Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The identification and avoidance of obstacles in unmanned farms is an important topic, and this article identifies obstacles in the field, which is of great significance. This detection model builds upon the YOLOv8 architecture with three main improvements. This is a subject which is a good fit for this journal.

But authors need improve part of the manuscript.  Here are some specific recommendations:

1. In 2.1. Related Work, this section is described too much and overlaps with the introduction. It is recommended to simplify it.

2. In section 2.2, please provide a detailed description of the effects of the added CBAM attention mechanism, BiFPN pyramid network, and WioUv3 on model improvement and recognition of obstacles such as pedestrians and utility poles.

3. In section 2.5, speed should be added, because this evaluation metric was applied in 3.2, the context should remain consistent.

4. Why choose YOLOv8 for research when YOLO series already has YOLOv9 and YOLOv10 versions?

5. The article points out: This dataset includes 1,800 images 413 of people, 1,160 images of harvesters, 253 images of tractors, 1,280 images of electric power 414 pylons, and 1,470 images of telegraph poles. But during the experiment, a large agricultural machinery was used behind the tractor. So, if unmanned operations enter the field, it is recommended to increase the dataset of obstacle types (such as graders, rotary tillers, sprayers, etc.).

6. The classification in the sample images of the dataset in the article is relatively rough, for example, there are many right front views of tractors. How accurate is the recognition of the front view? Suggest training and recognition in various directions of agricultural machinery. At the same time, it is suggested to supplement identification experiments on different lighting conditions, occlusion, distance, etc. in the field.

7. In Figure 10, was the tractor not detected in the bottom right corner of the image? Is it due to the long distance? Please explain.

8. There are both 640 * 640 and 640 × 640 in section 3.3, please make consistent modifications.

9. In Figure 14, the coordinate axis does not indicate coordinate information; The training cycle should be 200 cycles, why are there no other results around 150.

10. Line 497: "However, due to the relatively low speed of unmanned agricultural machinery operating in the field, the impact on detection speed can be considered negligible." Unmanned agricultural machinery operations will operate at relatively low speeds. Do you consider seasonal issues, cost, and work efficiency of crops?

11. Specific data should be provided in the conclusion to support the research results.

12. Some images are unclear or distorted, such as Figures 1, 2, 8, 9, 13 and 14. Please replace them.

13. Some images have overlapping text, such as Figures 10, 11 and 12. Please replace them.

 

Author Response

Point 1: In 2.1. Related Work, this section is described too much and overlaps with the introduction. It is recommended to simplify it.

 

Response 1: Thank you for pointing out the redundancy between the “Related Work” section and the “Introduction”. We agree that streamlining this section will improve the readability and coherence of the manuscript. In the revised manuscript, we have condensed the “Related Work” section by removing overlap and focusing on directly relevant research. Key details have been retained to provide the necessary background information without repeating information already covered in the Introduction. We have removed the introductory section on CBAM from 2.1.2 Attention Mechanism, as it is mentioned in the Introduction section and described in detail in 2.2.3 CBAM Attention Mechanism. We have removed the introduction of the early and less common loss function in 2.1.4 Loss Function, and removed the introduction of the WIou v3 loss function as it is mentioned in the introduction section and described in detail in 2.2.5 WIou v3 Loss Function. The above actions remove detailed background information and overlap with the introduction and shorten the length, while retaining the necessary references to key studies. We believe these changes will improve the clarity and relevance of the section. Thank you for your helpful feedback!

 

Point 2: In section 2.2, please provide a detailed description of the effects of the added CBAM attention mechanism, BiFPN pyramid network, and WioUv3 on model improvement and recognition of obstacles such as pedestrians and utility poles.

 

Response 2: Thank you for the insightful suggestion to describe in more detail how the added components (CBAM, BiFPN and WIoU v3) improve the model and their specific impact on obstacle recognition. We have revised Section 2.2 to describe in detail these mechanisms and their contribution to the detection of complex obstacles such as pedestrians and utility poles. This extended discussion will be included in Section 2.2.5 (lines 471-498) provide a clearer understanding of the contributions of these components. Thank you for your constructive feedback, which has helped us enhance the manuscript!

 

Point 3: In section 2.5, speed should be added, because this evaluation metric was applied in 3.2, the context should remain consistent.

 

Response 3: Thank you for pointing out the inconsistency in the assessment indicators between Section 2.5 and Section 3.2. We agree that the inclusion of the speed indicator in Section 2.5 is critical for consistency and a full understanding of the assessment framework. The manuscript has been revised to include a detailed explanation of the speed indicator in Section 2.5 (lines 574-590) to make it consistent with its application in Section 3.2. By adding this explanation, we aim to provide a consistent and comprehensive assessment framework throughout the manuscript. Thank you for your valuable feedback!

 

Point 4: Why choose YOLOv8 for research when YOLO series already has YOLOv9 and YOLOv10 versions?

 

Response 4: Thank you for your question regarding the choice of YOLOv8 for the study over newer versions such as YOLOv9 or YOLOv10. We considered this carefully during the study design process and made our decision for the following reasons: at the time of this study, YOLOv8 was the most mature and widely used version of the YOLO family. It has extensive documentation, community support, and best-in-class results in object detection tasks across a wide range of domains. yOLOv8 offers built-in flexibility for architectural modifications, making it suitable for integrating advanced components such as CBAM, BiFPN, and WIoU v3. Its modular design allows us to efficiently implement and test our proposed improvements, which may not be fully compatible or well documented in newer versions. The study began before YOLOv9 and YOLOv10 were released or widely adopted. YOLOv8 was the latest stable version available during the initial phase of this study. Transitioning to a newer version in the middle of the project may introduce additional variables and complexity that may delay research results. YOLOv8 has demonstrated state-of-the-art performance in terms of detection accuracy, speed, and scalability. Our research focused on enhancing these features in an agricultural setting rather than benchmarking the latest YOLO version. Our modifications to YOLOv8 (i.e., integration of CBAM, BiFPN, and WIoU v3) have resulted in significant improvements that ensure the model is competitive with or better than newer versions. We recognize that newer versions (e.g., YOLOv9 and YOLOv10) may bring more improvements. However, our selection of YOLOv8 is consistent with the study timeline, goals, and stable framework available for customization. To address this issue, we added the following explanation in Section 2.1 (lines 249-253): “Although newer versions of the YOLO family have been released (e.g., YOLOv9 and YOLOv10), this study utilizes YOLOv8 because it is mature, stable, and compatible with the proposed architectural enhancements. YOLOv8 offers advanced performance and customizability, making it a solid foundation for developing and validating improvements targeting agricultural obstacle detection.” We hope this explanation clarifies our reasons for choosing YOLOv8 and demonstrates its suitability for the research objectives. Thank you for your valuable feedback!

 

Point 5: The article points out: This dataset includes 1,800 images 413 of people, 1,160 images of harvesters, 253 images of tractors, 1,280 images of electric power 414 pylons, and 1,470 images of telegraph poles. But during the experiment, a large agricultural machinery was used behind the tractor. So, if unmanned operations enter the field, it is recommended to increase the dataset of obstacle types (such as graders, rotary tillers, sprayers, etc.).

Response 5: Thank you for your observations and suggestions to expand the dataset by adding other obstacle types such as graders, rototillers, and sprayers. We agree that adding these obstacles would further enhance the relevance and comprehensiveness of the dataset to real-world unmanned agricultural operations. Here is our response to your question: the current dataset is designed to capture a variety of common obstacles encountered in agriculture, such as pedestrians, harvesters, tractors, utility towers, and utility poles. These types were selected based on their prevalence and their potential impact on the operation of unmanned agricultural machinery. We recognize that the current dataset does not include certain other obstacle types, such as graders, rototillers, and sprayers, which are indeed important in certain agricultural environments. The lack of these obstacle types in the dataset is a limitation of this study, and we have now explicitly stated this in the revised manuscript. We have included this suggestion in the “Discussion” section of the manuscript (lines 833-839). Specifically, we plan to expand the dataset to include these additional obstacle types in future work. This expansion is intended to address unmanned operations in areas where larger, more specialized agricultural machinery is typically deployed. We appreciate your valuable feedback, which will guide us in enhancing the dataset and improving the practical applicability of the model. Thank you for your constructive suggestions!

 

Point 6: The classification in the sample images of the dataset in the article is relatively rough, for example, there are many right front views of tractors. How accurate is the recognition of the front view? Suggest training and recognition in various directions of agricultural machinery. At the same time, it is suggested to supplement identification experiments on different lighting conditions, occlusion, distance, etc. in the field.

 

Response 6: Thank you for your detailed feedback on the diversity of the dataset in terms of object viewpoints and environmental conditions. We appreciate your suggestion to incorporate multi-view training on agricultural machinery and to conduct additional experiments under a variety of field conditions. Here is our response and the corresponding actions we have taken: we recognize that the dataset contains a disproportionate number of frontal views of tractors and similar agricultural machinery, which may limit the model's ability to generalize to other viewpoints. Subsequent plans for this study will prioritize acquiring and annotating multi-view images to ensure robust recognition in all orientations. While the current study did evaluate the model under typical field conditions, we agree that additional testing under different lighting conditions (e.g., sunrise, sunset, overcast), occlusion levels, and varying object distances would provide a more comprehensive assessment of model performance. In future plans we will consider these factors and conduct relevant experiments that are designed to assess the robustness of the model under these scenarios. Examples include: lighting conditions: testing under bright sunlight, low light and shadow conditions; occlusion: simulating partial and severe occlusion from other objects (e.g., vegetation or machinery); and distance: measuring the detection accuracy of near and distant objects to assess the model's handling of scale variations.We have included these limitations in the Discussion section of the manuscript (lines 839-847), emphasizing the need to expand the dataset to include different views of the machine (e.g., rear, side) as well as to conduct performance tests in different scenarios. We believe that these future plans will address your concerns and significantly enhance the relevance and robustness of the study. Thank you for your constructive and insightful suggestions!

 

Point 7: In Figure 10, was the tractor not detected in the bottom right corner of the image? Is it due to the long distance? Please explain.

 

Response 7: Thank you for your observation of the undetected tractor in the lower right corner of Figure 10. We appreciate your attention to detail and agree that this needs clarification. Here is our explanation and action taken: the tractor detection failure in the lower right corner could indeed be due to the large distance between the object and the camera. At such a distance, the resolution of the object in the image becomes lower, which reduces the model's ability to extract the unique features needed for accurate detection. We have added this observation and explanation to the description of this image section of the manuscript (lines 605-608). In addition, we have added information in the Discussion section (lines 848-852) that indicates plans to improve the model through improved resolution processing or super-resolution techniques to achieve better long-range detection. Future work will explore techniques such as super-resolution algorithms or specialized remote detection modules to improve the model's ability to recognize objects at a distance. These enhancements will address current limitations and extend the applicability of the system to the larger agricultural field where obstacles may appear far from the camera. We hope this explanation addresses your concerns and demonstrates our commitment to improving the robustness of the model. Thank you for your valuable feedback!

 

Point 8: There are both 640 * 640 and 640 × 640 in section 3.3, please make consistent modifications.

 

Response 8: Thank you for pointing out the inconsistency in the image size notation in Section 3.3 (“640 * 640” vs. “640 × 640”). Consistency is essential for clarity and professionalism in scientific writing. We carefully reviewed the manuscript and made the necessary changes to ensure that “640 × 640” is used consistently throughout the text. Thank you for your detailed feedback, which helped us improve the clarity and presentation of the manuscript.

 

Point 9: In Figure 14, the coordinate axis does not indicate coordinate information; The training cycle should be 200 cycles, why are there no other results around 150.

 

Response 9: Thank you for pointing out the discrepancy in the number of elapsed times shown in Figure 9. During model training, the “patience” parameter is activated as part of the Early Stopping strategy. This parameter is used to monitor the performance of the model on the validation set and stops training if no significant improvement is observed on the specified epochs. In our case, patience was set to stop training when the validation performance did not improve after 20 epochs. Thus, although the maximum epoch was set to 200, training stopped after 150 epochs because no further improvement was observed. This mechanism is intended to prevent overfitting and improve training efficiency. To address your concerns, we will update the manuscript to include a detailed explanation of the PATIENCE parameter (lines 564-572) and explicitly mention this early stopping process. In addition, we will revise the captions of Fig. 9 (updated to Fig. 10) and Fig. 14 (updated to Fig. 15) to make it clear that the training is terminated before the maximum epochs are reached. We appreciate your attention to this detail and the opportunity to improve the clarity of the manuscript.

 

Point 10: Line 497: "However, due to the relatively low speed of unmanned agricultural machinery operating in the field, the impact on detection speed can be considered negligible." Unmanned agricultural machinery operations will operate at relatively low speeds. Do you consider seasonal issues, cost, and work efficiency of crops?

 

Response 10: Thank you for your insightful comments on the broader context of unmanned agricultural machinery operations, including seasonality issues, costs and crop efficiency. These factors are indeed critical in assessing the practical applicability of such systems. Below is our detailed response: Seasonal variations can significantly affect field conditions (e.g. soil hardness, vegetation density) and visibility (e.g. light levels during planting or harvesting seasons). The low operating speed of unmanned agricultural machinery ensures better control and safety, especially in challenging seasonal conditions. While unmanned machinery operates at slower speeds than traditional manual machinery, this trade-off often leads to improved accuracy and reduced crop damage, which can improve overall efficiency. Additionally, lower speeds reduce the likelihood of missing obstacle detections that could otherwise lead to costly repairs or operational delays. Using our system can reduce operational risk and increase machine efficiency, ultimately offsetting the upfront costs of deploying an advanced detection system. This expanded discussion will be included in the manuscript (lines 659-675) to provide a clearer understanding of the applicability of detection speeds, detection efficiencies, etc., to real-world agricultural production processes.We believe that this expanded discussion will provide a more comprehensive perspective on the actual deployment of unmanned agricultural systems. Thank you for raising these important points!

 

Point 11: Specific data should be provided in the conclusion to support the research results.

 

Response 11: Thank you for suggesting that specific data be included in the conclusions to better support the findings. We agree that providing quantitative results can improve the clarity and credibility of the study. The following is our response and the corresponding changes made to the manuscript: in this paper, we propose a field obstacle detection model CBBI-YOLO based on improved YOLOv8,which aims to solve the problem of real-time obstacle detection for unmanned agricultural machines in complex field environments. By introducing the CBAM attention mechanism in the Backbone layer of the baseline model, the feature expression is enhanced by combining the channel and spatial attention; the BiFPN feature pyramid is used in the neck network instead of the original PANet to improve the multi-scale feature fusion ability; and the loss function is replaced by WIoU v3 to improve the matching effect between the prediction frame and the real frame, so as to increase the model's detection accuracy and generalization Ability. The results show that the average accuracy of the improved model on the homemade dataset is 1.9 percentage points higher than that of YOLOv8, with an FPS of 52, which balances high accuracy and real-time performance, and provides technical support for obstacle detection of unmanned agricultural machines.

By integrating key quantitative results (e.g., average accuracy, detection speed) into the conclusions, the study's contribution is clearer and its practical significance is enhanced. This extended discussion will be included in the “Conclusion” section of the manuscript (lines 854-866) to better support the findings. Thank you again for your constructive feedback, which helps to improve the rigor and impact of our manuscript!

 

Point 12: Some images are unclear or distorted, such as Figures 1, 2, 8, 9, 13 and 14. Please replace them.

 

Response 12: Thank you for pointing out the clarity of some of the charts such as Figures 1, 2, 8, 9, 13 and 14. We appreciate your feedback, as clear and easy-to-read diagrams are essential for effectively communicating the results of a study. We have heeded your comments and have re-substituted these unclear or distorted images in the manuscript. We hope these improvements will address your concerns and significantly improve the visual quality and accessibility of the diagrams. Thank you again for your valuable feedback!

 

Point 13: Some images have overlapping text, such as Figures 10, 11 and 12. Please replace them.

 

Response 13: Thank you for pointing out the overlapping text in Figures 10, 11 and 12. Clear and correctly formatted visuals are critical to accurately communicating results, and we appreciate your attention to this issue. Due to the large number and dense distribution of detected obstacles in the images, there may be an overlap in the text of the obstacle detection boxes, which may not have been considered during the design of the baseline model. We have tried to separate the text content from the labeling box for a certain distance, but the visual effect achieved is not good, and we will continue to try other methods to improve this limitation. We have, however, replaced the images in this manuscript, improving the clarity to enhance the visual experience. Thank you again for your constructive feedback, which helps us to improve the clarity and visual quality of the paper.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents an exploration of improving the YOLOv8 model for target detection in complex field scenarios. The authors have integrated CBAM, BiFPN and WIoU v3 into YOLOv8, and the results of experiments show that the proposed model has an improvement in both the model’s detection accuracy and robustness. This paper is good, but i only have one small question  whether the dataset used in the experiments has gone through the preprocessing steps, if there are any, please provide the corresponding description.

Comments for author File: Comments.pdf

Author Response

Point 1: In line 408-410, the authors referred to five types of obstacles, the question is that these types of obstacles is insufficient for unmanned agricultural machinery in the field. Small animals or farm tools left in the field are also likely to be encountered.

 

Response 1: Thank you for your valuable comments. In our current study, we have selected five obstacle types (people, electric power towers, telegraph poles, tractors, and harvesters) as the primary detection targets based on preliminary field investigations and consultations with our subject matter collaborators, primarily because these objects are more frequent and potentially hazardous in agricultural operations, and have a greater impact on the safe operation of unmanned farm machinery. However, as you have pointed out, small animals or farm implements left in the field are also obstacles that may be encountered. For these types of obstacles, we will consider further expanding the dataset in future research and incorporating corresponding detection functions in the model to improve the generalizability and robustness of the system. Meanwhile, we will update the text (lines 833-852) in the revised version to clearly point out the limitations of the obstacle types in the current work and suggest that future work may consider covering more obstacle types. Thanks again for your valuable suggestions.

 

Point 2: A more detailed description of the dataset used in the experiments should be added. For example, the preprocessing steps, including data normalization, augmentation techniques should be clearly described.

 

Response 2: Thank you for your constructive comments on the dataset description. In the revised draft (lines 523-553), we will provide a more detailed explanation of the dataset used in the experiments. Specifically, we will detail the preprocessing steps, including：Data normalization: A description of how the input images are normalized (e.g., scaling pixel values to a specific range) to improve consistency in model training;

Data Enhancement Techniques: A comprehensive list of enhancement techniques, such as noise, brightness adjustment, that are employed to increase the diversity of the dataset. These methods are designed to improve the robustness of the model to variations in real-world scenarios; Dataset Characterization: More details about the composition of the dataset, such as the number of samples for each obstacle type, their distribution in the training, validation, and test sets. We believe that these additions will give the reader a clearer understanding of the experimental setup and improve its reproducibility. Thank you for pointing out this important aspect.

 

Point 3: Fig.13 and 14 are not aesthetically pleasing. The text in Fig.13 is too small to be clearly read, while the text in Fig.14 is too large.

 

Response 3: Thank you for your valuable comments on figures 13 and 14. We recognize that there are problems with the readability of the text in these figures. In the revised version, we will make the following adjustments to improve the aesthetics and clarity of the figures: Figure 14 (originally drafted as Figure 13): The text size will be increased to ensure readability. In addition, the overall layout will be reviewed to optimize the presentation of the diagram elements; Figure 15 (originally drafted as Figure 14): text size will be reduced to maintain consistency with the overall formatting of the manuscript. We will also ensure that the spacing and alignment of the elements in the figure are improved to improve the visual balance. We appreciate your feedback as it helps to improve the quality of our work.

 

Point 4: The parameter of epoch was set to 200, but the results after epoch 150 are not shown in the figures.(Fig.9 and 14)

 

Response 4: Thank you for pointing out the discrepancy in the number of elapsed times shown in Figure 9 and Figure 14. During model training, the “patience” parameter is activated as part of the Early Stopping strategy. This parameter is used to monitor the performance of the model on the validation set and stops training if no significant improvement is observed on the specified epochs. In our case, patience was set to stop training when the validation performance did not improve after 20 epochs. Thus, although the maximum epoch was set to 200, training stopped after 150 epochs because no further improvement was observed. This mechanism is intended to prevent overfitting and improve training efficiency. To address your concerns, we will update the manuscript to include a detailed explanation of the PATIENCE parameter (lines 564-572) and explicitly mention this early stopping process. In addition, we will revise the captions of Fig. 9 (updated to Fig. 10) and Fig. 14 (updated to Fig. 15) to make it clear that the training is terminated before the maximum epochs are reached.We appreciate your attention to this detail and the opportunity to improve the clarity of the manuscript.

 

Point 5: The discussion section of this paper is somewhat simplistic. A more in-depth and comprehensive discussion of the research results is needed. For example, introduce the comparison with the results of previous studies. In addition, the discussion of the research limitations can be added.

 

Response 5: Thank you for your valuable suggestions on the discussion section. We agree that a more in-depth and comprehensive discussion will make the manuscript better. In the revised version (lines 751-765, 833-852), we will make the following improvements: We will provide a detailed comparison of our results with similar studies in the literature. This will include a discussion of the advantages of our improved YOLOv8 model in terms of detection accuracy, speed, and robustness, as well as areas where other methods may outperform ours. Specific metrics will be used to highlight comparative performance; We will discuss the limitations of the study. For example, the model currently focuses on five main barrier types, which may not cover all potential situations in real-world agricultural environments. The effect of environmental conditions (e.g., lighting, weather) on model performance will also be addressed as a potential limitation. These additions will provide a more balanced and nuanced interpretation of our findings and align with the feedback from the improvement discussion section.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

I have reviewed your paper , would like to share a brief summary along with some comments and suggestions that I believe may enhance your work. Firstly, your paper presents the CBBI-YOLO model, which integrates the CBAM, BiFPN, and WIoU v3 loss function to improve obstacle detection in agricultural environments. Through a comprehensive dataset and rigorous experimental validation, the model demonstrates significant improvements in precision..etc.I have attached a PDF copy of your paper "agriculture-3310296-peer-review-v1" with my comments and suggestions.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Point 1: please ,stating the specific challenges faced in obstacle detection within agricultural environments

 

Response 1: Thank you for emphasizing the importance of discussing the specific challenges of obstacle detection in agricultural environments. In the revised draft, we will add a detailed subsection to address these challenges. Specifically, we will elaborate: The variety of obstacle types, ranging from large static objects to small dynamic objects, with varying detection difficulties. The influence of complex and variable backgrounds (e.g., vegetation and soil patterns) in agricultural fields. Challenges of environmental factors (including light and weather conditions) on detection accuracy. Real-time processing requirements for safe operation of unmanned agricultural machinery. Occlusion problems caused by crops or other objects in the field. Limitations due to insufficient diversity of datasets representing global agricultural conditions. We will integrate this discussion into the manuscript (lines 53-81) in order to provide the reader with a comprehensive understanding of the unique obstacles in the field. Thank you for your valuable suggestions, which will greatly enhance the clarity and depth of the manuscript.

 

Point 2: Clearly outline how your proposed improvements to YOLOv8 differ from existing models

 

Response 2: Thank you for pointing out the need to more clearly explain how our proposed improvements to YOLOv8 differ from the existing model. In the revised draft, we will clearly outline the changes made to YOLOv8 and compare them to related work. We will emphasize the following aspects: first, innovations in the model structure, such as enhanced feature extraction, these improvements are deeper and more adaptable compared to other models; second, algorithmic improvements, we utilize a new loss function, which helps the model deliver outstanding performance in complex environments; Third, the performance improvement, the ablation experiment can confirm that our model has more outstanding performance compared to the common existing models. This information will be included in the revised manuscript (lines 164-171) to provide a clear, structured comparison. We appreciate your suggestion as it will help clarify the novelty and significance of our work.

 

Point 3: CBAM and BiFPN，are briefly defined or explained. please ！

 

Response 3: Thank you for emphasizing the need for a clearer explanation of CBAM and BiFPN. in the revised draft, we will briefly define and explain these modules as follows: CBAM is an attention mechanism that applies channel and spatial attention sequentially to refine the feature representation. Channel Attention: focuses on identifying important feature channels by recalibrating the weights based on the global context. Spatial attention: highlights important spatial regions by utilizing contextual information in the feature map. These combined attention mechanisms enhance the model's ability to focus on relevant features while suppressing irrelevant features, making it particularly suitable for complex and chaotic environments such as agricultural fields. BiFPN is a state-of-the-art feature fusion network that improves the traditional feature pyramid network by introducing bidirectional connectivity and efficient weighting mechanisms. It allows information to flow in both top-down and bottom-up directions, ensuring better feature reuse and more robust multi-scale feature representation. In addition, learnable weights are used to prioritize important features, thus improving model performance without significantly increasing computational cost. BiFPN is particularly effective for detecting objects at different scales, a common challenge in agricultural environments. We will incorporate these explanations into the revised manuscript (lines 180-182, 194-196) to ensure a clear understanding for all readers. Thank you for your suggestions, which will improve the readability of the manuscript.

 

Point 4: why did you choose precision, recall and average precision as your primary metrics？

 

Response 4: Thank you for your insights on the selection of metrics for evaluation. Precision, Recall, and Average Accuracy (AP) were selected as the primary metrics for the following reasons: Precision quantifies how accurately the model recognizes real obstacles without incorrectly categorizing irrelevant objects, which is critical in agricultural environments where false positives (e.g., incorrectly categorizing vegetation as an obstacle) can lead to unwanted system outages. Recall measures the model's ability to detect all relevant obstacles, ensuring that no critical obstacles are missed, which is critical for the safe operation of unmanned agricultural machinery.

Average Precision (AP) provides a balanced assessment of the model's precision-recall trade-off at various confidence thresholds, providing a single, interpretable value for overall performance. This is particularly useful for comparing our model to other state-of-the-art methods. These metrics are widely recognized and widely used in the object detection community, making our results comparable to existing studies and benchmarks. By combining precision, recall, and AP, we ensure a comprehensive assessment of the model's effectiveness in terms of accuracy, sensitivity, and overall performance. We will emphasize this rationale more explicitly in the revised manuscript (lines 574-589) to clarify our choice of metrics.

 

Point 5: I suggest to you, considering the integration of the Related Work sections into the Introduction. This change could enhance the flow and coherence of the paper by providing a more comprehensive context for your research right from the beginning

 

Response 5: Thank you for your suggestion to integrate the related work sections into the introduction. We agree that combining these sections would enhance the flow of the paper and provide a more cohesive context for our research. In the revised version, we have made a small adjustment to this. We appreciate your feedback and will implement this change to improve the coherence and readability of the manuscript.

 

Point 6: How each modification to the YOLOv8 architecture (e.g. integration of CBAM. BiFPN, and Wlou v3) was implemented?

 

Response 6: Thank you for your questions regarding the implementation of the YOLOv8 architectural revision. In the revision, we will detail how each component is integrated and its role in improving model performance. Below is an overview of the key implementations: CBAM is integrated into the backbone of the YOLOv8 architecture after the convolutional layer responsible for feature extraction. The channel attention and spatial attention modules are applied sequentially to enhance the model's attention to the relevant feature maps. This integration helps the model to prioritize critical spatial regions and channels, thus improving its ability to detect obstacles in complex agricultural environments. BiFPN replaces the standard neck in YOLOv8 to allow bidirectional information flow across different feature scales. A weighted feature fusion mechanism is introduced, in which each feature map is assigned a learnable weight, allowing the network to adaptively prioritize more informative features. This modification improves the model's ability to detect obstacles at different scales and enhances robustness in multi-scale scenarios. The standard IoU losses in YOLOv8 are replaced in the loss function by WIoU v3. WIoU v3 introduces additional penalties for bounding box deviations in terms of scale, position, and aspect ratio to more accurately localize obstacles. The implementation involves updating the loss calculation to incorporate these additional terms, which improves precision and recall metrics. We will add a section (lines 343-350) in the Methods section dedicated to detailing the above implementation. Thank you for your suggestions, which will improve the clarity and completeness of the manuscript.

 

Point 7: How this attention mechanism specifically improves the detection capabilities in complex agricultural environments?

 

Response 7: Thank you for your questions regarding the impact of attention mechanisms such as CBAM on detection capabilities in complex agricultural environments. In the revised manuscript, we will elucidate the specific advantages CBAM offers and how it addresses the unique challenges of these environments: Agricultural fields often present highly textured and cluttered backgrounds, such as dense vegetation or heterogeneous soil patterns. CBAM's channel attention selectively emphasizes the most relevant feature channels and suppresses irrelevant channels, such as noise induced by background texture. Spatial attention further refines the focus to the most important areas within the feature map, such as obstacle boundaries, which help to distinguish objects from their surroundings. CBAM enhances its expressive power by focusing more on subtle but critical features. Together, these improvements enhance the robustness and accuracy of the model under the diverse and dynamic conditions typical of agricultural environments. We will demonstrate the detection capabilities of CBAM in complex agricultural environments through examples in the revised manuscript (lines 686-692).

 

Point 8: Can you discuss how this real-time capability can impact the deployment of unmanned agricultural machinery in the field.！！！

 

Response 8: Thank you for emphasizing the importance of discussing the real-time capabilities of the proposed system. In the revised draft, we will include a dedicated subsection to explain how real-time performance significantly affects the deployment of unmanned agricultural machinery in the field. Highlights of our discussion will include the fact that real-time obstacle detection enables unmanned agricultural machinery to quickly respond to potential hazards, such as moving animals, workers, or unexpected objects. This minimizes the risk of accidents and damage. In agricultural operations, processing delays can disrupt workflow, especially in dynamic environments where obstacles may suddenly appear. Real-time processing ensures that machinery runs smoothly and without unnecessary interruptions. The agricultural sector is inherently unpredictable, including factors such as changes in terrain, weather conditions and moving objects. Real-time detection allows machinery to quickly adapt to these changing conditions, maintaining performance and reliability. Efficient real-time decision-making reduces unnecessary stops, starts and route adjustments, optimizing fuel and energy consumption, which is especially important for large-scale agricultural operations. These benefits highlight the critical role of real-time detection in ensuring the utility, safety, and efficiency of unmanned agricultural machinery. We will integrate this discussion into the revised manuscript (lines 659-675) to emphasize the broader significance of our work.

 

Point 9: Figure 1 appears too complex consider simplifying it by focusing on the most critical components relevant to your paper.

 

Response 9: Thank you for your feedback on the complexity of Figure 1. We understand the importance of clarity and simplicity in effectively communicating key messages. Based on your suggestions, we will: redesign Figure 1 to focus on the most critical sections that are directly related to our research contributions and findings. Remove extraneous details that may distract from the central idea of the chart. Resize text, labels, and graphic elements to ensure that all components are easy to understand. Use a consistent color scheme and layout to improve visual clarity. Any omitted or simplified details will be described in the accompanying text to ensure that the diagram remains concise while still allowing access to the full context. Revised images will enhance the overall readability and coherence of the paper. We appreciate your suggestions, which will improve the quality of the manuscript.

 

Point 10: BiFPN is well-articulated, providing a more detailed explanation of how these components interact within the model could enhance understanding!!

 

Response 10: Thank you for your insightful comments on the need for a more detailed explanation of how model components interact. In response, we will revise the manuscript to provide a more comprehensive description of these interactions, focusing on how CBAM, BiFPN, and WIoU v3 can contribute to improving the functionality of the model. CBAM is integrated into the backbone network directly after the convolutional layer responsible for feature extraction. The channel attention mechanism dynamically emphasizes channels that are highly relevant to the detection task, while the spatial attention mechanism highlights key spatial regions, ensuring that only the most informative features are passed on to subsequent layers. This synergy enhances the detection of challenging objects, especially in cluttered agricultural environments. BiFPN is the neck of the model, connecting the backbone to the head by processing multi-scale feature maps. Features at different levels are fused bi-directionally (top-down and bottom-up) to optimize the contribution of each scale through a learnable weighting mechanism. This interaction enables the model to balance fine-grained details and high-level semantics, thus improving detection across different object sizes and scales. During training, WIoU v3 interacts directly with the predictions by penalizing deviations in bounding box position, scale, and aspect ratio.

It adjusts the loss pattern to encourage the model to generate more accurate and well-aligned bounding boxes. Improved feature extraction (via CBAM) and multi-scale fusion (via BiFPN) combine to ensure that bounding box refinement runs on high-quality features. Together, these components enhance the robustness and accuracy of the model. We will update the manuscript by adding a new section (Section 2.2.5) to elaborate on these interactions.

 

Point 11: How CBAM operates，specifically detailing the channel and spatial attention mechanisms?

 

Response 11: Thank you for your insightful suggestion to elaborate on the workings of CBAM, in particular the channeling and spatial attention mechanisms. In the revised draft, we will explain these processes in more detail to clarify their role in the model. The following is a summary of how each attentional mechanism operates: the channel attentional mechanism is designed to emphasize the most informative feature channels while suppressing less relevant ones. It operates as follows: a global representation of the feature map is generated by applying global average pooling and global maximum pooling in the spatial dimension. These two outputs are passed through a shared fully-connected layer, a network that captures channel dependencies. The outputs of the fully connected layer are combined and activation functions are used to generate channel attention weights. These weights are applied to the original feature map to enhance the importance of the relevant channels. Spatial attention is focused on identifying important regions within the feature map, helping the model to highlight regions associated with obstacles while ignoring background noise. Channel information is aggregated by applying global average pooling and global maximum pooling along the channel axes. The resulting two feature maps are connected and passed through a convolutional layer in order to generate a spatial attention map. The spatial attention map is activated using an S-function and applied to the feature map to enhance important spatial regions. By combining these two mechanisms, CBAM ensures that the model dynamically focuses on the most critical aspects of channels and spatial regions, significantly improving its ability to detect objects in a cluttered and complex agricultural environment. We will explain the process in detail in the paper (lines 369, 382) to enhance clarity.

 

Point 12: Can you discuss how well CBAM performs under varying conditions！

 

Response 12: Thank you for your insightful suggestion to discuss the performance of CBAM under different conditions. In the revised version, we will analyze in detail the performance of CBAM in different environments and operational scenarios to demonstrate its robustness and adaptability. The effectiveness of CBAM was evaluated under various conditions encountered in the agricultural field, including: the often cluttered conditions in agricultural fields, such as overlapping crops, machinery, and natural debris. CBAM's channel attention mechanism prioritizes features associated with obstacles while suppressing background noise. Obstacles in the field (e.g., electric power towers and people) vary widely in scale. CBAM works in concert with BiFPN to fuse multi-scale features to ensure accurate detection across a range of object sizes. Under these varying conditions, CBAM continues to improve detection metrics, including precision, recall, and average accuracy, demonstrating its adaptability to the dynamic and challenging environments typical of the agricultural sector. This detailed discussion will be added to the manuscript (lines 399-412) to address your concerns and your valuable feedback is appreciated.

 

Point 13: Please explanation of how WloU v3 differs from traditional loU-based loss functions would enhance understanding!

 

Response 13: Thank you for pointing out the need for a more detailed explanation of the difference between WIoU v3 and traditional IoU-based loss functions. We agree that providing this clarification will enhance the reader's understanding. Below is a summary of the key differences that will be included in the revised draft. WIoU v3 provides several improvements to address the limitations of traditional IoU-based loss functions, especially in object detection tasks involving agricultural obstacles. While traditional IoU-based loss functions mainly focus on the overlap region between predicted and real frames, WIoU v3 goes one step further by considering not only the overlap region, but also geometric information such as shape, orientation, etc., and dynamically adjusting the weights so that the model focuses on key details during the training process, which improves the accuracy of the model in object localization. This additional note will be included in the revised draft (lines 442-447), summarizing the differences and advantages of WIoU v3. We appreciate your suggestions, which will undoubtedly improve the clarity and depth of the paper.

Point 14: Can you Highlighting the variety in obstacle types (e.g.，people vehicles，power pylons)and how they represent real-world scenarios can strengthen the relevance of your dataset for practical applications!

 

Response 14: Thank you for emphasizing the importance of highlighting the diversity of obstacle types and their relevance to realistic scenarios. We appreciate this suggestion as it is in line with our goal of demonstrating the practical applicability of our dataset and methods. Below is an overview of how we will revise the manuscript to address this point. The dataset used in this study was carefully curated to cover a variety of obstacle types commonly found in agricultural settings. The inclusion of various obstacle types ensures that the model can be effectively generalized to real-world scenarios. The main obstacle categories are as follows:1people: workers or individuals who may inadvertently enter the field during operations. Detection is critical to ensure the safety of unmanned agricultural machinery deployments.2 Vehicles (Tractors and Harvesters): are critical to identify and avoid heavy machinery that is frequently operated in agricultural fields. These obstacles vary in size, shape, and movement patterns, posing a challenge to the detection system.3 Power Infrastructure (Electric power pylons and Telegraph Poles): These static obstacles must be avoided as they can cause serious operational disruption or damage. By integrating these different obstacle types, the dataset: Model the complexity and variability of real agricultural environments, including static, dynamic, large, small and irregularly shaped obstacles.

Enhance the robustness and generalization capabilities of the proposed model to make it more suitable for deployment in various agricultural scenarios. Ensure safety, efficiency and operational effectiveness by addressing the challenges posed by various obstacles. We will add this explanation to the manuscript (lines 500-522) and include examples of obstacle types in the dataset to further demonstrate their practical relevance. This addition will address your concerns by explicitly linking the diversity of obstacle types to the practical applicability of the dataset, thereby strengthening the overall contribution of the paper. Thank you for your valuable feedback.

 

Point 15: Ensure that the axes are Clearly labeled，and the units of measurement （if applicable） are indicated.

 

Response 15: Thank you for pointing out the need to more clearly label axes and include units of measurement where applicable. In the revision, we will ensure that all charts (especially performance indicator and assessment charts) meet the correct formatting standards. Below is an overview of the specific steps we will take to address this. Each axis will be clearly labeled to indicate the quantity it represents. Units will be clearly included where relevant. We have reviewed all figures in the manuscript to ensure that the formatting is consistent and professional across the board. This revision brings the presentation in line with standard scientific practice and improves overall clarity for the reader. Thank you again for your valuable feedback, which helps to improve the quality and readability of our manuscript.

 

Point 16: Compared to other studies that may have used smaller findings.

 

Response 16: Thank you for emphasizing the importance of comparing our findings to studies with smaller datasets or limited scope. We recognize that such comparisons can further illuminate the importance and robustness of our methods. In the revised draft (lines 766-780), our study is compared to existing studies, particularly those that use smaller datasets or address a narrower scope. Many previous agricultural obstacle detection studies have used datasets with limited diversity in terms of obstacle types or environmental conditions. For example, the main focus has been on detecting only static obstacles, but lacks dynamic obstacles such as moving vehicles or people. In contrast, our dataset includes a wider range of obstacle types (e.g., people, vehicles, electric power towers), providing a more com-prehensive benchmark for evaluating model performance. Past research has achieved high accuracy on small datasets, but performance degrades significantly when applied to more complex realistic scenarios. Our approach utilizes modules such as CBAM, BiFPN, and WIoU v3, which allow for better generalization under a variety of conditions. Unlike studies that validate models under controlled conditions, our approach is evaluated on field datasets that represent real agricultural environments. This enhances the practical relevance and deployment potential of our model for unmanned machinery. These comparisons highlight the advantages of our proposed approach in dealing with various obstacle types, challenging conditions, and realistic scenarios, making it a more robust and scalable solution for agricultural obstacle detection. This additional discussion will give the reader a clearer understanding of how our work builds on and goes beyond previous studies. Thank you for your valuable feedback, which helps to strengthen the contextual relevance of our findings.

 

Point 17: Please, clearly relate the findings of your experiments to the broader implications for obstacle detection in agricultural settings.

 

Response 17: Thank you for your valuable suggestions to better link our experimental results to broader implications in agricultural settings. We will revise the manuscript to explicitly discuss how our findings can help advance the field of obstacle detection and their potential impact on the deployment and efficiency of unmanned agricultural machinery. Our experimental results provide important insights and practical benefits to the field of agricultural obstacle detection: by achieving high precision and recall over a wide range of obstacles, our model reduces the likelihood of collisions or accidents involving people, agricultural machinery. This directly contributes to improving the safety of unmanned agricultural machinery in real-world deployments. Agricultural environments are inherently dynamic and cluttered, with obstacles varying in size, shape, and motion. Our model is equipped with Attention-Based Mechanisms (CBAM) and Fusion of Multiscale Features (BiFPN) and excels in detecting both static obstacles (e.g., utility poles) and dynamic obstacles (e.g., moving tractors). This adaptability ensures consistent performance across a wide range of field conditions, including changing light and weather. Accurate obstacle detection minimizes the risk of machine damage or operational disruption due to undetected obstacles. This means less downtime and lower maintenance costs, which improves the overall productivity of automated farming systems.The integration of WIoU v3 with the model improves localization accuracy, especially for small and partially obscured objects such as tools or animals. This capability is critical for large-scale agriculture, where timely and accurate detection ensures efficient operations over large areas. Our model has real-time performance that seamlessly integrates with the control system in the autonomous machinery, thus facilitating immediate decision making to avoid obstacles. This real-time capability is critical for deploying unmanned systems in rapidly changing field environments. By providing a reliable detection mechanism, our approach supports the broader goals of precision agriculture, where intelligent systems optimize resources, reduce waste and increase crop yields. These results demonstrate that our proposed approach not only advances obstacle detection technology, but also addresses real-world challenges facing agricultural automation. By bridging the gap between research and applications, our work lays the foundation for safer, more efficient, and more scalable smart agriculture solutions.

This revised manuscript (lines 799-832) will clearly articulate the broader implications of our findings and enhance the relevance of research to real-world applications. Thank you for your suggestions, which have greatly improved the depth of context of our manuscript.

 

Point 18: It would be beneficial to explicitly compare the performance metrics of your CBBI-YOLO model with those reported in recent studies.

 

Response 18: Thank you for suggesting an explicit comparison of the performance metrics of our proposed CBBI-YOLO model with the latest research. Such a comparison will give the reader a clearer picture of the progress made by our approach. We have revised the manuscript (lines 751-765) to include detailed comparisons of key performance metrics with relevant recent studies. We will provide a detailed comparison of our results with similar studies in the literature. This will include a discussion of the advantages of our improved YOLOv8 model in terms of detection accuracy, speed, and robustness, as well as areas where other methods may outperform ours. Specific metrics will be used to highlight comparative performance. This clear comparison shows that CBBI-YOLO outperforms existing methods in terms of accuracy, speed, and adaptability, making it a powerful solution for agricultural obstacle detection. We thank you for your suggestion as this analysis enhances the contextual relevance and impact of our study.

 

Point 19: Any limitations found in the studies you compare with and explain how your approach overcomes!

 

Response 19: Thank you for suggesting that we highlight the limitations of our comparative studies and explain how our proposed CBBI-YOLO model overcomes these challenges. We have revised the manuscript to discuss the limitations of the studies we compared and the advances our approach offers to address these gaps. While previous studies have made significant contributions to obstacle detection in agricultural environments, they exhibit certain limitations that our CBBI-YOLO model aims to overcome. Faster R-CNN relies on standard feature extraction techniques that may not effectively capture complex spatial relationships or fine-grained details. This can lead to poor de-tection performance, especially for small or partially occluded objects. We integrate CBAM (Channel and Spatial Attention Module) to enhance feature focusing and en-sure that key features are emphasized in the detection. BiFPN (Bidirectional Feature Pyramid Network) further improves multi-scale detection, especially for small and ir-regularly shaped obstacles. The traditional IoU-based loss function used in the YOLOv5 model is less effective when dealing with overlapping or elongated objects. This can lead to inaccurate bounding box predictions, especially for obstacles such as utility poles or harvesters. By implementing WIoU v3, we have improved the accuracy of bounding box localization, especially for challenging obstacle shapes and orienta-tions. Some models achieve high accuracy, but this comes at the cost of slower infer-ence times, limiting their applicability in real-time applications. This hinders practical deployment in autonomous agricultural machinery. Our model balances accuracy and speed with an FPS of 52, making it suitable for real-time deployment. By addressing these limitations, the CBBI-YOLO model introduced in this paper is a significant pro-gression in agricultural obstacle detection. It combines a diverse data representation, advanced feature extraction, improved localization and real-time functionality to pro-vide a robust and scalable solution.This discussion will be added to the manuscript (lines 781-798) to give the reader a clear understanding of how our work goes beyond the limitations of previous studies. Thank you for your insightful suggestions, which help to enhance the impact and clarity of our findings.

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors made some modifications, but only explained some issues or provided explanations in the discussion section without making substantive changes. It is recommended to make major revisions or additions.

 The following comments remain unresolved: 4, 5, 6, 7, 11 and 13. Not modifying and supplementing these contents will affect the overall quality of the paper. It is recommended to improve and supplement them.

Author Response

Point 1: The following comments remain unresolved: 4, 5, 6, 7, 11 and 13. Not modifying and supplementing these contents will affect the overall quality of the paper. It is recommended to improve and supplement them.

 

Response 1: Thank you very much for your valuable feedback. We have carefully reviewed the comments and made every possible effort to address them within the scope of our current study. However, due to certain constraints in our experimental environment and resources（Question 4: At the time of conducting this research, YOLOv8 was the most mature and widely adopted version in the YOLO series YOLOv8 provides built-in flexibility for architecture modifications, making it suitable for integrating advanced components such as CBAM, BiFPN, and WIoU v3. The research began before YOLOv9 and YOLOv10 were released or widely adopted. YOLOv8 was the latest stable version available during the initial phases of this study. Transitioning to a newer version mid-project could have introduced additional variables and complexity, potentially delaying the research outcomes. YOLOv8 already demonstrates state-of-the-art performance in terms of detection accuracy, speed, and scalability. The focus of our research was to enhance these capabilities for agricultural environments, not to benchmark the latest YOLO versions; Question 5: Expanding the dataset to include additional obstacle types is indeed an important suggestion. However, our current dataset is limited by the availability of labeled images for these specific obstacle types, as collecting and annotating agricultural field data is both time-consuming and resource-intensive. Moving forward, we plan to collaborate with relevant institutions to enlarge the dataset and include obstacles like graders, rotary tillers, and sprayers to improve model generalization. Question 6: The suggestion to include training for diverse viewing angles and varying conditions is highly valuable. While our current dataset predominantly focuses on certain standard views (e.g., front and side views), incorporating images from multiple angles and under challenging conditions (e.g., occlusion, varying lighting, and distances) would require substantial additional data collection. Unfortunately, due to limitations in field access and seasonal constraints during this study, we were unable to gather such data comprehensively. We will prioritize this aspect in future investigations. Question 7：We have thoroughly examined the issue. The tractor's non-detection is likely due to the combination of long distance and the resulting loss of image resolution, which impacts the model's ability to extract distinguishable features. Currently, we are limited by the resolution of our dataset images and the hardware available for collecting data. Addressing this would require additional high-resolution images and further adjustments to the model architecture to better handle such cases. This enhancement will be a key focus in our future research. Question 11: While we have aimed to provide as much specific data as possible, the inclusion of additional experiments to further reinforce our conclusions is constrained by our current dataset size and computational resources. For example, including a broader range of metrics for all scenarios would require additional field tests and annotation efforts, which exceed the timeframe of this study. Nonetheless, we will incorporate these aspects in subsequent work and expand the dataset and evaluations. Question 13: We have replaced the figures with clearer versions to the best of our ability. However, due to the complexity of some graphical representations and limitations in visualization software, minor overlapping may still occur in highly dense annotations. Further refinements will be considered in future publications as we upgrade our tools for figure preparation.), we were unable to fully resolve some of the issues. We kindly ask for your understanding and assure you that your suggestions will be a priority in our future work. We sincerely hope that this response clarifies our limitations and demonstrates our commitment to improving these aspects in future research. We value your insights and believe they will help us refine our work in subsequent studies. Thank you again for your understanding and constructive feedback.

Author Response File: Author Response.pdf