An artificial intelligence technology-based high-throughput plant phenotype data analysis platform

Abstract

The application discloses a high-throughput plant phenotype data analysis platform based on artificial intelligence technology, which comprises an image acquisition unit, a deep feature extraction unit, a phenotype feature extraction unit, a quality prediction unit and a comprehensive evaluation unit; deep features of plant images are extracted through a multi-scale convolutional neural network, adaptive multi-scale feature fusion is carried out through an attention mechanism, plant physiological and biochemical indexes are predicted based on a deep regression network, and a comprehensive phenotype evaluation index is calculated by using an entropy weight method-hierarchical analysis method combined weighting method; the application realizes rapid, accurate, non-destructive and systematic analysis of plant phenotypes, and provides an efficient and intelligent technical tool for plant breeding, quality evaluation and precision agriculture.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to a high-throughput plant phenotypic data analysis platform based on artificial intelligence technology. Background Technology

[0002] Plant phenotype refers to the set of observable traits exhibited by plants under specific environmental conditions, including multi-dimensional information such as morphological characteristics, physiological indicators, and biochemical components. High-throughput plant phenotyping technology is an important supporting tool for modern breeding, precision agriculture, and plant science research, and is of great significance for accelerating crop improvement and increasing agricultural production efficiency. Traditional plant phenotyping methods mainly rely on manual measurement and laboratory chemical analysis, which have the following prominent problems: First, manual measurement is highly subjective and inefficient, making it difficult to meet the needs of rapid screening of large-scale samples; second, destructive chemical analysis results in samples that cannot be reused, increasing breeding costs; third, single-indicator analysis cannot comprehensively reflect the overall phenotypic characteristics of plants, and there is a lack of systematic evaluation methods; fourth, there is a lack of effective data management and analysis tools, and a large amount of phenotypic data has not been fully explored and utilized.

[0003] In recent years, the rapid development of computer vision and deep learning technologies has provided new solutions for plant phenotypic analysis. Existing image-based plant phenotypic research mainly focuses on morphological feature extraction, such as leaf area, plant height, and color, but non-destructive prediction of intrinsic quality indicators such as chlorophyll content, soluble sugars, and vitamins remains challenging. Furthermore, existing methods often employ single-scale feature extraction, failing to fully utilize multi-level feature information; in multi-indicator prediction, there is a lack of collaborative optimization among tasks; and in terms of comprehensive evaluation, there is a lack of scientifically sound multi-indicator fusion strategies and visualization methods.

[0004] Therefore, there is an urgent need to develop a high-throughput plant phenotypic analysis platform to achieve rapid, accurate, non-destructive, and systematic analysis of plant phenotypic characteristics, providing technical support for modern breeding and precision agriculture. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-throughput plant phenotypic data analysis platform based on artificial intelligence technology. This platform addresses the problems of low efficiency, high cost, strong subjectivity, and lack of systematic evaluation methods in traditional plant phenotypic analysis methods. To achieve the above objectives, the specific technical solution of this invention is as follows:

[0006] This invention provides a high-throughput plant phenotypic data analysis platform based on artificial intelligence technology. The platform includes: an image acquisition unit, a deep feature extraction unit, a phenotypic feature extraction unit, a quality prediction unit, and a comprehensive evaluation unit.

[0007] The image acquisition unit is used to acquire image data of the plant sample, and the image data is represented as a tensor. ,in For the sample size, The number of image channels. and These represent the image height and width, respectively.

[0008] The deep feature extraction unit uses a multi-scale convolutional neural network to extract features from the image. This network includes... Feature extraction layer, the first The feature map of the layer is represented as ,in For the first Number of feature channels in the layer and is the spatial dimension of the feature map.

[0009] The phenotypic feature extraction unit extracts phenotypic feature vectors from multi-layer feature maps and calculates fused features through a multi-scale feature fusion mechanism.

[0010] The quality prediction unit predicts plant physiological and biochemical indicators based on a deep regression network, which uses a fully connected layer structure.

[0011] The comprehensive evaluation unit calculates the plant comprehensive phenotypic evaluation index based on multi-objective optimization theory. The formula for calculating the evaluation index is as follows: ;in For the first The comprehensive evaluation index of each sample For the first The standardized scores of each indicator are obtained using the following standardization method: ,in and The first The mean and standard deviation of each indicator across all samples. For the first The weighting coefficients of each indicator.

[0012] Furthermore, feature extraction functions Defined as: ;in For the first Convolution operations of layers, For batch normalization, For activation function, These are the weight parameters.

[0013] Furthermore, the method for calculating fused features through a multi-scale feature fusion mechanism is as follows: in, For multi-scale fusion functions; The adaptive weight vector satisfies the constraints. and ; The upsampling function unifies feature maps at different scales to the same dimension. For feature dimensions.

[0014] Furthermore, the multi-scale convolutional neural network employs a residual connection structure, the first... The formula for calculating the residual block of a layer is: ,in for The residual mapping function of the layer, It is an identity mapping.

[0015] Furthermore, when the input and output dimensions do not match, the identity mapping is used. Dimensional adjustment of convolution: , The projection weight matrix; the total number of network layers. The value ranges from 18 to 101.

[0016] Furthermore, the prediction model expression for the quality prediction unit is: ;in For the first The predicted value of each target indicator, For the number of target indicators, and For the first The weight matrix and bias vector of each prediction head.

[0017] Furthermore, the adaptive weight vector The calculation process, which utilizes an attention mechanism, includes the following steps:

[0018] a) Calculate the global average pooling result for each layer of features: ;

[0019] b) Calculate the attention score using a two-layer fully connected network: ,in and For the weights of the fully connected layer, To reduce the ratio, the value ranges from 4 to 16;

[0020] c) Softmax normalization of the attention scores yields adaptive weights: .

[0021] Furthermore, the deep regression network is trained using a multi-task learning strategy, and the total loss function is defined as:

[0022] ;in, For the first Mean squared error loss for each task: ; and The first The sample at the th The actual and predicted values ​​for each target; Assigning task weights, and then adaptively learning using an uncertainty-weighted method: ;in, For the first Learnable parameters for each task, For L2 regularization terms, is the regularization coefficient, and its value range is . to .

[0023] Furthermore, the index weight coefficients in the comprehensive evaluation unit The weighting method is determined through a combination of entropy weighting and analytic hierarchy process (AHP), specifically including:

[0024] Calculate the first Information entropy of each indicator: ;in For the first The sample at the th Calculate the entropy weight based on the proportion of each indicator: ;in For the first The information utility value of each indicator.

[0025] Furthermore, subjective weights are calculated using the analytic hierarchy process (AHP). Construct a judgment matrix Solve the characteristic equation: Find the largest eigenvalue The corresponding feature vectors, after normalization, are obtained as follows: ;but: ;in, is the combination coefficient, with a value range of [0.3, 0.7].

[0026] Furthermore, the plant phenotypic comprehensive evaluation report generated by the comprehensive evaluation unit includes: based on the sample ranking results, according to the comprehensive evaluation index: Sort in descending order and select the top-ranked items. Excellent samples, among which The value range is [10, 30].

[0027] Furthermore, a multi-indicator radar chart is used to display the performance of each sample. Standardized scores on each evaluation indicator The polygon area of ​​the radar chart is used as a visualization measure of the overall phenotypic performance; where the first... Polygon area of ​​a sample radar image : .

[0028] Furthermore, for each predicted value Calculate the 95% confidence interval: ;in This is the critical value. This is the standard error of the prediction.

[0029] Compared with the prior art, the beneficial effects of this invention are:

[0030] This invention employs an automated analysis process based on deep learning, enabling the phenotypic analysis of hundreds of plant samples within minutes. Compared to traditional manual measurement and chemical analysis methods, this significantly reduces labor and time costs. Through multi-scale convolutional neural networks and attention mechanisms, it fully extracts deep feature information from plant images, achieving high-precision prediction of multiple physiological and biochemical indicators. This platform can obtain intrinsic plant quality information without destroying the samples, making the samples reusable. It is particularly suitable for the evaluation of rare germplasm resources and breeding materials. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the composition of a high-throughput plant phenotypic data analysis platform based on artificial intelligence technology according to the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0033] Example

[0034] like Figure 1 The diagram shows the composition of a high-throughput plant phenotypic data analysis platform based on artificial intelligence technology according to the present invention. The platform includes: an image acquisition unit, a deep feature extraction unit, a phenotypic feature extraction unit, a quality prediction unit, and a comprehensive evaluation unit.

[0035] The image acquisition unit is used to acquire image data of the plant sample, and the image data is represented as a tensor. ,in For the sample size, The number of image channels. and These represent the image height and width, respectively.

[0036] In this embodiment, the image acquisition unit is configured as a standardized image acquisition system, including an industrial camera, an LED cold light source, and a black background. During the acquisition process, the plant sample is laid flat on the background, the light source angle is adjusted to 45 degrees to reduce specular reflection, and the camera is placed vertically downwards at a distance of 30 cm from the sample. The image acquisition parameters are set as follows: resolution 224×224 pixels (i.e., ... = =224), RGB three-channel ( =3), 32 samples are processed per batch ( =32). The acquired raw image data is organized in tensor form. This facilitates the batch processing of subsequent deep learning models.

[0037] Specifically, this embodiment uses a 5-megapixel industrial camera (model MV-CA050-10GC) equipped with a 6000K color temperature LED ring light source to ensure accurate color reproduction of the image. The background board is made of matte black PVC material, which effectively reduces background noise interference; the image acquisition software is developed based on the OpenCV library and supports automatic exposure adjustment, white balance correction, and distortion correction to ensure the stability and consistency of image quality.

[0038] The deep feature extraction unit uses a multi-scale convolutional neural network to extract features from the image. This network includes... Feature extraction layer, the first The feature map of the layer is represented as ,in For the first Number of feature channels in the layer and is the spatial dimension of the feature map.

[0039] Feature extraction function Defined as: ;in For the first Convolution operations of layers, For batch normalization, For activation function, These are the weight parameters.

[0040] The phenotypic feature extraction unit extracts phenotypic feature vectors from multi-layer feature maps and calculates fused features through a multi-scale feature fusion mechanism. The method for calculating fused features through the multi-scale feature fusion mechanism is as follows: in, For multi-scale fusion functions; The adaptive weight vector satisfies the constraints. and ; The upsampling function unifies feature maps at different scales to the same dimension. For feature dimensions.

[0041] The network uses ResNet-50 as its basic network architecture, which includes... =5 feature extraction layers, respectively corresponding to , , , and Layers. The dimensions of the feature maps in each layer are as follows: , , , , Through layer-by-layer convolution operations, the network gradually extracts semantic features from shallow low-level features such as edges and textures to deep semantic features.

[0042] In practical implementation, convolution operations Use a 3×3 convolution kernel with a stride of 1 or 2 (for downsampling layers) and padding of 1. Batch normalization. The features of each channel are standardized, and the calculation formula is as follows: ,in and The mean and standard deviation of this batch are given. This is a numerically stable term. Activation function. Using modified linear units ), introducing nonlinear capabilities.

[0043] Residual connection structures effectively alleviate the gradient vanishing problem in deep networks. When the input and output dimensions of layer l are the same, identity mapping is used. When the dimensions do not match (e.g., changes in the number of channels or a decrease in spatial resolution), a 1×1 convolutional projection is used. Adjusting the dimensions. This allows networks to be trained to 50 layers or even deeper (such as ResNet-101), while maintaining high accuracy and avoiding performance degradation.

[0044] Shallow features contain rich details (such as leaf edges and vein textures), while deep features contain high-level semantic information (such as overall shape and color distribution). Through adaptive weighted fusion, the system can automatically adjust the importance of each layer of features according to the specific task.

[0045] The adaptive weight vector The calculation process, which utilizes an attention mechanism, includes the following steps:

[0046] a) Calculate the global average pooling result for each layer of features: ;

[0047] b) Calculate the attention score using a two-layer fully connected network: ,in and For the weights of the fully connected layer, To reduce the ratio, the value ranges from 4 to 16;

[0048] c) Softmax normalization of the attention scores yields adaptive weights: .

[0049] The multi-scale convolutional neural network employs a residual connection structure, the first... The formula for calculating the residual block of a layer is: ,in for The residual mapping function of the layer, It is an identity mapping.

[0050] When the input and output dimensions do not match, the identity mapping is used. Dimensional adjustment of convolution: , The projection weight matrix; the total number of network layers. The value ranges from 18 to 101.

[0051] The quality prediction unit predicts plant physiological and biochemical indicators based on a deep regression network, which uses a fully connected layer structure.

[0052] The deep regression network is trained using a multi-task learning strategy, and the total loss function is defined as:

[0053] ;in, For the first Mean squared error loss for each task: ; and The first The sample at the th The actual and predicted values ​​for each target; Assigning task weights, and then adaptively learning using an uncertainty-weighted method: ;in, For the first Learnable parameters for each task, For L2 regularization terms, is the regularization coefficient, and its value range is . to .

[0054] The prediction model expression for the quality prediction unit is: ;in For the first The predicted value of each target indicator, For the number of target indicators, and For the first The weight matrix and bias vector of each prediction head.

[0055] For example, simultaneously predicting =6 physiological and biochemical indicators, including: chlorophyll content (SPAD value), soluble sugar content (mg / g), vitamins Content (mg / 100g), leaf area (cm²), dry matter content (%), and nitrogen content (%).

[0056] The deep regression network structure is a three-layer fully connected network: the first layer flattens the fused features F_fused and maps them to 512 dimensions; the second layer reduces the dimensionality to 256 dimensions; and the third layer outputs six prediction heads. Each prediction head is an independent linear layer with parameters... and .

[0057] The advantages of multi-task learning are: shared feature representation layers can learn common knowledge between tasks, improving data utilization efficiency; the correlation between tasks can act as a regularization mechanism, preventing overfitting; and uncertainty-weighted adaptive balancing of the importance of each task avoids simple tasks dominating the training process. The network is trained using the Adam optimizer with a learning rate of 1e-4, a batch size of 32, and 200 training epochs.

[0058] The comprehensive evaluation unit calculates the plant comprehensive phenotypic evaluation index based on multi-objective optimization theory. The formula for calculating the evaluation index is as follows: ;in For the first The comprehensive evaluation index of each sample For the first The standardized scores of each indicator are obtained using the following standardization method: ,in and The first The mean and standard deviation of each indicator across all samples. For the first The weighting coefficients of each indicator.

[0059] The index weight coefficient in the comprehensive evaluation unit The weighting method is determined through a combination of entropy weighting and analytic hierarchy process (AHP), specifically including:

[0060] Calculate the first Information entropy of each indicator: ;in For the first The sample at the th Calculate the entropy weight based on the proportion of each indicator: ;in For the first The information utility value of each indicator.

[0061] Subjective weights are calculated using the analytic hierarchy process. Construct a judgment matrix Solve the characteristic equation: Find the largest eigenvalue The corresponding feature vectors, after normalization, are obtained as follows: ;but: ;in, is the combination coefficient, with a value range of [0.3, 0.7].

[0062] The comprehensive evaluation report of plant phenotypic characteristics generated by the comprehensive evaluation unit includes: based on the sample ranking results, and according to the comprehensive evaluation index: Sort in descending order and select the top-ranked items. Excellent samples, among which The value range is [10, 30].

[0063] A multi-indicator radar chart is used to display the performance of each sample. Standardized scores on each evaluation indicator The polygon area of ​​the radar chart is used as a visualization measure of the overall phenotypic performance; where the first... Polygon area of ​​a sample radar image : .

[0064] For each predicted value Calculate the 95% confidence interval: ;in This is the critical value. This is the standard error of the prediction.

[0065] To verify the technical effectiveness of this invention, phenotypic analysis and quality evaluation of 120 tomato breeding materials were conducted using this platform. The specific implementation process and results are as follows:

[0066] Five fully expanded leaves were taken from each of 120 tomato breeding materials as analysis samples, for a total of 600 samples. Image acquisition was carried out according to a standardized procedure: the leaves were laid flat on a black background, and RGB images were acquired using an industrial camera under standard lighting conditions (6000K color temperature, 1000 lux illuminance) with a resolution of 224×224 pixels.

[0067] A pre-trained ResNet-50 network was used as the feature extraction backbone. This network was pre-trained on the ImageNet dataset and then fine-tuned using 1200 labeled tomato leaf images. During training, data augmentation strategies (random rotation ±15 degrees, random flipping, color jittering, etc.) were employed to expand the training set and prevent overfitting. After 200 rounds of training, the network's loss on the validation set converged and stabilized.

[0068] In the feature extraction stage, the network performs forward propagation on 600 samples, extracting 5 layers of feature maps. Visual analysis reveals that shallow features (…) The main focus is on leaf margins and vein texture, and mid-layer features. Pay attention to leaf morphology and color distribution, and deeper features ( (Capture overall semantic information.) The attention mechanism automatically calculates the weight allocation as follows: This indicates that mid-to-high-level features contribute more to tomato quality prediction. Multi-scale feature fusion yields a 2048-dimensional feature vector, which serves as the input for subsequent predictions.

[0069] Based on the fusion characteristics, the deep regression network simultaneously predicts six key quality indicators. The prediction results are compared with those obtained from laboratory chemical analysis as follows.

[0070] Chlorophyll content (SPAD value): Predicted range 32.5-58.3, correlation coefficient R² = 0.92 with SPAD-502 determination, root mean square error RMSE = 2.1 SPAD units, relative error 4.6%. Soluble sugar content (mg / g): Predicted range 25.8-67.4 mg / g, correlation coefficient R² = 0.89 with HPLC determination, root mean square error RMSE = 3.5 mg / g, relative error 7.8%. Vitamin C content (mg / 100g): Predicted range 18.2-42.6 mg / 100g, correlation coefficient R² = 0.87 with 2,6-dichlorophenolindophenol titration determination, root mean square error RMSE = 2.8 mg / 100g, relative error 9.2%. Leaf area (cm²): Predicted range 85.3-245.7 cm², correlation coefficient R²=0.95 with the LI-3100C leaf area meter, root mean square error RMSE=8.6 cm², relative error 5.1%. Dry matter content (%): Predicted range 8.5%-15.8%, correlation coefficient R²=0.88 with the drying and weighing method, root mean square error RMSE=0.65%, relative error 5.9%. Nitrogen content (%): Predicted range 2.8%-5.2%, correlation coefficient R²=0.90 with the Kjeldahl method, root mean square error RMSE=0.18%, relative error 5.4%.

[0071] The average prediction accuracy of the six indicators reached 90.2%, meeting the requirements for practical application. The prediction process took approximately 25 minutes, which is about 600 times more efficient than traditional chemical analysis methods (which require 2-3 days). Based on the six quality indicators, a combined weighting method of entropy weighting and analytic hierarchy process was used to calculate the comprehensive evaluation index, obtain subjective weights, and sort the materials in descending order of the comprehensive evaluation index. The top 20% (24 samples) of excellent materials were selected, with an average comprehensive evaluation index of 0.78, which is significantly higher than the overall average of 0.58 (t-test, p<0.001).

[0072] These 24 superior materials demonstrated excellent performance across all six indicators, particularly with soluble sugar and vitamin C content exceeding the average by 32% and 28%, respectively, indicating high nutritional quality and breeding value. Radar chart visualization analysis revealed that the top three materials (TM-087, TM-053, and TM-112) exhibited balanced performance across all indicators, with radar chart polygon areas of 4.82, 4.65, and 4.53, respectively, significantly larger than the average area of ​​2.87. These materials are recommended as key breeding parents for hybridization combinations.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-throughput plant phenotypic data analysis platform system based on artificial intelligence technology, characterized in that, The platform system includes: an image acquisition unit, a depth feature extraction unit, a phenotypic feature extraction unit, a quality prediction unit, and a comprehensive evaluation unit; The image acquisition unit is used to acquire image data of the plant sample, and the image data is represented as a tensor. ,in For the sample size, The number of image channels. and These are the image height and width, respectively. The deep feature extraction unit uses a multi-scale convolutional neural network to extract features from the image. This network includes... Feature extraction layer, the first The feature map of the layer is represented as ,in For the first Number of feature channels in the layer and The dimension of the feature map space; The phenotypic feature extraction unit extracts phenotypic feature vectors from multi-layer feature maps and calculates fused features through a multi-scale feature fusion mechanism. Feature extraction function Defined as: ;in For the first Convolution operations of layers, For batch normalization, For activation function, These are weight parameters; The method for calculating fused features using a multi-scale feature fusion mechanism is as follows: ;in, For multi-scale fusion functions; The adaptive weight vector satisfies the constraints. and ; The upsampling function unifies feature maps at different scales to the same dimension. For feature dimensions; The quality prediction unit predicts plant physiological and biochemical indicators based on a deep regression network, which adopts a fully connected layer structure. The comprehensive evaluation unit calculates the plant comprehensive phenotypic evaluation index based on multi-objective optimization theory. The formula for calculating the evaluation index is as follows: ;in For the first The comprehensive evaluation index of each sample For the first The standardized scores of each indicator are obtained using the following standardization method: ,in and The first The mean and standard deviation of each indicator across all samples. For the first The weighting coefficients of each indicator.

2. The platform system according to claim 1, characterized in that, The multi-scale convolutional neural network employs a residual connection structure, the first... The formula for calculating the residual block of a layer is: ,in for The residual mapping function of the layer, It is an identity mapping.

3. The platform system according to claim 2, characterized in that, When the input and output dimensions do not match, the identity mapping is used. Dimensional adjustment of convolution: , The projection weight matrix; the total number of network layers. The value ranges from 18 to 101.

4. The platform system according to claim 3, characterized in that, The prediction model expression for the quality prediction unit is: ;in For the first The predicted value of each target indicator, For the number of target indicators, and For the first The weight matrix and bias vector of each prediction head.

5. The platform system according to claim 4, characterized in that, The adaptive weight vector The calculation process, which utilizes an attention mechanism, includes the following steps: a) Calculate the global average pooling result for each layer of features: ; b) Calculate the attention score using a two-layer fully connected network: ,in and For the weights of the fully connected layer, To reduce the ratio, the value ranges from 4 to 16; c) Softmax normalization of the attention scores yields adaptive weights: .

6. The platform system according to claim 5, characterized in that, The index weight coefficient in the comprehensive evaluation unit The weighting method is determined through a combination of entropy weighting and analytic hierarchy process (AHP), specifically including: Calculate the first Information entropy of each indicator: ;in For the first The sample at the th Calculate the entropy weight based on the proportion of each indicator: ;in For the first The information utility value of each indicator; Subjective weights are calculated using the analytic hierarchy process. Construct a judgment matrix Solve the characteristic equation: Find the largest eigenvalue The corresponding feature vectors, after normalization, are obtained as follows: ;but: ;in, is the combination coefficient, with a value range of [0.3, 0.7].

7. The platform system according to claim 6, characterized in that, The comprehensive evaluation report of plant phenotypic characteristics generated by the comprehensive evaluation unit includes: based on the sample ranking results, and according to the comprehensive evaluation index: Sort in descending order and select the top-ranked items. Excellent samples, among which The value range is [10, 30].

8. The platform system according to claim 7, characterized in that, A multi-indicator radar chart is used to display the performance of each sample. Standardized scores on each evaluation indicator The polygon area of ​​the radar chart is used as a visualization measure of the overall phenotypic performance; where the first... Polygon area of ​​a sample radar image : ; For each predicted value Calculate the 95% confidence interval: ;in This is the critical value. This is the standard error of the prediction.

Images