Cauliflower phenotype analysis and prediction method based on field high-throughput phenotype system

Abstract

The invention discloses a cauliflower phenotype analysis and prediction method based on a field high-throughput phenotype system. According to the method, on the basis of RGB images and hyperspectral images of field cauliflower, through a cauliflower segmentation model, cauliflower, flower ball and leaf areas are automatically extracted, phenotype data and chlorophyll content are extracted, data of different growth periods are sorted according to a time sequence, a growth curve is drawn, and a regression model is constructed, so that the growth states of the cauliflower in different stages are predicted. Compared with the prior art, an effective non-contact scheme is provided for cauliflower phenotype analysis, phenotype parameters of cauliflower can be obtained without manual field on-site measurement, and corresponding culture measurement is combined, so that the cauliflower culture efficiency can be effectively improved, the culture cost can be reduced, and double friendliness of human resources and biological safety is realized.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, agricultural automation, computer vision technology and crop phenotype detection technology in the production field, and specifically relates to a cauliflower phenotype analysis and prediction method based on a field high-throughput phenotyping system. Background Art

[0002] In cauliflower cultivation, evaluating the growth status of cauliflower at each growth stage is crucial to improving economic benefits. Accurately drawing the growth curve of cauliflower is of great significance for formulating cultivation strategies and increasing yields. The growth status of cauliflower refers to the changes in phenotypes such as the length and width of the cauliflower, the color and length and width of the cauliflower, the shape of the leaves, and the chlorophyll content during the growth of the cauliflower. These data are visualized to draw a growth curve. This helps to analyze the growth pattern of cauliflower, predict growth trends, and discover potential health problems, thereby shortening the cultivation cycle and increasing yields.

[0003] A comprehensive analysis of existing technologies shows that the phenotypic analysis of traditional cauliflower mainly relies on manual observation and measurement. Although such methods have low technical requirements, they have certain limitations in efficiency and accuracy and are difficult to implement on a large scale. With the development of computer vision, image processing technology and machine learning algorithms, especially the breakthrough of deep learning in the field of image recognition, new solutions have been provided for plant phenotype extraction. This technology automatically extracts features from large amounts of image data, effectively reducing the reliance on manual data collection and organization, making real-time monitoring and long-term prediction possible. However, existing deep learning methods have difficulty in improving accuracy when processing targets of different scales at the same time. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] The technical problems to be solved by the present invention include the following aspects. The phenotypic analysis of traditional cauliflower often relies on manual observation and measurement, which is not only time-consuming and labor-intensive, but also has obvious limitations in the real-time and objectivity of the data. These methods are difficult to capture the impact of various complex environmental factors on cauliflower in its natural growth environment. Although the cauliflower phenotyping system that incorporates deep learning technology has made significant progress in reducing human intervention and improving the degree of prediction automation, they still face challenges in simultaneously identifying multi-scale targets such as cauliflower, flower heads and leaves, and it is difficult to fully achieve accurate target extraction and phenotypic analysis.

[0006] (II) Technical solution

[0007] In order to solve the above problems, the present invention provides the following technical solutions, and proposes a cauliflower phenotyping prediction method based on a field high-throughput phenotyping system. The specific solution is as follows.

[0008] A cauliflower phenotyping prediction method based on a field high-throughput phenotyping system, comprising:

[0009] Step S1, using an RGB camera and a hyperspectral camera to perform contactless image acquisition of cauliflower in the field in different time periods, obtaining RGB images and hyperspectral images, and the acquisition time span covers the entire growth period of the cauliflower;

[0010] Step S2, using the RGB image as a data set, constructing a segmentation model of the cauliflower plant, flower head, and leaves, and generating a region of interest;

[0011] Step S3, using the region of interest obtained in step S2, respectively calculating the phenotypic data of the whole cauliflower, the flower head and the leaves in the RGB image;

[0012] Step S4, combining the hyperspectral image acquired in step S1 with the leaf region of interest acquired in step S2, acquiring a hyperspectral image of the leaf, and screening the spectral image of a specific band to calculate the chlorophyll content;

[0013] Step S5, based on the phenotypic data obtained in step S3 and the chlorophyll content obtained in step S4, the data of different growth periods are sorted in chronological order, a growth curve is drawn, a regression model is constructed, and the phenotypic parameters of the cauliflower in the corresponding period can be predicted by inputting the growth time.

[0014] Preferably, the contactless image acquisition of cauliflowers in the field in step S1 is specifically achieved in the following manner: a gantry device equipped with an RGB camera and a push-broom hyperspectral camera is built, the gantry device is moved smoothly in the field, and cauliflowers in multiple areas are photographed from the top in sequence; the RGB camera first photographs the images of the ruler and the standard colorimetric card at the starting position, performs color correction, calculates the ratio of the actual length of the ruler to the length in the image, obtains the scale, and then photographs the cauliflower images piece by piece; before acquisition, the hyperspectral camera closes the shutter to photograph the dark frame for dark current correction, and then photographs the barium sulfate plate for radiation correction, and after the correction is completed, the cauliflower is hyperspectrally scanned row by row; the acquired RGB images and hyperspectral images are placed in corresponding folders according to the acquisition date.

[0015] Preferably, in step S2, the cauliflower plant, flower head, and leaf segmentation models are constructed to generate regions of interest, which are specifically implemented in the following manner:

[0016] Step S21, using labelme to mark the contours of the whole cauliflower, the flower head and the leaves, generate a marking file, and construct a segmentation data set, which is divided into a training set, a validation set and a test set at a ratio of 8:1:1;

[0017] Step S22, using Mosaic enhancement to randomly stitch the four images into a large image, scaling the image proportionally, and randomly intercepting it with a 640*640 selection box to obtain an enhanced image;

[0018] Step S23, input the enhanced image into YOLOGB-Seg, and the image size is 640*640*3; the YOLOGB-Seg is improved on the basis of YOLOV8-Seg, and is divided into three parts: a backbone feature extraction network Backbone, a neck feature enhancement network Neck, and a head feature prediction network Head: Backbone reduces the image size, increases the receptive field, and increases the number of channels through continuous convolution and downsampling, and finally outputs three effective feature maps P3, P4, and P5; the Neck part uses BiFPN instead of PAN of YOLOV8-Seg, and performs feature fusion on P3, P4, and P5 output by the backbone network to obtain feature maps H3, H4, and H5 with enhanced feature information; the Head part is a classifier and a regressor, which fuses the feature maps H3, H4, and H5 to obtain the position and category information of targets of different scales, and then classifies pixel by pixel in the detection frame containing the target to remove the background.

[0019] Preferably, the backbone network Backbone of YOLOGB-Seg in step S23 is specifically implemented in the following manner:

[0020] The backbone network Backbone consists of 5 logical layers, each of which contains downsampling and convolution operations. The image size is halved and the number of channels is doubled after each logical layer. The backbone network ends with spatial pyramid pooling SPPF to capture multi-scale features and output feature map F1. YOLOGB-Seg adds a GAM attention mechanism after SPPF to improve the feature expression ability of the network. GAM consists of channel attention and spatial attention. In the channel attention: first, the dimension of the feature map F1 is converted to H×W×C, where H is the height, W is the width, and C is the number of channels. The converted feature map is input into the multi-layer perceptron MLP, channel weights are assigned, and then the image dimension is restored to C×W×H. Finally, the channel attention matrix M is obtained through Sigmoid output. c ; In spatial attention: the number of channels is shrunk and expanded through 7×7 convolution, and finally the spatial attention matrix M is obtained through Sigmoid output s ; Channel attention matrix M c The intermediate feature map F2 is obtained by element-by-element multiplication with the feature map F1, and the output feature map P5 is obtained by multiplying the spatial attention matrix with the feature map F2. The calculation formula of this process is as follows:

[0021] F2=F1⊙CAM(F1)

[0022] P5=F2⊙SAM(F2)

[0023] Among them, CAM is channel attention, SAM is spatial attention, and ⊙ is the multiplication of corresponding elements of the matrix;

[0024] The feature maps P3 and P4 output by the 3rd and 4th logic layers of the backbone network Backbone and the output feature map P5 of GAM are transmitted to the Neck network as valid feature maps.

[0025] Preferably, the Neck network BiFPN of YOLOGB-Seg in step S23 is specifically implemented in the following manner:

[0026] BiFPN assigns weights to feature maps of different layers and performs feature fusion: after upsampling, feature map P5 is concatenated with weighted feature map P4 along the channel to obtain upsampled fused feature map O4; after upsampling, feature map O4 is concatenated with weighted feature map P3 along the channel to obtain upsampled fused feature map H3; after downsampling, H3 is concatenated with weighted feature map O4 along the channel to obtain downsampled fused feature map H4; after downsampling, H4 is concatenated with feature map P5 along the channel to obtain downsampled feature fusion feature map H5; H3, H4, and H5 are passed to the Head network as enhanced feature maps;

[0027] Among them, the BiFPN calculation process is as follows:

[0028] O4=cat(α4×P4, upsample(P5))

[0029] H3=cat(α3×P3, upsample(O4))

[0030] H4=cat(downsample(H3),β3×O4)

[0031] H5=cat(downsample(H4),P5)

[0032] Among them, cat() is the feature map channel splicing operation, the feature maps involved in the splicing have the same width and height, and the number of channels of the final output feature map is the sum of the number of channels of the input feature map; upsample() is the upsampling operation, the feature map is upsampled, the size is doubled, and the number of channels is halved; downsample() is the downsampling operation, the feature map is downsampled, the size is halved, and the number of channels is doubled; α i , β i It is the weight of the corresponding feature map, which changes dynamically during training to increase the focus on important features. i represents the logical layer where the input feature map is located.

[0033] Preferably, the phenotypic data of the whole cauliflower, the curd and the leaves in step S3 are obtained by the following steps:

[0034] Step S31, cauliflower plant phenotype extraction: obtain the minimum circumscribed rectangle of the cauliflower segmentation area, multiply the width and height of the rectangular frame by the obtained scale to obtain the top-view width and height of the cauliflower; use contourArea() to obtain the top-view area of ​​the cauliflower image and then multiply it by the square of the scale to obtain the top-view area of ​​the cauliflower;

[0035] Step S32, obtaining the curd phenotype: obtaining the minimum circumscribed rectangle of the curd segmentation area, multiplying the width and height of the rectangular frame by the scale to obtain the top view width and height of the curd; calculating the aspect ratio to evaluate the roundness of the curd top view cross section, the closer the ratio is to 1, the higher the curd roundness; the color of the curd is obtained by classification of the deep learning network YOLOGB-Seg, which is used to characterize the cauliflower variety;

[0036] Step S33, leaf shape and color phenotype extraction: contourArea() is used to multiply the square of the scale to calculate the true leaf area; the leaf shape is characterized by the aspect ratio, and the length and width of the leaf are obtained by multiplying the length and width of the minimum circumscribed rectangle of the leaf by the scale, and the leaves are divided into nearly circular, elliptical, oblong or lanceolate shapes according to different aspect ratios; the yellow-green ratio of the leaf is calculated by first converting the RGB image into an HSV image, obtaining the green area of ​​the leaf by setting a threshold, calculating the area of ​​the green area, and using the proportion of the green area in the total leaf area to characterize the yellow-green ratio of the leaf.

[0037] Preferably, the calculation of leaf chlorophyll content in step S4 is specifically achieved in the following manner: selecting sensitive bands based on the acquired hyperspectral image, extracting leaf hyperspectral images in combination with the segmented leaf areas, and calculating the chlorophyll content using the enhanced vegetation index EVI.

[0038] Preferably, the phenotypic data collation and growth curve drawing in step S5 are specifically implemented in the following manner: the images in the same folder are output to the same table file after the phenotypic data are obtained in steps S4 and S5, and the phenotypes of the same cauliflower are sorted by time and imported into a table through batch processing; the plant growth curve and the chlorophyll content transformation diagram are drawn with time as the horizontal axis and the top view length and width of the cauliflower and the chlorophyll content as the vertical axis to construct a cauliflower growth prediction model.

[0039] (III) Beneficial effects

[0040] Compared with the prior art, the present invention has at least the following positive technical effects.

[0041] Most existing cauliflower phenotyping methods are based on manual observation and measurement, which are not only overly dependent on labor, but also highly subjective in measurement, and cannot achieve high-throughput cauliflower phenotyping. The method proposed in the present invention uses a gantry device to obtain large-scale, high-precision field cauliflower RGB and hyperspectral images, uses a deep learning network YOLOGB-Seg to extract the region of interest in real time, obtains the target minimum bounding rectangle through opencv, calculates various cauliflower phenotypic parameters, and effectively eliminates subjective differences in phenotypic analysis.

[0042] Cauliflower phenotyping involves the extraction of different organs of cauliflower, such as flower heads and leaves. Existing cauliflower phenotyping methods mostly use a phased analysis method for targets of different scales, analyzing different targets step by step, which is slow, and it is difficult to improve the accuracy of simultaneous segmentation and extraction. The present invention introduces the BiFPN bidirectional feature fusion mechanism in the constructed segmentation model, fully integrates the semantic information of feature maps of different layers, retains more detailed information, effectively captures targets of different scales, and can accurately segment different targets such as cauliflower, flower heads, and leaves in cauliflower images, significantly improving the efficiency of phenotyping analysis.

[0043] The essence of the existing deep learning network to establish a regression model is to reveal the correlation between one or more independent variables and the dependent variable. Therefore, there are limitations when processing continuous time series data for a single individual, and it is difficult to achieve long-term multi-point prediction of plant growth performance. The present invention involves cauliflowers in different growth stages in the image acquisition stage, and collects data of the entire growth cycle of cauliflower at fixed time intervals. The time information and the phenotypic information extracted by the phenotypic system constitute a cauliflower growth prediction model, which expands the growth prediction of a single time point into a multi-point prediction of a continuous time series. A growth curve is drawn with the time series as the horizontal axis and the length, width, overhead area and chlorophyll content as the vertical axis, so as to predict the subsequent growth state changes of early cauliflower and comprehensively evaluate its growth performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention is a flow chart of the method.

[0045] Figure 2 This is the segmentation result diagram of the present invention.

[0046] Figure 3 This is the structural diagram of the segmentation model YOLOGB-Seg of the present invention.

[0047] Figure 4 This is the structure diagram of the global attention mechanism GAM of the present invention.

[0048] Figure 5 This is a diagram of leaf color classification results of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with the accompanying drawings and implementation examples.

[0050] The present invention discloses a cauliflower phenotyping prediction method based on a field high-throughput phenotyping system. Figure 1 Shown is a flow chart of extracting cauliflower phenotypes based on a field high-throughput phenotyping system provided in this embodiment, including the following steps.

[0051] Step S1, continuously collecting non-contact RGB images and hyperspectral images of cauliflower grown in the field in different time periods to obtain RGB and hyperspectral images.

[0052] Specifically, the steps of continuously acquiring non-contact RGB images and hyperspectral images of cauliflower grown in the field in different time periods in step S1 are as follows: (1) building a data acquisition platform, and the data acquisition hardware platform acquires cauliflower images by installing an RGB camera and a hyperspectral camera on a gantry device; (2) the RGB camera shoots a standard colorimetric card and a ruler to complete color correction and obtain a scale, the hyperspectral camera closes the shutter to shoot a dark frame to complete dark current correction, and opens the shutter to shoot a barium sulfate plate to complete radiation correction; (3) the camera shoots the cauliflower in the field row by row along the gantry direction, and the cauliflower is imaged every three days for drawing a growth curve.

[0053] Step S2, using the RGB image as a data set, constructing a cauliflower segmentation model, and applying the segmentation model to segment the cauliflower plant, flower head and leaf detection frames and regions of interest. The segmentation results are as follows: Figure 2 As shown in the figure, sub-image a is the segmentation result of the whole cauliflower plant. In this example, the cauliflower segmentation is binary classification. After the cauliflower area is detected, the cauliflower outline is separated pixel by pixel. Sub-image b is the segmentation result of the flower head. The flower head will generate category information while segmenting, which is used to distinguish different varieties of cauliflower. Sub-image c is the segmentation result of the leaves. The leaves belonging to the same cauliflower plant are grouped together and numbered internally to facilitate subsequent data traceability.

[0054] Specifically, the steps of constructing the cauliflower segmentation model in step S2 are as follows: (1) Mosaic data enhancement technology is used for image preprocessing (2) The YOLOGB-Seg cauliflower segmentation model is constructed. The Coco dataset is used for pre-training in the initial stage of the cauliflower segmentation model training to obtain the initial weight of the model; (3) The cauliflower dataset is used to further train the model in the fine-tuning stage of the model to obtain the segmentation model of cauliflower, flower head and leaves; (4) The model hyperparameters are set as follows: the weight file of YOLOV8 transfer learning on Coco is used, the optimizer is AdamW, the momentum parameter Momentum is set to 0.9, and the weight decay parameter Weight Decay is 0.0001. The initial learning rate is set to 0.01, and the learning rate decay factor lrf is set to 0.01. The entire training process lasted for 200 epochs, and the batch size was set to 4.

[0055] The YOLOGB-Seg model structure is as follows Figure 3 As shown in the figure, it consists of a feature extraction network Backbone, a feature enhancement network Neck and a feature prediction network Head. Backbone consists of 5 logical layers, each of which contains convolution and downsampling operations. After each logical layer, the size of the image is halved and the number of channels is doubled. A GAM attention mechanism is added after the fifth logical layer to suppress feature information irrelevant to the target and improve the model's expressiveness. Each feature layer will generate a feature map Pi, and the feature maps P3, P4, and P5 output by the last three layers are selected as effective feature maps to participate in the construction of the subsequent structure of the model; the Neck part uses BiFPN instead of YOLOV8-Seg's PAN, and performs feature fusion on P3, P4, and P5 output by the backbone network to obtain feature maps H3, H4, and H5 with enhanced feature information; the Head part consists of a classifier and a regressor. By fusing the features of H3, H4, and H5, the position and category information of targets of different scales are obtained, a detection frame is generated, and then each pixel in the detection frame is classified, the background area is removed, the target is segmented, and the regions of interest of cauliflower, flower head, and leaves are obtained.

[0056] The structure of the GAM attention mechanism is as follows Figure 4 As shown in Figure 1, GAM consists of channel attention and spatial attention. In channel attention, first, the dimension of the feature map F1 is converted to H×W×C, where H is the height, W is the width, and C is the number of channels. The converted feature map is input into the multi-layer perceptron MLP, channel weights are assigned, and then the image dimension is restored to C×W×H. Finally, the channel attention matrix M is obtained through Sigmoid output. c ; In spatial attention: the number of channels is shrunk and stretched through 7×7 convolution, and finally the spatial attention matrix M is obtained through Sigmoid output s ; Channel attention matrix Mc The intermediate feature map F2 is obtained by element-by-element multiplication with the feature map F1, and the output feature map F3 is obtained by multiplying the spatial attention matrix with the feature map F2.

[0057] The BiFPN structure is as follows Figure 3 The Neck part fuses features at different levels through top-down and bottom-up paths: the high-level feature map P5 is upsampled, and channel splicing and convolution operations are performed with the low-level feature map P4 to obtain a new feature map O4. O4 is upsampled again, and channel splicing and convolution operations are performed with P3 to obtain feature map H3; the low-level feature map H3 is downsampled, and channel splicing and convolution operations are performed with feature map P4 to obtain feature map H4. H4 is downsampled again, and channel splicing and convolution operations are performed with P5 to obtain feature map H5. A weighting mechanism is introduced in feature fusion, and the weights change dynamically during network training to adjust the importance of different features to target extraction.

[0058] In order to objectively evaluate the accuracy of cauliflower segmentation of the present invention, this exemplary embodiment calculated the mAP50 (mAP at an IoU threshold of 0.5), mAP50-95 (the average value of mAP within the threshold range of 0.5-0.95), precision Precision and recall Recall of pixel-level segmentation and manual pixel-level annotation of cauliflower, flower head and leaves. The results showed that the mAP50 of cauliflower segmentation was 95.69%, mAP50-95 was 61.04%, Precision was 96.73%, and Recall was 89.33%, the mAP50 of flower head segmentation was 98.08%, mAP50-95 was 74.25%, Precision was 98.63%, and Recall was 94.93%, and the mAP50 of leaf segmentation was 95.43%, mAP50-95 was 86.52%, Precision was 85.49%, and Recall was 87.59%. It shows that there is a high similarity between the model segmentation and the manually labeled segmentation. The size and position of different organs of cauliflower in the segmentation results are consistent with the true values, and the model has a high segmentation accuracy.

[0059] Step S3, using the region of interest obtained in step S2, respectively calculate the phenotypic data of the whole cauliflower, the flower head and the leaves in the RGB image.

[0060] Specifically, the phenotypic calculation steps of cauliflower, curd and leaves in step S3 are as follows: (1) The whole cauliflower, curd and leaves are processed step by step according to the region of interest obtained by YOLOGB-Seg; (2) The pixel value of the background area is set to 0, and then the image is converted into a grayscale image, a threshold is set, and the cauliflower area is extracted; (3) The minimum circumscribed rectangle of the cauliflower is obtained, and the width and height of the rectangular frame are multiplied by the scale as the top-view width and height of the curd, contourArea() is multiplied by the square of the scale to obtain the top-view area of ​​the cauliflower, and arcLength() is multiplied by the scale to obtain the top-view contour circumference of the cauliflower. The top-view length, width, top-view area and top-view contour circumference of the curd and leaves are calculated in the same way. In particular, the color features of the cauliflower and leaves are directly obtained through YOLOGB-Seg multi-classification, and the results are as follows: Figure 5 As shown, when the leaves are segmented, color category information is generated. In this example, the leaf colors are divided into three categories: green, deep green, and grey green. The category information is finally output to a table to intuitively represent the health of the leaves. The shape of the leaves is classified by calculating the aspect ratio. The classification standard of this embodiment is: aspect ratio ≤ 2 is nearly circular, 2 < aspect ratio ≤ 2.5 is elliptical, 2.5 < aspect ratio ≤ 3 is oblong, and aspect ratio > 3 is lanceolate; the steps for calculating the yellow-green ratio of the leaves are as follows: (1) Set the pixel value of the background area in the image to 0 and extract the cauliflower leaf area; (2) Convert the RGB image to an HSV image, obtain the green area of ​​the leaves by setting a threshold, and calculate the area of ​​the green area. In this embodiment, the proportion of the green area of ​​the leaves to the total area of ​​the leaves is used to represent the yellow-green ratio of the leaves.

[0061] Step S4, combining the hyperspectral image acquired in step S1 with the leaf region of interest acquired in step S2, acquiring a hyperspectral image of the leaf, and screening the spectral image of a specific band to calculate the chlorophyll content;

[0062] Specifically, the calculation steps of leaf chlorophyll content in step S4 are as follows: (1) Select the near infrared 830nm band, red light 625nm band, and blue light 480nm band of the hyperspectral image to generate a three-channel pseudo-color image; (2) Combine the RGB image processed by YOLOGB-Seg to set the pixel value of the background area to 0; (3) Calculate the leaf enhanced vegetation index EVI, the calculation formula is as shown in formula (1), where NIR is the reflectivity of the near infrared band, Red is the reflectivity of the red light band, Blue is the reflectivity of the blue light band, G is the gain coefficient, which is usually set to 2.5, C1 and C2 are aerosol drag coefficients, which are used for the red and blue light bands, respectively, and are usually set to 6 and 7.5, and L is the soil brightness coefficient, which is usually set to 1; (4) Establish a regression model based on the chlorophyll content measured in the field;

[0063]

[0064] Step S5, arranging the cauliflower phenotype data obtained in step S3 and step S4 in chronological order to obtain the cauliflower phenotypes at different growth stages, and drawing a growth curve.

[0065] Specifically, the phenotypic data are collated and the growth curve and chlorophyll content are drawn in step S5, and the specific implementation steps are as follows: (1) the phenotypic system outputs the phenotypic data obtained from the cauliflower of the same period to the same table file, and names the table with time; (2) the system imports the same cauliflower into a table in chronological order through batch processing, with the cauliflower's top view width and height and chlorophyll content as the vertical coordinates and time as the horizontal coordinate. The plant growth curve and chlorophyll content change diagram are updated at fixed time intervals, and a regression model is constructed. The phenotypic parameters of the cauliflower of the corresponding period can be predicted by inputting the growth time.

[0066] The specific examples described in the application are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific examples described in the present invention, or replace them in a similar manner, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A cauliflower phenotyping prediction method based on a field high-throughput phenotyping system, characterized in that ,include: Step S1, using an RGB camera and a hyperspectral camera to perform contactless image acquisition of cauliflower in the field in different time periods, obtaining RGB images and hyperspectral images, and the acquisition time span covers the entire growth period of the cauliflower; Step S2, using the RGB image as a data set, constructing a segmentation model of the cauliflower plant, flower head, and leaves, and generating a region of interest; Step S3, using the region of interest obtained in step S2, respectively calculating the phenotypic data of the whole cauliflower, the flower head and the leaves in the RGB image; Step S4, combining the hyperspectral image acquired in step S1 with the leaf region of interest acquired in step S2, acquiring a hyperspectral image of the leaf, and screening the spectral image of a specific band to calculate the chlorophyll content; Step S5, based on the phenotypic data obtained in step S3 and the chlorophyll content obtained in step S4, the data of different growth periods are sorted in chronological order, a growth curve is drawn, a regression model is constructed, and the phenotypic parameters of the cauliflower in the corresponding period can be predicted by inputting the growth time.

2. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 1, characterized in that: In the step S1, the contactless image acquisition of the cauliflower in the field is specifically achieved in the following manner: a gantry device equipped with an RGB camera and a push-broom hyperspectral camera is built, the gantry device is moved smoothly in the field, and the cauliflowers in multiple areas are photographed from the top in sequence; the RGB camera first photographs the images of the ruler and the standard colorimetric card at the starting position, performs color correction, calculates the ratio of the actual length of the ruler to the length in the image, obtains the scale, and then photographs the cauliflower images piece by piece; before acquisition, the hyperspectral camera closes the shutter to photograph the dark frame for dark current correction, and then photographs the barium sulfate plate for radiation correction, and after the correction is completed, the cauliflower is hyperspectrally scanned line by line; the acquired RGB image and the hyperspectral image are placed in corresponding folders according to the acquisition date.

3. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 1, characterized in that: In step S2, a segmentation model of cauliflower plants, flower heads, and leaves is constructed to generate regions of interest, which is specifically achieved in the following manner: Step S21, using labelme to mark the contours of the whole cauliflower, the flower head and the leaves, generate a marking file, and construct a segmentation data set, which is divided into a training set, a validation set and a test set at a ratio of 8:1:1; Step S22, using Mosaic enhancement to randomly stitch the four images into a large image, scaling the image proportionally, and randomly intercepting it with a 640*640 selection box to obtain an enhanced image; Step S23, inputting the enhanced image into YOLOGB-Seg, the image size is 640*640*3; The YOLOGB-Seg is improved on the basis of YOLOV8-Seg, and is divided into three parts: a backbone feature extraction network Backbone, a neck feature enhancement network Neck, and a head feature prediction network Head; Backbone reduces the image size, increases the receptive field, and increases the number of channels through continuous convolution and downsampling, and finally outputs three effective feature maps P3, P4, and P5; the Neck part uses BiFPN instead of PAN of YOLOV8-Seg, and performs feature fusion on P3, P4, and P5 output by the backbone network to obtain feature maps H3, H4, and H5 with enhanced feature information; the Head part is a classifier and a regressor, which fuses the feature maps H3, H4, and H5 to obtain the position and category information of targets of different scales, and then classifies pixel by pixel in the detection frame containing the target to remove the background.

4. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 3, characterized in that: The backbone network Backbone of YOLOGB-Seg in step S23 is specifically implemented in the following manner: The backbone network Backbone consists of 5 logical layers, each of which contains downsampling and convolution operations. The image size is halved and the number of channels is doubled after each logical layer. The backbone network ends with spatial pyramid pooling SPPF to capture multi-scale features and output feature map F1. YOLOGB-Seg adds a GAM attention mechanism after SPPF to improve the feature expression ability of the network. GAM consists of channel attention and spatial attention. In the channel attention: first, the dimension of the feature map F1 is converted to H×W×C, where H is the height, W is the width, and C is the number of channels. The converted feature map is input into the multi-layer perceptron MLP, channel weights are assigned, and then the image dimension is restored to C×W×H. Finally, the channel attention matrix M is obtained through Sigmoid output. c ; In spatial attention: the number of channels is shrunk and expanded through 7×7 convolution, and finally the spatial attention matrix M is obtained through Sigmoid output s ; Channel attention matrix M c The intermediate feature map F2 is obtained by element-by-element multiplication with the feature map F1, and the output feature map P5 is obtained by multiplying the spatial attention matrix with the feature map F2. The calculation formula of this process is as follows: F2=F1⊙CAM(F1) P5=F2⊙SAM(F2) Among them, CAM is channel attention, SAM is spatial attention, and ⊙ is the multiplication of corresponding elements of the matrix; The feature maps P3 and P4 output by the 3rd and 4th logic layers of the backbone network Backbone and the output feature map P5 of GAM are transmitted to the Neck network as valid feature maps.

5. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 3, characterized in that: The BiFPN of the Neck network of YOLOGB-Seg in step S23 is specifically implemented in the following manner: BiFPN assigns weights to feature maps of different layers and performs feature fusion: after upsampling, feature map P5 is weighted and spliced ​​along the channel with feature map P4 to obtain upsampled fused feature map O4; after upsampling, feature map O4 is weighted and spliced ​​along the channel with P3 to obtain upsampled fused feature map H3; after downsampling, H3 is weighted and spliced ​​along the channel with O4 to obtain downsampled fused feature map H4; after downsampling, H4 is spliced ​​along the channel with feature map P5 to obtain downsampled feature fusion feature map H5; H3, H4, and H5 are passed to the Head network as enhanced feature maps; the BiFPN calculation process is as follows: O4=cat(α4×P4, upsample(P5)) H3=cat(α3×P3, upsample(O4)) H4=cat(downsample(H3),β3×O4) H5=cat(downsample(H4),P5) Among them, cat() is the feature map channel splicing operation, the feature maps involved in the splicing have the same width and height, and the number of channels of the final output feature map is the sum of the number of channels of the input feature map; upsample() is the upsampling operation, the feature map is upsampled, the size is doubled, and the number of channels is halved; downsample() is the downsampling operation, the feature map is downsampled, the size is halved, and the number of channels is doubled; α i , β i It is the weight of the corresponding feature map, which changes dynamically during training to increase the focus on important features. i represents the logical layer where the input feature map is located.

6. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 1, characterized in that: The phenotypic data of the whole cauliflower, the curd and the leaves in step S3 are obtained by the following steps: Step S31, cauliflower plant phenotype extraction: obtain the minimum circumscribed rectangle of the cauliflower segmentation area, multiply the width and height of the rectangular frame by the obtained scale to obtain the top-view width and height of the cauliflower; use contourArea() to obtain the top-view area of ​​the cauliflower image, and multiply it by the square of the scale to obtain the top-view area of ​​the cauliflower; Step S32, obtaining the curd phenotype: obtaining the minimum circumscribed rectangle of the curd segmentation area, multiplying the width and height of the rectangular frame by the scale to obtain the top view width and height of the curd; calculating the aspect ratio to evaluate the roundness of the curd top view cross section, the closer the ratio is to 1, the higher the curd roundness; the color of the curd is obtained by classification of the deep learning network YOLOGB-Seg, which is used to characterize the cauliflower variety; Step S33, extracting leaf shape and color phenotype: using contourArea() and the square of the scale to calculate the real leaf area; the leaf shape is characterized by the aspect ratio, and the length and width of the leaf are obtained by multiplying the length and width of the minimum circumscribed rectangle of the leaf by the scale, and the leaves are divided into nearly circular, elliptical, oblong or lanceolate shapes according to different aspect ratios; To calculate the yellow-green ratio of leaves, the RGB image is first converted into an HSV image, the green area of ​​the leaves is obtained by setting a threshold, the area of ​​the green area is calculated, and the yellow-green ratio of the leaves is represented by the proportion of the green area in the total leaf area.

7. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 1, characterized in that: The calculation of leaf chlorophyll content in step S4 is specifically achieved in the following manner: selecting sensitive bands based on the acquired hyperspectral image, extracting leaf hyperspectral images based on the segmented leaf areas, and calculating the chlorophyll content using the enhanced vegetation index EVI.

8. The cauliflower phenotyping prediction method based on a field high-throughput phenotyping system according to claim 1, characterized in that: The phenotypic data collation and growth curve drawing in step S5 are specifically implemented in the following manner: the images in the same folder are output to the same table file after the phenotypic data are obtained in steps S4 and S5, and the phenotypes of the same cauliflower are sorted by time and imported into a table through batch processing; the plant growth curve and the chlorophyll content transformation diagram are drawn with time as the horizontal axis and the top view length and width of the cauliflower and the chlorophyll content as the vertical axis to construct a cauliflower growth prediction model.

Images