the small-hole imaging and scale principles, a person-tree scale height measurement model was constructed.

 3.  Tree height measurements were obtained. Using this method, the heights of three species in the validation set were extracted; the range of the absolute error was 0.02 m-0.98 m, and the range of the relative error was 0.20%-10.33%, with the RMSE below  0.43 m, the rRMSE below 4.96%, and the R2 above 0.93.

Conclusion: 

 The person-tree scale height measurement model proposed in this paper greatly improves the efficiency of tree height measurement while ensuring sufficient accuracy and provides a new method for the dynamic monitoring and investigation of urban forest resources.

I have a few comments on the article:

1. Could the authors comment on the measurement of tree height, depending on their height (dominant crown class, tall dominant trees, codominant and medium crown trees). Are there breaking points with this method in terms of the height of the trees and the accuracy of their measurement?

2. Has the structure of YOLO5 been expanded in any way?

3. Yolov5 has four different models: Yolov5s, Yolov5m, Yolov51 and Yolov5x. They all have the same components, but the difference between them is the parameter of depth multiple and width multiple. In the article, the authors do not mention which YOLO5 model they used. (was it YOLO5x?)

 

 

 

Author Response

Thanks for your comments and suggestion. Meanwhile, we have made corresponding revisions and explanations in the manuscript. Detailed responses are provided in specific comments.

Author Response File: Author Response.pdf