Cotton verticillium wilt grading method based on deep learning

Abstract

This invention discloses a deep learning-based method for classifying cotton Verticillium wilt, which relates to the technical field of cotton pest and disease control. The key technical solution includes the following steps: S1, data collection and labeling; S2, dataset preprocessing; S3, setting evaluation indicators; S4, model construction; which has the advantages of accurate and rapid identification.

Description

Technical Field

[0001] This invention relates to the field of cotton pest and disease control technology, and more specifically, to a deep learning-based method for classifying cotton Verticillium wilt. Background Technology

[0002] Verticillium wilt of cotton is a soil-borne disease. Long-term monoculture in cotton fields leads to the continuous accumulation of pathogens in the soil, causing the disease to worsen year by year and becoming one of the key factors restricting cotton yield. *Verticillium dahliae*, the main fungus causing Verticillium wilt in the field, exhibits high genetic diversity and strong environmental adaptability. *Verticillium dahliae* has a wide host range and strong pathogenicity. Under environmental stress, *Verticillium dahliae* produces microsclerotia to maintain its life form. These microsclerotia have a strong survival ability and can survive in the soil for many years, making the eradication of cotton Verticillium wilt even more difficult. Chemical control methods can only inhibit and prevent Verticillium wilt, which is costly and has limited effectiveness, making the prevention, control, and recurrence of cotton Verticillium wilt difficult.

[0003] Cotton disease resistance breeding, as a key aspect of cotton Verticillium wilt control, has received widespread attention and importance in recent years, with many excellent disease-resistant varieties being developed. However, the time and manpower costs involved in cotton disease resistance breeding are enormous. Therefore, obtaining an efficient method for classifying cotton Verticillium wilt is crucial. Cotton Verticillium wilt is a systemic vascular disease affecting the entire plant, and plant dissection results are considered a relatively accurate method for identifying and evaluating disease resistance. As a key aspect of disease resistance breeding research, the assessment of Verticillium wilt disease severity among different cotton varieties currently relies entirely on visual inspection by professionals, which is time-consuming and somewhat subjective. During the investigation, if a large number of plants are surveyed in the field, the researchers' judgment of the disease severity will vary depending on the number of plants being assessed. The criteria for judging the disease severity may also change in the early and later stages of the investigation, and different researchers may classify the same diseased plant into different disease severity levels.

[0004] Current deep learning-based disease research mainly focuses on disease classification and identification, with limited research on disease segmentation and grading. Furthermore, most studies use models with large parameter counts, high computational complexity, and insufficient ability to identify small lesions and segment boundaries. Data sets for cotton Verticillium wilt stem profile identification are extremely scarce. The color characteristics of cotton Verticillium wilt stem profiles vary across different regions and time periods, and research specifically targeting cotton stem profiles is relatively limited. Research on the grading of Verticillium wilt in cotton stem profiles can efficiently and accurately identify cotton germplasm resources with strong disease resistance, contributing to improved intelligent breeding and providing important references for subsequent variety improvement and breeding work. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a cotton Verticillium wilt grading method based on deep learning, which has the advantages of accurate and rapid identification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for grading cotton Verticillium wilt, comprising the following steps:

[0007] S1. Data Acquisition and Labeling

[0008] We collected phenotypic photos of cotton stem profiles in the field as a dataset and labeled them using five categories: background, cortex, xylem, pith, and lesions.

[0009] S2, Dataset Preprocessing

[0010] The YOLOv8 instance segmentation model was used to identify and segment cotton stalk profiles in images, and the dataset was divided into training and test sets in a 9:1 ratio.

[0011] S3. Setting Evaluation Indicators

[0012] Average pixel precision, average intersection-over-union ratio, average precision, and average recall are used as evaluation metrics to assess the model's coverage of element classification, predictions, and true labels, as well as the accuracy and coverage of lesion pixels.

[0013] S4, Model Building

[0014] The Unet model was improved by replacing the initial encoder of the Unet model with the first four stages of the VGG16 encoder and loading pre-trained parameters. A fusion channel was introduced, and a CBMA attention convolution module was added to train the model and complete the lesion proportion segmentation.

[0015] By adopting the above technical solutions, the YOLOv8 instance segmentation model is used to sequentially segment and arrange the cross-sections of photos taken in the same group, reducing the cropping process and removing irrelevant elements. By setting evaluation metrics, the performance of the model can be effectively evaluated, and information about the model's performance in pixel classification, disease identification, and coverage can be provided. Replacing the encoder structure with a VGG16 multi-layer convolutional module can achieve a larger receptive field and multi-scale feature pyramid generation, while preserving edge texture details and further improving the accuracy of semantic segmentation. By adding a CBAM module, the ability to focus on target regions and capture edge details is enhanced while retaining the powerful feature extraction capabilities of VGG16.

[0016] Furthermore, in step S1, the data annotation adopts a priority annotation method, and the shapes are sorted according to the priority of background < cortex < xylem < medulla < lesion.

[0017] By adopting the above technical solution, the labeling difficulties caused by the overlapping of many label categories are avoided. This allows high-priority labels to cover low-priority labels, intelligently handles multiple label overlaps, and completes the labeling of each tissue in the cotton stalk section in one go. It also effectively avoids labeling errors caused by label overlap and greatly reduces the amount of labeling and labeling time.

[0018] Furthermore, step S2 also includes a post-processing step, which involves performing operations on the cotton stalk profile segmented by the YOLOv8 instance segmentation model, including but not limited to uniform image extraction, sequential sorting, renaming, adding background settings, converting image formats, and unifying image sizes.

[0019] By adopting the above technical solutions, the problem of excessive memory usage can be alleviated.

[0020] Furthermore, step S2 adopts the training method of loading ImageNet pre-trained weights.

[0021] By adopting the above technical solutions, the convergence speed and generalization ability in small sample scenarios can be improved.

[0022] Furthermore, during step S4 training, all parameters of the VGG16 encoder are frozen first, and the CBAM and decoder are trained separately for 10-20 rounds. After the training parameters stabilize, all parameters of the VGG16 encoder are unfrozen.

[0023] By adopting the above technical solution, the conflict caused by the VGG16 encoder skipping the initial training through ImageNet pre-training can be avoided, thus ensuring the stability of the pre-trained features.

[0024] Furthermore, it also includes setting up an application interface, which is developed using the PyTorch deep learning framework.

[0025] By adopting the above technical solution, the complex segmentation model can be encapsulated into a user-friendly interface application. The system can automatically load images through an intuitive window interface and display the distribution of the area of ​​each component, as well as the stem cutting disease level and the corresponding grade standard.

[0026] Furthermore, in step S4, the formula for calculating the proportion of diseased spots on the cotton stalk cross-section is as follows:

[0027]

[0028] In the formula, Sz represents the proportion of lesions, Sd represents the area of ​​lesions in semantic segmentation, and Sx represents the area of ​​xylem in semantic segmentation.

[0029] By adopting the above technical solution, the pixel ratio of each tissue in the cotton stalk section, such as cortex, xylem, pith, and lesions, can be accurately extracted, thereby calculating the area. This replaces the previous method of manually identifying and judging the disease level of stalk sections, greatly reducing the subjectivity and error of human identification. The operation is simple, the process is complete, and it is easy to preserve evidence.

[0030] In summary, the present invention has the following beneficial effects: The deep learning-based cotton Verticillium wilt grading method of the present invention establishes a multi-location, multi-scene cotton Verticillium wilt stalk cutting image dataset, and completes the segmentation and cropping of the same group of photos and the removal of irrelevant elements through the YOLOv8 instance segmentation model, thereby improving the accuracy and efficiency of cotton Verticillium wilt grading; by setting evaluation indicators, it is helpful to evaluate the model's coverage of element classification, prediction and real labels, as well as the accuracy and coverage of lesion pixels; the present invention adopts the Unet model, replacing the encoder structure from a double-layer convolution stack to a VGG16 multi-layer convolution module, to achieve a larger receptive field and multi-scale feature pyramid generation, while preserving edge texture details, further improving the accuracy of semantic segmentation; in addition, the CBAM module is added, which enhances the ability to focus on key features such as cotton stalk cutting lesions and capture edge details while retaining the powerful feature extraction capabilities of VGG16. Attached Figure Description

[0031] Figure 1 This is a training curve diagram of the YOLOv8 instance segmentation model of the present invention;

[0032] Figure 2 This is a schematic diagram of dataset preprocessing according to the present invention; in the figure, (a) the original dataset, (b) the YOLOv8 model architecture, (c) normalization, and (d) the result of preprocessing.

[0033] Figure 3 This is a structural diagram of the VGG16 of the present invention;

[0034] Figure 4 This is the average intersection-union ratio (OCR) and average recall confusion matrix of the segmentation model of this invention compared with other models;

[0035] Figure 5 This is a comparison chart of the actual segmentation effects of the segmentation model of this invention with other models;

[0036] Figure 6 The present invention compares the disease index in different regions, including (a) Huyanghe City and (b) Shihezi City. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the embodiments.

[0038] This invention provides a deep learning-based method for grading cotton Verticillium wilt, comprising the following steps:

[0039] S1. Data Acquisition and Labeling

[0040] We collected phenotypic photos of cotton stem profiles in the field as a dataset and labeled them using five categories: background, cortex, xylem, medulla, and lesions.

[0041] The dataset consists of 396,370 phenotypic photographs of cotton stalk cross-sections taken in 2023 and 2024 at two locations in Awat County and Korla City, Xinjiang Uygur Autonomous Region, and in 2025 at five locations in Awat County, Korla City, Shawan City, Shihezi City, and Huyanghe City, Xinjiang Uygur Autonomous Region. The cotton stalk cross-sections were cut at a 45° angle, 15cm above the ground, using fruiting branches. Ten cross-sections were photographed at a time, using a single background to mask background noise and highlight disease characteristics, thus improving the accuracy of disease segmentation. In practice, a black background was consistently used to ensure consistency and comparability of the photographic results. The cross-sections were laid flat on a black cloth, with the camera lens kept as parallel as possible to the cross-section. After shooting, all images were screened, and images with blurriness or angular deviations were removed.

[0042] Five categories—background, cortex, xylem, medulla, and lesions—were used for labeling. Specifically, Labelme labeling software was used to label the cotton stalk cross-section. The severity of cotton Verticillium wilt disease in stalk cross-sections was graded based on the relative proportion of lesion area. To address the problem of numerous categories requiring labeling, overlapping lesions, and increased labeling difficulty and time, this invention employs a priority labeling strategy: after generating a JSON file, the code iterates through and parses the labeled data, sorting the shapes using a label priority mechanism: background < cortex < xylem < medulla < lesions. This ensures that high-priority labels cover low-priority labels, intelligently handles overlapping labels, and finally converts the data into a two-dimensional mask, generating a PNG format segmentation mask and visualization. This method allows for the labeling of all tissues in the cotton stalk cross-section in a single step, effectively avoiding label overlap errors and significantly reducing the amount of labeling work and time.

[0043] S2, Dataset Preprocessing

[0044] The YOLOv8 instance segmentation model was used to identify and segment cotton stalk profiles in images, and the dataset was divided into training and test sets in a 9:1 ratio.

[0045] To improve detection efficiency, focusing on segmenting and extracting the effective area ratio of lesions, cotton stalk cross-sections were photographed in groups of ten. However, semantic segmentation models often only identify and segment one object, and the colors of the outer skin (brown, green) of different cotton stalks can overlap with the cotton stalk cross-section (cortex: green; lesions: brown, black), leading to misidentification and poor segmentation results. To reduce image cropping time and remove irrelevant elements, the YOLOv8 instance segmentation model was used to identify and segment cotton stalk cross-sections in the images.

[0046] In addition, this invention performs unified extraction, sequential sorting, renaming, background setting, image format conversion, and image size standardization on cotton stalk cross-sections segmented by the YOLOv8 model to alleviate the problem of excessive GPU memory usage.

[0047] S3. Setting Evaluation Indicators

[0048] Average pixel precision, average intersection-over-union ratio, average precision, and average recall are used as evaluation metrics to assess the model's coverage of element classification, predictions and true labels, as well as the accuracy and coverage of lesion pixels.

[0049] This invention classifies different combinations of the true class of an image and the class predicted by the classifier into four cases: True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN). TP + FP + TN + FN = the total number of samples.

[0050] Mean Pixel Accuracy (MPA) is used to evaluate the overall performance of a model in pixel-level classification tasks. It is defined as the ratio between the number of correctly classified pixels in the prediction results and the total number of pixels in the image. The calculation formula is as follows:

[0051]

[0052] In the formula, N represents the number of categories, and Pi represents the predicted value in category i;

[0053] Mean Intersection over Union (MIoU) measures the degree of overlap between the segmentation results predicted by the model and the ground truth labels. The calculation process is as follows: First, the ratio of the intersection to the union of the predicted and ground truth regions (i.e., the IoU for that class) is calculated for each class. Then, the average IoU is taken for all classes. The IoU calculation formula is as follows:

[0054]

[0055] In the formula: TP represents true positives, FP represents false positives, TN represents true negatives, and FN represents false negatives, satisfying the relation: TP + FP + TN + FN = N (N is the total number of samples).

[0056] Mean Precision (MP) measures the accuracy of a model in pixel-level disease detection tasks. Specifically, it is defined as the proportion of samples that are actually positive out of those predicted as positive (i.e., disease pixels). A higher MP indicates fewer false positives where the model misclassifies healthy pixels as disease, and a more reliable prediction. The calculation formula is as follows:

[0057]

[0058] In the formula, N represents the number of categories, and Precisioni represents the precision of category i.

[0059] The formula for calculating single-class precision is as follows:

[0060]

[0061] Mean Recall (MR) is used to evaluate a model's ability to identify and cover real-world defect pixels. Specifically, it is defined as the proportion of all real-world defect pixels correctly predicted by the model. A higher MR indicates fewer defect pixels missed by the model and a more comprehensive detection capability. The calculation formula is as follows:

[0062]

[0063] In the formula, N represents the number of categories, and Recalli represents the recall rate of each category.

[0064] The formula for calculating single-class recall is as follows:

[0065]

[0066] S4, Model Building

[0067] The Unet model was improved by replacing the initial encoder of the Unet model with the first four stages of the VGG16 encoder and loading pre-trained parameters. A fusion channel was introduced, and a CBMA attention convolution module was added to train the model and complete the lesion proportion segmentation.

[0068] The conventional Unet model has a U-shaped overall structure, consisting of an encoder and a decoder. The encoder on the left mainly compresses the required image by downsampling, and the compression path is composed of two consecutive 3×3 convolutions, which can extract multi-scale features of the segmented object. The decoder on the right restores the spatial resolution by upsampling.

[0069] This invention replaces the initial encoder of the Unet model with the first four stages of the VGG16 encoder and loads pre-trained parameters. The encoder structure is replaced by a multi-layer VGG16 convolutional module instead of a stacked two-layer convolution, achieving a larger receptive field and multi-scale feature pyramid generation, while preserving edge texture details and further improving semantic segmentation accuracy. Furthermore, to improve convergence speed and generalization performance in small-sample scenarios, the training method is changed from random initialization to loading ImageNet pre-trained weights. When replacing the initial encoder of the Unet model with the first four stages of the VGG16 encoder, there is a significant difference in the number of output channels between the VGG16 encoder and the Unet decoder. By adjusting the decoder upsampling path, an additional upsampling stage is added to ensure feature image size matching.

[0070] This invention incorporates a CBAM module, which enhances target region focusing and edge detail capture capabilities while retaining the powerful feature extraction capabilities of VGG16. As a new module, CBAM requires initial training from scratch, which conflicts with the VGG16 encoder's skipped initial training via ImageNet pre-training. Simultaneous training of both modules would corrupt the randomly initialized attention parameters in the pre-training process. This invention addresses this by first freezing all parameters of the VGG16 encoder, then training CBAM and the decoder separately for 10-20 epochs. Once the training parameters are stable, all VGG16 encoder parameters are unfrozen to ensure the stability of the pre-trained features.

[0071] The formula for calculating the percentage of diseased spots on the cross-section of cotton stalks is as follows:

[0072]

[0073] In the formula, Sz represents the proportion of lesions, Sd represents the area of ​​lesions in semantic segmentation, and Sx represents the area of ​​xylem in semantic segmentation.

[0074] In addition, the system includes steps for setting up the application interface, which is developed using the PyTorch deep learning framework. The system can automatically load images through an intuitive window interface and display the distribution of the area of ​​each component, as well as the stem cutting disease level and its corresponding grading standard.

[0075] Model environment setup

[0076] This development was based on the PyTorch deep learning framework, using Python 3.9 and PyCharm Community Edition 2022.2.3 IDE, running Windows 11, with an AMD Ryzen 7 7840H CPU with Radeon 780M Graphics (3.80 GHz), an NVIDIA GeForce RTX 4060 GPU, and 32 GB of RAM. The model training optimizer used was the Adaptive Moment Estimation (Adam) algorithm, with an initial learning rate of 0.0001 (adaptively adjusted according to batch size, remaining at 0.0001 after adjustment), weight decay of 0, a momentum parameter of 0.9, and a cosine annealing strategy for learning rate scheduling. The minimum learning rate was 0.01 times the initial learning rate, and the training lasted for 100 epochs.

[0077] ablation experiment

[0078] This invention replaces the Unet encoder with VGG16 as the main component and introduces a CBAM module that integrates channels and spatial attention to enhance model performance and the ability to focus on key features such as cotton stalk lesions. It also reduces the interference of background and noise from outdoor photography on the final model segmentation results, thereby improving the overall accuracy of the model and the quality of segmentation.

[0079] Table 1 Comparison of ablation test results

[0080]

[0081] Note: Unet1 is the original Unet model, Unet2 is a model that replaces the Unet1 encoder with VGG16, and Unet3 is a model that adds the CBAM attention module to Unet2.

[0082] Table 1 shows that the Unet model with VGG16 as the main encoder, replacing the Unet encoder in this invention, improved mIou by 0.84 points, mPA by 0.74 points, mRecal by 0.23 points, and mPrecision by 0.74 points. The Unet model with the CBAM attention module added to the VGG16-based architecture improved mIou by 1.45 points, mPA by 0.74 points, mRecal by 1.45 points, and mPrecision by 1.07 points. This improvement indicates advancements in spatial consistency and boundary accuracy.

[0083] Comparison of different models

[0084] The Unet-CBAM model of this invention is compared with classic deep learning networks Deeplabv3+, PSPnet, and Hrnet. Quantitative performance metrics are shown in Table 2. Figure 4 The segmentation results further illustrate the subtle differences between these algorithms, and are presented as percentages of segmentation results for each profile ( Figure 5 The differences in the segmentation results of each model in terms of the detailed outlines of each component of the profile: background, cortex, xylem, medulla, lesions, etc. are highlighted.

[0085] Table 2 Evaluation of segmentation performance of different models

[0086]

[0087] from Figure 4 As can be seen, the Unet model of this invention significantly outperforms other models in mIoU and mRecall for background, cortex, xylem, medulla, and lesions, and exhibits balanced performance across all metrics. In particular, it achieves 93.79% in MPrecision, demonstrating its strong ability to accurately predict positive samples. Its mIoU and mRecall reach 87.69% and 93.67% respectively, slightly higher than the other three models' 91.71% and 95.24%, indicating that the Unet model of this invention outperforms the other three models in overall global pixel recognition performance.

[0088] from Figure 5 The segmentation results visually demonstrate the segmentation effects of different models, providing clear visual data support. The proportions of each component reflect the relative proportions of the cotton stem surface cortex, xylem, pith, and lesions. Comparative analysis shows that the improved Unet model of this invention outperforms other models in terms of the detail and reconstruction of disease edge contours. Specifically, the improved Unet model has advantages in the accuracy of depicting edge contours and the reconstruction of details, capable of finely depicting clear boundaries and revealing subtle textures and patterns within the diseased area, which is crucial for early disease detection and accurate classification. When faced with complex backgrounds, the improved Unet model exhibits better robustness, significantly reducing missegmentation and missed segmentation, and enhancing the model's segmentation accuracy and reliability in distinguishing between diseases and background noise.

[0089] from Figure 5 Comparing the lesion segmentation results of the stem profile of cotton Verticillium wilt at level 0 (N), it can be seen that the Unet model is more accurate in segmenting the edges of the pith and xylem, which is conducive to obtaining a more accurate lesion proportion. Compared with the lesion segmentation results of the stem profile of cotton Verticillium wilt at level 1, it can be seen that Unet is more precise in segmenting the shallow area of ​​the lesion, and the identified lesion area is relatively more complete.

[0090] Verification and Results

[0091] A survey of cotton Verticillium wilt disease was conducted in two locations in Shihezi and Huyanghe cities, Xinjiang Uygur Autonomous Region, in 2025. Leaf phenotypic severity was identified during the peak of the disease outbreak, and stem cross-section model identification was performed later. A five-level classification system was used for phenotypic severity identification in the experimental sites. Leaf phenotypic severity was used as the actual value, while stem cross-section severity identified by the method of this invention and manually identified stem cross-section severity were used as predicted values ​​for fitting analysis and frequency distribution. Figure 6 As shown, the results indicate that the deep learning-based cotton Verticillium wilt grading method of this invention is significantly consistent with the leaf phenotypic disease level identification results and the manual identification results. The stalk splitting disease level identified by the deep learning model is more accurate than that identified manually. The frequency distribution map shows that the cotton stalk splitting disease index is higher than that of the cotton leaf phenotypic disease level, which is consistent with the reality that cotton is more severely affected by Verticillium wilt over time.

[0092] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for grading cotton Verticillium wilt, characterized in that, Includes the following steps: S1. Data Acquisition and Labeling We collected phenotypic photos of cotton stem profiles in the field as a dataset and labeled them using five categories: background, cortex, xylem, pith, and lesions. S2, Dataset Preprocessing The YOLOv8 instance segmentation model was used to identify and segment cotton stalk profiles in images, and the dataset was divided into training and test sets in a 9:1 ratio. S3. Setting Evaluation Indicators Average pixel precision, average intersection-over-union ratio, average precision, and average recall are used as evaluation metrics to assess the model's coverage of element classification, predictions, and true labels, as well as the accuracy and coverage of lesion pixels. S4, Model Building The Unet model was improved by replacing the initial encoder of the Unet model with the first four stages of the VGG16 encoder and loading pre-trained parameters. A fusion channel was introduced, and a CBMA attention convolution module was added to train the model and complete the lesion proportion segmentation.

2. The deep learning-based cotton Verticillium wilt grading method according to claim 1, characterized in that, In step S1, the data annotation adopts a priority annotation method, and the shapes are sorted according to the priority of background < cortex < xylem < medulla < lesion.

3. The deep learning-based cotton Verticillium wilt grading method according to claim 1, characterized in that, Step S2 also includes a post-processing step, which involves performing operations on the cotton stalk profile segmented by the YOLOv8 instance segmentation model, including but not limited to unified image extraction, sequential sorting, renaming, adding background settings, converting image formats, and unifying image sizes.

4. The deep learning-based cotton Verticillium wilt grading method according to claim 3, characterized in that, Step S2 adopts the training method of loading ImageNet pre-trained weights.

5. The deep learning-based cotton Verticillium wilt grading method according to claim 1, characterized in that, During step S4 training, first freeze all parameters of the VGG16 encoder, and train the CBAM and decoder separately for 10-20 rounds. After the training parameters stabilize, unfreeze all parameters of the VGG16 encoder.

6. The deep learning-based method for grading cotton Verticillium wilt according to any one of claims 1-5, characterized in that, In step S4, the formula for calculating the percentage of diseased spots on the cotton stalk cross-section is as follows: ； In the formula, Sz represents the proportion of lesions, Sd represents the area of ​​lesions in semantic segmentation, and Sx represents the area of ​​xylem in semantic segmentation.

7. The deep learning-based cotton Verticillium wilt grading method according to claim 6, characterized in that, It also includes setting up an application interface, which is developed using the PyTorch deep learning framework.

Images