High-throughput image data driven crop breeding intelligent assistant system and decision method

Abstract

The application discloses a high-throughput image data driven crop breeding intelligent auxiliary system and decision method, comprising: deploying a perception array, acquiring microclimate parameters and soil physical and chemical state information in a target breeding area in real time, based on real-time data flow of the perception array, dynamically adjusting shooting frequency, imaging mode and spatial coverage range of an imaging unit through an image acquisition scheduler; collecting multi-view image sequences of each plant in the target crop population, including visible light, near-infrared and thermal infrared images, performing individual plant segmentation on the image sequences, binding with identity operation, and generating a data set of spatiotemporal continuous images in units of single plants; for the data set of each plant, constructing an individualized dynamic development trajectory model, updating model parameters through online learning, reflecting the response of the plant to environmental disturbance, and detecting singular phenotypes of trajectory data output by the trajectory model to identify abnormal development events.

Description

Technical Field

[0001] This invention relates to a high-throughput image data-driven intelligent auxiliary system for crop breeding, and more particularly to a high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding. Background Technology

[0002] In the field of crop breeding, with the development of precision agriculture and intelligent technologies, high-throughput image data acquisition has become a key means to improve breeding efficiency. Traditional breeding methods rely on manual observation and phenotypic measurement, which are time-consuming and easily affected by subjective factors. Modern technologies, on the other hand, use drones, sensors, and other equipment to acquire large-scale image data, providing a data foundation for crop growth monitoring and trait analysis. A common approach in existing technologies is to use static acquisition at equal time intervals, that is, to take images of crops at fixed time points to obtain static snapshots of the growth process. This approach, based on a preset sampling frequency, can collect data regularly, facilitating subsequent processing and analysis, and to some extent promoting the automation of the breeding process.

[0003] However, the static data collection scheme with equal time intervals has significant drawbacks. Because crop growth is a dynamic and non-linear process, phenotypic changes often accelerate or slow down under specific environmental conditions or growth stages, and data collection at fixed time intervals can easily lead to data sparsity or redundancy. During critical growth stages such as flowering or grain-filling, growth changes are rapid, and equal-interval collection may miss important turning points, thus affecting the accuracy of trait assessment. Furthermore, this scheme lacks flexibility and cannot adaptively adjust the collection frequency based on real-time environmental factors such as light, temperature, or water stress, resulting in data discontinuity and resource waste. Static collection also often ignores individual differences in crop growth, leading to insufficient data representativeness and limiting the precision of breeding decisions. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a high-throughput image data-driven intelligent auxiliary decision-making method for crop breeding, comprising the following steps:

[0006] S1: Deploy a sensing array to acquire real-time information on micro-meteorological parameters and soil physicochemical state within the target breeding area. Based on the real-time data stream from the sensing array, dynamically adjust the imaging frequency, imaging mode, and spatial coverage of the imaging unit through the image acquisition scheduler.

[0007] S2: Collect multi-view image sequences of each plant in the target crop population, including visible light, near infrared and thermal infrared images, perform individual plant segmentation on the image sequences, bind them with identity, and generate a dataset of spatiotemporal continuous images of individual plants.

[0008] S3: For each plant's dataset, construct an individualized dynamic development trajectory model, update the model parameters through online learning to reflect the plant's response to environmental disturbances, and perform singular phenotypic detection on the trajectory data output by the trajectory model to identify abnormal development events.

[0009] S4: Perform causal correlation analysis on the detected unusual phenotypic events and the corresponding environmental disturbance records to generate an environment-phenotypic response mapping map;

[0010] S5: Based on the mapping map and trajectory model, calculate the comprehensive breeding value score of each plant, output a list of plants with the highest comprehensive breeding value scores, and generate a decision report that includes key phenotypic events, environmental triggering conditions, and recommended hybridization combinations.

[0011] In a preferred embodiment of the present invention, in S1, the sensing array includes a distributed sensor node network;

[0012] The sensor node integrates a temperature and humidity sensor, a rain gauge, an anemometer, a quantum optical sensor, a soil moisture probe, and a multi-channel ion-selective electrode.

[0013] The sensor nodes communicate with the central data aggregation unit via a wide area network protocol, and the data sampling frequency is once per minute.

[0014] In a preferred embodiment of the present invention, in step S1, the image acquisition scheduler has a built-in environmental disturbance sensitivity threshold model; when the rate of change of micro-meteorological parameters or soil physicochemical state exceeds a preset threshold, the image acquisition frequency is automatically increased from once a day to once every two hours, and the high-resolution close-range imaging mode is activated simultaneously, while the spatial coverage is limited to the local area where the environmental disturbance occurs.

[0015] In a preferred embodiment of the present invention, in S2, the acquisition is achieved through an imaging unit, which includes an orbital moving platform and a multispectral camera array mounted thereon.

[0016] The multispectral camera array includes a visible light camera, a near-infrared camera, and a thermal imager; the track-mounted mobile platform is laid along the field furrows, with a moving speed of 0.1-1.5 m / s; the multispectral camera array is mounted via a gimbal, with a pitch angle adjustment range of -30° to +60° and a horizontal rotation range of ±180°.

[0017] In a preferred embodiment of the present invention, in S2, the individual plant segmentation adopts an instance segmentation network based on the improved U-Net architecture. The network introduces a dilated convolution module in the encoder stage to expand the receptive field of view, and embeds an attention gating mechanism in the decoder stage to suppress background noise.

[0018] The identity binding operation is implemented through a cross-frame feature matching algorithm, which calculates the cross-union ratio and principal component orientation consistency of plant outlines at adjacent time points. The cross-union ratio threshold is 0.7, and the principal component orientation consistency threshold is 15°.

[0019] In a preferred embodiment of the present invention, in S3, the trajectory model adopts a gated recurrent unit network, the network structure of which includes two hidden layers, each layer having 256 neurons;

[0020] The input to the trajectory model is a vector of plant morphological features arranged in chronological order, including plant height, leaf area index, ear projection area, number of tillers, and canopy temperature gradient.

[0021] Among them, plant height was obtained by fitting lidar point cloud, leaf area index was calculated by near-infrared and red light band reflectance, ear projection area was determined by the area of ​​high temperature region after thermal imaging segmentation, tiller number was obtained by counting the local curvature minimum points of the plant base contour, and canopy temperature gradient was defined as the difference between the thermal imaging temperature of the top and bottom of the plant.

[0022] In a preferred embodiment of the present invention, in step S3, the singular phenotype detection employs the isolated forest algorithm, using key turning points in the trajectory data as input samples. The key turning point is the time point where the absolute value of the first-order difference of the morphological feature vector exceeds twice the historical standard deviation of the feature. The isolated forest algorithm constructs 100 binary trees, and the anomaly detection threshold is set to an average path length of less than 2.5.

[0023] In a preferred embodiment of the present invention, in step S4, the causal association analysis employs the Granger causality test method, using a vector autoregression model to determine whether changes in environmental parameters significantly precede the occurrence of phenotypic mutations. The lag order of the vector autoregression model is determined by the Akaike information criterion, with a significance level of 5%.

[0024] The environment-phenotype response mapping map is stored in the form of a three-dimensional matrix, with row indexes corresponding to environment parameter types, column indexes corresponding to phenotypic feature types, and depth indexes corresponding to time delay steps.

[0025] A high-throughput image data-driven intelligent auxiliary system for crop breeding includes:

[0026] A sensing array is used to acquire real-time information on micrometeorological parameters and soil physicochemical state within the target breeding area;

[0027] An image acquisition scheduler is used to dynamically adjust the operating parameters of the imaging unit based on the output of the sensing array;

[0028] The imaging unit is used to acquire multi-view visible light, near-infrared and thermal infrared image sequences of each plant in the target crop population.

[0029] The module is used to perform individual plant segmentation and identity binding operations on the image sequence to generate a dataset of spatiotemporally continuous images based on individual plants.

[0030] The modeling module is used to build and update individualized dynamic developmental trajectory models for datasets of spatiotemporally continuous images of each plant.

[0031] The detection module is used to perform singular phenotype detection on the trajectory data output by the individualized dynamic developmental trajectory model and identify abnormal developmental events.

[0032] The analysis module is used to perform causal correlation analysis between singular phenotypic events and environmental disturbance records, and generate an environment-phenotypic response mapping map.

[0033] The scoring module is used to calculate the comprehensive breeding value score of each plant based on the environment-phenotype response mapping map and the individualized dynamic development trajectory model.

[0034] The decision-making module is used to output a list of plants with the highest comprehensive breeding value scores and a decision report.

[0035] In a preferred embodiment of the present invention, the construction module uses an instance segmentation network based on an improved U-Net architecture to segment individual plants and achieves identity binding through a cross-frame feature matching algorithm. The dilation rate of the dilated convolution module in the improved U-Net architecture is set to 2, 4, and 8 respectively, and the weight coefficient of the attention gating mechanism is determined by the product of channel attention and spatial attention. In the cross-frame feature matching algorithm, the intersection-union ratio threshold of the plant contour is set to 0.7, and the principal component orientation consistency threshold is set to 15°.

[0036] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0037] (1) This invention provides a high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding. By deploying a sensing array and an intelligent image acquisition scheduler, the system can dynamically adjust the data acquisition strategy according to real-time environmental changes. When the temperature, humidity and other parameters are detected to fluctuate drastically, the scheduler automatically increases the shooting frequency and guides the imaging unit to focus on the sensitive area. This mechanism is based on the correlation model between environmental disturbance and phenotypic response, realizing the transformation from fixed sampling to event-driven, and accurately capturing the key physiological and morphological changes caused by brief stress that are easily missed by traditional methods. Compared with the static acquisition scheme with equal time intervals used in the prior art, this invention effectively solves the problem of data acquisition blind spots caused by the asynchronous nature of crop development and the transient nature of the environment.

[0038] (2) This invention provides a high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding. By integrating a multispectral imaging unit, an improved instance segmentation network and a cross-temporal feature matching algorithm, the system achieves accurate identification and stable tracking of individual plants in complex field populations. The improved network uses dilated convolution to expand the receptive field of view to separate dense plants and suppresses background interference through an attention mechanism. The matching algorithm ensures the continuity of identity based on the spatiotemporal geometric features of the plants, thereby solving the problem of accurate segmentation and long-term tracking of non-standard plant types in heterogeneous populations. It produces a structured phenotypic image sequence with spatiotemporal alignment based on individual plants. Compared with traditional methods that rely on fixed parameters or standard templates, this invention significantly improves the completeness and accuracy of data acquisition in natural and variable environments. Furthermore, the generated high-quality individualized dataset makes it possible to conduct in-depth quantitative analysis of the unique growth dynamics of each material, thereby greatly enhancing the ability to analyze complex quantitative traits. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a three-dimensional structural diagram of a preferred embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0043] Application Overview:

[0044] The core contradiction of current high-throughput crop phenotyping technologies lies in the fundamental mismatch between the fixed data collection patterns and the dynamic changes in crop growth and the environment. Whether it is static data collection at equal time intervals or analysis methods that rely on preset templates, the underlying logic assumes that crop growth is uniform, synchronous, and predictable. However, in the actual field environment, crop growth is nonlinear and asynchronous, and its key phenotypic responses are often triggered by brief and specific environmental disturbances. This contradiction makes it difficult for existing technologies to accurately capture transient phenotypic events driven by environmental changes in highly heterogeneous natural populations, and also makes it impossible to characterize the unique and nonlinear adaptive trajectories of individual plants. This limits the efficiency and accuracy of analyzing the genetic basis of complex traits and screening superior germplasm from massive amounts of data.

[0045] This solution addresses the challenge of accurately and efficiently identifying individuals with superior stress resistance and adaptability, and understanding their trait formation mechanisms, within highly heterogeneous crop populations. It employs environmentally responsive adaptive data acquisition, precise individual plant identification and tracking, individualized dynamic developmental trajectory modeling, and environment-phenotype causal correlation analysis. The system monitors the environment in real-time using a sensing array and dynamically triggers imaging units for focused high-frequency acquisition, ensuring the capture of key phenotypic events. An improved instance segmentation and feature matching algorithm is then used to construct a spatiotemporal continuous profile for each plant. A gated recurrent unit network is employed to establish a dynamic developmental model for each individual, and the isolated forest algorithm is used to detect singular phenotypes. Finally, a Granger causality test is used to establish a quantitative mapping between environmental disturbances and phenotypic responses, and based on this, a multidimensional breeding value score for the plant is calculated. Compared to existing technologies, this solution represents a paradigm shift from passive recording to active perception, from population averaging to individual precision, and from data description to causal understanding, significantly improving the intelligence and scientific rigor of breeding pre-selection.

[0046] Example 1:

[0047] This invention provides a high-throughput image data-driven intelligent auxiliary decision-making method for crop breeding, comprising the following steps:

[0048] S1: Deploy a sensing array to acquire real-time information on micro-meteorological parameters and soil physicochemical state within the target breeding area. Based on the real-time data stream from the sensing array, dynamically adjust the imaging frequency, imaging mode, and spatial coverage of the imaging unit through the image acquisition scheduler.

[0049] S2: Collect multi-view image sequences of each plant in the target crop population, including visible light, near infrared and thermal infrared images, perform individual plant segmentation on the image sequences, bind them with identity, and generate a dataset of spatiotemporal continuous images of individual plants.

[0050] S3: For each plant's dataset, construct an individualized dynamic development trajectory model, update the model parameters through online learning to reflect the plant's response to environmental disturbances, and perform singular phenotypic detection on the trajectory data output by the trajectory model to identify abnormal development events.

[0051] S4: Perform causal correlation analysis on the detected unusual phenotypic events and the corresponding environmental disturbance records to generate an environment-phenotypic response mapping map;

[0052] S5: Based on the mapping map and trajectory model, calculate the comprehensive breeding value score of each plant, output a list of plants with the highest comprehensive breeding value scores, and generate a decision report that includes key phenotypic events, environmental triggering conditions, and recommended hybridization combinations.

[0053] It should be noted that this invention provides a high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding. It constructs a closed-loop intelligent system of environmental perception, dynamic acquisition, individual modeling, and intelligent decision-making. It captures the real growth dynamics of crops through high spatiotemporal resolution multimodal data and uses artificial intelligence technology to establish the causal relationship between environmental disturbances and phenotypic responses, ultimately achieving a precise quantitative assessment of the breeding value of plants.

[0054] This system monitors changes in microclimate and soil environment in real time through a high-density sensor network deployed in the field, and dynamically triggers the operating mode of image acquisition equipment based on this information. When an environmental disturbance that could trigger phenotypic mutations is detected, the system automatically increases the acquisition frequency and adjusts imaging parameters to ensure that key physiological events are captured. Computer vision algorithms are used to segment plant instances and track their identities across time points in the acquired multispectral images, creating a complete growth profile for each crop throughout its entire growth cycle.

[0055] This addresses the mismatch between fixed collection frequencies and asynchronous crop development patterns when applying existing high-throughput phenotyping technologies to highly heterogeneous wild closely related germplasm resources. Traditional methods cannot capture transient phenotypic events triggered by short-term environmental changes. Furthermore, static analysis models based on the population synchronous development assumption struggle to characterize the unique, nonlinear growth trajectories of individual plants. Standardized imaging and analysis procedures lack robustness to common unusual configurations in wild germplasm, such as ear deformities and tillering, leading to data invalidation.

[0056] The above approach achieves a paradigm shift from passive recording to active perception, enhancing the probability of capturing key phenotypic events through environment-responsive data acquisition. By establishing a dynamic developmental trajectory model at the individual plant level, the system can accurately characterize the adaptation strategies of each individual in complex environments, achieving a shift from population averages to precise individual analysis. The employed singular phenotypic detection and environment-phenotype causal association analysis effectively solve the problem of analyzing complex traits in wild germplasm resources. Finally, by integrating a multidimensional scoring system that considers stability, adaptability, and genetic potential, the system provides breeders with quantifiable and interpretable decision-making support, significantly accelerating the screening process for superior germplasm with broad adaptability and stress resistance, and providing a powerful intelligent auxiliary tool for crop breeding.

[0057] S1: Deploy a sensing array to acquire real-time information on micro-meteorological parameters and soil physicochemical state within the target breeding area. Based on the real-time data stream from the sensing array, dynamically adjust the imaging frequency, imaging mode, and spatial coverage of the imaging unit through the image acquisition scheduler.

[0058] In a preferred embodiment of the present invention, in S1, the sensing array includes a distributed sensor node network;

[0059] The sensor node integrates a temperature and humidity sensor, a rain gauge, an anemometer, a quantum optical sensor, a soil moisture probe, and a multi-channel ion-selective electrode.

[0060] The sensor nodes communicate with the central data aggregation unit via a wide area network protocol, and the data sampling frequency is once per minute.

[0061] In a preferred embodiment of the present invention, in step S1, the image acquisition scheduler has a built-in environmental disturbance sensitivity threshold model; when the rate of change of micro-meteorological parameters or soil physicochemical state exceeds a preset threshold, the image acquisition frequency is automatically increased from once a day to once every two hours, and the high-resolution close-range imaging mode is activated simultaneously, while the spatial coverage is limited to the local area where the environmental disturbance occurs.

[0062] It should be noted that in step S1, a distributed sensor node network deployed in the breeding fields comprehensively captures micro-meteorological parameters of the target area at a high sampling rate of once per minute, such as air temperature, humidity, wind speed, rainfall, solar radiation intensity, and soil physicochemical states, such as soil moisture content, nitrogen, phosphorus, and potassium content, and pH value. These sensor nodes aggregate the data to the central processing unit via a low-power wide-area network protocol, forming a continuous and complete digital stream of the field environment.

[0063] This environmental digital stream is input in real time into the environmental disturbance sensitivity threshold model built into the image acquisition scheduler. This model quantifies the rate of change through mathematical difference calculations. For example, for air temperature, its decision function can be expressed as: when the absolute value of the temperature change ΔT / Δt exceeds a preset threshold θ within a sliding time window Δt, e.g., 10 minutes... T， For example, a temperature of 1.5℃ / 10min is considered a significant environmental disturbance event. This model can be formally represented as a series of judgment conditions, for the i-th environmental parameter E. i If dE is satisfied i / dt>θ i If so, a higher-level acquisition command will be triggered.

[0064] Once the model determines that any change in any environmental parameter exceeds its preset threshold, the scheduler immediately issues an instruction to dynamically adjust the three key operating parameters of the high-throughput imaging unit, thereby achieving precise optimization of the acquisition strategy. The imaging frequency has been significantly increased from once a day to once every two hours. This leap in frequency ensures that rapid physiological responses of crops under environmental stress can be captured with high temporal resolution, such as stomatal closure, leaf wilting, or instantaneous changes in ear morphology, thus avoiding the omission of key phenotypic events.

[0065] Secondly, the imaging mode is simultaneously switched to a high-resolution close-up mode. This means that the imaging platform slows down its movement and adjusts the camera gimbal angle to capture detailed images of key parts of the plant, such as the spike and new leaves, to obtain richer morphological details. Finally, the spatial coverage is intelligently limited to the local area where environmental disturbances occur. This focusing greatly improves the efficiency of data acquisition and reduces the generation and processing burden of redundant data.

[0066] In summary, step S1, through a deep closed-loop linkage between environmental perception and image acquisition, fundamentally solves the problem in existing technologies where fixed acquisition frequencies cannot capture key phenotypic events triggered by short-term environmental changes. The active perception mechanism ensures that the acquired image data sequence is not a static snapshot at uniform time intervals, but rather a high-value dynamic record surrounding key nodes of crop physiological activity. This provides high-quality, timely data for subsequent steps to construct dynamic developmental trajectory models that accurately reflect the individual plant's response to environmental disturbances, as well as for accurately identifying unusual phenotypes and conducting causal correlation analysis.

[0067] S2: Collect multi-view image sequences of each plant in the target crop population, including visible light, near infrared and thermal infrared images, perform individual plant segmentation on the image sequences, bind them with identity, and generate a dataset of spatiotemporal continuous images of individual plants.

[0068] In a preferred embodiment of the present invention, in S2, the acquisition is achieved through an imaging unit, which includes an orbital moving platform and a multispectral camera array mounted thereon.

[0069] The multispectral camera array includes a visible light camera, a near-infrared camera, and a thermal imager; the track-mounted mobile platform is laid along the field furrows and moves at a speed of 1 m / s. The multispectral camera array is mounted via a gimbal with a pitch angle adjustment range of -30° to +60° and a horizontal rotation range of ±180°.

[0070] In a preferred embodiment of the present invention, in S2, the individual plant segmentation adopts an instance segmentation network based on the improved U-Net architecture. The network introduces a dilated convolution module in the encoder stage to expand the receptive field of view, and embeds an attention gating mechanism in the decoder stage to suppress background noise.

[0071] The identity binding operation is implemented through a cross-frame feature matching algorithm, which calculates the cross-union ratio and principal component orientation consistency of plant outlines at adjacent time points. The cross-union ratio threshold is 0.7, and the principal component orientation consistency threshold is 15°.

[0072] It should be noted that in step S2, high-throughput imaging and computer vision algorithms are used to convert the data into digital data that can be used for subsequent in-depth analysis. First, multimodal image data is acquired using engineered imaging units. Then, the raw image data is intelligently identified and tracked individually, thereby building a complete digital twin profile for each crop throughout its entire growth period.

[0073] The imaging unit addresses imaging challenges arising from the irregular plant architecture and uneven spatial distribution of wild-related germplasm resources. A track-mounted mobile platform, laid along the furrows, ensures the stability and repeatability of the imaging perspective. Its adjustable speed (0.1-1.5 m / s) allows it to flexibly adapt to different instructions issued by the scheduler in step S1, enabling rapid scanning at 1.2 m / s during routine surveys and slow, detailed acquisition at 0.3 m / s in high-resolution close-up mode.

[0074] The integrated gimbal provides a high degree of freedom, with a pitch range of -30° to +60° and a horizontal range of ±180°. This allows the multispectral camera array, including high-resolution visible light, near-infrared, and thermal imagers, to capture images of the plant from multiple preset angles. This effectively avoids feature loss caused by single-view occlusion and ensures comprehensive phenotypic information can be obtained even for unusual plant configurations such as bushy tillers and drooping spikes. This multimodal imaging strategy provides morphological information through visible light images, physiological state information such as leaf area index through near-infrared images, and direct reflection of plant stress conditions and canopy temperature through thermal infrared images.

[0075] After acquiring the original image sequence, each plant instance in the image is segmented and accurately associated with images at different time points, i.e., individual plant segmentation is bound to its identity. Individual plant segmentation adopts an instance segmentation network based on the improved U-Net architecture. The standard U-Net is an encoder-decoder structure network for biomedical image segmentation, but it has a limited field of view and is easily affected by background noise when faced with complex field backgrounds and densely intersecting plants.

[0076] The improvement of this invention lies in introducing a dilated convolution module in the encoder stage. Mathematically, for the input feature map F, the dilated convolution operation D(K,d) inserts d-1 zero values ​​between elements of the standard convolution kernel K to expand the receptive field without increasing the number of parameters. Specifically, dilated convolutions with dilation rates d of 2, 4, and 8 are used in the third, fourth, and fifth layers of the network, respectively. This allows the network to fuse a wider range of contextual information at deeper layers, thereby more accurately determining whether a pixel belongs to a plant or the background, which is particularly beneficial for separating closely adjacent plants.

[0077] An attention gating mechanism is embedded in the decoder stage, which automatically generates attention weights through learning. This is used to suppress irrelevant background regions. The process involves fusing the feature map g from the encoder and the upsampled feature map x from the decoder, and then generating a gating signal through a small network, such as a 1x1 convolution plus a Sigmoid function. Where W and b are learnable parameters. The final output is This mechanism enables the network to focus on the target plant area when restoring image resolution, effectively suppressing noise from soil, shadows, and weeds, and greatly improving the robustness and accuracy of segmenting irregular plants in natural environments.

[0078] After segmentation, identity binding is performed using a cross-frame feature matching algorithm to address the issue of correctly tracking the same plant at different time points. This algorithm bases matching on two geometric features: Intersection over Union (IoU) and principal component orientation consistency. For two plant contour masks A at adjacent time points t and t+1... t and B t+1 The formula for calculating its intersection-union ratio is: .

[0079] Simultaneously, principal component analysis (PCA) is performed on each contour to extract its first principal component direction vector. And calculate the angle θ between the two direction vectors. The matching criterion is set as follows: if IoU > 0.7 and θ < 15°, then A is considered a match. t and B t+1 This means that the plants belong to the same plant. This dual-threshold criterion ensures that the IoU ensures a high degree of overlap in the spatial location of the plants, while the principal component orientation consistency, which represents the main growth direction of the plants, ensures the continuity of their posture. The combination of the two can effectively cope with the morphological changes brought about by plant growth and the slight swaying caused by a breeze, ensuring the accuracy of individual identification tracking in a highly heterogeneous population.

[0080] In step S2, a spatiotemporally continuous, multimodal image dataset was generated, based on individual plants. The massive raw image data collected by the imaging unit in step S1 was processed into highly structured semi-finished data. This dataset serves as a reliable data source for step S3 to construct individualized dynamic developmental trajectory models.

[0081] S3: For each plant's dataset, construct an individualized dynamic development trajectory model, update the model parameters through online learning to reflect the plant's response to environmental disturbances, and perform singular phenotypic detection on the trajectory data output by the trajectory model to identify abnormal development events.

[0082] In a preferred embodiment of the present invention, in S3, the trajectory model adopts a gated recurrent unit network, the network structure of which includes two hidden layers, each layer having 256 neurons;

[0083] The input to the trajectory model is a vector of plant morphological features arranged in chronological order, including plant height, leaf area index, ear projection area, number of tillers, and canopy temperature gradient.

[0084] Among them, plant height was obtained by fitting lidar point cloud, leaf area index was calculated by near-infrared and red light band reflectance, ear projection area was determined by the area of ​​high temperature region after thermal imaging segmentation, tiller number was obtained by counting the local curvature minimum points of the plant base contour, and canopy temperature gradient was defined as the difference between the thermal imaging temperature of the top and bottom of the plant.

[0085] In a preferred embodiment of the present invention, in step S3, the singular phenotype detection employs the isolated forest algorithm, using key turning points in the trajectory data as input samples. The key turning point is the time point where the absolute value of the first-order difference of the morphological feature vector exceeds twice the historical standard deviation of the feature. The isolated forest algorithm constructs 100 binary trees, and the anomaly detection threshold is set to an average path length of less than 2.5.

[0086] It should be noted that in step S3, the spatiotemporal continuous image dataset of a single plant generated in step S2 is transformed into a deep understanding of the plant's growth and development dynamics. Phenotypic events with special value are selected from this dataset. By using a deep learning model suitable for processing time series, a virtual plant that can be dynamically updated and simulates the growth pattern of each plant is constructed. An anomaly detection algorithm is used to scan the continuous trajectory output by this virtual plant to locate special growth events that deviate from the normal pattern.

[0087] The individualized dynamic development trajectory model was constructed by employing a gated recurrent unit (GRU) network to model the unique growth sequence pattern of each plant. GRU is a variant of recurrent neural networks specifically designed for processing sequential data. It introduces update and reset gates to control the flow and forgetting of information. (Update gate z) t The new information h that determines the current moment t The formula for determining the extent to which the memory state is updated is as follows: Where σ is the sigmoid function h t-1 It is the hidden state from the previous moment, x t This is the current input.

[0088] Reset door r t This determines how much past information needs to be forgotten, and the formula is: The current candidate hidden state. Depend on Calculate the final current state. This mechanism enables GRU to effectively capture long-term dependencies in time series; for example, drought stress experienced by plants during the seedling stage may have a delayed effect on their performance during the heading stage.

[0089] In this invention, the GRU network is specifically designed to contain two hidden layers, each with 256 neurons. This depth and width provide it with sufficient expressive power to learn the complex, nonlinear growth patterns exhibited by highly heterogeneous wild germplasm resources. More importantly, the model employs an online learning approach. Whenever new phenotypic observation data arrives, for example, morphological features parsed from new images obtained based on the S1 trigger mechanism, the model uses this data to fine-tune its parameters through a backpropagation algorithm at each time step. This allows its internal state to continuously track and adapt to the plant's real-time response to environmental disturbances, such as a rainfall or a sudden temperature rise, thereby achieving a dynamic and precise characterization of the individualized developmental trajectory of each plant.

[0090] The input to this GRU model, namely the plant morphological feature vector, has five dimensions designed from the multimodal image data of the S2 step. Plant height is fitted using lidar point cloud data, providing three-dimensional structural information of the plant; leaf area index is calculated from near-infrared and red light reflectance based on specific vegetation indices, reflecting the photosynthetic potential of the canopy; spike projection area is determined through thermal imaging segmentation, utilizing the temperature difference that usually exists between the spike and leaves; tiller number is obtained by geometric analysis of the plant base contour and counting local curvature minima, solving the problem of counting dense tillers; and the canopy temperature gradient directly reflects the plant's water stress status. These five features constitute a comprehensive description of the plant's state from morphological, physiological, and response perspectives, providing the GRU model with high-quality, multi-dimensional input, enabling it to learn richer growth dynamics.

[0091] After obtaining the continuous developmental trajectory of each plant, the anomalous phenotype detection module begins its work, aiming to automatically identify anomalous individuals with breeding value, such as plants with special stress resistance, extremely high yield potential, or unique agronomic traits. This invention employs the Isolation Forest algorithm, an unsupervised algorithm suitable for anomaly detection in high-dimensional data. Its anomalous data points, due to their sparseness and diversity, are more easily isolated in the feature space. Isolation Forest constructs multiple binary trees (iTrees) by randomly selecting features and split values.

[0092] When constructing an iTree, subsamples are randomly drawn from the dataset, and features q and splitting values ​​p are recursively randomly selected until all samples are isolated, each sample occupies a leaf node, or the tree reaches its height limit. For a dataset containing N samples, T such iTrees are constructed to form a "forest".

[0093] During detection, the path length h(x) of each sample x is calculated, which is the number of edges traversed from the root node to the leaf node where the sample is located. Abnormal samples are easier to isolate, and their average path length E(h(x)) is usually shorter. This invention makes key application optimizations to the algorithm; instead of directly detecting the entire trajectory sequence, it focuses on key turning points in the trajectory, namely, the time points where the absolute value of the first-order difference of the morphological feature vector exceeds twice the historical standard deviation of that feature. It captures moments when plant growth rates change drastically; these moments often correspond to significant responses to environmental disturbances, such as growth stagnation under stress or rapid growth after recovery, and are the golden window for discovering unusual phenotypes.

[0094] The algorithm constructs 100 binary trees, using the feature vectors at key inflection points as input samples. Finally, it calculates the average path length of each sample across all trees and sets a threshold of <2.5 for an anomaly. This method effectively identifies individuals that deviate from the mainstream developmental pattern of the population, such as plants that recover extremely quickly after drought or those with abnormally rapid ear development.

[0095] The S3 step represents a leap in analysis, moving from static description to dynamic modeling and from population patterns to individual anomalies. It empowers the system to understand the unique growth process of each plant, rather than simply capturing a snapshot at a specific point in time. The GRU model elevates raw image-derived data into dynamic trajectories containing growth patterns and response modes, which forms the basis for stability and adaptability assessments.

[0096] The Isolation Forest algorithm can efficiently and automatically select the most valuable candidates from thousands of individual trajectories, providing a clear objective and high-quality input for the subsequent causal analysis in step S4 and the comprehensive scoring in step S5.

[0097] S4: Perform causal correlation analysis on the detected unusual phenotypic events and the corresponding environmental disturbance records to generate an environment-phenotypic response mapping map;

[0098] In a preferred embodiment of the present invention, in step S4, the causal association analysis employs the Granger causality test method, using a vector autoregression model to determine whether changes in environmental parameters significantly precede the occurrence of phenotypic mutations. The lag order of the vector autoregression model is determined by the Akaike information criterion, with a significance level of 5%.

[0099] The environment-phenotype response mapping map is stored in the form of a three-dimensional matrix, with row indexes corresponding to environment parameter types, column indexes corresponding to phenotypic feature types, and depth indexes corresponding to time delay steps.

[0100] It should be noted that in step S4, by deeply correlating the unusual phenotypic events detected in S3 with the high-resolution environmental disturbance data continuously recorded in S1, a knowledge graph that quantitatively describes the environmental stimulus-phenotypic response pattern is constructed using rigorous econometric methods, providing a scientific basis for understanding the plant's adaptation mechanism and assessing its breeding value.

[0101] This step employs Granger causality, which posits that if a change in an environmental parameter significantly improves the predictive power of future phenotypic changes, then statistically, that environmental parameter can be considered a Granger cause of the phenotypic change. This mature time series analysis method is tested, and its application is customized. The test relies on a vector autoregression (VAR) model. For a time series E containing environmental parameters... t and phenotypic characteristics time series P t For a binary system, the VAR model represents it as a system of linear equations: .

[0102] Here, p is the lag order, which determines how many time steps of historical information are considered. The core of the Granger causality test is to examine the lag term of the environmental parameter E in the second equation, i.e. Whether the joint significance is not zero. Specifically, first estimate a restricted model, that is, assume that E is not a Granger cause of P, and let all = 0, then estimate an unrestricted model, i.e., a lagged term including E. Then construct the F-statistic:

[0103] .

[0104] Among them, RSS r and RSS ur These are the sum of squared residuals for the restricted and unrestricted models, respectively, and T is the sample size. If the p-value corresponding to the calculated F-statistic is less than the preset significance level (set to 5% in this invention), the null hypothesis is rejected, and the environmental parameter E is considered to be a Granger cause of phenotype P.

[0105] In this invention, the method is optimized for two parameters. First, the lag order p is automatically determined using the Akaike Information Criterion (AIC). The formula for calculating the AIC criterion is as follows: Where k is the number of model parameters and L is the likelihood function value. The optimal lag order is chosen as p, which minimizes the AIC value. Different environmental factors, such as sudden temperature changes and persistent soil drought, have different time scales of influence on phenotypes, i.e., different lag effects. The AIC criterion can adaptively find the most suitable causal time window based on the data itself, avoiding the bias that may be caused by artificially setting the lag order.

[0106] Secondly, the data used in the analysis is based on the time point of each anomalous phenotypic event detected by S3, extracting environmental and phenotypic data from the 72 hours preceding its occurrence. This timeframe is based on crop physiology to ensure coverage of the complete process from the occurrence of environmental stress to the appearance of a visible phenotypic response.

[0107] This step generates a structured environment-phenotypic response mapping map, stored as a three-dimensional matrix. The three dimensions correspond to the environmental parameter type, phenotypic trait type, and time delay step, respectively. Each element in the matrix stores the Granger causality test F-statistic value for the corresponding triple: environmental factor, phenotypic trait, and time delay. This value quantifies the strength of the causal relationship. This map is no longer just scattered data points, but a systematic knowledge base that clearly shows which environmental factors significantly affect which phenotypic traits, and after what time delay this effect typically manifests.

[0108] Step S4 elevates the massive amounts of data acquired in the preceding steps—environmental data, image data, and trajectory data—to the level of mechanism understanding. By establishing causal relationships, the system not only knows that a particular plant recovered after drought, but can also quantitatively demonstrate that the decrease in soil moisture content is a Granger cause of the increased canopy temperature gradient (stress symptoms) and the slow recovery of leaf area index.

[0109] This deep understanding is the direct basis for the comprehensive scoring of breeding value in step S5, especially for calculating stability and adaptability scores. For example, a plant that can maintain a relatively stable internal phenotypic trajectory when the external environment fluctuates drastically will have a higher stability score; while a plant whose phenotypic trajectory can quickly return to the normal path after the environmental stress is removed will have a better adaptability score.

[0110] S5: Based on the mapping map and trajectory model, calculate the comprehensive breeding value score of each plant, output a list of plants with the highest comprehensive breeding value scores, and generate a decision report that includes key phenotypic events, environmental triggering conditions, and recommended hybridization combinations.

[0111] It should be noted that in step S5, the massive amounts of multimodal data, dynamic models, and causal relationships generated in the preceding steps are condensed into a quantifiable, interpretable, and directly guiding intelligent decision-making report for breeding practices. Through the constructed comprehensive scoring system, the breeding value of each plant in the population is comprehensively assessed, and based on this, a final plan including optimal germplasm selection and hybridization combination recommendations is generated, thus achieving a closed loop from data insight to breeding action.

[0112] This step comprehensively evaluates the breeding potential of plants from three core dimensions: stability, adaptability, and genetic potential. The design of these three dimensions directly stems from the previous steps, reflecting a high degree of coupling between the final decision-making and the whole-process analysis. The stability score quantifies the phenotypic fluctuation amplitude of plants in a variable environment, and its calculation formula is , where and are respectively the mean standard deviation and the mean of the mean values of multiple morphological characteristics of the plant during the whole growth period, such as plant height and leaf area index.

[0113] This calculation completely depends on the individualized dynamic development trajectory model constructed in Step S3. The continuous time-series data provided by this model is the basis for calculating reliable standard deviations. The adaptability score evaluates the recovery ability and final performance of plants after encountering stress, and is calculated by . Among them, the calculation of the recovery rate Recovery_Rate depends on the stress events and recovery time windows located by the environment-phenotype response mapping map generated in Step S4. For example, using the map to find the starting point and ending point of the phenotypic decline caused by soil drought, so as to accurately calculate the recovery rate. The final biomass is estimated by parameters such as plant height and leaf area index in the trajectory model.

[0114] The genetic potential score aims to identify individuals whose growth trajectories are similar to those of known excellent parents. Its core algorithm is dynamic time warping (DTW). DTW is a classic algorithm for measuring the similarity between two time series of different lengths. For the trajectory P=(p1,p2, ...,p m ) of the plant to be evaluated and the trajectory R=(r1,r2,...,r n ) of the reference excellent parent, DTW constructs an m×n distance matrix and finds a path W from (1,1) to (m,n), called the warping path, so that the cumulative distance of all points on the path is the smallest. Its mathematical expression is: .

[0115] Among them, is the Euclidean distance between two points, and K is the length of the warping path. The genetic potential score is set to (when DTW(P,R)<0.3, otherwise 0). This score highly depends on the accurate and individualized spatio-temporal continuous trajectories constructed in Steps S2 and S3. Without high-quality trajectory data, the DTW comparison will be meaningless.

[0116] Finally, the comprehensive breeding value score S comprehensive is the weighted sum of the above three scores: S comprehensive =w1S stability +w2S adaptation + w3S geneticIn this invention, the weighting coefficients are set to 0.4, 0.35, and 0.25, respectively. This allocation of weights reflects the priorities in breeding practice: while ensuring stable yields, pursuing adaptability to adverse environments, and simultaneously exploring potential superior genetic backgrounds.

[0117] Step S5 generates an operational breeding decision report. The system outputs a list of plants ranked in the top 5% by overall score. For each selected superior candidate plant, the report not only lists its score but also deeply integrates information from all previous steps: including key phenotypic events extracted from S3, such as a sudden increase in ear area on a certain day and its timestamp; environmental triggering conditions derived from causal analysis in S4, such as the event being caused by a specific light and temperature combination in the previous 48 hours; and recommended hybridization combinations based on genetic potential scores and intrapopulation genetic similarity analysis.

[0118] In summary, the S5 step integrates the information flow of the entire intelligent system, transforming the results of a series of technical processes, including environmental perception, image acquisition, individual trajectory modeling, and causal correlation analysis, into clear and explicit decision-making basis for breeders. This changes the traditional breeding dilemma of relying on experience, long cycles, and low efficiency. Through a data-driven approach, it enables the rapid and accurate discovery of hidden superior genotypes in highly heterogeneous populations and provides scientific guidance for hybridization, significantly improving the intelligence level and decision-making efficiency of crop breeding.

[0119] Example 2:

[0120] In this invention, for densely planted wheat or rice, a combined air-ground collection mode is proposed, and a multi-scale attention mechanism is used for plant segmentation in the later S2 stage to solve the problems of occlusion and false detection under high-density planting conditions.

[0121] S1: Deploy a sensing array and collect data in a coordinated air-to-ground manner.

[0122] In this embodiment, the sensing array and imaging unit in this step adopt a three-dimensional mobile acquisition mode with air-ground cooperation to meet the data acquisition needs of large-area breeding fields, specifically:

[0123] Aerial data acquisition utilizes drones equipped with high-resolution RGB and multispectral cameras to collect low-altitude orthophotos along a preset flight path, thereby obtaining macroscopic phenotypic data of the top of the crop canopy.

[0124] Ground-based supplementary data acquisition utilizes a track-mounted mobile platform laid along field furrows at a speed of 1.0 m / s, upon which a multispectral camera array is mounted. The multispectral camera array is mounted via a gimbal with a pitch adjustment range of -30° to +60° and a horizontal rotation range of ±180°. Side-view images are acquired below the canopy to supplement blind spots in the top view, obtaining detailed information on stems and lower leaves. All acquisition devices are integrated with RTK-GPS modules to ensure that each image has geospatial coordinates.

[0125] S2: Multi-scale attention segmentation mechanism segments plants.

[0126] In this step, addressing the problem of complex backgrounds and severe occlusion between ears of wheat and fruits in densely planted crops, this embodiment proposes a crop phenotypic segmentation network (MSA-CropNet) based on a multi-scale attention mechanism to segment individual plants. Specifically:

[0127] The network receives multimodal image input and extracts feature maps of the image layer by layer through the encoder. In this process, the network needs to preserve high-level semantic information, such as that this is a wheat plant, while not losing low-level detailed information, such as that this is the edge of a leaf tip.

[0128] Faced with the complex background interference of green light and leaves in the field, the network needs to distinguish between the crop itself and weeds and soil. After feature extraction, MSA-CropNet uses a feature filtering mechanism to evaluate the importance of each channel and each spatial location in the feature map, automatically suppressing the response value of background noise and enhancing the response value of the target crop.

[0129] To address the issue of adjacent plants easily sticking together under dense planting conditions, the network predicts the target category and location coordinates separately through decoupled detection heads, and uses a special rejection mechanism to forcibly separate prediction boxes with excessive overlap, thereby achieving accurate counting and outline delineation of individual plants or ears of grain.

[0130] In the above, the crop phenotypic segmentation network with multi-scale attention mechanism is a convolutional neural network. However, in traditional convolutional neural networks, when multiple samplings are performed to expand the receptive field of view, detailed information about crop edges is lost, resulting in uneven leaf edge segmentation and affecting the calculation accuracy of leaf area index (LAI). This invention uses dilated convolutions instead of standard convolutions in the last two deep layers of the backbone network of the convolutional neural network, namely ResNet. Specifically, the dilation rates are set to r=2, 4, and 8, which multiplies the receptive field of the network without reducing the feature map resolution, i.e., without downsampling. This allows the network to capture the global morphology of the entire plant, avoiding the fragmentation of large plants for identification, while retaining pixel-level features of small structures such as leaf tips and petioles, avoiding the missed detection of small objects, and significantly improving the segmentation ability of fine crop structures.

[0131] Furthermore, this invention embeds a custom dual attention module between the feature extraction network and the prediction head, which includes channel attention and spatial attention. This prevents green weeds from being mistaken for green crops in field images due to their similar colors, and avoids misidentification of light and shadow as fruits or dark leaves.

[0132] In the aforementioned channel attention, channel weights are obtained through global average pooling. Logically, this automatically identifies and amplifies crop-specific spectral feature channels, such as those sensitive to chlorophyll reflectance, while suppressing the weights of feature channels representing soil and shadows. Spatial attention, on the other hand, generates a spatial mask through convolution operations, forcing the network to focus on the foreground crop region of the image and ignoring the background region. The logical formula is as follows:

[0133] .

[0134] This logic formula describes the dynamic weighting process from raw input to refined output in a field mapping, aiming to simulate the human eye's ability to focus on target crops in complex field backgrounds.

[0135] This represents elements and multiplication, thereby giving the network the ability to focus like a human eye, in the first multiplication. Irrelevant feature signals, such as soil color signals, were filtered out; in the second multiplication... Noise in the background area, such as shadows between crop gaps, was filtered out.

[0136] This allows the system to automatically filter out background noise during the automatic identification process, even in environments with abundant weeds or uneven lighting, significantly reducing the false detection rate and ensuring that the segmented target is only the target crop itself.

[0137] The input feature map represents the original feature map extracted by the backbone network. In a field scene, this part of the data is mixed with various information, with crops as the foreground and soil, weeds, and shadows as the background. The signal strength of each information may be similar, such as weeds and crops being green.

[0138] The channel attention weights represent what the feature map focuses on. The network analyzes the feature channels through global pooling and generates a weight vector. When a channel responds strongly to chlorophyll reflectance or a specific texture, the network automatically assigns a high weight close to 1 to that channel; while for channels that respond to soil color or random noise, they are assigned a low weight close to 0.

[0139] Spatial attention weights indicate where the feature map focuses. After channel weighting, the network further analyzes the spatial location and generates a two-dimensional mask. The network identifies which pixel regions in the image belong to foreground entities and assigns them high weights, while which regions belong to background gaps and assign them low weights.

[0140] In this embodiment, a decoupled prediction head is also employed to ensure that in densely planted crop areas, such as wheat breeding areas, adjacent fruits / ears often physically contact or overlap, avoiding the situation where traditional networks identify two targets as one or the prediction box drifts. The loss function is defined as follows:

[0141] .

[0142] Furthermore, in the aforementioned loss function, exclusion terms for the true target are introduced respectively. ) and exclusion terms between prediction boxes ( The repulsion term for the real target is used to penalize predicted boxes that shift to neighboring non-corresponding real targets. The repulsion term between predicted boxes is used to create magnetic repulsion between predicted boxes from different individuals.

[0143] ;

[0144] .

[0145] The exclusion term of the above real target ( The formula aims to address the problem of prediction frames drifting to adjacent plants under dense planting conditions.

[0146] The prediction box represents a prediction target generated by the network, such as the bounding box of a wheat plant.

[0147] The ground truth bounding box represents the actual crop target corresponding to the predicted bounding box.

[0148] For non-matching nearby ground truth boxes, it represents all other real crop targets in the image besides the target ground truth box. In densely planted wheat fields, it refers to neighboring plants that are very easy to confuse with the target plant.

[0149] To determine the ratio of the intersection to the actual frame area, calculate the proportion of the overlap area between the predicted frame and the neighboring actual frames to the area of ​​the actual frame.

[0150] This is achieved through the exclusionary term of the real target ( The formula makes the prediction box... It should only cover its corresponding target. ,when Because of obstruction or blurring, it mistakenly drifted and covered the adjacent plant G, then The value of will increase, causing the loss function to... By minimizing this loss, the network is forced to learn to avoid nearby interfering targets, so that it can still draw a clear boundary line when two crops are close together, instead of framing them together.

[0151] Exclusion terms between prediction boxes ( The formula aims to solve the problem of overlapping prediction boxes caused by multiple object detections or multiple object adhesion in dense scenes.

[0152] Different prediction boxes represent two different prediction boxes generated by the network, and these two boxes correspond to different real targets, or one of them is a false detection.

[0153] The intersection-over-union ratio (IoU) is used to calculate the degree of overlap between two predicted bounding boxes.

[0154] It is a smoothed exponential function. It is a monotonically increasing penalty function. The greater the overlap of the input x, the greater the penalty value of the output.

[0155] In this way, by predicting the exclusion terms between boxes ( The formula makes predicted boxes belonging to different targets mutually exclusive. In areas where wheat ears or fruits are extremely dense, traditional networks tend to generate two highly overlapping boxes for two adjacent targets, leading to unclear counts or blurred boundaries. By introducing... And stipulate and For two different objects to be detected, when the overlap between the two is... When it is too high, This increases, thereby predicting the exclusion term between frames ( ) and exclusion terms between prediction boxes ( The synergistic addition of [a certain element] to the convolutional network creates a magnetic repulsion effect, forcing the dense prediction boxes to separate from each other and shrink to their respective target centers, thereby significantly improving the accuracy of single plant / single ear counting under heavy shading.

[0156] Example 3:

[0157] This invention also provides a high-throughput image data-driven intelligent auxiliary system for crop breeding, comprising:

[0158] The system comprises: a sensing array for real-time acquisition of micro-meteorological parameters and soil physicochemical state information within the target breeding area; an image acquisition scheduler for dynamically adjusting the operating parameters of the imaging unit based on the output of the sensing array; an imaging unit for acquiring multi-view visible light, near-infrared, and thermal infrared image sequences of each plant in the target crop population; a construction module for performing individual plant segmentation and identity binding operations on the image sequences to generate a dataset of spatiotemporal continuous images of individual plants; and a modeling module for constructing and updating an individualized dynamic development trajectory model for the dataset of spatiotemporal continuous images of each plant.

[0159] The detection module is used to detect singular phenotypes in the trajectory data output by the individualized dynamic development trajectory model and identify abnormal development events; the analysis module is used to perform causal correlation analysis between singular phenotype events and environmental disturbance records to generate an environment-phenotype response mapping map; the scoring module is used to calculate the comprehensive breeding value score of each plant based on the environment-phenotype response mapping map and the individualized dynamic development trajectory model.

[0160] The decision-making module is used to output a list of plants with the highest comprehensive breeding value scores and a decision report.

[0161] In this invention, the construction module uses an instance segmentation network based on an improved U-Net architecture to segment individual plants, and achieves identity binding through a cross-frame feature matching algorithm. The dilation rate of the dilated convolution module in the improved U-Net architecture is set to 2, 4, and 8 respectively, and the weight coefficient of the attention gating mechanism is determined by the product of channel attention and spatial attention. In the cross-frame feature matching algorithm, the intersection-union ratio threshold of the plant contour is set to 0.7, and the principal component orientation consistency threshold is set to 15°.

[0162] It should be noted that this invention provides a high-throughput image data-driven intelligent auxiliary system for crop breeding, which constructs an automated and intelligent closed loop from physical world perception to breeding decision generation. Through collaborative hardware and software algorithm modules, this system transforms fragmented field observation data into deep insights into the breeding value of individual crops, ultimately providing breeders with accurate and interpretable decision support.

[0163] The system's operation begins with a sensing array, which acts as the system's sensory nerves, collecting field microclimate and soil physicochemical parameters at a high frequency and in a distributed manner, providing a real-time digital twin of the environmental background for the entire system. This data is then transmitted in real time to an image acquisition scheduler, which serves as the system's decision-making center and incorporates an environmental disturbance sensitivity threshold model based on crop physiology.

[0164] When the rate of environmental change exceeds a preset threshold, it immediately sends instructions to the imaging unit to dynamically adjust its shooting frequency, imaging mode, and spatial coverage. This responsive acquisition mechanism changes the passive mode of traditional fixed-frequency acquisition, ensuring a high probability of capturing key phenotypic response events triggered by short-term environmental fluctuations, providing a high-value, highly relevant data source for subsequent analysis.

[0165] The multi-view, multispectral image sequences acquired by the imaging unit are fed into the core data processing chain. The construction module is responsible for instantiating and individualizing the original images. The instance segmentation is performed using a specially improved U-Net architecture. In the encoder stage, the network introduces dilated convolutional modules with dilation rates of 2, 4, and 8, which expands the network's receptive field of view in deep feature maps. This enables the network to better understand the global spatial relationships between plants and between plants and the background in complex field backgrounds, thereby accurately segmenting irregularly shaped plants that are closely adjacent or partially occluded.

[0166] The attention gating mechanism embedded in the decoder stage automatically generates a weight map by multiplying channel attention and spatial attention. This guides the network to focus on the target plant region when recovering pixel-level labels, effectively suppressing interference from background noise such as soil, weeds, and shadows, and greatly improving segmentation accuracy. Subsequently, a cross-frame feature matching algorithm is used for identity binding. This algorithm sets an intersection-union ratio (IU) threshold of 0.7 and a principal component orientation consistency threshold of 15°. This dual geometric feature criterion combines the spatial continuity of the plant and the continuity of its morphological growth direction, ensuring that images of the same plant at different time points can be accurately and stably associated, ultimately generating a spatiotemporally continuous image dataset based on individual plants.

[0167] After obtaining high-quality spatiotemporal datasets of individual plants, the modeling module utilizes a gated recurrent unit (GRU) network to construct a dynamic developmental trajectory model for each plant. The two-layer, 256-neuron hidden layer structure of the GRU network provides sufficient complexity to learn highly heterogeneous growth patterns. Its inherent update and reset gate mechanisms enable it to effectively capture long-term dependencies in time series, such as the delayed effects of early stress on later growth. More importantly, the model continuously updates its parameters through online learning, allowing it to dynamically track and reflect the plant's real-time response to environmental disturbances, thereby achieving a digital simulation of the unique life history of each individual.

[0168] Building upon this foundation, the detection module employs the Isolation Forest algorithm to scan the trajectory data. By calculating the average path length of key growth inflection points within a randomly divided set of binary trees, it efficiently identifies anomalous phenotypic events that deviate from the normal developmental pattern of the population. These events often represent potential desirable traits, such as specific stress resistance or high yield potential.

[0169] The analysis module employs the Granger causality test to conduct a rigorous causal correlation analysis between detected singular phenotypic events and corresponding environmental disturbance records. By constructing a vector autoregression model and using the Akaike information criterion to determine the optimal lag order, the system can quantitatively determine which changes in environmental parameters statistically significantly lead the occurrence of phenotypic mutations and determine their time delay. The generated environment-phenotypic response mapping map systematically stores the influence intensity of different environmental factors on different phenotypic characteristics at different time scales in the form of a three-dimensional matrix.

[0170] Finally, the scoring module integrates the results of all the aforementioned modules to construct a multi-dimensional comprehensive scoring model for breeding value. This model integrates the stability score calculated based on trajectory fluctuation, the fitness score calculated based on stress recovery ability, and the genetic potential score based on dynamic time warping algorithm and trajectory matching with superior parents. The final evaluation is formed by weighted summation.

[0171] The decision-making module then outputs a list of top-ranked plants and a detailed decision report. The report not only includes scores but also integrates key phenotypic events, environmental triggering conditions, and recommended hybridization combinations based on genetic complementarity analysis, providing breeders with comprehensive and in-depth scientific evidence for their final decision.

[0172] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A high-throughput image data-driven intelligent auxiliary decision-making method for crop breeding, characterized in that, Includes the following steps: S1: Deploy a sensing array to acquire real-time information on micro-meteorological parameters and soil physicochemical state within the target breeding area. Based on the real-time data stream from the sensing array, dynamically adjust the imaging frequency, imaging mode, and spatial coverage of the imaging unit through the image acquisition scheduler. S2: Collect multi-view image sequences of each plant in the target crop population, including visible light, near infrared and thermal infrared images, perform individual plant segmentation on the image sequences, bind them with identity, and generate a dataset of spatiotemporal continuous images of individual plants. S3: For each plant's dataset, construct an individualized dynamic development trajectory model, update the model parameters through online learning to reflect the plant's response to environmental disturbances, and perform singular phenotypic detection on the trajectory data output by the trajectory model to identify abnormal development events. S4: Perform causal correlation analysis on the detected unusual phenotypic events and the corresponding environmental disturbance records to generate an environment-phenotypic response mapping map; S5: Based on the mapping map and trajectory model, calculate the comprehensive breeding value score of each plant, output a list of plants with the highest comprehensive breeding value scores, and generate a decision report that includes key phenotypic events, environmental triggering conditions, and recommended hybridization combinations.

2. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S1, the sensing array includes a distributed sensor node network; The sensor node integrates a temperature and humidity sensor, a rain gauge, an anemometer, a quantum optical sensor, a soil moisture probe, and a multi-channel ion-selective electrode. The sensor nodes communicate with the central data aggregation unit via a wide area network protocol, and the data sampling frequency is once per minute.

3. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S1, the image acquisition scheduler has a built-in environmental disturbance sensitivity threshold model. When the rate of change of micro-meteorological parameters or soil physicochemical state exceeds the preset threshold, the image acquisition frequency is automatically increased from once a day to once every two hours, and the high-resolution close-range imaging mode is activated simultaneously, while the spatial coverage is limited to the local area where the environmental disturbance occurs.

4. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S2, data acquisition is achieved through an imaging unit, which is set up in the field for operation or in conjunction with aerial data acquisition. The imaging unit includes a track-mounted mobile platform and a multispectral camera array mounted on it. The multispectral camera array includes a visible light camera, a near-infrared camera, and a thermal imager; the track-mounted mobile platform is laid along the field furrows, with a moving speed of 0.1-1.5 m / s; the multispectral camera array is mounted via a gimbal, with a pitch angle adjustment range of -30° to +60° and a horizontal rotation range of ±180°.

5. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S2, individual plant segmentation is performed using an instance segmentation network based on an improved U-Net architecture, or using a multi-scale attention mechanism. The network introduces dilated convolutional modules in the encoder stage to expand the receptive field of view, and embeds an attention gating mechanism in the decoder stage to suppress background noise. The identity binding operation is implemented through a cross-frame feature matching algorithm, which calculates the cross-union ratio and principal component orientation consistency of plant outlines at adjacent time points. The cross-union ratio threshold is 0.7, and the principal component orientation consistency threshold is 15°.

6. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S3, the trajectory model uses a gated recurrent unit network, and the network structure contains two hidden layers, with 256 neurons in each layer. The input to the trajectory model is a vector of plant morphological features arranged in chronological order, including plant height, leaf area index, ear projection area, number of tillers, and canopy temperature gradient. Among them, plant height was obtained by fitting lidar point cloud, leaf area index was calculated by near-infrared and red light band reflectance, ear projection area was determined by the area of ​​high temperature region after thermal imaging segmentation, tiller number was obtained by counting the local curvature minimum points of the plant base contour, and canopy temperature gradient was defined as the difference between the thermal imaging temperature of the top and bottom of the plant.

7. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S3, the singular phenotype detection uses the isolated forest algorithm, taking key turning points in the trajectory data as input samples. The key turning point is the time point when the absolute value of the first difference of the morphological feature vector exceeds twice the historical standard deviation of the feature. The isolated forest algorithm constructs 100 binary trees, and the anomaly judgment threshold is set to an average path length of less than 2.

5.

8. The intelligent auxiliary decision-making method for crop breeding driven by high-throughput image data according to claim 1, characterized in that: In S4, the causal association analysis uses the Granger causality test, which uses a vector autoregression model to determine whether changes in environmental parameters significantly precede the occurrence of phenotypic mutations. The lag order of the vector autoregression model is determined by the Akaike information criterion, with a significance level of 5%. The environment-phenotype response mapping map is stored in the form of a three-dimensional matrix, with row indexes corresponding to environment parameter types, column indexes corresponding to phenotypic feature types, and depth indexes corresponding to time delay steps.

9. A high-throughput image data-driven intelligent auxiliary system for crop breeding, based on the decision-making method according to any one of claims 1-8, characterized in that, include: A sensing array is used to acquire real-time information on micrometeorological parameters and soil physicochemical state within the target breeding area; An image acquisition scheduler is used to dynamically adjust the operating parameters of the imaging unit based on the output of the sensing array; The imaging unit is used to acquire multi-view visible light, near-infrared and thermal infrared image sequences of each plant in the target crop population. The module is used to perform individual plant segmentation and identity binding operations on the image sequence to generate a dataset of spatiotemporally continuous images based on individual plants. The modeling module is used to build and update individualized dynamic developmental trajectory models for datasets of spatiotemporally continuous images of each plant. The detection module is used to perform singular phenotype detection on the trajectory data output by the individualized dynamic developmental trajectory model and identify abnormal developmental events. The analysis module is used to perform causal correlation analysis between singular phenotypic events and environmental disturbance records, and generate an environment-phenotypic response mapping map. The scoring module is used to calculate the comprehensive breeding value score of each plant based on the environment-phenotype response mapping map and the individualized dynamic development trajectory model. The decision-making module is used to output a list of plants with the highest comprehensive breeding value scores and a decision report.

10. The high-throughput image data-driven intelligent auxiliary system and decision-making method for crop breeding according to claim 1, characterized in that: The construction module uses an instance segmentation network based on the improved U-Net architecture to segment individual plants and achieves identity binding through a cross-frame feature matching algorithm. The dilation rate of the dilated convolution module in the improved U-Net architecture is set to 2, 4, and 8 respectively, and the weight coefficient of the attention gating mechanism is determined by the product of channel attention and spatial attention. In the cross-frame feature matching algorithm, the intersection-union ratio threshold of the plant contour is set to 0.7, and the principal component orientation consistency threshold is set to 15°.

Images