A method for calculating wheat emergence rate and uniformity at seedling stage

Abstract

The application discloses a method for calculating the emergence rate and uniformity of wheat seedlings, comprising the following steps: S1, labeling the outer rectangle of each initial image of wheat seedlings to obtain a target detection dataset; S2, constructing a target detection model based on YOLOv8, replacing the original path aggregation network structure with a bidirectional feature pyramid network and introducing a feature fusion mechanism; performing initial training; S3, performing contour labeling of the wheat seedlings to obtain an instance segmentation dataset; then constructing an instance segmentation model based on YOLOv11, replacing the original architecture with FasterNet architecture; training to obtain the number and relative leaf area of the wheat seedlings; and S4, calculating each uniformity index. The application is efficient and convenient for data collection and processing through a mobile device and an online website; phenotype data of the relative leaf area are proposed according to the image characteristics, and the seedling uniformity is evaluated accordingly, so that the result is more objective; and the calculation efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural technology, and more specifically to a method for calculating wheat emergence rate and seedling uniformity. Background Technology

[0002] As a staple crop for approximately 35% of the global population, wheat's yield stability is of paramount importance. Previous research has shown that wheat emergence rate and uniformity influence yield formation. Emergence rate affects the basic seedling number, which is the foundation for the number of ears per acre, thus impacting wheat yield. Seedling uniformity refers to the spatial distribution and growth uniformity of wheat seedlings. Uneven emergence leads to weak seedlings competing with strong seedlings for water and nutrients, reducing overall yield potential. Therefore, rapid and accurate assessment of emergence rate and seedling uniformity in different wheat varieties is crucial for a deeper understanding of the physiological mechanisms and genetic basis of wheat emergence, providing a foundation for wheat genetic improvement.

[0003] Currently, traditional seedling survey methods mainly rely on manual labor, using visual methods to count wheat seedlings and evaluate uniformity. This approach is inefficient and highly subjective. While the development of high-throughput phenotyping and deep learning technologies has significantly improved the efficiency of wheat phenotyping, existing methods still face challenges when dealing with the extremely small target of wheat seedlings. These challenges include high data acquisition difficulty, poor data accuracy, low identification accuracy, and high false positive and false negative rates, leading to significant deviations in seedling emergence rate calculations and hindering fair and objective evaluation of uniformity in subsequent seedling emergence stages. Furthermore, existing wheat uniformity evaluation systems, with their high requirements for overall spatial distribution and high computational complexity, cannot meet the requirements for rapid and effective high-throughput evaluation of wheat variety seedling uniformity. Summary of the Invention

[0004] To address the aforementioned problems, the purpose of this invention is to propose a method for calculating wheat emergence rate and seedling uniformity. This method allows for data collection and processing via mobile devices and online websites, making it more efficient and convenient. Furthermore, based on the characteristics of image data captured by mobile devices, a phenotypic data point—the relative leaf area of ​​wheat seedlings—is proposed, and this phenotypic data is used to evaluate the uniformity of wheat emergence, yielding more objective results. Two deep learning models are constructed based on the YOLO architecture. After architectural improvements using BiFPN and FasterNet, and through multiple training iterations, a final model capable of automatic and accurate operation is obtained, significantly improving the efficiency and accuracy of wheat emergence rate calculation. Subsequently, an effective uniformity evaluation index is constructed to effectively assess the emergence rate and uniformity of different varieties, providing assistance for wheat variety selection.

[0005] This was achieved through the following technical solutions:

[0006] A method for calculating wheat emergence rate and seedling uniformity includes the following steps:

[0007] S1. Obtain multiple initial images of wheat seedlings at the emergence stage. Using the rectangular bounding box annotation method, annotate each initial image with a bounding rectangle of wheat seedlings to obtain the target detection dataset. From the target detection dataset, divide it into a training set and a validation set.

[0008] S2. Construct an object detection model based on YOLOv8, replacing the original path aggregation network PANet structure in YOLOv8 with a bidirectional feature pyramid network BiFPN, and introducing a feature fusion mechanism into BiFPN. Specifically, when introducing the feature fusion mechanism, firstly, establish bidirectional connections between each path in the top-down and bottom-up directions of BiFPN, forming each bidirectional path; remove each node with only a single input edge in each bidirectional path, and add additional connections between the input and output nodes of each layer to treat each bidirectional path as a feature network layer; repeat the process of adding additional connections multiple times; introduce learnable weights on each feature network layer for weighted fusion of different feature network layers; then, add a detection layer P2 to the Neck of BiFPN to predict the position and size of the bounding box corresponding to each wheat seedling; input the training set from the object detection dataset partitioned in step S1 into the object detection model for initial training to obtain the initial training results.

[0009] S3. Using contour annotation, wheat seedling contours are annotated for each image in the initial training results to obtain an instance segmentation dataset. Then, an instance segmentation model is built based on YOLOv11, replacing the original architecture in YOLOv11 with the FasterNet architecture. The replaced instance segmentation model includes several FasterNet Blocks, and each FasterNetBlock includes a PConv partial convolutional layer and two PWConv pointwise convolutional layers. The instance segmentation dataset is input into the instance segmentation model for training to obtain the number of wheat seedlings and the relative leaf area of ​​each wheat seedling. The number of wheat seedlings is taken as the number of wheat seedlings, and the number of wheat seedlings is divided by the number of wheat seeds sown to obtain the wheat emergence rate.

[0010] S4. Based on the relative leaf area of ​​each wheat seedling, calculate the corresponding feature data, including the standard deviation, coefficient of variation, and entropy value of the relative leaf area; normalize each feature data; based on the normalized feature data, calculate the corresponding uniformity index: Uniformity = 1 / StdDev + 1 / CV + Entropy; where Uniformity is the uniformity index, StdDev is the standard deviation, CV is the coefficient of variation, Entropy is the entropy value, and / is the division symbol; store the wheat emergence rate and each uniformity index as the final result.

[0011] Preferably, in step S1, when acquiring multiple initial images of wheat seedlings, the images of each planting plot in the CK-treated and LN-treated plots are first captured using a mobile device. Each planting plot corresponds to a wheat variety. Each plot corresponds to a folder, and each folder is used to store each result image of the corresponding CK-treated plot or each result image of the corresponding LN-treated plot. Each result image is then transmitted as an initial image to an online data annotation and augmentation website for annotation. By creating image copies with different treatments, the scale and diversity of the dataset are effectively expanded, providing richer data for model training. Furthermore, the use of mobile devices and online websites for data processing reduces the tedious operations of traditional manual processing.

[0012] Preferably, during the annotation of the bounding rectangle for wheat seedlings in step S1, a corresponding txt file is generated for each initial image. Each txt file includes the category of each wheat seedling, the four points of the bounding rectangle corresponding to each wheat seedling, and the length and width information of each corresponding bounding rectangle. The four points of the bounding rectangle corresponding to each wheat seedling and the length and width information of each corresponding bounding rectangle can accurately define the actual position and area occupied by each wheat seedling in the image, facilitating subsequent segmentation.

[0013] Preferably, after dividing the training set and validation set in step S1, the method further includes data augmentation of the training set, which includes rotation, translation, contrast enhancement, and noise addition. Data augmentation can improve the model's generalization ability and robustness.

[0014] Preferably, in step S1, the training set and validation set are divided in an 8:2 ratio; the amount of data in the training set after augmentation is at least three times the amount of data in the training set before augmentation. The 8:2 division ratio ensures sufficient training data, while the data augmentation to more than three times significantly increases the number of training samples, which helps to prevent model overfitting.

[0015] Preferably, in step S2, when replacing the original path aggregation network PANet structure with a bidirectional feature pyramid network BiFPN and introducing a feature fusion mechanism, the method further includes conducting a first ablation experiment using the validation set in the object detection dataset. This is used to perform a first comparative evaluation of the object detection model built on YOLOv8 and the object detection model after introducing the feature fusion mechanism. A second ablation experiment is then conducted, using the training results of the validation set in the object detection model after introducing the feature fusion mechanism as input. This is used to perform a second comparative evaluation of the instance segmentation models before and after the architecture replacement. Through ablation experiments, the effectiveness of the model structure improvement can be scientifically verified, and the accuracy of obtaining the final seedling number and uniformity indicators can be guaranteed.

[0016] Preferably, the evaluation metrics for the first comparative evaluation include precision, recall, average precision, confusion matrix, and average detection time; the evaluation metrics for the second comparative evaluation include average precision, average intersection-over-union ratio, average segmentation time, and model size. This comprehensive evaluation across multiple dimensions, including precision, speed, and time, allows for a holistic assessment of model performance from different perspectives, ensuring the objectivity and practicality of the evaluation results.

[0017] Preferably, during the first or second ablation experiment, if the corresponding comparative evaluation results do not meet the corresponding evaluation indicators, the process returns to step S2 to build a target detection model based on YOLOv8 and iterates until the corresponding evaluation indicators are met. This iterative optimization process ensures that the model performance meets the predetermined standards, thereby guaranteeing the reliability and effectiveness of the final solution.

[0018] Preferably, in step S3, when annotating the wheat seedling contours, a corresponding txt file II is generated for each image in the target detection dataset. Each txt file II includes the category of each wheat seedling and the coordinates of each vertex of the irregular polygon in the contour of each wheat seedling. Using the vertex coordinates of the irregular polygons to annotate the contours can more accurately depict the shape of the wheat seedlings, providing a high-quality data source for the instance segmentation task.

[0019] Preferably, the normalization process in step S4 is a linear normalization process, which linearly normalizes each feature data to the [0,1] interval.

[0020] The beneficial effects of this invention compared to the prior art are:

[0021] The technical solution of this invention allows for data collection and processing using mobile devices and online websites, making it more efficient and convenient. Furthermore, based on the characteristics of image data captured by mobile devices, it proposes the phenotypic data of relative leaf area of ​​wheat seedlings and evaluates the uniformity of wheat seedlings using this phenotypic data, obtaining more objective results. It also constructs two deep learning models based on the YOLO architecture, and after architectural improvements such as BiFPN and FasterNet, obtains a final model that can run automatically and accurately after multiple training iterations, greatly improving the efficiency and accuracy of wheat emergence rate calculation. Subsequently, an effective uniformity evaluation index is constructed to effectively assess the emergence rate and uniformity of different varieties, providing assistance for wheat variety selection. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for calculating wheat emergence rate and seedling uniformity;

[0023] Figure 2 This is a detailed flowchart with examples of a method for calculating wheat emergence rate and seedling uniformity;

[0024] Figure 3 A comparison chart showing the results of an object detection model and an instance segmentation model;

[0025] Figure 4 R represents the difference between the calculated and measured emergence rate. 2 The diagram shows the correlation between the calculated and measured emergence rates, where R² represents the correlation between the calculated and measured emergence rates, and RMSE represents the root mean square error.

[0026] Figure 5 This is a comparison chart of detection results in a real-world scenario;

[0027] Figure 6 This is a comprehensive graph for phenotypic parameter analysis. Detailed Implementation

[0028] The following will be combined with the present invention. Figures 1 to 6 The technical solutions in the embodiments of the present invention will be described in detail below.

[0029] like Figure 1 The diagram shown is a flowchart of a method for calculating wheat emergence rate and seedling uniformity; as shown... Figure 2 The diagram shown is a detailed flowchart with examples of a method for calculating wheat emergence rate and seedling uniformity; combined with... Figure 1 and Figure 2As shown, when calculating the emergence rate and seedling uniformity of wheat, the wheat planting plots are photographed by a mobile device, and then annotated and processed on an online website. By combining two models with improved structures, each wheat seedling is accurately identified, and the emergence rate and uniformity of wheat seedlings are calculated accurately and conveniently.

[0030] The method specifically includes the following steps:

[0031] S1. Obtain multiple initial images of wheat seedlings at the emergence stage. Using the rectangular bounding box annotation method, annotate each initial image with a bounding rectangle of wheat seedlings to obtain the target detection dataset. From the target detection dataset, divide it into a training set and a validation set.

[0032] In this embodiment, when acquiring multiple initial images of wheat seedlings, images of each planting plot in both the CK (conventional) and LN (low-nitrogen) treatment plots are first captured. Each planting plot corresponds to a specific wheat variety. Each plot corresponds to a separate folder, which stores either the result image for the CK treatment plot or the result image for the LN treatment plot. Each result image is then transmitted as an initial image to an online data annotation and augmentation website for annotation. By creating image copies of different treatments, the scale and diversity of the dataset are effectively expanded, providing richer data for model training. Furthermore, utilizing mobile devices and online websites for data processing reduces the tedious operations of traditional manual processing.

[0033] For example, when photographing wheat planting plots of 420 varieties, images can be taken manually using a handheld smartphone. Data collection is done from a uniform angle and height, ensuring each shot fully covers one variety's plot. The dataset is then renamed because image names from mobile devices cannot be directly correlated with varieties. Therefore, a table mapping variety names is created, with at least two columns: one for the numbering starting from 1, and the other for the variety name. Since there are 420 varieties, the numbers range from 1 to 420. Next, the dataset names are corrected. Because two treatments (CK and LN) are used, each repeated three times, there are six plots in total, numbered CK1, CK2, CK3, LN1, LN2, and LN3. The data collected from each plot is stored in folders, and the image names within each folder are changed to correspond to the variety's number according to the collection order. This completes the initial image acquisition.

[0034] In this embodiment, during the annotation of the bounding rectangles for wheat seedlings in step S1, a corresponding txt file is generated for each initial image. Each txt file includes the category of each wheat seedling, the four points of the bounding rectangle corresponding to each wheat seedling, and the length and width information of each corresponding bounding rectangle. Since object detection relies on candidate boxes, the candidate boxes must completely surround the entire wheat seedling during data annotation. The candidate boxes are then matched with the bounding rectangles corresponding to each wheat seedling. In this way, the four endpoints of the bounding rectangles and the length and width information of each corresponding bounding rectangle can be clearly known, thereby accurately defining the actual position and area occupied by each wheat seedling in the image, which is convenient for subsequent segmentation.

[0035] When splitting the training and validation sets, an 8:2 ratio is used. For example, if 1280 original images are used in the object detection dataset, 1024 images will be used for the training set and 256 images for the validation set. Furthermore, data augmentation is performed on the training set, including rotation, translation, contrast enhancement, and noise addition. The augmented training set should be at least three times the size of the unaugmented data. Data augmentation improves the model's generalization ability and robustness; the 8:2 split ensures sufficient training data, and augmentation to more than three times significantly increases the number of training samples, helping to prevent overfitting.

[0036] S2. Construct an object detection model based on YOLOv8, replace the original path aggregation network PANet structure in YOLOv8 with the bidirectional feature pyramid network BiFPN, and introduce a feature fusion mechanism in the bidirectional feature pyramid network BiFPN.

[0037] When introducing the feature fusion mechanism, a bidirectional connection is first established between each path in the top-down and bottom-up bidirectional feature pyramid network BiFPND, forming each bidirectional path. This allows for more complete information flow and fusion between features of different scales. Then, each node with only a single input edge in each bidirectional path is removed, and additional connections are added between the input and output nodes of each layer, treating each bidirectional path as a feature network layer. This process of adding additional connections is repeated multiple times to optimize cross-scale connections between different layers. Subsequently, learnable weights are introduced into each feature network layer for weighted fusion of different feature network layers.

[0038] Then, a detection layer P2 can be added to the neck of the BiFPND network to improve the model's feature extraction and fusion capabilities for small targets. This is suitable for effectively predicting the position and size of the bounding box for each wheat seedling, a small target. The added detection layer retains more detailed information and has a smaller receptive field, making it suitable for capturing local features. The training set from the target detection dataset partitioned in step S1 is then input into the target detection model for initial training to obtain the initial training results.

[0039] In this embodiment, when replacing the original path aggregation network PANet structure with the bidirectional feature pyramid network BiFPN and introducing a feature fusion mechanism in step S2, a first ablation experiment is also conducted using the validation set in the object detection dataset. This experiment serves as a first comparative evaluation of the object detection model built on YOLOv8 and the object detection model after introducing the feature fusion mechanism. Through the ablation experiment, the effectiveness of the model structure improvement can be scientifically verified, and the accuracy of obtaining the final seedling number and uniformity indicators can be guaranteed.

[0040] In the first comparative evaluation, the corresponding evaluation metrics included precision, recall, average precision, confusion matrix, and average detection time. Precision refers to the proportion of samples predicted as wheat seedlings that are actually wheat seedlings; recall refers to the proportion of actual wheat seedling samples that were correctly predicted by the model of this application; the confusion matrix is ​​used to show the comparison between the prediction results and the actual situation. A comprehensive evaluation from multiple dimensions, including precision, speed, and time, can comprehensively measure the model performance from different perspectives, ensuring the objectivity and practicality of the evaluation results.

[0041] Specifically, in the first ablation experiment conducted at the university, with the gradual addition of the improved modules, the convergence speed of the training process accelerated, the oscillations gradually decreased, and no overfitting was observed. This indicates that BiFPN, the Small Object Detection (SOD) layer, and high-resolution image input collectively enhanced the feature extraction and feature fusion capabilities, thereby improving the overall model accuracy and verifying the effectiveness of the improved method. The detection time slightly decreased from 0.156s to 0.152s, and the detection accuracy improved to a maximum of 0.929. Combined with subsequent... Figure 3 The AP (mean accuracy) curve further shows that the model proposed in this application has the smallest oscillation amplitude, the smoothest curve, and no overfitting phenomenon. This provides high-quality input for the second-stage segmentation model and subsequent relative leaf area calculation.

[0042] S3. Using the contour annotation method, wheat seedling contours are annotated for each image in the initial training results to obtain the instance segmentation dataset. Then, an instance segmentation model is built based on YOLOv11, replacing the original architecture in YOLOv11 with the FasterNet architecture. The replaced instance segmentation model includes several FasterNet Blocks. Each FasterNetBlock includes a PConv partial convolutional layer and two PWConv pointwise convolutional layers.

[0043] After the architecture replacement is completed, the training results of the validation set in the object detection model after the feature fusion mechanism are used as input for a second ablation experiment. This second ablation experiment is used to conduct a comparative evaluation of the instance segmentation models before and after the architecture replacement. The evaluation metrics for this second comparative evaluation include average accuracy, average intersection-over-union ratio (IoU), average segmentation time, and model size. Specifically, after introducing FasterNet, the average segmentation time decreased significantly from 0.225s to 0.127s, while maintaining similar accuracy. This fully demonstrates the lightweight advantage of FasterNet and verifies the effectiveness of the improved method.

[0044] Furthermore, if the corresponding comparative evaluation results do not meet the corresponding evaluation indicators during the first or second ablation experiment—for example, if the average segmentation time in the second ablation experiment does not meet the pre-set requirements—the process returns to step S2 to build a target detection model based on YOLOv8 and iterates until the corresponding evaluation indicators are met. This iterative optimization process ensures that the model performance meets the predetermined standards, thereby guaranteeing the reliability and effectiveness of the final solution.

[0045] Then, the instance segmentation dataset is input into the instance segmentation model for training to obtain the number of wheat seedlings and the relative leaf area of ​​each wheat seedling. The number of wheat seedlings is taken as the wheat emergence rate. The number of wheat seeds sown is data known at the time of sowing in each planting plot. Therefore, dividing the wheat emergence rate by the number of wheat seeds sown gives the wheat emergence rate.

[0046] In this embodiment, during the wheat seedling contour annotation in step S3, a corresponding txt file II is generated for each image in the target detection dataset. Each txt file II includes the category of each wheat seedling and the coordinates of each vertex of the irregular polygon in the contour of each wheat seedling. Using the vertex coordinates of the irregular polygon to annotate the contour can more accurately depict the shape of the wheat seedling, providing a high-quality data source for the instance segmentation task.

[0047] like Figure 3The image shows a comparison of the results of an object detection model and an instance segmentation model, and also compares them with other known models. Here, P represents precision, R represents recall, Detect Time represents detection time, AP represents average precision, AP@0.5 represents average precision at an intersection-union ratio of 0.5, mIoU represents average intersection-union ratio, and Segment Time represents segmentation time. First, a comparison of the AP curves of the first-stage detection model is shown (…). Figure 3 (a) Specifically, it shows a comparison of the AP curves of YOLOv5, YOLOv8, YOLOv10, YOLOv11, and the object detection model of this application, demonstrating that the object detection model of this application achieves the best accuracy. Next is a comparison of the AP curves of the two-stage segmentation models ( Figure 3 Section b) also includes comparisons with some known models. Specifically, it shows a comparison of the AP curves of the instance segmentation model in the second stage of this application with the currently known Mask R-CNN segmentation model, YOLOv8 segmentation model, and YOLOv10 segmentation model. It can be seen that the accuracy of the lightweight model in this application is basically the same as that of the unlightweight model. Finally, there is a comprehensive evaluation comparison chart ( Figure 3 (c) The model of this application was compared with the known model again on multiple evaluation indicators, which verified that the model of this application achieved the best detection accuracy and the second best segmentation accuracy, and the detection time was greatly reduced.

[0048] S4. Based on the relative leaf area of ​​each wheat seedling, calculate the corresponding feature data, including the standard deviation, coefficient of variation, and entropy value of the relative leaf area; perform linear normalization on each feature data to the interval [0,1]; based on the linearly normalized feature data, calculate the corresponding uniformity index: Uniformity = 1 / StdDev + 1 / CV + Entropy; where Uniformity is the uniformity index, StdDev is the standard deviation, CV is the coefficient of variation, Entropy is the entropy value, and / is the division symbol; store the wheat emergence rate and each uniformity index as the final result.

[0049] like Figure 4 The figure shows the R-value of a calculated germination rate versus a measured value. 2 The diagram illustrates the correlation between the calculated and measured emergence rates, where R² represents the correlation between the calculated and measured rates, and RMSE represents the root mean square error. Figure 5 The image shown is a comparison of detection results in a real-world scenario. Figure 4 a) and ( Figure 5 a) corresponds to YOLOv5, ( Figure 4 b) and ( Figure 5 b) corresponds to YOLOv8, ( Figure 4 c) and ( Figure 5 c) corresponds to YOLOv10, ( Figure 4 d) and ( Figure 5 d) corresponds to YOLOv11, ( Figure 4 e) and ( Figure 5 e) corresponds to the two-stage model with the improved structure in this application. The red box represents missed detections, and the yellow ellipse represents duplicate detections. After completing the calculation in step S4, the same image data can be used to detect other known conventional models to compare the actual detection results. To verify the effectiveness of the two-stage model with the improved architecture in this method, it was compared with other classic object detection and instance segmentation models. All models were trained and validated on the same training and validation sets, and the detection time (in seconds) and model size (in MB) were also included. Experimental results show that the two-stage model of this method achieves the highest detection accuracy (0.929), the second highest segmentation accuracy (0.976), and the lowest single-image computation time (0.279s). These results indicate that compared with other models, the two-stage model with the improved architecture of this method shows significant advantages in detection and segmentation accuracy, computation speed, and time.

[0050] like Figure 6 The figure shown is a comprehensive graph of phenotypic parameter analysis. This graph was calculated using data from 420 wheat varieties. Under both CK and LN treatments, emergence rate data for 420 varieties and evenness data for 153 natural population varieties were obtained. Emergency represents emergence rate, Uniformity represents evenness, Cluster 1 represents varieties with high emergence rate and high evenness, and Cluster 2 represents the remaining variety categories. Based on these four sets of data, a heatmap was first drawn (…). Figure 6 (a) visually demonstrates the differences in germination rate and uniformity of the same variety under different treatments; subsequently, a linear fitting plot and a frequency distribution histogram were drawn. Figure 6 (b), and generated a box plot with significance analysis ( Figure 6 Finally, a hierarchical cluster analysis was performed on the four sets of data, and a hierarchical cluster dendrogram was constructed. Figure 6 (d) was used to screen varieties that combine high emergence rate and high uniformity. To ensure the validity of the classification results, a significant difference analysis was performed on the two phylogenetic clustering results, and a histogram with error bars was plotted. Figure 6 (e). Under normal treatment, the median emergence rate and uniformity were 0.63 and 5.64, respectively; under low nitrogen treatment, the median emergence rate and uniformity were 0.69 and 6.03, respectively. No significant correlation was found between emergence rate and uniformity under either treatment condition (p<0.001). Actual analysis and verification show that the emergence rate and uniformity indicators calculated by this method are highly accurate.

[0051] In summary, this invention utilizes mobile devices and online websites for data collection and processing, making it more efficient and convenient. Furthermore, based on the characteristics of image data captured by mobile devices, it proposes the phenotypic data of relative leaf area of ​​wheat seedlings and uses this phenotypic data to evaluate the uniformity of wheat seedlings, yielding more objective results. It also constructs two deep learning models based on the YOLO architecture, and through architectural improvements such as BiFPN and FasterNet, and after multiple training iterations, obtains a final model that can run automatically and accurately, greatly improving the efficiency and accuracy of wheat emergence rate calculation. Subsequently, an effective uniformity evaluation index is constructed to effectively assess the emergence rate and uniformity of different varieties, providing assistance for wheat variety selection and demonstrating significant progress.

[0052] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for calculating wheat emergence rate and seedling uniformity, characterized in that, Includes the following steps: S1. Obtain multiple initial images of wheat seedlings at the emergence stage. Using the rectangular bounding box annotation method, annotate each initial image with a bounding rectangle of wheat seedlings to obtain the target detection dataset. From the target detection dataset, divide it into a training set and a validation set. S2. Construct an object detection model based on YOLOv8, replacing the original path aggregation network PANet structure in YOLOv8 with a bidirectional feature pyramid network BiFPN, and introducing a feature fusion mechanism into BiFPN. Specifically, when introducing the feature fusion mechanism, firstly, establish bidirectional connections between each path in the top-down and bottom-up bidirectional feature pyramid network BiFPN to form each bidirectional path; remove each node with only a single input edge in each bidirectional path, and add additional connections between the input and output nodes of each layer to treat each bidirectional path as a feature network layer; repeat the process of adding additional connections multiple times; introduce learnable weights on each feature network layer for weighted fusion of different feature network layers; then, add a detection layer P2 to the Neck of the bidirectional feature pyramid network BiFPN to predict the position and size of the bounding box corresponding to each wheat seedling. Input the training set of the object detection dataset divided in step S1 into the object detection model for initial training to obtain the initial training results. S3. Using the contour annotation method, wheat seedling contours are annotated for each image in the initial training results to obtain the instance segmentation dataset. Then, an instance segmentation model is built based on YOLOv11. The original architecture in YOLOv11 is replaced with the FasterNet architecture. The replaced instance segmentation model includes several FasterNet Blocks. Each FasterNetBlock includes a PConv partial convolutional layer and two PWConv pointwise convolutional layers. The instance segmentation dataset is input into the instance segmentation model for training to obtain the number of wheat seedlings and the relative leaf area of ​​each wheat seedling. The number of wheat seedlings is taken as the number of wheat seedlings that have emerged. The number of wheat seedlings that have emerged is divided by the number of wheat seeds sown to obtain the wheat emergence rate. S4. Based on the relative leaf area of ​​each wheat seedling, calculate the corresponding feature data, including the standard deviation, coefficient of variation, and entropy value of the relative leaf area; normalize each feature data; based on the normalized feature data, calculate the corresponding uniformity index: Uniformity = 1 / StdDev + 1 / CV + Entropy; where Uniformity is the uniformity index, StdDev is the standard deviation, CV is the coefficient of variation, Entropy is the entropy value, and / is the division symbol; The wheat emergence rate and each uniformity index were stored as the final result.

2. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, In step S1, when acquiring multiple initial images of wheat seedlings, the first step is to use a mobile device to capture images of each planting plot in the CK and LN treatment plots. Each planting plot corresponds to a wheat variety. Each plot corresponds to a folder, and each folder is used to store each result image of the corresponding CK treatment plot or the corresponding LN treatment plot. Each result image is then transmitted as an initial image to an online data annotation and augmentation website for annotation.

3. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, When annotating the bounding rectangle of wheat seedlings in step S1, a corresponding txt file is generated for each initial image. Each txt file includes the category of each wheat seedling, the four points of the bounding rectangle corresponding to each wheat seedling, and the length and width information of each corresponding bounding rectangle.

4. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, After dividing the training set and validation set in step S1, the training set is further augmented with data augmentation, which includes rotation, translation, contrast enhancement and noise addition.

5. The method for calculating wheat emergence rate and seedling uniformity according to claim 4, characterized in that, In step S1, the training set and validation set are divided in an 8:2 ratio; the amount of data in the training set after augmentation is at least three times the amount of data in the training set before augmentation.

6. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, In step S2, when the original path aggregation network PANet structure is replaced with the bidirectional feature pyramid network BiFPN and a feature fusion mechanism is introduced, the first ablation experiment is also conducted using the validation set in the object detection dataset. This is used to conduct a first comparative evaluation of the object detection model built on YOLOv8 and the object detection model after the feature fusion mechanism is introduced. The training results of the validation set in the target detection model after the feature fusion mechanism are used as input to conduct a second ablation experiment, which is used to conduct a second comparative evaluation of the instance segmentation model before and after the architecture replacement.

7. The method for calculating wheat emergence rate and seedling uniformity according to claim 6, characterized in that, The evaluation metrics for the first comparative evaluation include precision, recall, mean precision, confusion matrix, and mean detection time; the evaluation metrics for the second comparative evaluation include mean precision, mean intersection-union ratio, mean segmentation time, and model size.

8. The method for calculating wheat emergence rate and seedling uniformity according to claim 7, characterized in that, If the corresponding comparative evaluation results do not meet the corresponding evaluation indicators during the first or second ablation experiment, the process returns to step S2 to build a target detection model based on YOLOv8 and repeats until the corresponding evaluation indicators are met.

9. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, In step S3, when annotating the contours of wheat seedlings, a corresponding txt file II is generated for each image in the target detection dataset. Each txt file II includes the category of each wheat seedling and the coordinates of each vertex of the irregular polygon in the contour of each wheat seedling.

10. The method for calculating wheat emergence rate and seedling uniformity according to claim 1, characterized in that, The normalization process in step S4 is a linear normalization process, which linearly normalizes each feature data to the [0,1] interval.

Images