A new vegetable variety DUS / VCU multi-modal evaluation system

Abstract

The application discloses a kind of vegetable new varieties DUS / VCU multi-modal evaluation system, including rule configuration module, task generation module, data governance module, approximate variety search module, DUS auxiliary determination module, VCU prediction module and report review module.The application will DUS / VCU standard rule be converted into computable task, and phenotype image, depth data, spectral data, environmental information and genotypic information are used for approximate variety search, DUS auxiliary determination and VCU performance prediction, to improve test efficiency, objectivity and result consistency.

Description

Technical Field

[0001] This invention belongs to the technical fields of plant variety testing, seed industry detection, plant phenomics, agricultural artificial intelligence, and crop cultivation value evaluation, specifically relating to a DUS / VCU multimodal evaluation system for new vegetable varieties. DUS stands for Distinctness, Uniformity, and Stability; VCU stands for Value for Cultivation and Use. Background Technology

[0002] Before new vegetable varieties are authorized, registered, approved, and promoted, their distinctness, uniformity, and stability need to be tested, and their yield, quality, resistance, adaptability, and marketability need to be evaluated in conjunction with their cultivation value. Current DUS (Distinctive Features in Crop Testing) tests typically follow crop testing guidelines, with technicians observing and recording data from plants, leaves, flowers, fruits, pods, bulbs, and other organs during a specified growth period. VCU (Various Features in Crop Testing) tests, on the other hand, require continuous field data collection at multiple locations and over multiple years.

[0003] The above process still relies heavily on manual surveys in practice. The collection period, organs, trait grades, and judgment criteria vary significantly among different crops, and recording biases can easily arise for the same trait across different years, locations, equipment, and observers. With the increase in the number of candidate varieties and the expansion of the reference variety library, the workload for screening similar varieties, locating differential traits, and verifying cross-environmental consistency has increased significantly.

[0004] Machine vision, UAV remote sensing, 3D phenotype, spectral detection and molecular marker technologies can be used for crop trait measurement or variety-assisted analysis, but most methods are limited to single organ identification, single image index measurement or single crop scene classification, and fail to organize the standard trait rules, data collection tasks, evidence images, reference variety libraries, algorithm versions and judgment conclusions in the test guidelines into a traceable business loop.

[0005] Therefore, this invention provides an intelligent evaluation system that can be reused among multiple vegetable crops. It transforms the DUS / VCU standard rules into computable tasks and uses phenotypic images, depth data, spectral data, environmental information, and genotype information for similar variety retrieval, DUS-assisted judgment, and VCU performance prediction, thereby improving testing efficiency, objectivity, and result consistency. Summary of the Invention

[0006] To address the shortcomings of traditional DUS / VCU testing, which relies on manual observation, is time-consuming, subjective, and lacks consistency across years; varies significantly among different vegetable crops in terms of trait regularity, sampling sites, and sampling times, making it difficult to establish a unified intelligent testing process; relies heavily on expert experience for candidate and approximate reference variety selection, making it difficult to fully utilize historical sample and standard variety libraries; and suffers from the significant environmental influences on yield, quality, resistance, adaptability, and marketability in VCU evaluation, making it difficult to reliably evaluate using a single indicator, this invention provides a multimodal DUS / VCU evaluation system for new vegetable varieties.

[0007] This invention's multimodal evaluation system maps trait clauses, data collection targets, algorithm features, and decision thresholds in the testing guidelines into machine-executable tasks and writes the results back into the candidate variety testing evidence chain. The core of this invention lies in using a standard trait dictionary to drive the data collection tasks and decision-making process, organizing trait clauses, data collection targets, quality control, algorithm features, similar variety retrieval, and reporting fields for different crops into a single closed loop.

[0008] The system is designed with crop adapters for various vegetable crops, including tomatoes, cucumbers, cabbages, green beans, cowpeas, and onions. Each adapter includes the collection period, organ category, trait grade conversion rules, quality control thresholds, review trigger conditions, and model parameters, enabling the same system to reuse the collection and judgment framework across different crops.

[0009] This application provides a DUS / VCU multimodal evaluation system for new vegetable varieties, including the following modules:

[0010] Rule configuration module: Used to build a standard trait dictionary.

[0011] Task generation module: Automatically generates data collection tasks and their numbers based on a standard trait dictionary.

[0012] Data governance module: Processes the collected data and outputs quality control results, organ categories, organ masks, and standardized trait characteristics.

[0013] Similar Variety Search Module: Used to search the reference variety database and obtain a list of similar varieties.

[0014] DUS Auxiliary Judgment Module: Based on a list of similar varieties, auxiliary indicators of specificity, uniformity, and stability are calculated to obtain a list of differential traits.

[0015] VCU prediction module: VCU comprehensive score obtained based on VCU multi-task model.

[0016] Report review module: Used to generate standardized reports, which include raw data, collection task number, collection time, phenotypic images, list of similar varieties, list of differential traits, and VCU comprehensive score.

[0017] In one possible implementation, the rule configuration module specifically implements the following: constructing a standard trait dictionary. The crop code, trait code, trait name, trait type, collection organ, growth period, collection result type, algorithm features, grade conversion rules, and verification rules are associated and stored in the standard trait dictionary.

[0018] In one possible implementation, the task generation module is specifically implemented as follows: based on the crop code, test number, candidate variety, test plot and current growth stage, it reads the organs, angles, quantities, quality requirements and data types to be collected from the standard trait dictionary, and automatically generates collection tasks and numbers; based on the generated collection tasks, it obtains the corresponding raw data, including phenotypic images, depth data, spectral data, environmental sequence characteristics and field yield and quality records.

[0019] In one possible implementation, the data governance module is specifically implemented as follows: performing clarity, exposure, color, scale, and background quality control on the acquired phenotypic images, depth data, and spectral data; after the quality control is qualified, performing organ detection, organ segmentation, and trait extraction in sequence, and outputting the quality control results, organ categories, organ masks, and standardized trait features, wherein the standardized trait features include image traits, depth traits, and spectral features.

[0020] In one possible implementation, the similar variety retrieval module is specifically implemented as follows: constructing a multimodal vector from manually recorded traits, standardized trait features, environmental sequence features, external genotype detection results, and field yield and quality records; retrieving the top K similar varieties in the reference variety library based on vector similarity, and reordering them according to key traits in the testing guidelines, commercial confounding factor, and phylogenetic relationship to obtain a list of similar varieties.

[0021] Among them, artificial traits are derived from testers according to... Recorded phenotypic grades; external genotyping results are derived from molecular testing of candidate variety samples or from established databases.

[0022] In one possible implementation, the VCU prediction module is specifically implemented as follows: standardized trait characteristics, environmental sequence characteristics, external genotype detection results, and field yield and quality records are input into the VCU multi-task model; the VCU multi-task model includes a phenotypic image encoder, an environmental sequence encoder, a genotype encoder, and a cross-modal fusion output layer, and outputs a comprehensive VCU score obtained by combining yield score, quality score, resistance score, stability score, and marketability score.

[0023] In one possible implementation, the phenotypic image encoder takes standardized trait features and field yield and quality records as input, and includes image sub-branches, standardized trait numerical sub-branches, and branch fusion units.

[0024] The image sub-branch employs a CNN visual encoder, including:

[0025] Basic backbone: Lightweight ResNet or ConvNeXt are used.

[0026] Feature extraction layer: Composed of convolutional blocks, batch normalization (BN), and global average pooling (GAP), outputting a high-dimensional visual embedding vector for a single image.

[0027] Image feature projection layer: A fully connected FC layer is used to map visual features to a unified latent space dimension D to obtain the image embedding.

[0028] The standardized trait numerical sub-branch adopts a Tabular numerical encoder. For discrete qualitative traits, an embedding layer is used; for continuous quantitative traits, a normalization layer and a multilayer perceptron (MLP) are used; finally, the standardized numerical trait embedding of dimension D is output through the numerical feature fusion FC layer.

[0029] The branch fusion unit is used to add the image embedding and the standardized numerical trait embedding element by element and then normalize them to obtain a unified phenotypic latent vector, which serves as the final output of the phenotypic image encoder.

[0030] In one possible implementation, the environment sequence encoder takes environment sequence features as input and processes them sequentially through a sequence preprocessing layer, a temporal backbone layer, a sequence global aggregation layer, and a dimension alignment FC layer.

[0031] Sequence preprocessing layer: includes missing value imputation, standardization, and one-dimensional convolutional short-time local feature extraction.

[0032] The temporal backbone layer adopts a bidirectional LSTM to capture environmental interactions between different growth stages and different pilot sites through multi-head self-attention; and distinguishes growth time nodes such as sowing, grain filling, and harvesting through position encoding.

[0033] Sequence Global Aggregation Layer: Compresses the entire environmental time series into a fixed-length vector through global average pooling in the time dimension.

[0034] Dimension-aligned FC layer: Maps the fixed-length vector output by the sequence global aggregation layer to the latent space dimension D, and outputs the environment latent vector.

[0035] In one possible implementation, the genotype encoder takes external genotype detection results as input and processes them sequentially through a site embedding layer, a genotype global modeling layer, a whole genome aggregation layer, and a dimension alignment FC layer.

[0036] Site embedding layer: Each genotype site is independently embedded, encoding discrete genotypes into low-dimensional site vectors.

[0037] Genotype global modeling layer: One-dimensional convolution Conv1D is used to capture linkage imbalance features of adjacent markers.

[0038] Whole genome aggregation layer: Global pooling and integration of genetic information from all loci.

[0039] Dimension-aligned FC layer: Maps the output of the whole genome aggregation layer to the latent space dimension D, and outputs the genotype latent vector with the same dimension as other branches.

[0040] In one possible implementation, the cross-modal fusion output layer receives the output vectors of three encoders, obtains five sub-scores (yield, quality, resistance, stability, and marketability) through the cross-modal fusion unit and the multi-task prediction head branch, and finally performs weighted aggregation through the VCU comprehensive score calculation module to output the final VCU comprehensive score.

[0041] The cross-modal fusion unit employs cross-modal cross-attention fusion to output interactive fusion features.

[0042] The multi-task prediction head includes five independent parallel regression prediction heads. Each prediction head has a uniform structure, consisting of an FC hidden layer, ReLU activation, Dropout regularization, and an output sigmoid, which outputs a single standardized score.

[0043] The beneficial effects of this invention are as follows:

[0044] (1) This invention enables the unified configuration, automatic generation and reuse of DUS / VCU testing processes for various vegetable crops through a standard trait dictionary and crop adapter.

[0045] (2) This invention forms a traceable chain of varieties, plots, plants, organs, images, traits, model versions and reports through multimodal acquisition and data governance.

[0046] (3) This invention improves the efficiency of reference variety screening and reduces the risk of missing similar varieties by using similar variety vector retrieval and key trait reordering.

[0047] (4) This invention improves the consistency, interpretability and verification efficiency of test conclusions by using DUS multi-task assisted judgment and VCU multi-environment prediction. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall system structure according to an embodiment of the present invention.

[0049] Figure 2 This is a flowchart illustrating the construction process of a standard trait dictionary according to an embodiment of the present invention.

[0050] Figure 3 This is a flowchart illustrating the multimodal data acquisition process according to an embodiment of the present invention.

[0051] Figure 4 This is a flowchart of the similar variety retrieval and DUS determination process in an embodiment of the present invention. Detailed Implementation

[0052] The present invention will be described below with reference to the accompanying drawings and embodiments. The described embodiments are used to explain the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions for the acquisition device, model structure, threshold configuration, and data storage method.

[0053] In this implementation, the system establishes test tasks using the test number as the primary key. Multiple sample records are generated for each candidate variety across different plots, years, and locations. These sample records are further bound to individual plant numbers, organ numbers, image numbers, environmental numbers, algorithm model versions, and manual review records.

[0054] like Figure 1 As shown in the figure, this application provides a DUS / VCU multimodal evaluation system for new vegetable varieties, including the following modules:

[0055] Rule configuration module: such as Figure 2 As shown, the rule configuration module imports crop testing guidelines, DUS / VCU testing rules, and internal testing procedures into the standard rule layer and parses them into a standard trait dictionary. The parsing operations include extracting trait codes, harvested organs, reproductive stage, harvesting angle, trait type, grade conversion rules, algorithm features, and verification rules. This dictionary is continuously called during harvesting task generation, trait extraction, DUS determination, and report generation, serving as the basis for rule configuration and judgment. Standard Trait Dictionary Includes the following fields.

[0056]

[0057] Task generation module: Automatically generates data collection tasks and their numbers based on a standard trait dictionary.

[0058] like Figure 3As shown, the data acquisition process first generates acquisition tasks based on candidate varieties, experimental plots, and crop adapters. An experimental plot is the smallest observation unit for the same candidate variety at the same location, year, replicate, and cultivation treatment; a crop adapter is a set of rules in a standard trait dictionary corresponding to the crop code. The system matches the acquisition organs, angles, quantities, and quality requirements according to the crop code and growth stage, thereby generating acquisition tasks. Based on the generated acquisition tasks, corresponding raw data are obtained, including phenotypic images, depth data, spectral data, environmental sequence characteristics, and field yield and quality records.

[0059] Data governance module: Processes the collected data and outputs quality control results, organ categories, organ masks, and standardized trait characteristics. Specifically:

[0060] Quality control is performed on the acquired phenotypic images, depth data, and spectral data, including sharpness detection, exposure detection, color correction, scale correction, and background consistency detection. After passing quality control, organ detection, organ segmentation, and trait extraction (the organs include plants, leaves, flowers, fruits, pods, bulbs, or rhizomes) are performed sequentially. The organ detection model first obtains the target bounding box, and then the instance segmentation model obtains the organ contour and mask. In one embodiment, the organ detection model is a convolutional neural network or a Transformer detector, and the instance segmentation model is a Mask R-CNN or a semantic segmentation network. The segmentation result formula is expressed as follows:

[0061]

[0062] Where O represents the set of organ detection and segmentation results in the same image, This represents the i-th organ object. Represents the target box. Indicates organ masking. Indicates organ category, Indicates the confidence level of the detection. Indicates the individual plant number. Indicates the organ number.

[0063] organ masking The corresponding pixel regions are used to calculate color, morphology, texture, 3D structure, and growth dynamics features. Color features are calculated from the RGB, HSV, and CIELAB statistical values ​​and color difference within the mask; morphological features are calculated from the area, perimeter, major axis, minor axis, aspect ratio, roundness, and symmetry of the mask contour; texture features are obtained from the gray-level co-occurrence matrix, local binary pattern, and deep network embedding features; and 3D structure and growth dynamics features are obtained by fitting depth data, multi-time point images, or environmental sequences, respectively.

[0064] Similar Variety Search Module: Used to search the reference variety database and obtain a list of similar varieties.

[0065] In one possible implementation, the approximate variety retrieval module is specifically implemented as follows:

[0066] like Figure 4 As shown, the approximate variety retrieval module constructs a multimodal vector for each candidate variety. This vector includes manually recorded traits, standardized trait characteristics, environmental sequence characteristics, external genotype detection results, and field yield and quality records. Among these, the manually recorded traits are derived from the testers' observations... Recorded phenotypic grades; external genotyping results are derived from molecular testing of candidate variety samples or from established databases.

[0067]

[0068] Where V represents the multimodal vector of the candidate variety. Indicates traits recorded manually. Represents standardized trait characteristics. This indicates field yield and quality records. Indicates environmental sequence characteristics, This indicates the results of external genotyping testing.

[0069] The similarity between candidate variety A and reference variety B can be calculated using the following formula:

[0070]

[0071] in, , , , As weight, and The system retrieves the top K reference varieties with the highest similarity from the vector index of the reference variety library based on the aforementioned similarity, forming an initial list of similar varieties. Subsequently, the initial list is reordered based on key trait differences in the testing guidelines, commercial confounding factors, and phylogenetic relationships to obtain a final list of similar varieties. Modal weights for varieties that have not passed quality control or are not applicable to the current crop are reset to 0 and do not participate in the ranking of similar varieties for the current sample.

[0072] DUS Auxiliary Judgment Module: Based on the list of similar varieties, auxiliary indices of specificity, uniformity, and stability are calculated to obtain a list of differential traits. Among them, the specificity index is used to locate the differential traits of candidate varieties relative to similar varieties, the uniformity index is used to locate outliers within candidate varieties, and the stability index is used to locate traits that are unstable across years or locations.

[0073] Specificity determination includes differences in trait grades and differences in continuous variables. For quantitative trait j, the system calculates the standardized differences between candidate variety A and reference variety B. If the difference exceeds a preset threshold, and the difference remains consistent in the same direction across different replicates, years, or locations, then trait j is marked as a differential trait supporting specificity and added to the differential trait list.

[0074]

[0075] in, and Let J represent the mean values ​​of candidate variety A and reference variety B on quantitative trait j, respectively. and They represent the corresponding variances, and These represent the sample size; The larger the value, the more stable the difference in trait j between the two varieties.

[0076] Consistency determination is achieved through the coefficient of variation, outlier distance, and proportion of heteromorphic plants within candidate varieties. Plants exceeding preset variation or outlier thresholds are marked as internal outliers. Stability determination is achieved through a cross-year, cross-location mixed linear model, treating year, location, and replication as environmental or random effects to estimate trait stability. The above statistical calculations belong to a reusable conventional statistical framework in DUS testing; the improvement of this invention lies in binding it to data collection tasks, evidence images, similar variety retrieval, and report traceability. In consistency determination, the coefficient of variation threshold, outlier distance threshold, and heteromorphic plant proportion threshold are determined by the corresponding crop... In the configuration and stability assessment, the year, location, and repeatability settings are linked to the test number, and the output stability risk items are written back to the list of differential traits and the manual review list.

[0077] VCU prediction module: VCU comprehensive score obtained based on VCU multi-task model.

[0078] The VCU multi-task model receives standardized trait features, environmental sequence features, external genotype detection results, and field yield and quality records. The VCU multi-task model includes a phenotypic image encoder, an environmental sequence encoder, a genotype encoder, and a cross-model fusion output layer. The phenotypic image encoder extracts plant growth, canopy structure, and organ status; the environmental sequence encoder extracts temperature, humidity, light, and soil changes; the genotype encoder extracts genetic background information; and the cross-model fusion output layer outputs a comprehensive VCU score, which includes yield, quality, resistance, stability, and marketability scores.

[0079] The VCU overall score is calculated using the following formula:

[0080]

[0081] in, This indicates the overall VCU score. , , , and These represent yield score, quality score, resistance score, stability score, and marketability score, respectively. For the corresponding weights, and The weights for different crops are configured by the standard trait dictionary or VCU evaluation procedure.

[0082] Different crops can be assigned different weights. For example, tomatoes can have their quality and marketability weighted higher, cucumbers can have their marketable fruit rate and disease resistance weighted higher, and cabbage can have its shape, firmness, and crack resistance weighted higher. The weights for different crops are read from the VCU evaluation procedure field in the standard trait dictionary, rather than being selected by the operator on the spot. For example, tomatoes can have their quality and marketability weighted higher, cucumbers can have their marketable fruit rate and disease resistance weighted higher, and cabbage can have its shape, firmness, and crack resistance weighted higher, and the corresponding weight version is recorded in the report.

[0083] In one possible implementation, the phenotypic image encoder takes standardized trait features and field yield and quality records as input, and includes image sub-branches, standardized trait numerical sub-branches, and branch fusion units.

[0084] The image sub-branch employs a CNN visual encoder, including:

[0085] Basic backbone: Lightweight ResNet or ConvNeXt (adapted to small samples of agricultural field images).

[0086] Feature extraction layer: Composed of convolutional blocks, batch normalization (BN), and global average pooling (GAP), outputting a high-dimensional visual embedding vector for a single image.

[0087] Image feature projection layer: A fully connected FC layer is used to map visual features to a unified latent space dimension D (aligned with the environment and genotype embedding dimensions) to obtain image embedding.

[0088] The standardized trait numerical sub-branch adopts a Tabular numerical encoder. For discrete qualitative traits, an embedding layer (such as disease resistance level and maturity classification) is used; for continuous quantitative traits, a normalization layer and a multilayer perceptron (MLP) are used; finally, the standardized numerical trait embedding of dimension D is output through the numerical feature fusion FC layer.

[0089] The branch fusion unit is used to add the image embedding and the standardized numerical trait embedding element by element and then normalize them to obtain a unified phenotypic latent vector H_phen, which serves as the final output of the phenotypic image encoder.

[0090] In one possible implementation, the environment sequence encoder takes environment sequence features as input and processes them sequentially through a sequence preprocessing layer, a temporal backbone layer, a sequence global aggregation layer, and a dimension alignment FC layer.

[0091] Sequence preprocessing layer: includes missing value imputation, standardization, and one-dimensional convolutional short-time local feature extraction.

[0092] The temporal backbone layer adopts a bidirectional LSTM to capture environmental interactions between different growth stages and different pilot sites through multi-head self-attention; and distinguishes growth time nodes such as sowing, grain filling, and harvesting through position encoding.

[0093] Sequence Global Aggregation Layer: Compresses the entire environmental time series into a fixed-length vector through global average pooling in the time dimension.

[0094] Dimension-aligned FC layer: Maps the fixed-length vector output by the sequence global aggregation layer to the latent space dimension D, and outputs the environment latent vector H_env.

[0095] In one possible implementation, the genotype encoder takes external genotype detection results as input and processes them sequentially through a site embedding layer, a genotype global modeling layer, a whole genome aggregation layer, and a dimension alignment FC layer.

[0096] Site embedding layer: Each genotype site is independently embedded, encoding discrete genotypes into low-dimensional site vectors.

[0097] Genotype global modeling layer: One-dimensional convolution Conv1D is used to capture linkage imbalance features of adjacent markers.

[0098] Whole genome aggregation layer: Global pooling and integration of genetic information from all loci.

[0099] Dimension-aligned FC layer: Maps the output of the whole genome aggregation layer to the latent space dimension D, and outputs the genotype latent vector H_gen with the same dimension as other branches.

[0100] In one possible implementation, the cross-modal fusion output layer receives the output vectors H_phen, H_env, and H_gen from three encoders, obtains five sub-scores (yield, quality, resistance, stability, and marketability) through the cross-modal fusion unit and the multi-task prediction head branch, and finally performs weighted aggregation through the VCU comprehensive score calculation module to output the final VCU comprehensive score.

[0101] The cross-modal fusion unit employs cross-modal cross-attention fusion, using one branch as the Query and the other two branches as the Key and Value, respectively, to calculate cross-modal interactive attention, capture the interaction effects of "genotype, environment, and phenotype", and output the interactive fusion feature H_fuse.

[0102] The multi-task prediction head includes five independent parallel regression prediction heads. Each prediction head has a uniform structure, consisting of an FC hidden layer, ReLU activation, Dropout regularization, and an output sigmoid, which outputs a single standardized score.

[0103] Report review module: Used to generate standardized reports, which include raw data, collection task number, collection time, phenotypic images, list of similar varieties, list of differential traits, and VCU comprehensive score.

[0104] Each conclusion in the report can be traced back to the corresponding data collection task, image number, organ number, algorithm version, and review record. For samples where the model confidence is lower than a preset threshold, the distance to the key trait judgment boundary is less than a preset tolerance, the similarity between similar varieties is higher than a preset threshold, or the conclusions of different modalities are inconsistent, the system does not directly give a final conclusion. Instead, it generates a review task according to the review rules in the standard trait dictionary and saves the manual review results to the review record table.

[0105] The manual review task should include at least the triggering rule, the traits involved, the evidence image number, the similar variety number, the model version, the suggested review method, and the review conclusion field. After the review conclusion is bound to the collection task number, it enters the report and interface layer and can be used as the basis for subsequent threshold adjustment or model update.

[0106] Example

[0107] Example 1 is a DUS-assisted test for a new tomato variety. The system collects images of plant growth habits, leaves, and inflorescences during the flowering period, and longitudinal and transverse cross-sectional images of the fruit during the commercial ripening period. The algorithm segments the fruit outline, calculates the fruit height, maximum width, aspect ratio, roundness, shoulder shape, apex shape, and ripe fruit color, and converts the continuous indicators into DUS trait grades.

[0108] Example 2 is a DUS-assisted test for a new cucumber variety. The system acquires images of the front, side, and end of the cucumber, segments the cucumber body, tubercles, stripes, and pedicel areas, and calculates the cucumber length, diameter, length-to-diameter ratio, curvature, tubercle density, tubercle size, peel color, and stripe length ratio.

[0109] Example 3 is a combined VCU evaluation of cabbage. The system collects top view, side view and head cross-section of the plant, calculates the head diameter, height, height-to-width ratio, head shape index, outer leaf coverage and compactness proxy index, and combines maturity period, splitting risk and multi-environment yield to form a VCU score.

[0110]

[0111] Effect verification

[0112] The invention was validated using multi-crop, multi-location, and multi-year trial data. Candidate and reference varieties were numbered according to a unified data collection task, and manual measurement results and expert judgment results were used as controls.

[0113]

[0114] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0116] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A DUS / VCU multimodal evaluation system for new vegetable varieties, characterized in that, Includes the following modules: Rule configuration module: used to build a standard trait dictionary Task generation module: Automatically generates data collection tasks and their numbers based on a standard trait dictionary; Data governance module: processes the collected data and outputs quality control results, organ categories, organ masks, and standardized trait characteristics; Similar Variety Retrieval Module: Used to search the reference variety database and obtain a list of similar varieties; DUS auxiliary judgment module: Based on the list of similar varieties, calculate the auxiliary indicators of specificity, uniformity and stability to obtain a list of differential traits; VCU prediction module: VCU comprehensive score obtained based on VCU multi-task model; Report review module: Used to generate standardized reports, which include raw data, collection task number, collection time, phenotypic images, list of similar varieties, list of differential traits, and VCU comprehensive score.

2. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 1, characterized in that, The rule configuration module is implemented as follows: Constructing a standard trait dictionary. The crop code, trait code, trait name, trait type, collection organ, growth period, collection result type, algorithm features, grade conversion rules, and verification rules are associated and stored in the standard trait dictionary.

3. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 1, characterized in that, The task generation module is implemented as follows: Based on the crop code, test number, candidate variety, test plot and current growth stage, it reads the organs, angles, quantities, quality requirements and data types to be collected from the standard trait dictionary, and automatically generates collection tasks and numbers; based on the generated collection tasks, it obtains the corresponding raw data, including phenotypic images, depth data, spectral data, environmental sequence characteristics and field yield and quality records.

4. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 1, characterized in that, The data governance module is implemented as follows: it performs clarity, exposure, color, scale, and background quality control on the acquired phenotypic images, depth data, and spectral data; after the quality control is qualified, it performs organ detection, organ segmentation, and trait extraction in sequence, and outputs the quality control results, organ categories, organ masks, and standardized trait features, which include image traits, depth traits, and spectral features.

5. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 1, characterized in that, The similar variety retrieval module is implemented as follows: It constructs a multimodal vector from manually recorded traits, standardized trait features, environmental sequence features, external genotype detection results, and field yield and quality records; it retrieves the top K similar varieties in the reference variety library based on vector similarity, and reorders them according to key traits in the testing guidelines, commercial confounding factor, and phylogenetic relationship to obtain a list of similar varieties; Among them, artificial traits are derived from testers according to... Recorded phenotypic grades; external genotyping results are derived from molecular testing of candidate variety samples or from established databases.

6. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 1, characterized in that, The VCU prediction module is implemented as follows: standardized trait characteristics, environmental sequence characteristics, external genotype detection results, and field yield and quality records are input into the VCU multi-task model; the VCU multi-task model includes a phenotypic image encoder, an environmental sequence encoder, a genotype encoder, and a cross-modal fusion output layer, and outputs a comprehensive VCU score obtained from a comprehensive yield score, quality score, resistance score, stability score, and marketability score.

7. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 6, characterized in that, The phenotypic image encoder takes standardized trait features and field yield and quality records as input, and includes image sub-branch, standardized trait numerical sub-branch and branch fusion unit; The image sub-branch employs a CNN visual encoder, including: Basic backbone: Lightweight ResNet or ConvNeXt are used; Feature extraction layer: Composed of convolutional blocks, BN and GAP global average pooling, outputting a high-dimensional visual embedding vector for a single image; Image feature projection layer: A fully connected FC layer is used to map visual features to a unified latent space dimension D to obtain image embedding; The standardized trait numerical sub-branch adopts a Tabular numerical encoder. For discrete qualitative traits, an embedding layer is used; for continuous quantitative traits, a normalization layer and a multilayer perceptron (MLP) are used; finally, the standardized numerical trait embedding of dimension D is output through the numerical feature fusion FC layer. The branch fusion unit is used to add the image embedding and the standardized numerical trait embedding element by element and then normalize them to obtain a unified phenotypic latent vector, which serves as the final output of the phenotypic image encoder.

8. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 6, characterized in that, The environment sequence encoder takes environment sequence features as input and processes them sequentially through a sequence preprocessing layer, a temporal backbone layer, a sequence global aggregation layer, and a dimension alignment FC layer. Sequence preprocessing layer: includes missing value imputation, normalization, and one-dimensional convolutional short-time local feature extraction; The temporal backbone layer employs a bidirectional LSTM to capture environmental interactions between different reproductive stages and pilot sites through multi-head self-attention; and distinguishes reproductive time-series nodes through position encoding. Sequence Global Aggregation Layer: Compresses the entire environmental time series into a fixed-length vector through global average pooling in the time dimension; Dimension-aligned FC layer: Maps the fixed-length vector output by the sequence global aggregation layer to the latent space dimension D, and outputs the environment latent vector.

9. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 6, characterized in that, The genotype encoder takes the external genotype detection results as input and processes them sequentially through the site embedding layer, the genotype global modeling layer, the whole genome aggregation layer, and the dimension alignment FC layer. Locus embedding layer: Each genotype locus is independently embedded, encoding discrete genotypes into low-dimensional locus vectors; Genotype global modeling layer: One-dimensional convolution Conv1D is used to capture linkage imbalance features of adjacent markers; Whole-genome aggregation layer: Globally pools and integrates genetic information from all loci; Dimension-aligned FC layer: Maps the output of the whole genome aggregation layer to the latent space dimension D, and outputs the genotype latent vector with the same dimension as other branches.

10. The DUS / VCU multimodal evaluation system for new vegetable varieties according to claim 6, characterized in that, The cross-modal fusion output layer receives the output vectors of three encoders, obtains five sub-scores (yield, quality, resistance, stability, and marketability) through the cross-modal fusion unit and the multi-task prediction head branch, and finally performs weighted aggregation through the VCU comprehensive score calculation module to output the final VCU comprehensive score. The cross-modal fusion unit employs cross-modal cross-attention fusion to output interactive fusion features; The multi-task prediction head includes five independent parallel regression prediction heads. Each prediction head has a uniform structure, consisting of an FC hidden layer, ReLU activation, Dropout regularization, and an output sigmoid, which outputs a single standardized score.

Images