Research on Small-Target Detection of Flax Pests and Diseases in Natural Environment by Integrating Similarity-Aware Activation Module and Bidirectional Feature Pyramid Network Module Features

Comments 1: Make sure of using terms such as "tiny targets" versus "small targets" throughout the manuscript to avoid confusion.

Response 1: Thank you for pointing this out. I agree with this suggestion. After being reminded by the expert, I looked into the differences between the two terms and changed all "tiny targets" to "small targets" in the paper. "Tiny" emphasizes being extremely small, almost to the point of being barely perceptible or difficult to handle. "Tiny targets" are usually used to describe objects that are minuscule in size and pose significant challenges in detection, identification, or processing. In the field of object detection, it often refers to objects with very few pixels and indistinct visual features, such as cells. On the other hand, "small" is a more general term for "little". "Small targets" generally refer to objects that are slightly larger in size compared to "tiny targets". They are small but not to the extreme of being barely perceptible or difficult to handle. In the context of object detection, it has a broader scope, describing smaller objects that are relatively easier to identify and process. Therefore, for the description of flax pests and diseases, "small targets" is more appropriate.

Comments 2: Figure 6, some texts are in Chinese language. Please revise.

Response 2: After being reminded by the expert, I have revised Figure 6 by changing the Chinese text into English. (Page 17,Line668)

Comments 3: The conclusion mentions future research, try to suggest for integrating emerging technologies if possible.

Response 3: Thank you for pointing this out. I agree with this suggestion. I have integrated emerging technologies in the conclusion section, as follows:

With the vigorous development of edge computing technology, we will fully leverage its advantages by migrating some computing tasks from the cloud to edge devices closer to the data source or users, such as edge servers or smart gateways set up in the fields. This enables the YOLOv8 model to operate more efficiently on these devices, reducing data transmission latency and further enhancing the timeliness and accuracy of real - time monitoring. Meanwhile, we will train and test the model using flax pest and disease datasets from different regions to verify its universality and applicability. In addition, we will explore the combination of the model with the emerging hyperspectral imaging technology for pest and disease detection tasks of other crops. Hyperspectral imaging can capture more abundant spectral information of crops, providing multi - dimensional data input for the model, thus offering more powerful technical support for intelligent and precise agricultural management.

Comments 4: The includes precision, recall, and mAP etc. Is it possible to further analysis on processing time per image or frames per second (FPS), as these are critical for real-time applications?

Response 4: Thank you for pointing this out. I agree with this suggestion. I have added the processing time for analyzing each image in Table 2. The details are as follows:

Comments 1: We recommend rewriting the introduction, the paragraphs used are too long, this makes it extremely difficult to read and interpret.

Response 1: Thank you for pointing this out. I agree with this suggestion. After being reminded by the expert, I have made some cuts to the introduction and shortened the length of paragraphs. I hope it will be clearer for the expert to read this time.

Comments 2: Include in the introduction relevant information regarding the detection of pests using current techniques, this can help visualize the current problem.

Response 2: After being reminded by the expert, I have added relevant information about the detection of pests using the current technology (YOLOv8) in the introduction.(Page3,Line82-88, Page 4,Line122-127 and Line 138-143)

Comments 3: Figure 1 requires precision in its description, it would be recommended that you use letters a, b, c... to identify and describe some special characteristic of each of the photographs.

Response 3: Thank you for pointing this out. I agree with this suggestion. I have added letters a, b, c, and d to Figure 1, and provided some descriptions for each photo. The details are as follows:

(a) Yellow leaf disease: The flax is affected by yellow leaf disease. Its most prominent pathological feature is the yellowing of leaves. This yellowing may start from the leaf margins or local areas and gradually spread to the entire leaf. At the plant level, the growth of diseased flax is inhibited. The plant height may be lower than that of normal plants, with fewer branches, and the overall appearance is rather weak. Moreover, as the disease worsens, the yield and quality of flax will be severely damaged.

(b) Wilt disease: It can occur from the seedling stage to the adult plant stage. In the seedling stage, the young stems wilt and fall over, and the leaves turn yellow and wither. When the stem is diseased, it shows a gray - brown or brown color, shrinks inward, rots at the base, wilts, and falls over and dies. The roots turn gray - brown and sometimes constrict. In adult plants, the diseased plants are short and yellowed, wilt from the top down, the roots are damaged and turn brown. When the diseased stem is cut open, the vascular bundles turn brown, and in severe cases, the whole plant wilts and dies.

(c) Cotton bollworm: The cotton bollworm is a distinctive pest. It is 30 - 42 millimeters in length. Its head is yellow and covered with brown reticular patterns. There are 12 tubercles on its body segments, with long setae. The crochets of its abdominal feet are in a double - row median band. Its body color is variable, such as light green, green, yellow - white, light red, etc., and can change according to the environment of the host plant, which serves as an excellent camouflage.

(d) Colasposoma dauricum: The adult Colasposoma dauricum is about 6 millimeters in length and about 3 millimeters in width, with an oval - shaped body. Its antennae are filiform, slightly longer than half of its body length. The end segments are thicker and slightly flattened. Its compound eyes are prominently protruding.(Page 5,Line210-233)

Comments 4: Figure 4. BiFPN feature network architecture, it would be recommended to include an explanatory legend.

Response 4: Thank you for pointing this out. I agree with this suggestion. I have added an explanation of the BiFPN feature network architecture in Figure 4, as follows:

To overcome the limitations of the PANet structure, this study introduces the BiFPN structure to improve the neck of the original YOLOv8n. The BiFPN structure constructs a more flexible and efficient feature fusion framework. It performs feature stacking in channels, which is similar to integrating feature information from different perspectives in the same dimension. This enables the convergence of features such as pest and disease textures, plant morphologies, and leaf veins contained in each channel, enhancing the richness of features. During the feature fusion process, weight information is taken into account. That is, through a carefully designed weight - allocation strategy, according to the criticality of feature maps for flax pest and disease detection, feature maps in areas with obvious pest and disease features are assigned higher weights, while those in background areas have relatively lower weights. This highlights the information of great value for detection, enabling the model to focus on key features and improving detection accuracy. Moreover, it realizes two - way cross - scale feature fusion. The two - way cross - scale characteristic breaks the limitation of traditional feature fusion, which only occurs between adjacent scales. For example, when processing feature maps with dual - input paths, if these feature maps have the same size, an additional path is introduced from the features of the backbone network to achieve fusion with the feature maps in the PAN (Feature Pyramid Network), as shown in Figure 4. This fusion method allows leap - frogging information interaction between feature maps of different scales. For instance, the information of tiny pest and disease targets in low - level detailed feature maps can be quickly integrated with the information about the overall morphology of pests and diseases in high - level semantic feature maps. This enables the model to understand both local details and overall concepts simultaneously, greatly improving the model's recognition sensitivity and accuracy for pest and disease targets of different scales in flax images. It effectively solves the problem of difficult multi - scale target detection in flax pest and disease detection, providing strong technical support for accurate pest and disease monitoring and control. (Page 6,Line367-393)

Comments 5: Figure 5. EIoU Diagram, it is very low resolution, there is no adequate explanation in the text, it should improve.

Response 5: Thank you for pointing this out. I agree with this suggestion. I have increased the resolution of Figure 5 EIOU and provided a more comprehensive explanation, as follows:

First, EIOU conducts a more detailed decomposition and optimization of the loss function calculation. Although CIOU takes into account the aspect ratio, in complex situations such as the irregular deformation of flax leaves caused by pests and diseases, its effect on bounding box regression constraints is not ideal. EIOU calculates and optimizes the loss function by dividing it into overlap loss, center distance loss, and aspect ratio loss, which is conducive to accurately adjusting each dimension of the bounding box during training. Secondly, affected by shooting angles, plant growth postures, etc., the aspect ratio of targets in flax pest and disease images varies greatly. The aspect ratio loss calculation method of CIOU is insufficiently adaptable to this. By introducing a new aspect ratio penalty term, EIOU can better adapt to the diversity of target aspect ratios, reduce detection errors caused by inaccurate aspect ratios, and improve the positioning accuracy of pest and disease targets. (Page 10,Line428-439)

Comments 6: In the discussion of results, it is required to use references; it must be discussed by comparing the results obtained.

Response 6: After being reminded by the expert, I have cited relevant references in the results discussion section to support the obtained findings.

Comments 1: In terms of structure, sections 1 and 2 are clear (Introduction and Materials and Methods). Which section best identifies the results? Which section represents the Discussions? The latter section should contain a minimum of comparison with similar works. Since line 344 there are no more references.

Response 1: First of all, I would like to express my gratitude to the experts for their affirmation and suggestions. Prompted by the experts, I have provided more detailed explanations for both the tables and figures. Secondly, as the introduction part is crucial for presenting the identification results, I have added the findings of this study at the end of the introduction. Meanwhile, I have also included a minimal comparison and discussion with similar works in the introduction. Finally, regarding the issue of "no more citations starting from line 344", I have cited relevant references in the results discussion section to support the obtained results.

The detailed modifications in the PDF are as follows.

Comments 2: Keywords: “BiFPN”…”SimAM”…”Small Target Detection”…are already present in the title. Change them.

Response 2: After being reminded by the expert, I have modified the keywords. The revised keywords are as follows: Pest and Disease Detection; YOLOv8; Data Augmentation; Feature Integration; Attention Mechanism; EIOU Loss Function

Comments 3: Line 136: YOLOv8…Delete space

Response 3: After being reminded by the expert, I have removed that space. (Page 3,Line155)

Comments 4: Figure 1: Reduce font size.

Response 4: After being reminded by the expert, I have modified the font size of Figure 1. (Page 4,Line209)

Comments 5: Figure 2: Reduce font size.

Response 5: After being reminded by the expert, I have modified the font size of Figure 2. (Page 7,Line283)

Comments 6: Line 206: Improve caption.

Response 6: After being reminded by the expert, I have improved the caption (Page 7,Line283)

Comments 7: Line 254: is shown in Eq. (1). Modify throughout the text.

Response 7: After being reminded by the expert, I have changed the representation of all formulas in the full text to the form of Eq. ().

Comments 8: Lines 256-270: Is font size different?

Response 8: After being reminded by the expert, I have modified the font size of this part.

Comments 9: Line 292: Improve caption.

Response 9: After being reminded by the expert, I have improved the caption (Page 10,Line404)

Comments 10: Figure 5: See the previous comment.

Response 10: After being reminded by the expert, I have modified the font size of Figure 2. (Page 12,Line451)

Comments 11: Line 410: Table 1…baseline models”. What are they?

Response 11: Thank you for pointing this out. I agree with this suggestion. I have explained the "baseline models" in Table 1 as follows:

The baseline models of YOLOv8 refer to the original YOLOv8 model architecture without specific improvements or optimizations. Among them, YOLOv8n is a lightweight model with high speed but relatively low accuracy, suitable for scenarios with limited resources and high real - time requirements. YOLOv8s balances accuracy and speed, making it suitable for conventional detection tasks. YOLOv8m has a moderate scale and complexity, with higher accuracy than YOLOv8s but slightly slower speed, and is applicable to scenarios with certain accuracy requirements. The YOLOv8l model is large - sized, featuring high accuracy but slow speed, and is used in scenarios where high accuracy is required but real - time performance is not overly stringent. YOLOv8x is the largest, boasting the highest accuracy and the slowest speed, and is employed in scenarios with extremely high accuracy demands. (Page 14,Line532-543)

Comments 12: Line 436: Table 2…mainstream models”. What are they? 

Response 12: Thank you for pointing this out. I agree with this suggestion. I have explained the " mainstream models" in Table 2 as follows:

Faster R - CNN consists of RPN (Region Proposal Network) and Fast R - CNN. It first ex-tracts features, then RPN generates candidate regions, which are further processed by Fast R - CNN. It has high accuracy and is used for high - precision detection, but its speed is slow.SSD (Single Shot MultiBox Detector) is based on a convolutional network. It uses multi - scale feature maps and default boxes for detection, balancing speed and accuracy. It is suitable for scenarios with real - time requirements, yet it has limitations in handling specific targets. YOLOv5 has components such as the input end. It divides the grid for training and outputs the category and location. It is fast, has good accuracy, is widely ap-plied, and is flexible.YOLOv8n is a lightweight version of YOLOv8. It makes predictions based on grid - division, is suitable for resource - constrained devices and simple scenari-os, but its accuracy is inferior to that of larger models. (Page 14,Line560-571)

Comments 13: Line 443: Table 3…attention mechanisms”. What are they?

Response 13: Thank you for pointing this out. I agree with this suggestion. I have explained the " attention mechanisms " in Table 3 as follows:

SA is the Spatial Attention mechanism. It focuses on weighting the spatial location infor-mation of an image, helping to capture the location and distribution features of the target. CBAM integrates channel and spatial attention. It weights the channels first and then the spatial dimensions, comprehensively exploring features to improve detection accuracy. ECA is the Efficient Channel Attention mechanism. It efficiently calculates channel dependencies, highlights the features of key channels, and controls the computational cost. SimAM is a simple parameter - free module. It automatically learns the weights of neurons, precisely focusing on feature information to facilitate detection. (Page 15,Line587-595)

Comments 14: Line 463: “…Table 4.”

Response 14: After being reminded by the expert, I have changed "table4" to "Table4". (Page 16,Line618)

Comments 15: Line 464: Table 4…Improve description.

Response 15: Thank you for pointing this out. I agree with this suggestion. I have improved the description of Table 4, as follows:

For pest and disease detection, the addition of the SimAM attention mechanism en-hances the model's ability to recognize small insects and focus on leaf lesions. As a result, the precision of the model increases by 1.1%. The BiFPN module optimizes the fusion of lesion features at different scales, improving the accurate identification of pests of different sizes, and the precision of the model is increased by 2.9%. The EIoU loss function im-proves the accuracy of lesion location and the precision of pest location, enhancing the universality and accuracy of identification, and the recall of the model increases by 1.6%. The introduction of the four - detection - head further improves the overall performance of the model. Especially when detecting multiple diseases and pests, by increasing the number of detection heads, the model can more effectively capture and distinguish differ-ent targets, and the recall of the model is increased by 1.7%. The results of these ablation experiments fully verify the effectiveness of each module in improving the performance of the YOLOv8n model for flax pest and disease detection.When these four modules are in-tegrated, the model reaches the highest value in various performance indicators. Com-pared with the original model, the accuracy (P%) of the model is increased by 2.0%, the recall (R%) is increased by 2.7%, and the mean average precision (mAP%) is increased by 2.1%. This further demonstrates the synergistic effect of these modules in improving the performance of the model for flax pest and disease detection. This not only improves the detection efficiency but also optimizes the implementation of prevention and control measures, which is a great benefit to agricultural producers and helps to reduce losses caused by pests and diseases. (Page 16,Line620-640)

Comments 16: Line 520: Improve description.

Response 16: Thank you for pointing this out. I agree with this suggestion. I have improved the description of Figure 8, as follows:

The recognition results of the first and second groups of images in the vertical direction show that the original model has false detections of flax pests and diseases, while the im-proved model with the SimAM attention mechanism is more accurate in detection. Com-paring the third and fourth groups of images, the original model has missed detections and misdetections. After adding the BiFPN feature fusion mechanism and the EIOU loss function, the improved model has a higher accuracy in identifying pests and diseases under complex backgrounds. Observing the last group of images, the original model has inaccurate detections and false detections for small and medium - sized targets. The im-proved model with four detection heads is more sensitive to the detection of small and medium - sized targets, effectively improving the overall detection effect. (Page 19,Line692-708)

Comments 17: Line 522: PyQt5 ? What is it?

Response 17: Thank you for pointing this out. I agree with this suggestion. I have added an explanation of PyQt5, as follows:

PyQt5 is a GUI (Graphical User Interface) programming framework for the Python language. Developed based on the Qt library, it combines the powerful functionality of Qt with the simplicity and ease - of - use of Python. With PyQt5, developers can easily create various interactive application interfaces. It comes with a rich component library, enabling the effortless construction of interface elements such as buttons, text boxes, and image display areas. Moreover, it supports the signal - slot mechanism, which can efficiently handle user interaction events. (Page 20,Line712-718)

Comments 18: Line 530: Improve description.

Response 18: Thank you for pointing this out. I agree with this suggestion. I have improved the description of Figure 9, as follows:

PyQt5 for real-time detection of sesame diseases and pests. The system interface is shown in Figure 9. This system has several important functions. Firstly, it allows users to select a pre - trained pest and disease identification model, providing a precise technical founda-tion for subsequent detection work. Secondly, once the user loads an image into the system and clicks the detection button, the system can quickly and accurately display the cap-tured images of pests and diseases with their positions located. At the same time, it will also present the number of detected targets and the number of frames that can be pro-cessed per second. These data are of great significance for evaluating the detection effi-ciency and effectiveness. (Page 20,Line720-727)

Comments 19: Line 537: “to address…”.

Line 543: “to improve…”.

Line 548: “to further…”.

Line 553: “to improve…”

Response 19: After being reminded by the expert, I have revised the conclusion section in light of the content of this study. Also, I have corrected the formatting issues as per the expert's requirements.

Comments 1: In Figure 1, the authors showed the image of yellow and wilt disease. For "Wilt Disease", it looks like already dead. So the evaluation of disease is necessary, how heavy of the diseases?

Response 1: Thank you for pointing this out. Prompted by the expert, I have consulted the literature and gained a certain understanding of the disease index of wilt disease.

On the one hand, when observing the appearance of plants, apart from checking whether the plants exhibit a wilted state similar to death, we also need to pay attention to the yellowing and withering of leaves. Specifically, we should note the proportion of the yellowed and withered leaf area to the total leaf area of the whole plant. If most of the leaves are severely yellowed and withered, it usually indicates a relatively serious disease condition. At the same time, we should also check whether there are phenomena such as discoloration, atrophy, and necrosis in the stems and branches. The wider the range of the lesions, the more serious the disease is likely to be.

On the other hand, it is also important to analyze from the perspective of physiological indicators. For example, by detecting the water content of plants, plants with severe diseases usually experience severe water loss, and their water content will be significantly lower than that of healthy plants. In addition, analyzing the changes in the content of internal nutrients in plants, such as the content of nitrogen, phosphorus, potassium and other elements, is also helpful. After the plant is infected, the absorption and transportation of these elements may be hindered, and the abnormal content of these elements can also indicate the severity of the disease to a certain extent.

Furthermore, judging the severity of the disease according to the development stage of the disease is also a viable method. If only a small number of leaves on the plant begin to turn yellow and wilt, which are initial symptoms, then the disease is relatively mild. Conversely, if most of the tissues of the whole plant have become necrotic and are even approaching death, it indicates that the disease has reached a very serious stage.

In addition, by counting the proportion of diseased plants in the entire planting area, the severity of the disease can also be reflected indirectly. If only a few scattered plants are diseased, it may mean that the disease is in the local infection stage and the severity is still within a controllable range. However, if a large number of plants show obvious symptoms of verticillium wilt, it indicates that the disease has become relatively serious in this area and corresponding prevention and control measures are urgently needed.

Finally, we think the verticillium wilt (Wilt Disease) shown in Figure 1 presents an appearance of seemingly dead, and it is difficult to directly determine the severity of the disease. It is necessary to comprehensively evaluate the specific severity of verticillium wilt based on multiple indicators and situations. After the improved YOLOv8n in this study identifies a diseased plant, manual in - depth detection and judgment are required. If it is judged to be mild, methods such as applying more organic fertilizers and bio - fertilizers can be used for treatment. If it is judged to be in a severe stage, it can be removed to prevent affecting the growth of healthy plants.

Comments 2: For the "Cotton Bollworm", one is yellow, another one is green. Does it mean there are two varieties or stages of this warm? How about the percentage?

Response 2: Thank you for pointing this out. Prompted by the expert, I've consulted the literature and gained some understanding of the cotton bollworm. The variation in the color of cotton bollworms may stem from individual differences. Factors such as food sources and environmental conditions can have an impact. Consumption of plant leaves with different nutrient components or pigment contents, as well as changes in environmental conditions like temperature, humidity, and light, can all lead to differences in pigment formation and deposition. Although some differences in the growth stage are generally within a certain range and not typically distinct as yellow or green, special external stimuli can cause color changes to deviate from the normal pattern. Gene mutations or genetic diversity can also result in some individuals having special color phenotypes. In Figure 1, all are the larval stages of the cotton bollworm. After verifying the dataset of this study, the ratio of yellow to green cotton bollworms is 4:6.

Comments 3: The "Albumentation" was used in the augmentation. Based on reference 25, it looks like a toolbox to improve the processing speed, which can " keep GPUs utilization during training close to 100%". So the parameters or details of augmentation are not clear.

Response 3: Thank you for pointing this out. I agree with this suggestion. I have added the parameters and details of "Albumentation", as follows:

Albumentations is an efficient image processing library specifically designed for data augmentation. It offers a rich set of image transformation functions, each equipped with multiple practical parameters. For instance, in the random fog transformation, there are parameters like "fog_coef_lower" and "fog_coef_upper". These parameters are used to control the concentration range of the fog. They can add fog to different positions of the image and blur the background. By adjusting these two parameters, one can precisely simulate the image effects under different levels of hazy weather. In the rain transfor-mation, the "rain_type" parameter is used to specify the type of rain, such as "drizzle" (light rain) or "heavy" (heavy rain). Meanwhile, the "rain_amount" parameter can adjust the amount of rainfall, thus simulating various rainy - day images. This enhances the adaptability of the dataset to rainy - day environments and improves the model's detec-tion accuracy in rainy conditions. The random sunlight transformation utilizes the "sun_flare_brightness" parameter in combination with natural light and light - adjust-ment techniques. This parameter can set the brightness of the sun flare, thereby simulating images with random sun flares and enriching the diversity of the dataset. These elaborate image transformations provide the model with more comprehensive and realistic training samples, significantly enhancing its recognition ability in complex environments. These transformation functions can be flexibly combined to meet complex and diverse augmentation requirements. (Page 6,Line247-265)

Comments 4: Section 3, the Yolov8 was improved based on which version? The improvement was not described clearly.

Response 4: Thank you for pointing this out. After being reminded by the expert, I have described in Section 3 which version of YOLOv8 was the basis for the improvement, and the description of the improvements follows immediately, in Sub - sections 3.3 - 3.5. The specific modifications are as follows:

The original YOLOv8n has a relatively small number of network parameters and a low computational cost. Therefore, the object of this study is the lightweight and highly adaptable YOLOv8n. The specific experiments are presented in Section 4, specifically 4.3 Comparison of YOLOv8 Baseline Models. (Page 8,Line316-320)

Comments 5: Section 3.2 and Section 3.3, titles are the same.

Response 5: After being reminded by the expert, I have revised the title of Section 3.3. The titles is “Weighted Bidirectional Feature Pyramid Network “(Page 9,Line354)

Comments 6: Because the precisions, recalls, and mAPs are quite close in this manuscript. The statistic analysis is necessary to prove the improvement if it is significant. I strongly ask you to submit this result (for example, SPSS result files) to check in the future.

Response 6: Thank you for pointing this out. After being reminded by the expert, I have uploaded the result files, namely results.csv before the improvement and results1.csv after the improvement.

Comments 7: Table 4, the "t/s" is ? Testing time?

Response 7: Dear expert, I'm sorry that the expression in this part is not clear. In Table 4, "t/s" indicates that the difference in test time between the improved YOLOv8n model in this study and the baseline model is not significant and is within an appropriate range.

Comments 8: Figure 7. The colormap meaning is necessary.

Response 8: Thank you for pointing this out. I agree with this suggestion. I have provided an explanation of the meaning of the color map in Figure 7, as follows:

A heatmap is a type of chart that presents data information in an intuitive and visual way. It uses variations in color intensity to map the magnitude or density distribution of data values. In a heatmap, darker colors typically represent higher data values or areas with a more dense data distribution, while lighter colors correspond to lower data values or relatively sparser areas. (Page 18,Line675-680)

Comments 9: Figure 8 and the connected contents, I didn't get the points. How to observe "improved model is more sensitive to subtle disease and pest symptoms"?

Response 9: Thank you for pointing this out. I agree with this suggestion. After being reminded by the expert, I have increased the resolution of the images, making the confidence level values visible. Additionally, I have added relevant explanations on how to observe the issue of "the improved model being more sensitive to subtle pest and disease symptoms", as follows:

Figure 8 showcases a comparison of the performance between the original YOLOv8 model and the improved model in sesame disease and pest detection. The recognition re-sults of the first and second groups of images in the vertical direction show that the origi-nal model has false detections of flax pests and diseases, while the improved model with the SimAM attention mechanism is more accurate in detection. Comparing the third and fourth groups of images, the original model has missed detections and misdetections. Af-ter adding the BiFPN feature fusion mechanism and the EIOU loss function, the improved model has a higher accuracy in identifying pests and diseases under complex back-grounds. Observing the last group of images, the original model has inaccurate detections and false detections for small and medium - sized targets. The improved model with four detection heads is more sensitive to the detection of small and medium - sized targets, ef-fectively improving the overall detection effect. (Page 24,Line692-702)

Comments 1: Line256-314 font size

Response 1: Thank you for pointing this out. After being reminded by the expert, I have correctly modified the font size of this paragraph.

Comments 2: Line390,Line394, Line403font size

Response 2: Thank you for pointing this out. After being reminded by the expert, I have correctly modified the font size(Page 13,Line517, Line521, Line530)

Comments 3: Line493 font size

Response 3: Thank you for pointing this out. I agree with this suggestion. I have revised Figure 6 by changing the Chinese text into English.(Page 17,Line668)

Comments 4: Line531:in the work lacks discussion of the results,this must be supplemented.Conclusions should refer to research,are too general.

Response 4: Thank you for pointing this out. After being reminded by the expert, I have revised the conclusion section in light of the content of this study, as follows:

This paper addresses the issue of sesame disease and pest detection by proposing a detection model based on an improved YOLOv8. Through optimization and innovation of the traditional YOLOv8 network model, A dataset of flax pests and diseases considering different environmental conditions and weather conditions has been established.precise detection of sesame diseases and pests in complex environments and at different scales has been achieved. The following summarizes the work and conclusions of this paper:

1) For the task of detecting flax pests and diseases, this study has made improvements to the YOLOv8n model in multiple aspects. An improved BiFPN module was introduced in the feature fusion layer, effectively resolving the interference problem in multi - scale target detection. The SimAM attention mechanism was introduced, enhancing the model's abil-ity to extract features of flax pests and diseases, enabling it to adaptively focus on im-portant features and suppress irrelevant information. The EIOU loss function was adopt-ed to replace CIOU, optimizing the bounding box regression effect and improving the ac-curacy of prediction boxes. A four - detection - head structure was designed, improving the detection effect and speed for different types of flax pests and diseases. The accuracy, re-call rate, and mAP of the improved model are 92.1%, 91.6%, and 94.5% respectively, ena-bling precise identification of flax pests and diseases.

2) When comparing the improved YOLOv8n model with mainstream models such as Faster R - CNN, SSD, YOLOv5, and the original YOLOv8n, our model performs better in terms of accuracy, recall rate, and mAP. This indicates that the improvement measures proposed in this study have significantly enhanced the competitiveness of the model in the field of flax pest and disease detection. It can more effectively cope with complex flax pest and disease detection scenarios, providing more reliable technical support for pest and disease monitoring and prevention in the flax - growing industry.

3) A real - time detection system for flax pests and diseases was designed and developed based on Pyqt5. This system integrates the improved YOLOv8n model and can achieve rapid and convenient detection of flax pests and diseases. The development of this system further expands the application scenarios of the model, helping flax growers to promptly detect and address pest and disease problems, reducing the impact of pests and diseases on the yield and quality of flax. (Page 24,Line736-766)

Comments 5: References editing

Response 5: After being reminded by the expert, I have editing the references.

Comments 1: The resolution of Figure 5 minimizes the effort made in this improved version of the authors. It is recommended to substantially improve this topic.

Response 1: Thank you for pointing this out. After being reminded by the expert, I have significantly increased the resolution of Figure 5. Its file size has been adjusted from the original 2.1MB to 14.1MB, aiming to provide you with a clearer view when you review it. （page12,Line471）

Comments 1: Which section best identifies the results the Discussions? Section 4 maybe ? If the answer is yes the authors could title section 4 “Experimental design, results and discussion”.

- Are there similar works that have addressed the issue and/or used the models (or some of the models) seen in the article, or that have similar methodology ? If the answer is yes, a comparison of results and a minimum of comments are needed. If the answer is no, it should be written down and that would add more value to this paper.

Response 1: First of all, I would like to express my gratitude to the experts for their affirmation and suggestions. I agree with this suggestion. I have named Section 4 "Experimental Design, Results and Discussion" (page 13, Line514). Secondly, I have added the subsequent questions to the introduction part as follows:

Although YOLOv8 performs outstandingly in general object detection, in the complex field environment of flax, its ability to detect the tiny features of pests and diseases (such as early - stage disease spots and subtle pest traces) has limitations. Affected by lighting changes, a cluttered background, and the diverse forms of pests and diseases, the rates of missed detections and false detections increase. Especially, the recall rate for early - stage or less - obvious pests and diseases drops significantly. At the same time, there are a wide variety of flax pests and diseases with subtle features, which poses challenges to the generalization and feature - discrimination capabilities of the model. Currently, although it's not certain that all research has been comprehensively covered, no similar solutions or methods have been found in the field of flax pest and disease detection. Therefore, this research is pioneering in the academic sense, promising to fill the gap, bring an innovative perspective to this technology, help improve the yield and quality of flax, and is of great significance to agricultural production practices and the sustainable development of the agricultural industry. (page 4, Line 161 - 174)

The detailed modifications in the PDF are as follows.

Comments 2: Line 324: “…is shown in Eq. (1):…”. Delete dot after parenthesis.

Line 330: “The calculations of the mean and variance are shown in Eq. (2).and Eq. (3).  …“. Delete dots after

parenthesis.  “The calculations of the mean and variance are shown in Eq. (2) and Eq. (3), …”.

Line 338: “…is shown in Eq. (4).:   “. See the previous comment.

Line 420: is shown in Eq. (6):

Response 2: After being reminded by the expert, I have delete dot after parenthesis.

Comments 3: Line 384: “…Bidirectional Feature Pyramid Network (BiFPN) structure”

Response 3: After being reminded by the expert, I have changed this. (page 10, Line 425)

Comments 4: Line 525: Table 1…baseline models”. What are they? My comment was much simpler than authors thought...I think the authors misinterpreted. In any case, the changes made improved the text. Going back to the requested change (which I still ask for) is: “Comparison of YOLOv8 baseline models: Yolov8n, Yolov8s, Yolov8m, Yolov8l, Yolov8x”.

Line 558: Table 2…mainstream models”. What are they? My comment was much simpler than authors thought...I think the authors misinterpreted. In any case, the changes made improved the text. Going back to the requested change (which I still ask for) is: “…of mainstream models Faster R-CNN, SSD…”.

Line 573: Table 3…attention mechanisms”. What are they? See the previous comment.

Response 4: First of all, I'm truly sorry. I misunderstood your meaning before. After being reminded by the expert, I've made revisions according to your requirements in this version.

Table 1. Comparison of YOLOv8 baseline models: Yolov8n, Yolov8s, Yolov8m, Yolov8l, Yolov8x

Table 2. Comparison of mainstream models : Faster R-CNN, SSD, YOLOv5, YOLOv8n, Improved YOLOv8n

Table 3. Comparison of different attention mechanisms: SA, CBAM, ECA, SimAM

(Page 14,Line565, Page 15,Line598, Page 15,Line614)

Comments 5: Line 593: “…Table_4.” Add space

Response 5: After being reminded by the expert, I have add space. (Page 16,Line634)

Comments 1: About the leaf disease, firstly, I didn't find any information about the yellow leaf disease in sesame. Please give the name of the virus or fungi of this disease in Latin. Secondly, for the wild disease in sesame, the damage level is necessary, the author could follow the standard (DB 41/T 2483-2023), section 5.1.5.2 to evaluate the level of disease.

Response 1: First of all, I'm extremely sorry that in the first - version revision, I didn't mark the revised parts in red text. In the following version, I have marked the revised parts in red. Once again, thank you, expert, for your reminder. I will strictly follow this rule in the future. Secondly, the expert has considered everything very thoroughly! I have re - consulted the materials and found that the yellow - leaf disease of flax is mainly caused by Fusarium oxysporum. Additionally, abiotic factors such as a lack of elements (such as iron, zinc, nitrogen, etc.), soil compaction, poor drainage leading to root hypoxia and rot, or damage from pesticides and fertilizers can also cause the yellowing of flax leaves. Finally, regarding the issue of evaluating the disease severity, I agree with the expert's opinion and have evaluated the disease severity in accordance with the given reference, as follows:

Occurrence Grades of Flax Yellow Leaf Disease:

Grade 0: No yellowing of leaves, and the plant grows healthily.

Grade 1: A small number of lower leaves turn yellow, with the proportion of yellow leaves less than 10%. The plant growth is basically normal.

Grade 2: The lower and some middle leaves turn yellow, with the proportion of yel-low leaves ranging from 11% to 30%. The plant height is 10% - 20% shorter, and the num-ber of branches slightly decreases.

Grade 3: The middle and most of the lower leaves turn yellow, with the proportion of yellow leaves ranging from 31% to 50%. The plant height is 21% - 40% shorter, the number of branches decreases by 20% - 30%, and some plants show mild wilting.

Grade 4: Most of the leaves turn yellow, with the proportion of yellow leaves greater than 51%. The plant height is more than 41% shorter, the number of branches decreases by more than 31%, and many plants show obvious wilting or some die

Occurrence Grades of Flax Wilt Disease:

Grade 0: The plant shows no symptoms, grows normally, and there are no abnor-malities in leaves, stems, and roots.

Grade 1: A few (1 - 2) lower leaves turn slightly yellow without wilting. The plant grows basically normally. The base of the stem slightly changes color, and a small num-ber of fine roots necrotize. The yield loss is about within 5%.

Grade 2: 10% - 25% of the lower leaves turn yellow and show mild wilting. There are light brown lesions at the base of the stem, and some lateral roots necrotize. The growth of the plant is inhibited, and the yield loss is 5% - 15%.

Grade 3: One - third to one - half of the leaves turn yellow, wilt, and some wither. Dark brown lesions at the base of the stem encircle about half of it. Many lateral roots and some main roots necrotize. The growth is obviously hindered, and the yield loss is 15% - 35%.

Grade 4: More than half of the leaves turn yellow, wilt, and a large number wither. The lesions at the base of the stem nearly encircle it. The vascular bundles turn brown, most of the roots necrotize, the growth stops, and the yield loss is over 35%, or even a total crop failure. (Page 5,Line216-262)

Comments 2: The authors answered the rate between yellow and green worms was 4:6, this should be shown in the manuscript. This is not a remote-sensing journal, it is a journal in agronomy. The author should follow the usual way in the agronomy field.

Response 2: Thank you for pointing this out. Prompted by the expert, I have added the proportion of yellow and green cotton bollworms, as follows:

Among them, all the specimens in Figure 1 are in the larval stage of the cotton bollworm. After verifying the dataset of this study, the ratio of yellow to green cotton bollworms is 4:6. (Page 6,Line268-270)

Comments 3: Line 250-273, about the parameters in augmentation. Just as the authors mentioned, there are so many parameters. However, the values of these parameters still are missing. Only the names and descriptions were given. The value is the basic information for repeatability.

Response 3: Thank you for pointing this out. I agree with this suggestion. I have added the corresponding parameters along with their specific values., as follows:

Albumentations is an efficient image processing library specifically designed for data augmentation. It offers a rich set of image transformation functions, with each function equipped with multiple practical parameters. For example, in the random fog transformation, the “fog_coef_lower” parameter is set to 0.1 and the “fog_coef_upper” parameter is set to 0.5. These parameters are used to control the concentration range of the fog, enabling the addition of fog at different positions in the image and blurring the background. By adjusting these two parameters, one can precisely simulate the image effects in different levels of hazy weather. In the rain transformation, the “rain_type” parameter is used to specify the type of rain. Setting it to “drizzle” can simulate a light rain, while setting it to “heavy” can simulate a downpour. At the same time, setting the “rain_amount” parameter to a value between 0.3 and 0.8 can adjust the amount of rainfall, thus simulating various rainy - day images. This enhances the adaptability of the dataset to rainy - day environments and improves the model's detection accuracy in rainy conditions. Additionally, the random sunlight transformation utilizes the “sun_flare_brightness” parameter in combination with natural light and light - adjustment techniques. Setting this parameter to a value between 1.5 and 3.0 can set the brightness of the sun flare, thereby simulating images with random sun flares and enriching the diversity of the dataset. These elaborate image transformations provide the model with more comprehensive and realistic training samples, significantly enhancing its recognition ability in complex environments. These transformation functions can be flexibly combined to meet complex and diverse augmentation requirements. (Page 7,Line288-307)

Comments 4: In Table 2, firstly, the meaning of "t/s" is still missing. I guess it is "predict time" with the unit as "second". Secondly, I asked the author to do a statistical analysis to prove if the improvement is significant", and provide the original data. However, the data provided was training data in Figure 6.

The author should repeat several times data separation, augmentation, training, and validation. Then, the author will get some precision, recall, and others. The t-test should be done in the final step.

Response 4: Regarding the question you raised, I'm very sorry that I didn't clarify the meaning of "t/s" before. Your guess that it represents "prediction time" with the unit of "seconds" is correct. As for the statistical analysis part, I indeed had a misunderstanding before, and I apologize again. The data I provided previously was the training data in Figure 6, not the precision, recall, and other data obtained after multiple processes of data separation, augmentation, training, and validation as you requested. Next, I will repeat the processes of data separation, augmentation, training, and validation multiple times as per your suggestion, and collect the corresponding precision, recall, and other data. Therefore, the results obtained will fluctuate within a certain range. I have already modified the content of Table 4. And after completing these steps, I will conduct a t - test and provide you with the original data to prove whether the improvement is significant. The details are as follows:

Table 4. Results of ablation experiments

(Page 16,Line639)

4.8 Analysis of Significant Differences

In the research on flax pest and disease detection, to determine whether there are significant differences in the impact of the improved YOLOv8n model on the detection results of flax diseases and pests, the Paired t - test was adopted. For the data collected from flax pest and disease experiments, the null hypothesis of the Paired t - test is that there is no difference in the recognition results of the two methods. If N represents the number of tests, (i=1,2,…,N) is the recognition accuracy of the basic model at different iteration times, is the recognition accuracy of the improved model at different iteration times, and = - then the t - statistic is calculated as in Eq. (10):

（The specific formula cannot be displayed here. It is shown in my attachment. Please check the attachment, dear expert.）

In this study, the t - statistics calculated for flax diseases and pests are t_disease = 6.2108 and t_pest = 5.6813 respectively. When N =12 and the significance level α= 0.05 , the critical value is 2.2011. For both flax diseases and pests, since their t - statistics, 6.2108 and 5.6813 are both greater than the critical value of 2.2011, the null hypothesis is rejected. This indicates that in the detection of flax pests and diseases, there is a significant difference between the improved model and the basic model. That is, the improved model can significantly enhance the detection effect of flax pests and diseases.

 (Page 19,Line715-731)

Comments 5: Based on the authors' provided data, figure 6 is after smoothing. Please give the smooth parameter. It's possible to show in the title.

Response 5: I'm very sorry for not providing the "smooth" parameter before. I have now supplemented the specific value of the "smooth" parameter, which is shown in the title. The specific modification is as follows:

(a) mAP@0.5/%comparision curve (smooth=0.1)     

(b) Loss comparision curve (smooth=0.1)

 (Page 17,Line668)

Comments 6: In Figure 7, the author can not suppose that all readers understand the color meaning, especially this is an agronomy journal.

Response 6: Thank you for pointing this out. After being reminded by the expert, I have provided an introduction to the colors of the heatmap. The heatmap in this study uses the rainbow color scheme, as follows:

Regarding the use of colors in heatmaps, there are several common color - coding schemes. Take the rainbow color scheme for example. It gradually transitions from cool colors (such as blue) to warm colors (such as red). The blue end represents lower data values or sparse data distribution areas. As the color progresses towards red, the data values gradually increase or the data distribution becomes denser [42]. Another example is the Red - Green color scheme. Red is used to mark higher data values or dense distribution areas, while green corresponds to lower data values or sparse areas. Sometimes, based on the Red - Green scheme, yellow is added as a transition to form the Red - Yellow - Green (RYG) color scheme. Yellow lies between red and green, further refining the levels of data values or distribution densities. Through these different color - mapping mechanisms, users can quickly identify patterns, trends, and outliers in the data without having to analyze each value one by one, thus rapidly obtaining key information from large amounts of data. The heatmap in this study uses the rainbow color scheme. (Page 17,Line677-689)