Selection of a Navigation Strategy According to Agricultural Scenarios and Sensor Data Integrity
, Mário Luiz Tronco 1,†, Vitor Akihiro Hisano Higuti 1, Andres Eduardo Baquero Velasquez 1, Mateus Valverde Gasparino 1, Handel Emanuel Natividade Peres 1, Rodrigo Praxedes de Oliveira 1, Vivian Suzano Medeiros 1, Rouverson Pereira da Silva 2 and Marcelo Becker 1Round 1
Reviewer 1 Report
The author investigates navigation strategy selection based on agricultural scenarios and sensor data integrity. The author has made a complete analysis of various characteristics of various sensors equipped with AgBot, such as GNSS, camera, and LiDAR. However, these sensors may be vulnerable to limitations, such as low accuracy of GNSS for navigation under canopy, sensitivity of cameras for outdoor lighting and platform vibration. And the occlusion problem of using LiDAR. Then, according to these problems, feature collection data is extracted from GNSS, image and point cloud through experiments, and the feasibility of using each sensor is determined after analysis, and the selection vector indicating its feasibility is created. Different sensors and their technologies can be reasonably used in different situations. Finally, it can be shown that the method in this paper has higher performance and lower calculation cost. And the most appropriate sensors can be selected at a low cost for a specific agricultural environment.
The review comments are as follows:
1. Table 3 in line 127 should be changed to Table 2.
2. Table 2 in line 135 should be changed to Table 3.
3. The third and fourth points mentioned in line 164-167 are the lack of time for light to enter the camera and the lack of time for the camera to focus. Here, it is necessary to explain the reason for the lack of time, whether the speed of the robot is set too fast or the hardware reason.
4. The LRT hypothesis test described in line 201 lacks a definition and what are the advantages of choosing the LRT hypothesis test?
5. Lines 268-270 of the article explain the Occlusion classes are not able to use Lidar for navigation, so they choose to use camera for navigation, and the Occlusion classes is an ideal class for camera navigation. But why does it say in line 276-277 that LiDAR navigation is used?
6. It is not explained in line 279 why it is suggested to use the policy selection vector in formula 10, and how such formulas are obtained?
7. Line 368 should state which technique improves accuracy by 4% and requires doubling processing time; Or which technique improves accuracy by 4% relative to which technique requires doubling processing time.
8. Line 382 should explain why it is negligible after 10-3.
Author Response
Reviewers’ Response Letter
Manuscript ID: agronomy-2254039
First of all, we would like to thank the reviewers for their time and effort spent in
reviewing the manuscript and for their constructive and valuable comments,
which allowed us to improve the quality of the paper. We have addressed all the
comments and modified the paper accordingly. The specific comments are
addressed below.
1. Table 3 in line 127 should be changed to Table 2.
Answer: The error was corrected in the revised manuscript.
2. Table 2 in line 135 should be changed to Table 3.
Answer: The error was corrected in the revised manuscript.
3. The third and fourth points mentioned in line 164-167 are the lack of time for
light to enter the camera and the lack of time for the camera to focus. Here, it is
necessary to explain the reason for the lack of time, whether the speed of the
robot is set too fast or the hardware reason.
Answer: The amount of light entering the transducer depends on the cultivation
conditions. When the distance between the crop lines is larger, more light enters
the environment, and it takes time for the sensor to adjust the transducer opening accordingly. Conversely, when the lines are closer, less light enters. This
adjustment process can affect the accuracy of the measurements. We added a
paragraph to clarify this point in the revised manuscript (lines 174-179).
4. The LRT hypothesis test described in line 201 lacks a definition and what are
the advantages of choosing the LRT hypothesis test?
Answer: The Likelihood Ratio Test (LRT) is a statistical tool used to compare a full model with a restricted model. In our case, the full model includes all the
variables described, while the Stepwise model removes the variables that are not
statistically significant, resulting in a restricted model with fewer predictor
variables. One advantage of the Likelihood Ratio Test is that it allows the
comparison of models with different numbers of hyperparameters. Furthermore, it has a low computational cost, making it a practical and efficient tool for model
comparison. W We added a paragraph to clarify this point in the revised
manuscript (lines 213-220).
5. Lines 268-270 of the article explain the Occlusion classes are not able to use
Lidar for navigation, so they choose to use camera for navigation, and the
Occlusion classes is an ideal class for camera navigation. But why does it say in
line 276-277 that LiDAR navigation is used?
Answer: The Occlusion class can occur in both sensors, but they are actually
different classes despite having the same name. Our definition of Occlusion is
when the obstruction happens within the crop, and therefore, we can only rely on either LiDAR or camera for navigation. When the camera is occluded, we switch to LiDAR navigation, and when the LiDAR is occluded, we switch to camera navigation. To make the distinction clearer, we decided to split the original Occlusion class into two new classes: LiDAROcclusion and CameraOcclusion. This helps us better understand which sensor is being affected and facilitates the navigation decision-making process.
6. It is not explained in line 279 why it is suggested to use the policy selection
vector in formula 10, and how such formulas are obtained?
Answer: We use the feasibility matrix to assign weights to each of the classes,
and the Classification Vector to indicate the probability of each class being true.
However, since different sensor classes may favor the use of specific sensors, we
also calculate the Strategy Selection Vector to determine the best sensor to use in each situation. We added this information to further clarify the process in lines
297-298.
7. Line 368 should state which technique improves accuracy by 4% and requires
doubling processing time; Or which technique improves accuracy by 4% relative
to which technique requires doubling processing time.
Answer: The Decision Tree (DT) technique has a higher accuracy rate of 74.44%
compared to the Multinomial Logistic Regression (MLR) technique, which has an
accuracy rate of 70.93%. However, the DT technique takes twice as long to
process the data as the MLR technique. We attempted to clarify this point in lines 381-383.
8. Line 382 should explain why it is negligible after 10-3.
Answer: After the value of 10-3, there is a decrease in the accuracy of the model, this is an indication that the model is overfitting based on the training data. We attempted to clarify this point in lines 397-398.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
I would like to thank you for the opportunity to review your valuable paper. I also understand the need for this research due to the position of sugarcane in your country. However, there are many errors in the descriptions in the text. Corrections should be made based on the points I have pointed out.
Comments for author File: Comments.pdf
Author Response
Reviewer 2
Answer: We sincerely appreciate the reviewer's comments. All the errors pointed
out by the reviewer were corrected in the revised manuscript. Also, an extensive
english review was performed throughout the text.
Author Response File: Author Response.pdf
