Developing Forest Road Recognition Technology Using Deep Learning-Based Image Processing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors and editors. I have carefully read the manuscript of the scientific article "Developing Forest Road Recognition Technology Using Deep Learning-Based Image Processing", proposed for publication in the scientific journal "Forests (ISSN 1999-4907)". 

The article is devoted to an urgent topic, corresponds to the subject of the magazine and is of great interest to readers. 

At the same time, there are a number of comments on the article. 

1. Figure 1. Give the scale, the coordinate grid, specify the north arrow. 

2. Figure 4. It is not clear what is shown in Figure 4. I recommend that you add symbols to it. 

3. Table 1. How were meteorological parameters taken into account in the study? 

4. There is no "Discussion" section in the article. There is no comparison of the results obtained with previously published ones with other methods.

5. The limitations of the study are not described.

Author Response

Dear Editors and Reviewers,

 

We appreciate you and the reviewers for the careful comments and suggestions to improve the quality of our manuscript. We have uploaded the 1^st revised manuscript. Here is a list of the reviewer comments and our responses. The listed line numbers below are corresponding to the numbering in the manuscript when “Accept All Changes” is applied.

 

1. Figure 1. Give the scale, the coordinate grid, specify the north arrow. 

• Thanks for the reviewer’s comment. Added scale, coordinates, grid, and north arrow to Figure 1.

2. Figure 4. It is not clear what is shown in Figure 4. I recommend that you add symbols to it. 

• Thanks for the reviewer’s comment. The Yolov8 structure has been replaced with a picture that is clearly visible at once.

3. Table 1. How were meteorological parameters taken into account in the study? 

• Thanks for the reviewer’s reminder. The meteorological parameter taken into consideration when shooting driving videos is the average cloud cover. This is because the contrast and saturation of the forest road change depending on the amount of light source. Added related information to lines 90-97.

4. There is no "Discussion" section in the article. There is no comparison of the results obtained with previously published ones with other methods.

• Thanks for the reviewer’s comment. Discussion items were added, and related previous studies were compared with this study. Added relevant information to lines 257-288.

5. The limitations of the study are not described.

• Thanks for the reviewer’s comment. In the Discussion section, we have updated the limitations of the study and mentioned future research directions to overcome the limitations. Added relevant information to lines 278-288

 

Reviewer 2 Report

Comments and Suggestions for Authors

The study presents a methodology for road segmentation and determination of road centerline using The YOLOv8 segmentation technique. The data used is obtained by the camera integrated into the vehicle. Since the data were taken on similar dates, generalization remains limited, as it is not clear how the results can be obtained in case of seasonal differences. However, the results obtained determine the segmentation with high accuracy. Another limitation is that the road centers are not determined properly. The authors should consider other recommendations and answer the questions given below;

 

 

1.     In section 2.1: it is mentioned that data were recorded under various weather conditions, however, table 2 indicated that the image collection dates are very close and the conditions are similar. Thus, the statement should be revised. In the discussion part, the seasonal effect should be mentioned.

 

2.     Figure 7 provides the loss function and epoch information, but no information is provided on how the intervals were selected. What are the best epoch and loss values? What are the corresponding accuracies? Is there any information on processing time?

 

3.     From the results, it appears that the method does not properly detect road centerlines. The green lines indicate the border, and red and yellow lines indicate the center between the green lines. If the road is oriented to the right or left, the road's center line should be determined in a way that it passes through the center in that direction.

 

4.     In the study, the data set was evaluated by separating 80, 10, 10. However, the results may be dependent on this separation, so why was k-fold cross-validation not used?

 

5.     1.     The discussion part does not discuss the results considering the previous similar studies that used very high resolution airborne, UAV data.

 

Author Response

Dear Editors and Reviewers,

 

We appreciate you and the reviewers for the careful comments and suggestions to improve the quality of our manuscript. We have uploaded the 1^st revised manuscript. Here is a list of the reviewer comments and our responses. The listed line numbers below are corresponding to the numbering in the manuscript when “Accept All Changes” is applied.

 

1. In section 2.1: it is mentioned that data were recorded under various weather conditions, however, table 2 indicated that the image collection dates are very close and the conditions are similar. Thus, the statement should be revised. In the discussion part, the seasonal effect should be mentioned.

 

• Thanks for the reviewer’s reminder. The meteorological parameter taken into consideration when shooting driving videos is the average cloud cover. This is because the contrast and saturation of the forest path change depending on the amount of light source. Added related information to lines 90-97. In the discussion section, factors affecting the image were introduced and ways to overcome them were presented. Added related information to lines 278-288.

 

2. Figure 7 provides the loss function and epoch information, but no information is provided on how the intervals were selected. What are the best epoch and loss values? What are the corresponding accuracies? Is there any information on processing time?

 

• Thanks for the reviewer’s comment. In this study, the highest epoch was 172, and the loss value was 0.22. At this time, the accuracy (recall rate) is 0.97, and the total training time is about 3 hours. The variables used for training are listed in lines 145-148.

 

3. From the results, it appears that the method does not properly detect road centerlines. The green lines indicate the border, and red and yellow lines indicate the center between the green lines. If the road is oriented to the right or left, the road's center line should be determined in a way that it passes through the center in that direction.

 

• Thanks for the reviewer’s comment. Green lines represent the boundaries of both ends of the forest road, and the centers of both ends are represented by yellow lines. The red line represents the horizontal axis center of the image. When the road curves to the left, as shown in Figure 9(b), you can see that the yellow line is to the left of the red line. Also, since Figure 9(c) is a road that curves to the right, you can see that the yellow line is to the right of the red line. You can. Added additional information about each color line to lines 184-187.

 

4. In the study, the data set was evaluated by separating 80, 10, 10. However, the results may be dependent on this separation, so why was k-fold cross-validation not used?

 

• Thanks for the reviewer’s comment. When constructing the data set in this study, K-fold cross-validation was not considered because the data set was randomly mixed. Table.1 shows cross-validation with k-fold (K=5) using the same data set. It is judged to show consistent model performance as it has a similar mean value and low standard deviation as the existing model.

 

Table 1

Fold	mAP	Recall	F1	Precision
Fold 1	0.979	0.965	0.967	0.969
Fold 2	0.972	0.943	0.940	0.938
Fold 3	0.957	0.910	0.921	0.932
Fold 4	0.978	0.914	0.937	0.962
Fold 5	0.980	0.966	0.961	0.957
Average	0.973	0.940	0.945	0.952
Standard Deviation	0.009	0.024	0.017	0.014

 

 

5.   The discussion part does not discuss the results considering the previous similar studies that used very high resolution airborne, UAV data.

 

• Thanks for the reviewer’s comment. A section on the detection of forest roads and autonomous driving using ultra-high-resolution aerial images and drones has been added to the discussion. Added relevant information to lines 258-288.

 

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors replied to the questions properly and improved the manuscript with new statements. However, the detection of road centerline is still not clear. For example, in Figure 10a, the centerline is straight due to the road geometry. But in Figure 10b the road is curved, but the centerline is still straight. This situation cannot be defined as “road centerline detection”.

Author Response

Dear Editors and Reviewers,

 

We appreciate you and the reviewers for the careful comments and suggestions to improve the quality of our manuscript. We have uploaded the 2^nd revised manuscript. Here is a list of the reviewer comments and our responses. The listed line numbers below are corresponding to the numbering in the manuscript when “Accept All Changes” is applied.

 

Reviewer #2

1. The authors replied to the questions properly and improved the manuscript with new statements. However, the detection of road centerline is still not clear. For example, in Figure 10a, the centerline is straight due to the road geometry. But in Figure 10b the road is curved, but the centerline is still straight. This situation cannot be defined as “road centerline detection”.

• Thanks for the reviewer’s comment. We meant that the road centerline on the image is just the extension of a center position between both side ends of road. This indicates the lateral offset of the forwarder at the current location and the amount and direction of steering is determined As your comment road center detection sounds more reasonable. So, in the main manuscript the phrase ‘forest road center line’ was changed to ‘forest road center.’ Thank you so much for your thorough review.