Alfalfa cold resistance evaluation system based on deep learning

Abstract

The invention discloses a deep learning-based cold resistance evaluation system for medicago sativa L., and the system comprises a data generation module which is used for generating a phenotypic image and corresponding physiological data of medicago sativa L. under low-temperature stress through a diffusion model embedded with plant low-temperature response physical constraints; the evaluation model module is used for extracting cold resistance characteristics from the image and physiological data by adopting a causal-driven dual-channel adaptive network; and the decision module comprises a hierarchical model distillation unit and a federal reinforcement learning unit, and the hierarchical model distillation unit and the federal reinforcement learning unit realize joint training of model compression and decision strategy optimization through an edge-cloud collaborative architecture, output a cold-resistant decision and realize visualization through an augmented reality interface. The method can effectively solve the core problems of traditional medicago sativa cold resistance assessment in the aspects of data generation, model generalization, decision-making efficiency and the like.

Description

Technical Field

[0001] The present invention relates to the field of evaluation systems, and particularly to a forage alfalfa cold tolerance trait evaluation system based on deep learning. Background Art

[0002] Forage alfalfa is an important forage variety in the genus Medicago. Its cold tolerance refers to the ability of the plant to maintain normal physiological activities, resist frost damage, and survive the winter smoothly under low temperature stress (such as winter low temperature and cold snap in spring). This trait directly affects its planting adaptability, yield stability, and sustainable utilization value in temperate and cold regions; The evaluation of forage alfalfa cold tolerance quantifies its survival ability, physiological stability, and environmental adaptability under low temperature stress through multi-dimensional data and technical means, providing a scientific basis for variety breeding, field management, and cold resistance decision-making. When evaluating the cold tolerance of forage alfalfa, a forage alfalfa cold tolerance trait evaluation system will be used.

[0003] Existing forage alfalfa cold tolerance trait evaluation systems have problems such as high cost and long cycle of real low temperature stress experiments, irreversible damage to plants that may be caused by extreme low temperature conditions, limited data diversity, inability to reflect cell-level damage relying solely on images, and being easily misled by environmental noise relying solely on physiological index data. Therefore, a forage alfalfa cold tolerance trait evaluation system based on deep learning is proposed. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: how to solve the problems existing in the existing forage alfalfa cold tolerance trait evaluation system, such as high cost and long cycle of real low temperature stress experiments, irreversible damage to plants that may be caused by extreme low temperature conditions, limited data diversity, inability to reflect cell-level damage relying solely on images, and being easily misled by environmental noise relying solely on physiological data, and provide a forage alfalfa cold tolerance trait evaluation system based on deep learning.

[0005] The present invention solves the above technical problems through the following technical solutions. The present invention includes: A data generation module for generating phenotypic images of forage alfalfa under low temperature stress and corresponding physiological data through a diffusion model embedded with physical constraints of plant low temperature response; An evaluation model module for extracting cold tolerance features from images and physiological data by using a causal-driven dual-channel adaptive network; A decision-making module, including a hierarchical model distillation unit and a federated reinforcement learning unit. The hierarchical model distillation unit and the federated reinforcement learning unit realize joint training of model compression and decision-making strategy optimization through an edge-cloud collaborative architecture, output cold resistance decisions, and visualize them through an augmented reality interface.

[0006] Furthermore, the model generation process of the physical constraint diffusion is as follows: Among them, is the noisy image at the t-th step, is the noise scheduling coefficient, is the random noise that follows the standard normal distribution; The objective function of the reverse denoising process includes a physical constraint term: Among them, is the input temperature gradient, S is the generated blade morphology, k is the cell membrane permeability constant measured through experiments, is the chlorophyll-temperature calibration function, and are the weight coefficients, is the generated chlorophyll content SPAD value, is the partial derivative of the blade morphology with respect to time.

[0007] Furthermore, the chlorophyll-temperature calibration function is constructed through the following steps: Measure the SPAD values at different temperatures T in the laboratory and fit a quadratic polynomial using the least squares method; Embed the fitting function into the diffusion model to constrain the generated data of physiological indicators to conform to biological laws; , and are the coefficients of the quadratic polynomial fitted by the least squares method.

[0008] Furthermore, the evaluation model module uses a causal-driven dual-channel, including: Image processing channel: Extract local features of the blade based on Vision Transformer and calculate the self-attention weights: ; In the formula, , and are the query (Query), key (Key), and value (Value) matrices, is the dimension of the feature vector; Physiological data channel: The dynamic weight allocation network adjusts the weights according to the sensor confidence and optimizes the learnable parameters through the backpropagation algorithm; The output features of the image processing channel and the physiological data channel are concatenated into a feature matrix F after dimension alignment.

[0009] Furthermore, the causal-driven dual-channel adaptive network fuses the feature through the following process: Construct a causal graph G from historical data using the ICECI causal discovery algorithm to identify the causal relationships among environmental temperature, genetic markers, and phenotypic characteristics; The cross-channel causal attention output is: ; In the formula, is the input matrix after splicing image features and physiological features; , and are the projection matrices of query, key, and value; is the feature dimension; is the binary mask matrix (0 / 1 values) generated based on the causal graph G.

[0010] Furthermore, the hierarchical model distillation process of the hierarchical model distillation unit of the decision module includes: The loss function of the cloud teacher model is: ; In the formula, is the cross-entropy loss (classification task loss), is the knowledge distillation loss (teacher-student feature similarity loss), γ: the weight coefficient for balancing the two types of losses The edge student model optimizes the feature matching loss by the gradient descent method: ; In the formula, and are the feature outputs of the teacher model and the student model; is the weight coefficient of the task loss (such as classification or regression loss).

[0011] Furthermore, the reward function of the federated reinforcement learning unit is defined as: ; Among them, is the change in cold tolerance score, is the predicted yield value, is the operating cost, α, β, and γ are weight coefficients, and proximal policy optimization (PPO) algorithm is used for policy update.

[0012] Furthermore, the visualization interface is implemented through the following steps: Render the cold tolerance heat map in real time, and the color mapping function is: ; Among them, and is the cold tolerance classification threshold calibrated based on experimental data.

[0013] The present invention has the following advantages compared with the prior art: The alfalfa cold tolerance trait evaluation system based on deep learning generates high-fidelity low-temperature stress data through a physically constrained diffusion model, reduces the dependence on real environment experiments, significantly shortens the data acquisition cycle, reduces labor and equipment costs. The causality-driven dual-channel network suppresses environmental interference and false feature associations, improves the prediction stability of the model in complex field environments, and enhances the decision-making credibility through attention visualization. The hierarchical model distillation and lightweight deployment support low-latency response of field devices. Combining the dynamic policy update of federated reinforcement learning, it realizes real-time adjustment and multi-objective analysis of cold resistance measures. The federated learning architecture ensures local processing of farmer data, and gradient sparsification and differential privacy technologies prevent sensitive information leakage, promoting cross-ranch collaborative model training. The interface intuitively displays the cold tolerance heat map and recommended measures, reduces the technical usage threshold, assists farmers in quickly locating problem areas and performing precise interventions. The data generation with physical equation constraints and the chlorophyll-temperature calibration function ensure cross-modal consistency of image, physiological, and environmental data, conforming to the laws of plant physiology, making this system more worthy of popularization and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is the system block diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The following details the embodiments of the present invention. These embodiments are implemented on the premise of the technical solution of the present invention, and provide detailed implementation manners and specific operation processes. However, the protection scope of the present invention is not limited to the following embodiments.

[0016] As Figure 1 shown, this embodiment provides a technical solution: An alfalfa cold tolerance trait evaluation system based on deep learning, including: A data generation module for generating alfalfa phenotypic images and corresponding physiological data under low-temperature stress through a diffusion model embedded with physical constraints of plant low-temperature response; An evaluation model module for extracting cold tolerance features from image and physiological data by using a causality-driven dual-channel adaptive network; A decision-making module, including a hierarchical model distillation unit and a federated reinforcement learning unit. The hierarchical model distillation unit and the federated reinforcement learning unit achieve joint training of model compression and decision-making strategy optimization through an edge-cloud collaborative architecture, output cold resistance decisions and realize visualization through an augmented reality interface.

[0017] The generation process of the physically constrained diffusion model is as follows: Among them, is the noisy image at the t-th step, is the noise schedule coefficient, is the random noise that follows the standard normal distribution; The objective function of the reverse denoising process contains a physical constraint term: Among them, is the input temperature gradient, S is the generated leaf morphology, k is the cell membrane permeability constant measured through experiments in the laboratory, is the chlorophyll-temperature calibration function, and are the weight coefficients, is the generated SPAD value of chlorophyll content, is the partial derivative of the leaf morphology with respect to time; Generate images and physiological data that are highly consistent with real low-temperature experiments through a physical constraint diffusion model, and solve the defect that traditional data augmentation methods (such as rotation and cropping) cannot simulate the morphological changes of plants under low-temperature stress.

[0018] Synchronously generate images and physiological indicators (SPAD value, electrolyte permeability), ensure that the data conforms to biological laws, and avoid manual annotation errors.

[0019] Eliminate the need for low-temperature experiments in the real environment, and reduce equipment loss and experimental cycle.

[0020] For example, to evaluate the cold tolerance of a certain alfalfa germplasm material at -10°C, but lack real low-temperature experimental conditions.

[0021] At this time, input the multi-spectral alfalfa image at room temperature (25°C), and generate the output with the target temperature gradient (from 25°C to -10°C) to simulate the leaf morphology image at -10°C (leaf wilting, waterlogging, browning, etc.); Corresponding physiological data: the SPAD value drops from 55 to 32, and the electrolyte permeability rises from 15% to 68%; Directly use the generated data to train the cold tolerance evaluation model, without the need for real low-temperature experimental equipment. By adjusting the temperature gradient, batch generate data with different stress intensities to accelerate the screening of varieties / germplasm materials.

[0022] The chlorophyll-temperature calibration function is constructed through the following steps: Measure the SPAD value at different temperatures T in the laboratory, and use the least squares method to fit a quadratic polynomial; Embed the fitting function into the diffusion model to constrain the generated data of physiological indicators to conform to biological laws; , and are the coefficients of the quadratic polynomial fitted by the least squares method; The dynamic weight allocation network automatically adjusts the weights of physiological data according to sensor confidence, suppresses the influence of noise data on the model, and improves the prediction stability in complex field environments. The cross-channel causal attention layer is based on a predefined causal graph and only fuses features causally related to cold tolerance (such as leaf electrolyte permeability, cell membrane stability, concentration of osmoregulatory metabolites, etc.), avoiding the introduction of spurious associations (such as soil color, light angle). The decision-making basis of the model is visualized through an attention heatmap (such as focusing on the crown or stubble of alfalfa rather than the background area), assisting researchers in verifying biological rationality.

[0023] For example, in a certain farmland, the confidence of the temperature and humidity sensor decreases due to rain interference, and it is necessary to evaluate the cold tolerance of alfalfa.

[0024] Input image: Visible light image of alfalfa leaves taken by a drone; Physiological data: Chlorophyll sensor data affected by noise (confidence c = 0.3) Dynamic weight allocation: Weight w of physiological data phy =σ(0.3×Wadapt)≈0.1 (low weight) Weight w of image data img =1−w phy ≈0.9 (high weight) Causal attention fusion: The model gives priority to relying on image features (such as leaf wilting, waterlogging, browning, etc.) for prediction, ignores physiological data with low confidence, and outputs a cold tolerance score of 62 (medium cold tolerance), recommending local covering and heat preservation measures.

[0025] Furthermore, the evaluation model module is used to adopt a causal-driven dual-channel including: Image processing channel: Extract local features of the leaf based on Vision Transformer and calculate the self-attention weight: ; In the formula, , and are the query, key, and value matrices, is the dimension of the feature vector; Physiological data channel: The dynamic weight allocation network adjusts the weight according to the sensor confidence and optimizes the learnable parameters through the backpropagation algorithm; The output features of the image processing channel and the physiological data channel are concatenated into a feature matrix F after dimension alignment. Using Vision Transformer to extract local leaf features (such as phenotypic information like leaf color change, curling degree, browning area, etc.), which directly reflects the visible damage under low-temperature stress and conforms to the intuitive judgment logic of human beings on the plant state.

[0026] Internal relevance of physiological data: Synchronously analyzing physiological indicators such as cell membrane permeability, chlorophyll content (SPAD value), and osmotic regulators (soluble sugar content) to reveal the deep impact of low temperature on plant cell structure and metabolism, and making up for the deficiency that images only reflect apparent phenomena.

[0027] Fusion advantage: By dimension alignment and feature splicing, integrating phenotypic and physiological cross-modal information, avoiding the one-sidedness of a single data type, and more comprehensively describing the internal mechanism of cold tolerance traits.

[0028] Fine-grained extraction of local features: Based on the self-attention mechanism of Transformer, it can adaptively focus on key leaf regions (such as low-temperature sensitive parts like leaf tips and edges), suppress the interference of background noise (such as soil and weeds), and improve the pertinence of feature extraction.

[0029] By calculating the global self-attention weights, capturing the co-variation between different regions of the leaf (such as the correlation between vein distribution and water transport), which is suitable for analyzing complex phenotypic patterns.

[0030] Generalization advantage: The Transformer architecture has stronger robustness to image scaling, rotation and other transformations, and can adapt to image data with different shooting angles or resolutions in the field.

[0031] Sensor confidence-driven weight adjustment: Dynamically allocate weights according to the real-time confidence of sensors (such as the signal stability and data rationality of temperature and humidity sensors), automatically reduce the weight of low-confidence data (such as abnormal SPAD values caused by moisture in rainy days), and avoid noise misleading the model decision.

[0032] End-to-end optimization: Jointly optimize the weight parameters and the model prediction target through the backpropagation algorithm, so that the physiological data channel can adapt to the changes in the field environment and improve the data reliability in complex scenarios.

[0033] Reducing the dependence on manual preprocessing: There is no need to manually set fixed thresholds to filter data. The model automatically realizes soft filtering of noise, reduces the data preprocessing cost, and improves the system deployment efficiency.

[0034] Processing images and physiological data independently in two channels, avoiding the computational bottleneck of traditional serial architectures, and enabling low-latency feature extraction on edge devices (such as field cameras and sensor nodes).

[0035] Feature Dimension Alignment Strategy: Unify the output dimensions of the dual channels through lightweight operations such as linear projection or interpolation, reduce the computational overhead during splicing, and support fast multi-modal data fusion at the edge with limited computing power.

[0036] The causal-driven dual-channel adaptive network processes the fused features through the following process: Use the ICECI causal discovery algorithm to construct a causal graph G from historical data and identify the causal relationships among environmental temperature, gene markers, and phenotypic characteristics; The cross-channel causal attention output is: ; In the formula, is the input matrix after splicing the image features and physiological features; , and are the projection matrices for query, key, and value; is the feature dimension; is the binary mask matrix (0 / 1 values) generated based on the causal graph G; Construct a causal graph through the ICECI causal discovery algorithm to clarify the causal dependencies among environmental temperature, gene markers, and phenotypic characteristics (such as leaf morphology and physiological indicators), avoid the model learning spurious associations (such as environmental noise or irrelevant features), and enhance the biological rationality of feature fusion.

[0037] Use the binary mask matrix (M) generated by the causal graph to filter out non-causal associated features, and only retain the cross-channel information directly related to the cold tolerance of Medicago varia (such as the leaf structure features in the image and the cell membrane permeability in the physiological data), reduce the interference of redundant information on the model, and improve the purity of feature expression.

[0038] The causal attention mechanism visually displays the key features relied on by the model during decision-making (such as the association between leaf wilting, waterlogging degree, and chlorophyll content under low-temperature stress) through visualizing the attention weights, facilitating researchers to verify the biological correctness of the model logic and meeting the requirements for interpretability in the agricultural field.

[0039] The hierarchical model distillation process of the hierarchical model distillation unit of the decision module includes: The loss function of the cloud teacher model is: ; In the formula, is the cross-entropy loss (classification task loss), is the knowledge distillation loss (teacher-student feature similarity loss), γ: the weight coefficient for balancing the two types of losses The marginal student model optimizes the feature matching loss through gradient descent: ; In the formula, and are the feature outputs of the teacher model and the student model; is the weight coefficient of the task loss (such as classification or regression loss).

[0040] The complex knowledge of the cloud teacher model is compressed into the marginal student model through knowledge distillation (KD loss), significantly reducing the number of model parameters and computational complexity (such as reducing the number of Transformer layers and lowering the feature dimension), enabling it to run on edge devices with limited computing power (such as field sensors and drone terminals). Example: The cloud teacher model contains 12 layers of Transformer, and only 4 layers are retained after distillation of the marginal student model, with the inference speed increased by 5 times, adapting to embedded GPUs (such as NVIDIA Jetson Nano).

[0041] Cross-device compatibility: The student model can be deployed in various edge forms (such as handheld detection devices, field gateways, and agricultural machinery terminals), without relying on high-bandwidth network connections, and supporting real-time evaluation in offline or weak network environments.

[0042] The teacher model learns the global cold tolerance pattern using large-scale historical data (cross-farm cold tolerance data) in the cloud. The student model directly inherits the generalization ability of the teacher model through the feature matching loss (ℒstudent), avoiding the inefficiency of zero training on edge devices. Scenario: The teacher model is trained based on more than 20,000 alfalfa samples in the cloud, and the marginal student model only needs to be fine-tuned with a small amount of local data to adapt to the soil and climate conditions of a specific farm.

[0043] Joint optimization reduces overfitting: The student model is optimized at the edge by combining local task losses (such as the cold tolerance classification loss of Ranch A), retaining both the universality of cloud knowledge and adapting to regional data features (such as the leaf morphology differences of specific varieties), improving the robustness of the model.

[0044] Real-time response ability: Edge devices directly process real-time field data (such as drone aerial images and real-time physiological indicators of sensors) without uploading to the cloud, reducing the latency from "seconds" to "milliseconds", meeting the immediate needs of cold protection measures (such as temporary nighttime warming). Comparison: Traditional cloud inference requires waiting for data upload (about 500ms) + model calculation (800ms), with a total latency exceeding 1 second; edge inference only requires model calculation (150ms), which is suitable for dynamic monitoring (such as minute-level evaluation during cold wave warnings).

[0045] Communication cost savings: Reducing data upload traffic (only transmitting feature differences instead of raw data) can lower the costs of using cellular networks or satellite communications on farms, especially in remote areas.

[0046] Convenience of model iteration: The cloud teacher model can be updated regularly (such as incorporating data on new cold-resistant varieties), and through distillation, it can be quickly synchronized to all edge student models without manual upgrade for each device, supporting unified deployment in large-scale planting production areas. Process: Research institutions release new cold resistance assessment criteria → The cloud teacher model is updated → Edge devices automatically download the distilled student model at night → Directly apply the new assessment logic the next day.

[0047] Federated learning compatibility: Combining federated reinforcement learning, edge devices can participate in model training locally (only uploading gradient update data instead of raw data), protecting farmers' privacy while aggregating data from multiple parties and enhancing the generalization ability of the global model.

[0048] The reward function of the said federated reinforcement learning unit is defined as: ; where is the change in cold resistance score, is the predicted yield value, is the operating cost, and α, β, and γ are weight coefficients. The policy update uses the Proximal Policy Optimization (PPO) algorithm; The reward function combines the change in cold resistance score (ΔH), the predicted yield value (Y), and the operating cost (C) to achieve the global optimum of "cold resistance effect - production revenue - input cost".

[0049] Example: When a cold snap comes, the system automatically compares two scenarios: "covering with insulation" (high cost but significant increase in ΔH) and "spraying antifreeze agent" (low cost but limited increase in ΔH), and selects the strategy with the highest net profit according to the preset weights (such as α = 0.6, β = 0.3, γ = 0.1) (such as giving priority to covering plots of high-value varieties).

[0050] Flexibility of dynamic weight adjustment: Farmers can manually adjust α / β / γ according to factors such as seasons and market prices (such as increasing the β weight during the production increase season), or automatically optimize the weights through reinforcement learning to adapt to diverse needs.

[0051] Localized data processing: Each farm's edge device only uploads policy gradient update data (instead of raw planting data), ensuring that sensitive information such as farmers' soil characteristics and variety genes does not leave the local area, meeting the requirements of agricultural data privacy protection. Scenario: When multiple farms jointly train a cold resistance model, federated learning allows the model to absorb low-temperature response patterns in different regions (such as extreme low-temperature data from northern farms and late spring cold data from southern farms), while avoiding the risk of "revealing variety characteristics due to data sharing".

[0052] Cross - farm collaboration evolution: Through the Federated Averaging (FedAvg) mechanism, the cloud collects the policy optimization experiences of each edge node, generates a global cold - resistance policy, and then distributes it to each farm for iterative update, forming a virtuous cycle of "data does not leave the household, and models co - evolve".

[0053] Online learning and feedback loop: Utilizing the efficient update ability of the Proximal Policy Optimization (PPO) algorithm, the system can dynamically adjust cold - resistance measures according to real - time monitoring data (such as sudden drops in night temperature and changes in soil moisture), achieving a minute - level response of "monitoring - evaluation - decision - execution". Comparison: Traditional solutions rely on preset thresholds (such as starting heating when the temperature < 0°C), and cannot adapt to the differences in cold - tolerance thresholds of different varieties; Federated reinforcement learning can learn the personalized critical temperatures of each variety through historical data (such as variety A starts to suffer from frost damage at - 2°C, and variety B shows damage only at - 4°C), enabling precise intervention.

[0054] Long - term policy optimization: Through the cumulative reward mechanism, the model can learn the cold - resistance laws across seasons (such as the correlation between the low - temperature duration in the previous winter and the cold - tolerance during the greening period of the following year), and optimize long - term planting strategies (such as applying potassium fertilizer in advance in autumn to enhance winter cold - tolerance).

[0055] Automated decision - making replaces empiricism: Avoiding misjudgments caused by farmers' lack of professional knowledge (such as waste of costs caused by over - reliance on a single cold - resistance measure), the algorithm quantitatively analyzes and recommends the optimal solution through algorithm, which is especially suitable for the large - scale management of new - type business entities (such as family farms, cooperatives). Case: A certain cooperative manages 1000 mu of alfalfa. The system automatically generates differentiated cold - resistance solutions according to the soil temperature and variety distribution of each plot, saving 30% of the operation cost compared with manual decision - making, and at the same time increasing the cold - tolerance compliance rate by 25%.

[0056] Interpretable policy output: The decisions generated by reinforcement learning can be associated with specific reward items (such as "This recommendation is to cover the plastic film, because ΔH increases by 4 points, the expected yield increase is 5%, and the cost increases by 200 yuan / mu"), which is convenient for farmers to understand and trust the algorithm's suggestions.

[0057] The said visualization interface is implemented through the following steps: Render the cold - tolerance heat map in real - time, and the color mapping function is: ; wherein, and are the cold - tolerance grading thresholds calibrated based on experimental data.

[0058] The cold tolerance distribution of Medicago polymorpha in the field is dynamically presented through a color-coded heat map. Farmers can quickly identify high-risk areas (such as plots with red warnings), accurately associate the plot locations with geographical information, assist farmers in implementing differentiated cold resistance measures for different risk-level areas (red / yellow / green), avoid blind operations across the board, and clarify the risk levels based on a three-level color classification (red / yellow / green) with thresholds, guiding the allocation of resources according to priorities, improving decision-making efficiency. The visual interface lowers the threshold for using technology, and non-professionals can understand the evaluation results just by colors, reducing the cost of interpreting complex data.

[0059] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0060] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0061] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A cold tolerance trait evaluation system for alfalfa based on deep learning, characterized in that, Including: A data generation module for generating phenotypic images and corresponding physiological data of Medicago varia under low temperature stress by embedding a diffusion model with physical constraints of plant low temperature response; An evaluation model module for extracting cold tolerance features from images and physiological data using a causally-driven dual-channel adaptive network; A decision-making module, including a hierarchical model distillation unit and a federated reinforcement learning unit. The hierarchical model distillation unit and the federated reinforcement learning unit perform joint training for model compression and decision-making strategy optimization through an edge-cloud collaborative architecture, output cold resistance decisions, and visualize them through an augmented reality interface.

2. The alfalfa cold tolerance trait evaluation system based on deep learning according to claim 1, characterized in that: The model generation process of the physical constraint diffusion is as follows: Among them, is the noisy image at the t-th step, is the noise schedule coefficient, is the random noise following the standard normal distribution; The objective function of the reverse denoising process includes a physical constraint term: Among them, is the input temperature gradient, S is the generated leaf morphology, and k is the cell membrane permeability constant measured in the laboratory. is the chlorophyll-temperature calibration function. and is the weight coefficient. is the generated SPAD value of chlorophyll content. is the partial derivative of leaf morphology with respect to time.

3. The alfalfa cold tolerance trait evaluation system based on deep learning according to claim 2, wherein: The chlorophyll-temperature calibration function is constructed through the following steps: Measure the SPAD value at different temperatures T in the laboratory and fit a quadratic polynomial using the least squares method; Embed the fitting function into the diffusion model to ensure that the physiological indicators of the generated data conform to biological laws; , and are the coefficients of the quadratic polynomial fitted by the least squares method.

4. The alfalfa cold tolerance trait evaluation system based on deep learning according to claim 1, wherein: The evaluation model module is used to adopt a causally-driven dual-channel including: An image processing channel: Extract local leaf features based on Vision Transformer and calculate self-attention weights: ； In the formula, , and are the query, key, and value matrices, is the dimension of the feature vector; A physiological data channel: The dynamic weight allocation network adjusts weights according to sensor confidence and optimizes learnable parameters through the backpropagation algorithm; The output features of the image processing channel and the physiological data channel are concatenated into a feature matrix F after dimension alignment.

5. The evaluation system for the cold tolerance trait of alfalfa based on deep learning according to claim 4, characterized in that: The causally-driven dual-channel adaptive network fuses the features through the following process: Use the ICECI causal discovery algorithm to construct a causal graph G from historical data and identify the causal relationships between environmental temperature, gene markers, and phenotypic features; The cross-channel causal attention output is: ； Wherein, is the input matrix after splicing the image feature and the physiological feature; , and is the projection matrix for query, key, and value. is the feature dimension; is a binary mask matrix generated based on the causal graph G.

6. The evaluation system for the cold tolerance trait of alfalfa based on deep learning according to claim 1, characterized in that: The hierarchical model distillation process of the hierarchical model distillation unit of the decision-making module includes: The loss function of the cloud teacher model is: ； Wherein, is the cross-entropy loss, is the knowledge distillation loss, and γ is the weight coefficient for balancing the two types of losses The edge student model optimizes the feature matching loss through the gradient descent method: ； In the formula, and are the feature outputs of the teacher model and the student model; The weight coefficient for task loss.

7. The evaluation system for the cold tolerance trait of alfalfa based on deep learning according to claim 1, wherein: The reward function of the federated reinforcement learning unit is defined as: ； Among them, is the change amount of cold tolerance score, is the yield prediction value, is the operating cost, α, β and γ are weight coefficients, and the policy update adopts the proximal policy optimization algorithm.

8. The alfalfa cold tolerance trait evaluation system based on deep learning according to claim 1, characterized in that: The visualization interface is implemented through the following steps: Render a real-time cold tolerance heat map, and the color mapping function is: ； Among them, and are the cold tolerance classification thresholds calibrated based on experimental data.

Images