Cauliflower multi-character detection method based on high-throughput phenotypic technology

Abstract

The invention provides a cauliflower multi-character detection method based on a high-throughput phenotype technology. The method specifically comprises the steps that a photosensitive sensor is integrated in an imaging unit of a field high-throughput phenotype platform, when the photosensitive sensor detects that the illumination intensity is within a preset range, an RGB imaging unit is remotely triggered to conduct image data collection on broccoli, and meanwhile a distance measuring sensor is triggered to upload real-time camera height data; a pre-trained YOLOv8-seg model is adopted to carry out cauliflower plant and ball-flower identification on the preprocessed collected image, and plant leaf colors, leaf shapes, petioles and ball-flower colors are classified into different categories; based on the plant area and the ball-flower area obtained through identification and segmentation, plant width and leaf area parameters and ball-flower transverse diameter, longitudinal diameter and ball-flower area parameters are extracted. The broccoli multi-character data can be efficiently and accurately obtained, the breeding work is helped to reduce manual operation, ambient light interference and personal errors are reduced, and the reliability and consistency of character detection are improved.

Description

Technical Field

[0001] This invention belongs to the field of cauliflower breeding technology, specifically relating to a method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology. Background Technology

[0002] Cauliflower (Brassica oleracea var. botrytis L.) is widely cultivated globally due to its high nutritional and medicinal value. The plant morphology and head morphology directly and significantly influence cauliflower quality, and cultivating superior cauliflower traits is key to improving yield and varietal quality. In traditional breeding research, trait measurements of different cauliflower varieties are primarily performed manually, which suffers from high subjectivity, destructiveness, susceptibility to light conditions, and low efficiency, making it difficult to meet the needs of large-scale breeding and phenomics research.

[0003] To achieve targeted improvement of cauliflower traits, breeders need to conduct precise phenotypic analysis on multi-genotype germplasm resources. Conventional field testing methods rely on manual labor and have inherent drawbacks: firstly, destructive sampling leads to compromised sample integrity; secondly, the combined effects of limitations in manual operation and environmental variables (fluctuations in light intensity, deviations in observation location) create a risk of systematic bias; and thirdly, the lack of standardized data collection procedures and reproducibility assurance mechanisms makes it difficult to meet the engineering requirements of modern molecular breeding for data homogeneity and high-throughput phenomics technologies.

[0004] High-throughput phenotyping technology, as an important tool in modern agricultural research, achieves rapid and non-destructive acquisition of plant population phenotypic characteristics by integrating computer vision, sensor networks, and intelligent computing models. This technology can systematically capture diverse phenotypic features such as plant morphology, color attributes, and spectral information, and combine this with multispectral image analysis technology to achieve digital characterization of color features and phenotypic parameters. However, in the field of cauliflower genetic improvement and phenotypic analysis, the traditional model of manual observation is still widely relied upon. The application of multi-trait collaborative detection technology based on intelligent phenotyping acquisition systems remains significantly lacking, especially in the areas of feature extraction algorithm optimization and standardized detection process construction, which urgently require breakthrough research. Summary of the Invention

[0005] The purpose of this invention is to propose a multi-trait detection method for cauliflower based on high-throughput phenotyping technology. Utilizing RGB imaging technology and a convolutional neural network algorithm, this method automatically detects and quantifies phenotypic characteristics of cauliflower plants, including plant width, leaf area, leaf color, leaf shape, petiole, head diameter (bottle width and length), head area, and head color, through field image acquisition and intelligent analysis. This method can efficiently and accurately acquire data on cauliflower plant and head morphology and color, helping to reduce manual operations in breeding work, minimize environmental light interference and human error, and improve the reliability and consistency of trait detection.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology, specifically including the following steps: S1. A photosensitive sensor is integrated into the imaging unit of the field high-throughput phenotyping platform. When the photosensitive sensor detects that the light intensity is within the preset range, the RGB imaging unit is remotely triggered to collect image data of cauliflower. At the same time, the distance measuring sensor is triggered to upload real-time camera height data for scale calculation. S2. Preprocess the acquired images; S3. Use a pre-trained YOLOv8-seg model to identify cauliflower plants and flower heads in the pre-processed image, and generate detection boxes, masks and labels based on the identified plant and flower head regions; For the identified plant and flower head regions, the classification and detection functions of the pre-trained YOLOv8-seg model are used to classify the plant leaf color, leaf shape, and petiole into different categories, and the flower head color into different categories. S4. Based on the plant area and flower head area obtained by recognition and segmentation, the detection box is used for localization and mask pixel analysis, and the unit conversion is completed using the scale bar. The plant width and leaf area parameters are extracted, and the flower head transverse diameter, longitudinal diameter and flower head area parameters are extracted. S5. Perform statistical analysis and visualization output on the test results of each trait, and store all test data.

[0007] Preferably, when the photosensor detects a light intensity of 120~150 μmol / (m²), 2 When the range is within ·s), the RGB imaging unit is remotely triggered to acquire image data of mature cauliflower, and at the same time, the ranging sensor is triggered to upload real-time camera height data.

[0008] Preferably, the RGB imaging unit acquires top-view images of the cauliflower plant and flower head, with each image capturing at least one complete cauliflower plant.

[0009] Preferably, the photosensor is a VEMSEE light sensor.

[0010] Preferably, the ranging sensor is a VXTE laser ranging sensor.

[0011] Preferably, the preprocessing of the acquired image includes noise reduction, background segmentation, and color correction, specifically employing Gaussian filtering-based noise reduction, HSV color space-based background segmentation, and standard color correction.

[0012] Preferably, S4 specifically includes: The scale is calculated based on the real-time camera height data uploaded by the range sensor: Scale = Camera height / Camera focal length; Using OpenCV library functions, the plant width and flower head horizontal and vertical diameter parameters are calculated by combining the length and width pixel information of the plant and flower head detection boxes with the scale bar. Using OpenCV library functions, leaf area and flower head area parameters are calculated by combining the pixel information of plant and flower head masks with a scale.

[0013] Preferably, a pre-trained YOLOv8-seg model is used to classify plant leaf color, leaf shape, and petiole into different categories. The leaf color categories include light green leaves, green leaves, gray-green leaves, and dark green leaves. The leaf shape categories include nearly round leaves, oval leaves, elliptical leaves, oblong leaves, and lanceolate leaves. The petiole categories include leaves with petioles and leaves without petioles.

[0014] Preferably, a pre-trained YOLOv8-seg model is used to classify the flower ball color into different categories, including snow-white flower ball, yellow-white flower ball, orange flower ball, green flower ball, and purple flower ball.

[0015] Preferably, S5 specifically includes: Data analysis and results output: Based on the row and column of the image acquisition, the plants and flower heads are numbered, and statistical analysis is performed on the data of each trait; the analysis results are output in a visual form, including Excel spreadsheets and histograms. Data storage and application: The cauliflower detection data is stored in a database, including raw image data, preprocessing results, detection results of various traits, and various statistical indicators; the database adopts a relational architecture design, which supports fast query and batch export.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention employs high-throughput phenotyping technology combined with RGB imaging and YOLOv8-seg instance segmentation algorithm to achieve intelligent and batch detection of multiple traits in cauliflower, reducing manual operations in traditional trait measurement work, improving the accuracy, reliability and diversity of trait data, and providing a method for acquiring high-throughput phenotypic information of multiple traits in cauliflower. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the operation process of the present invention; Figure 2 This is a visual schematic diagram of the imaging unit of the improved field high-throughput phenotyping platform of the present invention. Figure 3 This is a schematic diagram of the visualization of cauliflower trait information according to the present invention, where a is the original image of the plant in the rosette stage, b is the segmentation effect of the plant instance in the rosette stage, c is the original image of the plant in the mature stage, and d is the segmentation effect of the plant instance in the mature stage. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The technical solution of the present invention will be described in detail below.

[0019] This invention proposes a method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology, referencing... Figure 1 Specifically, it includes the following steps: S1. A photosensitive sensor is integrated into the imaging unit of the field high-throughput phenotyping platform. When the photosensitive sensor detects that the light intensity is within the preset range, the RGB imaging unit is remotely triggered to acquire image data of cauliflower. At the same time, the distance sensor is triggered to upload real-time camera height data for scale calculation. The acquired cauliflower images are transmitted to the central computer in real time via wireless network and the preprocessing process is automatically started. S2. Preprocess the acquired images, including noise reduction, background segmentation, and color correction. S3. A pre-trained YOLOv8-seg model is used to identify cauliflower plant CP (Cauliflower plantt) and cauliflower curd CC (Cauliflower curd) in the pre-processed image, accurately locating the cauliflower plant and cauliflower curd regions, and generating detection boxes, masks and labels based on the identified plant and cauliflower curd regions. For the identified plant and flower head regions, the classification and detection functions of the pre-trained YOLOv8-seg model are used to classify the plant leaf color, leaf shape, and petiole into different categories, and the flower head color into different categories. S4. Based on the plant area and flower head area obtained by recognition and segmentation, image processing is performed by detection box localization and mask pixel analysis, combined with OpenCV library functions, and unit conversion is performed using a scale bar to extract plant width and leaf area parameters, as well as flower head transverse diameter, longitudinal diameter and flower head area parameters. S5. Perform statistical analysis and visualization output on the test results of each trait, and store all test data.

[0020] This invention trains current classic instance segmentation networks (including YOLOv8-seg, Mask R-CNN and YOLACT) and evaluates and analyzes the trained cauliflower trait detection models. Finally, the YOLOv8-seg model is selected as the optimal detection network. Therefore, the YOLOv8-seg instance segmentation algorithm is adopted in this invention.

[0021] The image dataset for model training is constructed as follows: First, the acquired digital images of cauliflower were converted into XML data format. Python was chosen as the development language, and the Labelme annotation tool was used to annotate the polygonal outlines of the cauliflower plants and flower heads in the images, while simultaneously implementing classification annotation based on chromaticity features. The data samples were stratified according to a training set:validation set:test set ratio of 8:1:1.

[0022] The model training process is as follows: A uniform input image resolution of 640×640 pixels is set, the batch size is configured to 4, and the number of parallel worker threads is 8. YOLOv8-seg.pt pre-trained weights are loaded for transfer learning. An adaptive learning rate strategy is employed during training, with an initial learning rate set to 0.01. Gradient updates are optimized using the Adam optimizer, and the full training cycle is fixed at 300 epochs to achieve optimal model performance. In the model evaluation phase, the Intersection over Union (IoU) threshold and the object detection confidence threshold are defined as 0.5, serving as performance evaluation benchmarks.

[0023] In this embodiment, when the photosensor detects a light intensity of 120~150 μmol / (m²), 2 When the range is within ·s), the RGB imaging unit is remotely triggered to acquire image data of mature cauliflower, and at the same time, the ranging sensor is triggered to upload real-time camera height data.

[0024] In this embodiment, the RGB imaging unit acquires top-view images of cauliflower plants and flower heads. The camera shooting interval is set according to the planting density via the X and Y axes of the field high-throughput phenotyping platform to ensure that each image captures at least one complete cauliflower plant.

[0025] In this embodiment, the photosensor is a VEMSEE light sensor, which can be networked with mobile devices to monitor changes in the environmental spectrum in real time. The light intensity detected by the photosensor is in the range of 120~150 μmol / (m2·s). Within this range, the colors of the cauliflower leaves and flower heads on the colorimetric card are consistent with the colors of the images captured by the high-throughput phenotyping platform.

[0026] In this embodiment, the ranging sensor is a VXTE laser ranging sensor, which can transmit real-time camera height data to the server to obtain a conversion scale.

[0027] In this embodiment, the preprocessing stage employs advanced image processing algorithms, including Gaussian filtering-based noise reduction, HSV color space-based background segmentation, and standard color card-based color correction, to ensure the accuracy of subsequent analysis.

[0028] In this embodiment, the size trait extraction stage in step S4 employs a mask prediction method based on instance segmentation: First, calculate the scale based on the real-time camera height data uploaded by the range sensor: Scale = Camera height / Camera focal length; Using OpenCV library functions, the plant width and flower head horizontal and vertical diameter parameters are calculated by combining the length and width pixel information of the plant and flower head detection boxes with the scale bar. Using OpenCV library functions, leaf area and flower head area parameters are calculated by combining the pixel information of plant and flower head masks with a scale.

[0029] In this embodiment, the classification stage of graded traits in S4 employs a fine-grained classification method based on deep learning. The preprocessed image is input into a pre-trained YOLOv8-seg model, which accurately classifies plant leaf color into four professional categories: light green (GB), green (G), gray-green (GG), and dark green (DG). Plant leaf shape is classified into five professional categories: nearly round (NC), ovate (Ov), elliptical (E), oblong (Ob), and lanceolate (L). Plant petiole type is classified into two professional categories: with petiole (P) and without petiole (NP). Flower head color is classified into five professional categories: snow white (SWC), yellowish white (YWC), orange (OC), green (GC), and purple (PC). Strict confidence thresholds (0.5) and IoU thresholds (0.5) are set during the classification process to ensure the reliability of the classification results. The system also automatically records the confidence score of each classification result, providing a reference for subsequent analysis. This classification method allows for direct identification and classification of cauliflower plant leaf color, leaf shape, petiole type, and flower head color during the detection process.

[0030] In this embodiment, S5 specifically refers to: Data Analysis and Results Output: Plants and flower heads are numbered based on the row and column of image acquisition, and statistical analysis is performed on the data for each trait. The analysis results are output in a visual format, including Excel spreadsheets and histograms. Specifically, the system generates detailed histograms showing the distribution of leaf color, leaf shape, petiole, and flower head color, using built-in statistical analysis algorithms based on the row and column information of the image acquisition. Simultaneously, it generates parameters in Excel spreadsheet format for the total plant width, leaf area, flower head transverse diameter, flower head longitudinal diameter, and the maximum, minimum, average, standard deviation, and coefficient of variation of the flower head area. These charts visually demonstrate the distribution and proportion of different trait categories in the field, providing researchers with important decision-making support.

[0031] Data Storage and Application: Cauliflower test data is stored in a database, including raw image data, preprocessing results, test results for various traits, and various statistical indicators to support subsequent variety comparison, genetic analysis, and breeding selection. The database adopts a relational architecture, supporting fast querying and batch export. Researchers can view test results in real time through dedicated client software and generate various analysis reports as needed. The system also provides data comparison functions, facilitating comparative analysis of different varieties and different growth stages, providing strong data support for breeding work.

[0032] The following is a specific embodiment: like Figure 2 The diagram shows the overall structure of the field high-throughput phenotyping platform, where the image represents the imaging unit integrating a VEMSEE photosensor and a VXTE laser rangefinder. The rangefinder and acquisition system are automatically triggered when the ambient light intensity detected by the VEMSEE light sensor reaches 135 μmol / (m²·s). The YOLOv8-seg instance segmentation algorithm was applied to analyze the images. In this case study, ten cauliflower images at different growth stages and with different leaf colors, shapes, petioles, and flower head colors were used in the experiment. The analysis results are as follows... Figure 3 As shown, complete cauliflower plants or flower heads were identified and segmented in each image, and their leaf color, leaf shape, petiole, and flower head color type were classified. The classification labels were placed in the upper left corner of the detection box. At the same time, the plant and flower head size data were extracted and converted and placed on the detection box and the mask.

[0033] The system automatically compiled the test results, and the leaf color classification results were as follows: GB category 10.0%, G category 30.0%, GG category 10.0%, and DG category 50.0%. The leaf shape classification results are as follows: NC class accounts for 10.0%, Ov class accounts for 10.0%, E class accounts for 40.0%, Ob class accounts for 30.0%, and L class accounts for 10.0%. The petiole classification results are as follows: P type accounts for 60.0%, and NP type accounts for 40.0%. The color classification results for the flower balls are as follows: SWC class accounts for 20.0%, YWC class accounts for 20.0%, OC class accounts for 20.0%, GC class accounts for 20.0%, and PC class accounts for 20.0%.

[0034] It also generates corresponding histograms of leaf color, leaf shape, petiole, and flower head color distribution. The extraction results of the size traits of each plant and flower head (plant width, leaf area, flower head transverse diameter, flower head longitudinal diameter, flower head area) will be automatically numbered according to the row and column positions of the image acquisition and exported as an Excel spreadsheet. At the same time, it automatically calculates the maximum, minimum, average, standard deviation, and coefficient of variation of a single sample.

[0035] The results of the plant width analysis in the example were as follows: maximum value = 174.83 cm, minimum value = 93.22 cm, average value = 136.82 cm, standard deviation = 27.25, and coefficient of variation = 18.89%. The leaf area analysis results are as follows: maximum value = 18798.49 cm², minimum value = 5549.06 cm², mean value = 11602.30 cm², standard deviation = 4149.25, coefficient of variation = 33.93%; The results of the flower head diameter analysis are as follows: maximum value = 28.07 cm, minimum value = 23.03 cm, mean value = 24.78 cm, standard deviation = 1.95, coefficient of variation = 7.04%; The results of the head diameter analysis are as follows: maximum value = 27.74 cm, minimum value = 22.19 cm, mean value = 24.85 cm, standard deviation = 2.21, coefficient of variation = 7.95%; The results of the flower head area analysis are as follows: maximum value = 589.09 cm2, minimum value = 376.73 cm2, average value = 466.09 cm2, standard deviation = 77.25, coefficient of variation = 14.82%.

[0036] All detection data (raw images, phenotypic classification data, and size trait statistics) are automatically archived and stored via an integrated storage architecture. Simultaneously, client-side visualization platforms support real-time monitoring of data assets and one-click export of multi-dimensional analysis reports.

[0037] In summary, this invention presents a high-throughput phenotypic intelligent detection method for cauliflower, utilizing RGB visual imaging and convolutional neural networks to achieve non-destructive detection of nine cauliflower traits. This method integrates a spectral sensing module and a mobile terminal interaction system into a field phenotyping platform to construct an intelligent detection system. Its core features include: a dynamic ambient light sensing sensor that automatically triggers image acquisition and processing when a specific spectral range is detected; wireless transmission technology enabling real-time monitoring of image acquisition quality and parameter extraction on the mobile terminal; and the development of a multi-dimensional feature analysis model based on convolutional neural network algorithms in deep learning to simultaneously acquire nine phenotypic data points: plant width, leaf area, leaf color, leaf shape, petiole type parameters, and flower head morphology (horizontal and vertical diameters, projected area, and color).

[0038] Compared to traditional detection techniques, this invention first eliminates illumination interference through an environmental spectral adaptive correction mechanism, ensuring the reliability of color feature extraction. Secondly, it employs machine vision technology to replace manual measurement, overcoming the limitations of incomplete morphological parameter acquisition and subjective errors in traditional methods. Simultaneously, it establishes an automated, end-to-end mobile detection process, significantly improving data acquisition efficiency. Through multi-dimensional analysis of plant images using convolutional neural networks, it achieves precise quantification of cauliflower phenotypic features.

[0039] This invention provides a standardized detection method for evaluating cauliflower germplasm resources, supports continuous phenotypic monitoring during the breeding process, and its cloud-based data management system provides an analytical framework for phenomics research. Practical application shows that this method effectively reduces the cost of trait analysis and accelerates the breeding process while ensuring detection accuracy, providing an innovative technical solution for smart agriculture and precision breeding.

[0040] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology, characterized in that, Specifically, the following steps are included: S1. A photosensitive sensor is integrated into the imaging unit of the field high-throughput phenotyping platform. When the photosensitive sensor detects that the light intensity is within the preset range, the RGB imaging unit is remotely triggered to collect image data of cauliflower. At the same time, the distance measuring sensor is triggered to upload real-time camera height data for scale calculation. S2. Preprocess the acquired images; S3. Use a pre-trained YOLOv8-seg model to identify cauliflower plants and flower heads in the pre-processed image, and generate detection boxes, masks and labels based on the identified plant and flower head regions; For the identified plant and flower head regions, the classification and detection functions of the pre-trained YOLOv8-seg model are used to classify the plant leaf color, leaf shape, and petiole into different categories, and the flower head color into different categories. S4. Based on the plant area and flower head area obtained by recognition and segmentation, the detection box is used for localization and mask pixel analysis, and the unit conversion is completed using the scale bar. The plant width and leaf area parameters are extracted, and the flower head transverse diameter, longitudinal diameter and flower head area parameters are extracted. S5. Perform statistical analysis and visualization output on the test results of each trait, and store all test data.

2. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, When the photosensor detects a light intensity of 120~150 μmol / (m 2 When the range is within ·s), the RGB imaging unit is remotely triggered to acquire image data of mature cauliflower, and at the same time, the ranging sensor is triggered to upload real-time camera height data.

3. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The RGB imaging unit captures top-view images of the cauliflower plant and flower head, with each image capturing at least one complete cauliflower plant.

4. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The photosensor is a VEMSEE light sensor.

5. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The ranging sensor used is a VXTE laser ranging sensor.

6. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The preprocessing of the acquired images includes noise reduction, background segmentation, and color correction. Specifically, it employs noise reduction based on Gaussian filtering, background segmentation technology based on the HSV color space, and color correction based on a standard color chart.

7. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, Specifically, S4 is: The scale is calculated based on the real-time camera height data uploaded by the range sensor: Scale = Camera height / Camera focal length; Using OpenCV library functions, the plant width and flower head horizontal and vertical diameter parameters are calculated by combining the length and width pixel information of the plant and flower head detection boxes with the scale bar. Using OpenCV library functions, leaf area and flower head area parameters are calculated by combining the pixel information of plant and flower head masks with a scale.

8. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The pre-trained YOLOv8-seg model was used to classify plant leaf color, leaf shape, and petiole into different categories. The leaf color categories included light green leaves, green leaves, gray-green leaves, and dark green leaves. The leaf shape categories included nearly round leaves, oval leaves, elliptical leaves, oblong leaves, and lanceolate leaves. The petiole categories included petiolate leaves and petioleless leaves.

9. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to claim 1, characterized in that, The pre-trained YOLOv8-seg model was used to classify the flower ball colors into different categories, including snow-white flower balls, yellow-white flower balls, orange flower balls, green flower balls, and purple flower balls.

10. The method for detecting multiple traits of cauliflower based on high-throughput phenotyping technology according to any one of claims 1-9, characterized in that, Specifically, S5 is: Data analysis and results output: Based on the row and column of the image acquisition, the plants and flower heads are numbered, and statistical analysis is performed on the data of each trait; the analysis results are output in a visual form, including Excel spreadsheets and histograms. Data storage and application: The cauliflower detection data is stored in a database, including raw image data, preprocessing results, detection results of various traits, and various statistical indicators; the database adopts a relational architecture design, which supports fast query and batch export.

Images