HR-YOLOv8: A Crop Growth Status Object Detection Method Based on YOLOv8

Round 1

Reviewer 1 Report

The authors presented a YOLOv8-based improved object detection method for crop growth status evaluation. The method includes three improvements: a DCHRA mechanism, HR-FPN inspired by HR-Net, and a combined loss function. The experiment results showed that the performance outperformed the YOLO series, and the number of parameters decreased notably. 

This work's performance improvement is noteworthy, but some drafting and structural problems remain. The specific comments are as follows.

1. There should be a cited reference at the first introduction of the dataset Oilpalmuav.

2. The figure captions, such as Fig. 4 and Fig.5, should have enough explanatory descriptions and at least the algorithm each graph uses.

3. The Fig.4 and Fig.5 is not clear enough. Especially in Fig.5, readers can hardly tell the difference between these four results.

4. The results are only repeated in the conclusion part, and there is no hypothesis or explanatory discussion section of the overall effect. In particular, the authors need to discuss the contribution of each innovation point in the ablation experiment to the overall performance and explain the improvement, especially why HR-FPN can achieve the performance improvement of the task proposed in this paper with such a reduction in the number of parameters. In addition, it is best to conduct a complete permutation and combination of experiments in the ablation experiment. 

5. In Line 302, use "the number of parameters" instead of "Para". "Para" is not a term. 

6. FLOP, FLOPS and FLOPs are different; the authors should pay attention to their usage. 

7. Line 352 has a repeated "number of parameters".

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The article "HR-YOLOv8: A Crop Growth Status Object Detection Method Based on YOLOv8" presents a new approach called HR-YOLOv8 to detect the growth status of crops using object detection methods; the proposed model incorporates a dual self-attention mechanism and a feature fusion module to improve the detection of small targets in crop images obtained from UAVs. The study discusses the limitations of existing methods and highlights the potential of deep learning techniques, particularly the YOLO model family, to automate the identification of crop growth stages.

 The proposed HR-YOLOv8 model represents a design of the feature fusion module (HR-FPN), the dual self-attention mechanism (DCHRA) and the loss function (IS-IoU). Based on the "oilpalmuav” dataset, the improved performance of HR-YOLOv8 was demonstrated compared to the reference models. The results show significant improvements in accuracy, recall and average accuracy, with fewer parameters and comparable or better computational efficiency. To improve the paper here are some remarks:

1. The introduction should be more specific

2. formulate the proposed methodology's specific objectives and goals at the manuscript's beginning. Ensure that readers clearly understand the issue being addressed and the intended contributions

3. The document does not explicitly mention a flowchart illustrating the proposed architecture, which integrates the HR-FPN module, the DCHRA mechanism and the IS-IoU function. A visual representation of the architecture, such as a flowchart or diagram, would have provided a clear and concise overview of how these components are integrated into the HR-YOLOv8 model

4. Although the architecture is evaluated on the oilpalmuav dataset, the generalizability of HR-YOLOv8 to various crop types and environmental conditions is not widely discussed

5. Although architecture is evaluated, the generalizability of HR-YOLOv8 to various types of cultures and classes is not widely discussed

6. The comparison of the architecture with other state-of-the-art algorithms is limited to a few specific models

7. While the architecture improves accuracy and precision, the efficiency and scalability of the model in real-world applications could be further explored. The balance between efficiency and precision, especially regarding several parameters and computational efficiency, is crucial for practical deployment in precision agriculture

8. The study addresses the potential problem of using non-homogeneous classes to test the proposed model to recognize the growth state of crops. The inclusion of classes such as "healthy palms", "poorly managed palms", "smaller palms", and "yellow palms" raises concerns about the comparability of these categories, as they encompass various attributes such as health status, color and size. Propose a more homogeneous categorization; this would allow a more targeted and specific analysis of the model's performance in detecting and classifying a particular attribute without the confusing influence of unrelated attributes

9. The study does not include visualization of outcome curves for measures such as precision, recall and F1 scores, nor comparison with other architectures via visual representations to show the model's behaviour during training

10. The document does not have a specific discussion section, which usually allows the results to be interpreted and contextualized against existing literature and state-of-the-art studies

11. The conclusion section looks more like a summary than a proper conclusion. Please provide a substantial critique of the work and open-ended questions regarding possible research directions

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The paper has been improved