Dendrobium breeding method based on deep learning

Abstract

The application discloses a kind of based on deep learning's dendrobium breeding method, which comprises the following steps: deploying multi-source sensor periodically automatic acquisition dendrobium plant from seedling to mature period image data, pre-processing to the image data collected;Double-branch semantic segmentation model is constructed, and stem and leaf are accurately segmented;EfficientNet or Vision Transformer based image classification model is constructed, and the phenological stage such as flower bud differentiation period, initial flowering period, flowering period is identified;Regression model based on 3D-CNN or Spectral-Attention mechanism is constructed, and the inversion model of spectral characteristics and key biochemical indexes such as leaf chlorophyll content, polysaccharide content is established, and the above model is trained and optimized;The multidimensional phenotype data output by each model is integrated, and a multi-trait comprehensive selection index is constructed;Each single plant in breeding population is comprehensively scored and ranked using mathematical model.The application realizes the automatic, objective and accurate collection and analysis of the whole cycle, multi-dimensional and high-throughput phenotype information of dendrobium.

Description

Technical Field

[0001] This invention belongs to the field of plant breeding technology and relates to a Dendrobium breeding method based on deep learning. Background Technology

[0002] Dendrobium is a precious Chinese medicinal herb and ornamental plant. Its breeding work is of great significance for improving the quality, yield and stress resistance of medicinal materials. Traditional Dendrobium breeding methods mainly rely on hybridization and molecular marker-assisted selection, but these methods have many bottlenecks. Breeders mainly rely on visual observation and experience to subjectively evaluate and screen the phenotypic characteristics of Dendrobium (such as stem thickness, internode length, leaf morphology, flower color, flowering period, etc.), select parents for hybridization, and then select superior individual plants from a large number of offspring. At present, the following problems exist: (1) strong subjectivity: artificial evaluation is easily affected by environmental and human subjective factors, lacks objective and unified quantitative standards, and has low accuracy; (2) long cycle: Dendrobium has a long growth cycle. It takes several years or even longer from hybridization to the stable expression of offspring traits, resulting in extremely low breeding efficiency; (3) low efficiency: it is impossible to deal with large-scale populations, and can only select a few key traits, making it difficult to conduct comprehensive analysis of multiple traits.

[0003] With the rapid development of deep learning technology, methods for plant phenotypic analysis and recognition are constantly improving. However, in the context of Dendrobium breeding, achieving automated, high-throughput acquisition and accurate analysis of multi-dimensional phenotypic information covering the entire growth cycle from seedling to maturity, encompassing morphology, physiology, and biochemistry, remains a significant challenge. Especially in actual planting environments, Dendrobium plants often exhibit complex phenomena such as stem adhesion, overlapping leaves, and mutual occlusion, making image-based phenotypic feature extraction highly susceptible to errors. Existing technologies struggle to effectively address these issues, resulting in low-quality phenotypic data that cannot support scientific breeding decisions. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, this invention provides a deep learning-based Dendrobium breeding method that can effectively improve the efficiency and accuracy of phenotypic information collection and analysis.

[0005] The technical solution adopted in this invention is as follows:

[0006] A deep learning-based method for Dendrobium breeding includes the following steps:

[0007] S1, Multi-source image data acquisition and preprocessing, the specific steps are as follows:

[0008] S101 deploys multi-source sensors in the Dendrobium breeding nursery to automatically collect image data of Dendrobium plants from seedling to maturity on a regular basis;

[0009] The image data includes at least RGB images and 3D point cloud data;

[0010] S102, preprocess the collected image data to construct a standardized Dendrobium multimodal dataset;

[0011] The preprocessing includes image cropping, rotation, scaling, noise reduction, background segmentation, image alignment, and data augmentation;

[0012] S2, the construction and training of deep learning models for key phenotypic traits, the specific steps are as follows:

[0013] S201, for the morphological traits of Dendrobium plant height, stem diameter, number of leaves and leaf area, the preprocessed RGB image and 3D point cloud data are input into the trained bi-branch semantic segmentation model. The model accurately segments the stems and leaves of Dendrobium by fusing RGB features and 3D geometric features, and calculates the morphological parameters of the stems and leaves.

[0014] S202 constructs a multimodal deep learning model based on EfficientNet or Vision Transformer. The model extracts texture, color, and morphological features of flower buds, petals, and inflorescence regions, and enhances the representation of key features using an attention mechanism, enabling automatic identification of phenological stages such as flower bud differentiation, initial flowering, and full bloom. Simultaneously, flower color and flower shape classification branches are set up to automatically distinguish flower color and shape; further, by combining continuous time-series images or three-dimensional morphological parameters, the developmental status of Dendrobium flowering can be predicted.

[0015] S203, for the initial screening of biochemical traits for growth vigor and prediction of potential medicinal components, combines hyperspectral image data to construct a regression model based on 3D-CNN or Spectral-Attention mechanism, and establishes an inversion model of spectral characteristics and key biochemical indicators such as leaf chlorophyll content and polysaccharide content.

[0016] S204, the image dataset preprocessed in S102 is divided into training set, validation set and test set, which are used to train and optimize the models in S201, S202 and S203 until the model reaches the predetermined accuracy on the test set.

[0017] S3, Comprehensive Evaluation and Decision-Making of Breeding Value, the specific steps are as follows:

[0018] S301, integrate the multi-dimensional phenotypic data output by each model in step S2, normalize the phenotypic indicators such as plant height, stem diameter, leaf area, flowering period grade, chlorophyll content and polysaccharide content by range standardization, and use the analytic hierarchy process, entropy weight method or machine learning method to determine the weight of each indicator, construct a comprehensive phenotypic scoring function, and form a multi-trait comprehensive selection index.

[0019] S302 uses a mathematical model to comprehensively score and rank each individual plant in the breeding population;

[0020] The mathematical model described is a weighted summation method or a machine learning ranking algorithm;

[0021] S303 allows breeders to quickly select top-ranked superior plants as parents for the next generation or directly enter regional trials.

[0022] Furthermore, the dual-branch semantic segmentation model described in step S201 above is an encoder-decoder structure, including an RGB branch, a point cloud branch, and a feature fusion module;

[0023] The RGB branch is used to extract color and texture features from RGB images; the point cloud branch is used to extract three-dimensional geometric features from three-dimensional point cloud data; and the feature fusion module is used to fuse the three-dimensional geometric features with the color and texture features at at least one level of the decoder, so as to use the three-dimensional geometric features to help solve the problem of stem adhesion and overlap in RGB images.

[0024] Furthermore, the feature fusion module described above is a geometric attention fusion module, which generates an attention weight map based on the three-dimensional geometric features extracted from the point cloud branch, and uses the attention weight map to modulate the color and texture feature maps extracted from the RGB branch to enhance the boundary features of different stem individuals in spatial structure.

[0025] Furthermore, the loss function used for training the bi-branch semantic segmentation model described in step S201 above is a composite loss function, including: standard cross-entropy loss for pixel-level classification; boundary loss for enhancing the accuracy of segmentation boundaries; and adhesion penalty loss for penalizing different stalk instances that are incorrectly predicted as the same connected region; the composite loss function = α * standard cross-entropy loss + β * boundary loss + γ * adhesion penalty term, where α, β, and γ are hyperparameters.

[0026] Furthermore, the three-dimensional point cloud data described in step S101 above is spatially aligned with the RGB image through calibration and registration.

[0027] Furthermore, the multi-source sensors mentioned in step S101 above include a visible light RGB camera, a hyperspectral camera, a multispectral camera, and a 3D laser scanner.

[0028] Furthermore, the morphological parameters of the stem mentioned in step S201 above include stem volume and spatial distribution information.

[0029] Furthermore, in step S301 above, the weights of different phenotypic data in the multi-trait comprehensive selection index are dynamically adjusted according to the phenological period identified in S202.

[0030] The advantages and beneficial effects of this invention are as follows:

[0031] 1. This invention achieves automatic, objective, and accurate collection and analysis of phenotypic information of Dendrobium officinale throughout its entire growth cycle, in multiple dimensions, and with high throughput, through the automatic acquisition of multi-source sensors and the construction of deep learning models. This breakthrough overcomes the bottleneck of traditional manual sampling and significantly improves breeding efficiency.

[0032] 2. The deep learning model constructed in this invention can accurately analyze the morphological, physiological and biochemical characteristics of Dendrobium, overcoming the shortcomings of traditional manual subjective judgment which is easily interfered with by subjective factors, and significantly improving the accuracy and stability of phenotypic identification.

[0033] 3. This invention addresses the characteristics of Dendrobium plant stems, which grow densely and are prone to adhesion and occlusion in two-dimensional images. It employs a bi-branch semantic segmentation model. This model, by fusing two-dimensional image information with geometric spatial information provided by three-dimensional point clouds, can effectively distinguish visually adhered individual stems, significantly reducing under-segmentation and over-segmentation in image segmentation and improving the accuracy of stem and leaf segmentation. Based on the accurate segmentation results, the reliability of subsequently calculated stem morphological parameters (such as length, diameter, and quantity) is higher, and measurement errors are significantly reduced, providing a more accurate data foundation for breeding decisions. Furthermore, combined with three-dimensional point cloud data, this invention can directly calculate three-dimensional morphological parameters that are difficult to obtain from traditional two-dimensional images, such as the true volume of the stems and their distribution information in three-dimensional space, achieving a more comprehensive and accurate quantitative analysis of Dendrobium plant morphology.

[0034] 4. This invention is based on the prediction of biochemical traits using a hyperspectral model, which allows for the screening of potential plant qualities in the early stages of growth without destructive sampling, thus shortening the breeding cycle by about 30%-50% and saving a lot of manpower and resources.

[0035] 5. This invention can simultaneously perform parallel analysis on multiple key traits and optimize them in a comprehensive selection index, solving the problem of neglecting one aspect for another in traditional breeding, forming a more scientific and precise breeding selection mechanism, and effectively improving breeding efficiency and success rate. Attached Figure Description

[0036] Figure 1 This diagram illustrates the detection and segmentation results of Dendrobium leaves according to the present invention. The diagram shows the recognition performance of the model used in this invention on leaf instances under conventional single-plant or sparse planting conditions.

[0037] Figure 2This diagram illustrates the detection and segmentation results of the present invention for a densely overlapping Dendrobium leaves scenario. The diagram shows the detection results of the model used in this invention in a complex scenario where leaves overlap and are spatially close, with the model successfully identifying 5 leaf instances while maintaining a high confidence level. Detailed Implementation

[0038] Data preparation: Collect a sufficient amount of data.

[0039] In traditional Dendrobium breeding, the evaluation of phenotypic characteristics is mainly dominated by manual observation, which makes the evaluation results susceptible to interference from environmental conditions and subjective human factors, and lacks objective and unified quantitative standards. At the same time, Dendrobium has a long growth cycle, and it takes several years from hybridization to the stable expression of traits in offspring, resulting in low breeding efficiency. In addition, in complex planting environments, the frequent occurrence of adjacent leaves, overlapping, shading, and adhesion makes it difficult to automatically, objectively, and accurately collect and analyze multi-dimensional phenotypic information, thus hindering the implementation of high-throughput breeding technology.

[0040] For example, in the specific implementation of Dendrobium breeding nurseries, when the plants are in the vigorous growth period, the stems are often partially covered by densely growing leaves, making it impossible to accurately locate the measurement position when manually measuring stem diameter due to leaf obstruction. Furthermore, in the flowering prediction stage, subtle morphological changes during flower bud differentiation are easily overlooked by observers, leading to delays in phenological period judgment. As a result, breeders find it difficult to obtain accurate data on plant height, leaf area, and flowering period in a timely manner, affecting subsequent selection of hybrid parents and individual plant screening decisions.

[0041] If the above problems are not solved, the Dendrobium breeding process will be hindered, making it impossible to achieve comprehensive analysis and evaluation of multiple traits, resulting in a longer development cycle for superior varieties and limited improvement in the quality and yield of medicinal materials. At the same time, biases in subjective evaluation may lead to incorrect screening, increase the ineffective consumption of breeding resources, and thus affect the overall effectiveness of the breeding system.

[0042] In response, this application proposes a deep learning-based method for Dendrobium breeding, comprising the following steps:

[0043] This invention provides a deep learning-based method for Dendrobium breeding, comprising the following steps:

[0044] S1, Multi-source image data acquisition and preprocessing, the specific steps are as follows:

[0045] S101 deploys multi-source sensors in the Dendrobium breeding nursery to automatically collect image data of Dendrobium plants from seedling to maturity on a regular basis;

[0046] The image data includes at least RGB images and 3D point cloud data;

[0047] S102, preprocess the collected image data to construct a standardized Dendrobium multimodal dataset;

[0048] The preprocessing includes image cropping, rotation, scaling, noise reduction, background segmentation, image alignment, and data augmentation;

[0049] S2, the construction and training of deep learning models for key phenotypic traits, the specific steps are as follows:

[0050] S201, for the morphological traits of Dendrobium plant height, stem diameter, number of leaves and leaf area, the preprocessed RGB image and 3D point cloud data are input into the trained bi-branch semantic segmentation model. The model accurately segments the stems and leaves of Dendrobium by fusing RGB features and 3D geometric features, and calculates the morphological parameters of the stems and leaves.

[0051] S202 constructs a multimodal deep learning model based on EfficientNet or Vision Transformer. The model extracts texture, color, and morphological features of flower buds, petals, and inflorescence regions, and enhances the representation of key features using an attention mechanism, enabling automatic identification of phenological stages such as flower bud differentiation, initial flowering, and full bloom. Simultaneously, flower color and flower shape classification branches are set up to automatically distinguish flower color and shape; further, by combining continuous time-series images or three-dimensional morphological parameters, the developmental status of Dendrobium flowering can be predicted.

[0052] S203, for the initial screening of biochemical traits for growth vigor and prediction of potential medicinal components, combines hyperspectral image data to construct a regression model based on 3D-CNN or Spectral-Attention mechanism, and establishes an inversion model of spectral characteristics and key biochemical indicators such as leaf chlorophyll content and polysaccharide content.

[0053] S204, the image dataset preprocessed in S102 is divided into training set, validation set and test set, which are used to train and optimize the models in S201, S202 and S203 until the model reaches the predetermined accuracy on the test set.

[0054] S3, Comprehensive Evaluation and Decision-Making of Breeding Value, the specific steps are as follows:

[0055] S301, integrate the multi-dimensional phenotypic data output from each model in step S2, normalize the phenotypic indicators such as plant height, stem diameter, leaf area, flowering period grade, chlorophyll content, and polysaccharide content using range standardization, and determine the weights of each indicator using the analytic hierarchy process, entropy weight method, or machine learning method to construct a comprehensive phenotypic scoring function, forming a multi-trait comprehensive selection index; among which, the chlorophyll content and polysaccharide content of leaves are determined by conventional laboratory chemical analysis methods;

[0056] S302 uses a mathematical model to comprehensively score and rank each individual plant in the breeding population;

[0057] The mathematical model described is a weighted summation method or a machine learning ranking algorithm; as an example, the mathematical model for the comprehensive score of a single plant can apply a support vector machine or neural network ranking algorithm, for example, outputting a ranking score based on feature vectors, thereby achieving an objective quantitative evaluation;

[0058] S303 allows breeders to quickly select top-ranked superior plants as parents for the next generation or directly enter regional trials.

[0059] The three-dimensional point cloud data described in step S101 above is spatially aligned with the RGB image through calibration and registration.

[0060] The multi-source sensors mentioned in step S101 above include a visible light RGB camera, a hyperspectral camera, a multispectral camera, and a 3D laser scanner.

[0061] The above technical solutions enable the automatic, objective, and precise collection and analysis of Dendrobium phenotypic characteristics, effectively overcoming the problems of strong subjectivity, long breeding cycles, and low efficiency in traditional methods. The multi-source data acquisition mechanism ensures data continuity and reliability, avoiding biases from manual evaluation; the fusion application of deep learning models solves the feature recognition challenges caused by factors such as adjacent leaves, overlap, shading, and adhesion in complex planting environments; and the multi-dimensional comprehensive evaluation mechanism improves the scientific rigor and efficiency of screening decisions by quantifying breeding value. As a result, the objectivity of the breeding process is enhanced, the capacity for large-scale population processing is significantly improved, and efficient technical support is provided for Dendrobium breeding.

[0062] Specifically, in some embodiments of this application, a bi-branch semantic segmentation model is proposed to segment Dendrobium stems and leaves. However, in its implementation, the segmentation boundary is blurred due to the adhesion and overlap of stems in the RGB image. It is difficult to accurately distinguish the physical boundary of adjacent stems by relying solely on two-dimensional image features, resulting in deviations in the calculation of stem morphological parameters, which in turn affects the objectivity and accuracy of breeding decisions.

[0063] In response, this application further proposes a dual-branch semantic segmentation model in step S201 as an encoder-decoder structure, including an RGB branch, a point cloud branch, and a feature fusion module; wherein, the RGB branch is used to extract color and texture features from RGB images; the point cloud branch is used to extract three-dimensional geometric features from three-dimensional point cloud data; and the feature fusion module is used to fuse three-dimensional geometric features with color and texture features at at least one level of the decoder, so as to use three-dimensional geometric features to help solve the problem of stem adhesion and overlap in RGB images.

[0064] The encoder-decoder structure refers to a neural network architecture design, which can be implemented using a symmetrical encoder-decoder layout, with the aim of processing input data of different modalities separately. The RGB branch refers to a sub-network used to process RGB images, which can be implemented using a convolutional neural network, with the aim of extracting color and texture information from the image. The point cloud branch refers to a sub-network used to process 3D point cloud data, which can be implemented using a PointNet architecture, with the aim of extracting 3D geometric features from the point cloud. The feature fusion module refers to a component used to integrate multimodal features, which can be implemented using feature concatenation or weighted summation, with the aim of fusing RGB features and geometric features to improve segmentation accuracy. In practical applications, feature fusion at at least one level of the decoder means selecting a specific level for feature fusion during the decoder's upsampling process, which can be implemented by fusing at multiple levels, with the aim of using high-resolution feature maps for accurate boundary segmentation.

[0065] Specifically, the solution in this application processes RGB images and 3D point cloud data in parallel through RGB branches and point cloud branches, extracting color texture features and 3D geometric features respectively; the feature fusion module integrates geometric features into RGB features at a specific level of the decoder, and uses 3D spatial information to correct boundary ambiguities in 2D images. For example, it enhances the distinguishability of stem boundary pixels through geometric features, so that the model prioritizes the 3D geometric structure during segmentation, thereby achieving accurate separation in the stem adhesion area and ensuring the generation of accurate segmentation results.

[0066] As a preferred embodiment, the solution of this application is implemented as follows: the RGB branch uses ResNet-50 as the backbone network to extract RGB image features, and the point cloud branch uses PointNet++ to process 3D point cloud data to extract geometric features; the feature fusion module fuses the geometric features with the RGB features in the second layer of the decoder by channel splicing.

[0067] Through the above solution, this application effectively solves the problem of blurred segmentation boundaries caused by stem adhesion and overlap in RGB images, improves the accuracy of stem morphology parameter calculation, and thus enhances the objectivity and precision of breeding decisions.

[0068] Specifically, in some embodiments of this application, a feature fusion module is proposed to fuse RGB features and three-dimensional geometric features to solve the problem of stem adhesion and overlap. However, in its implementation, since Dendrobium plants often exhibit dense adhesion, spatial overlap and blurred boundaries among individual stems in their natural growth environment, ordinary feature fusion methods are difficult to effectively utilize three-dimensional geometric information to accurately distinguish the boundary regions of adjacent stems. This leads to the semantic segmentation model generating incorrect connections in the adhesion region, failing to accurately separate different individual stems, and thus affecting the calculation accuracy of key morphological parameters such as stem volume and spatial distribution, ultimately restricting the reliability of breeding decisions.

[0069] In response, this application further proposes that the feature fusion module is a geometric attention fusion module, which generates an attention weight map based on the three-dimensional geometric features extracted from the point cloud branches, and uses the attention weight map to modulate the color and texture feature maps extracted from the RGB branches to enhance the boundary features of different individual stems in the spatial structure.

[0070] Among them, the geometric attention fusion module refers to a feature fusion structure based on spatial attention mechanism, which can be implemented using self-attention mechanism or convolutional attention module, with the aim of dynamically focusing on the boundary region between individual stems; specifically, the attention weight map refers to the weight distribution map generated by nonlinear transformation of three-dimensional geometric features, which can be implemented using heatmap output by sigmoid activation function, with the aim of identifying weak points at the boundary; in practical applications, modulation refers to the weighting operation on RGB feature map, which can be implemented using element-wise multiplication or channel weighting, with the aim of enhancing the distinguishability of boundary pixels.

[0071] Specifically, the solution in this application first generates an attention weight map based on the 3D geometric features extracted from the point cloud branches through a geometric attention fusion module. This weight map can accurately reflect the spatial boundary positions between individual stems. Subsequently, the color and texture feature maps extracted from the RGB branches are modulated using this weight map to enhance the feature response of the boundary regions. Finally, during the feature fusion process, this dynamic modulation mechanism based on geometric structure effectively distinguishes the physical boundaries of adjacent stems, avoiding misclassification of multiple individual stems as a single connected region, thereby ensuring that the semantic segmentation model accurately divides different individual stems during the decoding stage.

[0072] As a preferred embodiment, the solution of this application is specifically implemented as follows: The geometric attention fusion module can be specifically implemented as a sub-network containing a fully connected layer and a sigmoid activation function. The sub-network receives the three-dimensional geometric feature map output by the point cloud branch and generates an attention weight map that matches the size of the RGB feature map. Subsequently, the weight map is applied to the color and texture feature maps output by the RGB branch through element-wise multiplication to highlight the feature expression of the stem boundary region.

[0073] Through the above scheme, this application can effectively improve the accuracy of stem segmentation in the adhesion area, reduce erroneous connection, and ensure more accurate calculation of key morphological parameters such as stem volume and spatial distribution, thereby providing reliable data support for Dendrobium breeding decisions.

[0074] In some of the embodiments described above in this application, a bi-branch semantic segmentation model is proposed to accurately segment Dendrobium stems and leaves. However, in its implementation, since Dendrobium plant stems often exhibit adhesion and overlap during growth, the standard loss function only focuses on pixel-level classification accuracy and cannot effectively distinguish the boundaries and instance relationships of adjacent stems. This leads to the model incorrectly predicting different stems as the same connected region, which in turn affects the calculation accuracy of morphological parameters such as stem volume and spatial distribution, ultimately reducing the reliability of breeding evaluation.

[0075] In response, this application further proposes that the loss function used for training the bi-branch semantic segmentation model in step S201 is a composite loss function, including: standard cross-entropy loss for pixel-level classification; boundary loss for enhancing the accuracy of segmentation boundaries; and adhesion penalty loss for penalizing different stalk instances that are incorrectly predicted as the same connected region; the composite loss function = α*standard cross-entropy loss + β*boundary loss + γ*adhesion penalty term, where α, β, and γ are hyperparameters.

[0076] Among them, the standard cross-entropy loss is a loss function used to evaluate pixel-level classification accuracy. It can be implemented by combining the softmax function with cross-entropy calculation, aiming to ensure that each pixel is correctly classified as a stem or background. The boundary loss is a loss function specifically designed for segmenting boundary regions. It can be implemented by extracting the boundary using the Sobel operator and then calculating the difference between the predicted and the true boundary, aiming to enhance the prediction accuracy of boundary pixels. The adhesion penalty loss is a loss function used to penalize the adhesion of stem instances. It can be implemented by identifying the adhesion region through connected component analysis and then applying a penalty term, aiming to distinguish the individual boundaries of adjacent stems. The composite loss function is the overall loss formed by weighted combination of multiple loss functions. It can balance the contributions of different losses by adjusting the values ​​of hyperparameters α, β, and γ, aiming to synergistically optimize classification, boundary, and instance separation capabilities.

[0077] Specifically, the proposed solution employs a composite loss function design to simultaneously optimize pixel-level classification accuracy, boundary region prediction precision, and instance separation capability during model training. Standard cross-entropy loss provides basic classification capability, ensuring the distinction between stems and background; boundary loss focuses on prediction errors of boundary pixels, reducing boundary blurring caused by lighting variations or leaf occlusion; adhesion penalty loss identifies and penalizes cases where different stems are incorrectly predicted as the same connected region, forcing the model to learn to distinguish the individual boundaries of adjacent stems. These three factors are dynamically balanced through a weighted combination of hyperparameters α, β, and γ, enabling the model to effectively handle the adhesion and overlap issues of Dendrobium stems.

[0078] As a specific implementation method, when training the bi-branch semantic segmentation model, the boundary loss is calculated based on the difference between the predicted segmentation map and the real boundary map using the Canny edge detection algorithm; the adhesion penalty loss is applied by analyzing the difference between the number of connected regions in the predicted result and the real annotation; and the hyperparameters α, β, and γ are adjusted according to the performance of the validation set to optimize the segmentation accuracy of the model on the test set.

[0079] The above technical solutions effectively solve the problem of segmentation errors caused by the adhesion of Dendrobium stems, improve the calculation accuracy of stem volume and spatial distribution parameters, and thus enhance the reliability of breeding evaluation.

[0080] In practical applications, some of the embodiments described above in this application propose morphological parameters of stems to quantify the morphological characteristics of Dendrobium. However, in the process of implementation, since Dendrobium stems often exhibit adhesion, overlap, and shading phenomena in the growth environment, relying solely on basic parameters such as plant height and stem diameter cannot fully capture the three-dimensional structural features of the stems, resulting in distorted phenotypic data and thus affecting the accuracy and reliability of breeding value assessment.

[0081] In this regard, this application further proposes that the morphological parameters of the stem in step S201 include stem volume and spatial distribution information.

[0082] Specifically, stem volume refers to the geometric volume occupied by the Dendrobium stem in three-dimensional space. It can be achieved using voxelization statistical methods or convex hull volume calculation algorithms based on three-dimensional point cloud data. The purpose is to accurately quantify the biomass accumulation state and overall development degree of the stem, avoiding distortion of stem diameter measurement values ​​caused by adhesion between adjacent stems. Spatial distribution information can be understood as the spatial position coordinates, principal direction vector, and spatial topological distribution pattern of the stem in the three-dimensional coordinate system. It can be achieved using point cloud covariance matrix eigenvalue decomposition or centerline skeleton extraction technology. The purpose is to characterize the relative spatial relationship between individual stems and enhance the distinguishability of the boundaries of adhered areas.

[0083] Specifically, in step S201, the dual-branch semantic segmentation model first fuses RGB images and 3D point cloud data to complete pixel-level segmentation of the stem region. Subsequently, the system extracts the 3D geometric structure of the stem based on the segmentation results, obtains the stem volume parameters by calculating the spatial occupancy volume of the point cloud data, and analyzes the spatial distribution characteristics of the point cloud to generate the main direction vector and position distribution features. The stem volume parameters provide a quantitative basis for the biomass dimension, while the spatial distribution information characterizes the spatial configuration features of the stem. The synergistic effect of the two enables the system to distinguish physically adhered stem individuals, effectively solving the problem of morphological parameter extraction deviation caused by overlapping occlusion.

[0084] As a preferred embodiment, the solution of this application is implemented as follows: After the bi-branch semantic segmentation model outputs the stalk segmentation mask, the system divides the point cloud data of each stalk instance into voxel grids, counts the number of non-empty voxels, and calculates the stalk volume in combination with the physical size of the voxels; at the same time, the main direction vector of the stalk is obtained by decomposing the eigenvectors of the point cloud covariance matrix. This vector indicates the direction of the stalk's growth axis and can be used to identify the spatial separation boundary of adjacent stalks.

[0085] Through the above-mentioned scheme, this application effectively overcomes the interference of Dendrobium stem adhesion and overlap in complex growth environment, making the three-dimensional structural characterization of morphological traits more complete and accurate, significantly improving the reliability of phenotypic data, and thus providing a more objective quantitative basis for breeding value assessment.

[0086] In Dendrobium breeding, how to achieve automatic, objective, and accurate collection and analysis of phenotypic information throughout the entire growth cycle, in multiple dimensions, and with high throughput remains an urgent problem to be solved. Especially in complex growing environments, plant phenotypic characteristics are often affected by a variety of factors (such as adjacent leaves, overlap, shading, adhesion, etc.), posing a significant challenge to accurate identification and evaluation.

[0087] Specifically, in some of the embodiments described above in this application, a multi-trait comprehensive selection index was proposed to comprehensively evaluate the breeding value of Dendrobium officinale individual plants. However, in its implementation, the index weights were set in a fixed manner, failing to consider the dynamic changes in phenological stages throughout the entire growth cycle of Dendrobium officinale. This resulted in the inability to accurately reflect the relative importance of key traits (such as flowering-related traits or vegetative growth traits) at different growth stages (such as flower bud differentiation stage, initial flowering stage, etc.), causing the comprehensive evaluation results to deviate from actual breeding needs and affecting the accuracy and efficiency of screening superior individual plants.

[0088] In this regard, this application further proposes that in step S301, a multi-trait comprehensive selection index is constructed, and the weights of different phenotypic data in the index are dynamically adjusted according to the phenological period identified in S202.

[0089] Among them, the multi-trait comprehensive selection index is a mathematical model for quantitatively evaluating the breeding value of a single Dendrobium plant. It can be implemented using a weighted summation method or a machine learning ranking algorithm, with the aim of comprehensively considering the importance weights of breeding objectives such as yield, quality, resistance, and maturity. The phenological stages identified by S202 refer to the phenological stages such as flower bud differentiation, initial flowering, and full bloom identified by image classification models based on EfficientNet or VisionTransformer. It can be implemented based on the analysis of RGB images and three-dimensional point cloud data, with the aim of providing precise time points in the Dendrobium growth process. Dynamic weight adjustment refers to changing the contribution ratio of different phenotypic data in the index in real time according to the currently identified phenological stages. It can be implemented using preset rules or adaptive algorithms, with the aim of focusing the evaluation on the key traits of the current growth stage.

[0090] Specifically, the scheme of this application uses the image classification model constructed in step S202 to identify the phenological stage of Dendrobium in real time, and uses the identification results as the basis for dynamically adjusting the weight of the multi-trait comprehensive selection index; at key stages such as flower bud differentiation, the weight of physiological traits such as flowering period prediction and flower morphology classification is automatically increased, while the weight of unrelated traits is reduced; thereby ensuring that the comprehensive selection index can be adaptively optimized with the growth process and accurately reflect the priority changes of breeding objectives under different phenological stages.

[0091] As a preferred embodiment, the specific implementation of the scheme in this application is as follows: During the flower bud differentiation period of Dendrobium, the system adjusts the weight of phenotypic data such as flowering period prediction and flower morphology classification to a higher level based on the phenological period information identified by S202, while reducing the weight of vegetative growth traits such as stem volume and spatial distribution information; during the full bloom period, the weight of flower morphology classification is further strengthened to ensure that the comprehensive selection index focuses on the current core breeding objectives.

[0092] Through the above scheme, this application can adaptively adjust the weight of the multi-trait comprehensive selection index according to the dynamic changes of Dendrobium growth stage, effectively solve the evaluation bias problem caused by fixed weight, and significantly improve the accuracy of screening superior single plants and the timeliness of breeding decisions.

[0093] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for Dendrobium breeding based on deep learning, characterized in that: Includes the following steps: S1, Multi-source image data acquisition and preprocessing, the specific steps are as follows: S101 deploys multi-source sensors in the Dendrobium breeding nursery to automatically collect image data of Dendrobium plants from seedling to maturity on a regular basis; The image data includes at least RGB images and 3D point cloud data; S102, preprocess the collected image data to construct a standardized Dendrobium multimodal dataset; The preprocessing includes image cropping, rotation, scaling, noise reduction, background segmentation, image alignment, and data augmentation; S2, the construction and training of deep learning models for key phenotypic traits, the specific steps are as follows: S201, for the morphological traits of Dendrobium plant height, stem diameter, number of leaves and leaf area, the preprocessed RGB image and 3D point cloud data are input into the trained bi-branch semantic segmentation model. The model accurately segments the stems and leaves of Dendrobium by fusing RGB features and 3D geometric features, and calculates the morphological parameters of the stems and leaves. S202 constructs a multimodal deep learning model based on EfficientNet or Vision Transformer. The model extracts the texture, color, and morphological features of flower buds, petals, and inflorescence regions, and enhances the feature expression of key parts by combining an attention mechanism. This enables automatic identification of phenological stages such as flower bud differentiation, initial flowering, and full bloom. At the same time, flower color classification and flower shape classification branches are set to automatically distinguish flower color and shape. Furthermore, by combining continuous time-series images or three-dimensional morphological parameters, the developmental status of Dendrobium flowering can be predicted. S203, for the initial screening of biochemical traits for growth vigor and prediction of potential medicinal components, combines hyperspectral image data to construct a regression model based on 3D-CNN or Spectral-Attention mechanism, and establishes an inversion model of spectral characteristics and key biochemical indicators such as leaf chlorophyll content and polysaccharide content. S204, the image dataset preprocessed in S102 is divided into training set, validation set and test set, which are used to train and optimize the models in S201, S202 and S203 until the model reaches the predetermined accuracy on the test set. S3, Comprehensive Evaluation and Decision-Making of Breeding Value, the specific steps are as follows: S301, integrate the multi-dimensional phenotypic data output by each model in step S2, normalize the phenotypic indicators such as plant height, stem diameter, leaf area, flowering period grade, chlorophyll content and polysaccharide content by range standardization, and use the analytic hierarchy process, entropy weight method or machine learning method to determine the weight of each indicator, construct a comprehensive phenotypic scoring function, and form a multi-trait comprehensive selection index. S302 uses a mathematical model to comprehensively score and rank each individual plant in the breeding population; The mathematical model described is a weighted summation method or a machine learning ranking algorithm; S303 allows breeders to quickly select top-ranked superior plants as parents for the next generation or directly enter regional trials.

2. The Dendrobium breeding method based on deep learning according to claim 1, characterized in that: The dual-branch semantic segmentation model described in step S201 is an encoder-decoder structure, including an RGB branch, a point cloud branch, and a feature fusion module; The RGB branch is used to extract color and texture features from RGB images; the point cloud branch is used to extract three-dimensional geometric features from three-dimensional point cloud data; and the feature fusion module is used to fuse the three-dimensional geometric features with the color and texture features at at least one level of the decoder.

3. The Dendrobium breeding method based on deep learning according to claim 2, characterized in that: The feature fusion module is a geometric attention fusion module, which generates an attention weight map based on the three-dimensional geometric features extracted from the point cloud branch, and uses the attention weight map to modulate the color and texture feature maps extracted from the RGB branch to enhance the boundary features of different stem individuals in spatial structure.

4. The Dendrobium breeding method based on deep learning according to claim 1, characterized in that: The loss function used for training the bi-branch semantic segmentation model in step S201 is a composite loss function, including: standard cross-entropy loss for pixel-level classification; boundary loss for enhancing the accuracy of segmentation boundaries; and adhesion penalty loss for penalizing different stalk instances that are incorrectly predicted as the same connected region. The composite loss function is: α * standard cross-entropy loss + β * boundary loss + γ * adhesion penalty term, where α, β, and γ are hyperparameters.

5. The Dendrobium breeding method based on deep learning according to claim 1, characterized in that: The three-dimensional point cloud data described in step S101 is spatially aligned with the RGB image through calibration and registration.

6. The Dendrobium breeding method based on deep learning according to any one of claims 1 or 5, characterized in that: The multi-source sensors mentioned in step S101 include a visible light RGB camera, a hyperspectral camera, a multispectral camera, and a 3D laser scanner.

7. The Dendrobium breeding method based on deep learning according to claim 1, characterized in that: The morphological parameters of the stem mentioned in step S201 include stem volume and spatial distribution information.

8. The Dendrobium breeding method based on deep learning according to claim 1, characterized in that: In step S301, a multi-trait comprehensive selection index is constructed, and the weights of different phenotypic data in the index are dynamically adjusted according to the phenological period identified in S202.

Images