A deep learning-based cold tolerance evaluation system for alfalfa

Abstract

This invention discloses a deep learning-based cold tolerance assessment system for alfalfa. The system comprises a data generation module, which generates phenotypic images and corresponding physiological data of alfalfa under cold stress using a diffusion model embedded with physical constraints on the plant's low-temperature response; an evaluation model module, which extracts cold tolerance characteristics from the images and physiological data using a causal-driven, dual-channel adaptive network; and a decision module, comprising a hierarchical model distillation unit and a federated reinforcement learning unit. These units implement joint training for model compression and decision strategy optimization through an edge-cloud collaborative architecture, outputting cold tolerance decisions and visualizing them through an augmented reality interface. This invention effectively addresses the core issues of traditional alfalfa cold tolerance assessment, including data generation, model generalization, and decision efficiency.

Description

Technical Field

[0001] The present invention relates to the field of evaluation systems, and in particular to a cold-resistance evaluation system for alfalfa based on deep learning. Background Art

[0002] Alfalfa is an important forage species in the genus Alfalfa. Its cold tolerance refers to the plant's ability to maintain normal physiological activities, resist frost damage, and successfully survive the winter under low-temperature stress (such as winter low temperatures and late spring cold snaps). This trait directly affects its adaptability, yield stability, and sustainable utilization value in temperate and cold regions.

[0003] The cold resistance assessment of hybrid alfalfa is to quantify 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 selection, field management and cold resistance decision-making. When conducting the cold resistance assessment of hybrid alfalfa, the hybrid alfalfa cold resistance trait assessment system will be used.

[0004] Existing cold tolerance evaluation systems for mixed alfalfa have problems such as high cost and long cycle of real low temperature stress experiments, extreme low temperature conditions may cause irreversible damage to plants, limit data diversity, rely on images that cannot reflect damage at the cellular level, and rely only on physiological indicator data that are easily misled by environmental noise. Therefore, a cold tolerance evaluation system for mixed alfalfa based on deep learning is proposed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to solve the problems of the existing alfalfa cold resistance trait evaluation system, which has the following problems: the real low temperature stress experiment is costly and has a long cycle, and extreme low temperature conditions may cause irreversible damage to plants, limit data diversity, rely on images that cannot reflect cellular level damage, and rely only on physiological data that is easily misled by environmental noise. A deep learning-based alfalfa cold resistance trait evaluation system is provided.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions, which include:

[0007] A data generation module is used to generate phenotypic images and corresponding physiological data of alfalfa under low temperature stress by embedding a diffusion model with physical constraints on plant low temperature response;

[0008] An evaluation model module that uses a causal-driven two-channel adaptive network to extract cold tolerance features from image and physiological data;

[0009] The decision-making module includes 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 strategy optimization through an edge-cloud collaborative architecture, output cold resistance decisions and visualize them through an augmented reality interface.

[0010] Furthermore, the model generation process of the physical constraint diffusion is as follows:

[0011]

[0012] in, is the t-th step noise image, is the noise scheduling coefficient, is random noise that obeys the standard normal distribution;

[0013] The objective function of the inverse denoising process contains physical constraints:

[0014]

[0015] in, is the input temperature gradient, S is the generated leaf morphology, 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 chlorophyll content SPAD value, is the partial derivative of the leaf morphology over time.

[0016] Furthermore, the chlorophyll-temperature calibration function Build it by following these steps:

[0017] The SPAD values ​​at different temperatures T were measured in the laboratory, and a quadratic polynomial was fitted using the least squares method;

[0018] The fitting function Embedded diffusion model constrains the generated data of physiological indicators to conform to biological laws;

[0019] 、 and are the coefficients of the quadratic polynomial fitted by the least squares method.

[0020] Furthermore, the evaluation model module is used to adopt a causal-driven dual channel including:

[0021] Image processing channel: Extract local features of leaves based on Vision Transformer and calculate self-attention weights:

[0022] ;

[0023] Where, , and is the query, key, and value matrix, is the dimension of the feature vector;

[0024] Physiological data channel: A dynamic weight allocation network adjusts weights based on sensor confidence and optimizes learnable parameters through a back-propagation algorithm;

[0025] The output features of the image processing channel and the physiological data channel are aligned in dimension and then spliced ​​into a feature matrix F.

[0026] Furthermore, the causal-driven dual-channel adaptive network enters and exits the fusion feature through the following process:

[0027] Use the ICECI causal discovery algorithm to construct a causal graph G from historical data to identify the causal relationship between environmental temperature, genetic markers and phenotypic characteristics;

[0028] The cross-channel causal attention output is:

[0029] ;

[0030] Where, It is the input matrix after the image features and physiological features are concatenated;

[0031] 、 and is the projection matrix of query, key, and value;

[0032] is the feature dimension;

[0033] It is a binary mask matrix (0 / 1 value) generated based on the causal graph G.

[0034] Furthermore, the hierarchical model distillation unit of the decision module performs a hierarchical model distillation process including:

[0035] The loss function of the cloud teacher model is:

[0036] ;

[0037] Where, 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

[0038] The marginal student model optimizes the feature matching loss via gradient descent:

[0039] ;

[0040] Where, and Feature output for the teacher model and the student model;

[0041] is the weight coefficient of the task loss (such as classification or regression loss).

[0042] Furthermore, the reward function of the federated reinforcement learning unit is defined as:

[0043] ;

[0044] in, is the change in cold resistance score, is the output forecast value, is the operation cost, α, β and γ are weight coefficients, and the proximal policy optimization (PPO) algorithm is used for policy update.

[0045] Furthermore, the visual interface is implemented by the following steps:

[0046] Real-time rendering of cold resistance heat map, the color mapping function is:

[0047] ;

[0048] in, and It is the cold tolerance classification threshold calibrated based on experimental data.

[0049] Compared with the existing technology, the present invention has the following advantages: the deep learning-based alfalfa cold resistance trait evaluation system generates high-fidelity low temperature stress data through a physical constrained diffusion model, reduces dependence on real environment experiments, significantly shortens the data acquisition cycle, and reduces manpower and equipment costs. The causal-driven dual-channel network suppresses environmental interference and false feature associations, improves the model's prediction stability in complex field environments, and enhances decision credibility through attention visualization. Hierarchical model distillation and lightweight deployment support low-latency response of field equipment. Combined with the dynamic strategy update of federated reinforcement learning, real-time adjustment of cold resistance measures and multi-objective analysis are achieved. The federated learning architecture ensures local processing of farmer data. Gradient sparsification and differential privacy technology prevent the leakage of sensitive information and promote cross-pasture collaborative model training. The interface intuitively displays cold resistance heat maps and recommended measures, lowers the threshold for technology use, and assists farmers in quickly locating problem areas and performing precise interventions. Data generation constrained by physical equations and chlorophyll-temperature calibration functions ensure cross-modal consistency of image, physiological, and environmental data, conforms to the laws of plant physiology, and makes the system more worthy of promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0051] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0052] like Figure 1 As shown, this embodiment provides a technical solution: a system for evaluating the cold resistance of alfalfa based on deep learning, comprising:

[0053] A data generation module is used to generate phenotypic images and corresponding physiological data of alfalfa under low temperature stress by embedding a diffusion model with physical constraints on plant low temperature response;

[0054] An evaluation model module that uses a causal-driven two-channel adaptive network to extract cold tolerance features from image and physiological data;

[0055] The decision-making module includes 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 strategy optimization through an edge-cloud collaborative architecture, output cold-resistance decisions, and visualize them through an augmented reality interface.

[0056] The model generation process of the physical constraint diffusion is as follows:

[0057]

[0058] in, is the t-th step noise image, is the noise scheduling coefficient, is random noise that obeys the standard normal distribution;

[0059] The objective function of the inverse denoising process contains physical constraints:

[0060]

[0061] in, is the input temperature gradient, S is the generated leaf morphology, 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 chlorophyll content SPAD value, is the partial derivative of leaf morphology over time;

[0062] Through the physical constrained diffusion model, images and physiological data that are highly consistent with real low-temperature experiments are generated, solving the defect that traditional data enhancement methods (such as rotation and cropping) cannot simulate plant morphological changes under low-temperature stress.

[0063] Synchronously generate images and physiological indicators (SPAD value, electrolyte permeability) to ensure that the data conforms to biological laws and avoid manual labeling errors.

[0064] No need for low-temperature experiments in real environments, reducing equipment loss and experimental cycle.

[0065] For example, it is necessary to evaluate the cold resistance of a certain alfalfa germplasm material at -10℃, but there is a lack of real low-temperature experimental conditions.

[0066] In this case, a multispectral alfalfa image at room temperature (25°C) is input, and the target temperature gradient (from 25°C to -10°C) is generated as output to simulate the leaf morphology image at -10°C (leaf wilting, waterlogging, browning, etc.).

[0067] Corresponding physiological data: SPAD value decreased from 55 to 32, and electrolyte permeability increased from 15% to 68%;

[0068] The generated data can be directly used to train the cold tolerance assessment model without the need for real low-temperature experimental equipment. By adjusting the temperature gradient, data with different stress intensities can be generated in batches to accelerate the screening of varieties / germplasm materials.

[0069] The chlorophyll-temperature calibration function Build it by following these steps:

[0070] The SPAD values ​​at different temperatures T were measured in the laboratory, and a quadratic polynomial was fitted using the least squares method;

[0071] The fitting function Embedded diffusion model constrains the generated data of physiological indicators to conform to biological laws;

[0072] 、 and are the coefficients of the quadratic polynomial fitted by the least squares method;

[0073] A dynamic weight allocation network automatically adjusts physiological data weights based on sensor confidence, suppressing the impact of noisy data on the model and improving prediction stability in complex field environments. A cross-channel causal attention layer, based on a predefined causal graph, incorporates only features causally related to cold tolerance (such as leaf electrolyte permeability, cell membrane stability, and osmotic regulatory metabolite concentrations), avoiding the introduction of spurious associations (such as soil color and light angle). Attention heat maps visualize the basis for model decisions (e.g., focusing on alfalfa root suckers or stubble rather than background areas), assisting researchers in verifying biological plausibility.

[0074] For example, if the confidence level of temperature and humidity sensors in a certain farmland decreases due to rain interference, the cold tolerance of alfalfa needs to be evaluated.

[0075] Input image: A visible light image of alfalfa leaves taken by a drone;

[0076] Physiological data: Chlorophyll sensor data corrupted by noise (confidence level c=0.3)

[0077] Dynamic weight allocation:

[0078] Physiological data weight w phy =σ(0.3×Wadapt)≈0.1 (low weight)

[0079] Image data weight w img =1−w phy ≈0.9 (high weight)

[0080] Causal Attention Fusion:

[0081] The model prioritizes image features (leaf wilting, water stains, browning, etc.) for prediction, ignores low-confidence physiological data, outputs a cold-hardiness score of 62 (moderate cold-hardiness), and recommends local covering and insulation measures.

[0082] Furthermore, the evaluation model module is used to adopt a causal-driven dual channel including:

[0083] Image processing channel: Extract local features of leaves based on Vision Transformer and calculate self-attention weights:

[0084] ;

[0085] Where, , and is the query, key, and value matrix, is the dimension of the feature vector;

[0086] Physiological data channel: A dynamic weight allocation network adjusts weights based on sensor confidence and optimizes learnable parameters through a back-propagation algorithm;

[0087] The output features of the image processing channel and the physiological data channel are aligned in dimension and then spliced ​​into a feature matrix F;

[0088] Vision Transformer is used to extract local leaf features (such as leaf color changes, curling degree, browning area and other phenotypic information), which directly reflects the visible damage under low temperature stress and conforms to the human intuitive judgment logic of plant status.

[0089] Intrinsic correlation of physiological data: Synchronous analysis of physiological indicators such as cell membrane permeability, chlorophyll content (SPAD value), and osmotic regulators (soluble sugar content) reveals the profound impact of low temperature on plant cell structure and metabolism, making up for the deficiency of images that only reflect superficial phenomena.

[0090] Fusion advantages: Through dimension alignment and feature splicing, phenotypic and physiological cross-modal information is integrated to avoid the one-sidedness of a single data type and more comprehensively describe the intrinsic mechanism of cold-resistance traits.

[0091] Refined extraction of local features: Based on the Transformer's self-attention mechanism, it can adaptively focus on key areas of leaves (such as leaf tips, edges, and other low-temperature sensitive areas), suppress the interference of background noise (such as soil and weeds), and improve the targeted feature extraction.

[0092] By calculating the global self-attention weight, it captures the coordinated changes between different regions of the leaf (such as the relationship between vein distribution and water transport), which is suitable for analyzing complex phenotypic patterns.

[0093] Generalization advantage: The Transformer architecture is more robust to transformations such as image scaling and rotation, and can adapt to image data of different shooting angles or resolutions in the field.

[0094] Sensor confidence-driven weight adjustment: Dynamically assign weights based on the real-time confidence of sensors (such as signal stability and data rationality of temperature and humidity sensors), and automatically downgrade low-confidence data (such as abnormal SPAD values ​​caused by moisture on rainy days) to prevent noise from misleading model decisions.

[0095] End-to-end optimization: The back-propagation algorithm is used to jointly optimize weight parameters and model prediction targets, enabling physiological data channels to adapt to changes in the field environment and improving data reliability in complex scenarios.

[0096] Reduce reliance on manual preprocessing: There is no need to manually set fixed thresholds to filter data. The model automatically performs soft filtering of noise, reducing data preprocessing costs and improving system deployment efficiency.

[0097] The dual channels independently process images and physiological data, avoiding the computational bottlenecks of traditional serial architectures and enabling low-latency feature extraction on edge devices (such as field cameras and sensor nodes).

[0098] Feature dimension alignment strategy: Unify the dual-channel output dimensions through lightweight operations such as linear projection or interpolation, reduce the computational overhead during splicing, and support rapid multimodal data fusion at the edge with limited computing power.

[0099] The causal-driven two-channel adaptive network integrates features in and out through the following process:

[0100] Use the ICECI causal discovery algorithm to construct a causal graph G from historical data to identify the causal relationship between environmental temperature, genetic markers and phenotypic characteristics;

[0101] The cross-channel causal attention output is:

[0102] ;

[0103] Where, It is the input matrix after the image features and physiological features are concatenated;

[0104] 、 and is the projection matrix of query, key, and value;

[0105] is the feature dimension;

[0106] is a binary mask matrix (0 / 1 value) generated based on the causal graph G;

[0107] A causal graph is constructed using the ICECI causal discovery algorithm to clarify the causal dependency between environmental temperature, genetic markers, and phenotypic characteristics (such as leaf morphology and physiological indicators), preventing the model from learning false associations (such as environmental noise or irrelevant features) and improving the biological rationality of feature fusion.

[0108] The binary mask matrix (M) generated by the causal graph is used to filter out non-causally related features, retaining only the cross-channel information directly related to the cold resistance of alfalfa (such as leaf structure characteristics in images and cell membrane permeability in physiological data), reducing the interference of redundant information on the model and improving the purity of feature expression.

[0109] The causal attention mechanism visualizes attention weights to intuitively display the key features that the model relies on when making decisions (such as the relationship between leaf wilting and waterlogging under low temperature stress and chlorophyll content), making it easier for researchers to verify the biological correctness of the model logic and meeting the agricultural field's demand for explainability.

[0110] The hierarchical model distillation unit of the decision module performs a hierarchical model distillation process including:

[0111] The loss function of the cloud teacher model is:

[0112] ;

[0113] Where, 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

[0114] The marginal student model optimizes the feature matching loss via gradient descent:

[0115] ;

[0116] Where, and Feature output for the teacher model and the student model;

[0117] is the weight coefficient of the task loss (such as classification or regression loss).

[0118] Through knowledge distillation (KD loss), the complex knowledge of the cloud-based teacher model is compressed into the edge student model, significantly reducing the number of model parameters and computational complexity (e.g., reducing the number of Transformer layers and feature dimensions). This allows it to run on edge devices with limited computing power (such as field sensors and drone terminals). For example, the cloud-based teacher model contains 12 Transformer layers, while the edge student model retains only 4 layers after distillation, resulting in a 5x increase in inference speed and compatibility with embedded GPUs (such as the NVIDIA Jetson Nano).

[0119] Cross-device compatibility: The student model can be deployed on a variety of edge forms (such as handheld detection devices, field gateways, and agricultural machinery terminals) without relying on high-bandwidth network connections, supporting real-time evaluation in offline or weak network environments.

[0120] The teacher model in the cloud leverages large-scale historical data (cross-farm cold tolerance data) to learn global cold tolerance patterns. The student model directly inherits the teacher model's generalization capabilities through a feature matching loss (ℒstudent), avoiding the inefficiency of training from scratch on edge devices. Scenario: The teacher model is trained in the cloud using over 20,000 alfalfa samples. The student model at the edge only requires fine-tuning with a small amount of local data to adapt to the soil and climate conditions of a specific farm.

[0121] Joint optimization reduces overfitting: The student model is optimized at the edge by combining local task losses (such as the cold resistance classification loss of Pasture A). This not only retains the universality of cloud-based knowledge, but also adapts to regionalized data characteristics (such as the differences in leaf morphology of specific varieties), thereby improving model robustness.

[0122] Real-time responsiveness: Edge devices directly process real-time field data (such as drone aerial images and real-time sensor physiological indicators) without uploading to the cloud. This reduces latency from seconds to milliseconds, meeting the immediate needs of cold-weather mitigation measures (such as temporary nighttime heating). In contrast, traditional cloud-based inference requires data upload (approximately 500ms) and model calculation (800ms), resulting in a total latency exceeding 1 second. Edge inference only requires model calculation (150ms), making it suitable for dynamic monitoring (such as minute-by-minute assessments during cold wave warnings).

[0123] Communication cost savings: Reduced data upload traffic (transmitting only feature differences rather than raw data) reduces the cost of cellular or satellite communications for farms, especially in remote areas.

[0124] Convenient model iteration: The cloud-based teacher model can be regularly updated (for example, to incorporate data on new cold-resistant varieties) and quickly synchronized to all edge student models through distillation, eliminating the need for manual device-by-device upgrades and supporting unified deployment across large-scale planting and production areas. The process: A research institution releases new cold-hardiness assessment standards → the cloud-based teacher model is updated → edge devices automatically download the distilled student model at night → the new assessment logic is directly applied the next day.

[0125] Federated learning compatibility: Combined with federated reinforcement learning, edge devices can participate in model training locally (uploading only gradient update data rather than original data), protecting farmers' privacy while aggregating data from multiple parties and improving the generalization ability of the global model.

[0126] The reward function of the federated reinforcement learning unit is defined as:

[0127] ;

[0128] in, is the change in cold resistance score, is the output forecast value, is the operation cost, α, β and γ are weight coefficients, and the policy update adopts the proximal policy optimization (PPO) algorithm;

[0129] The reward function integrates the change in cold resistance score (ΔH), yield prediction value (Y), and operating cost (C) to achieve the global optimality of "cold resistance effect-production income-input cost".

[0130] Example: When a cold wave comes, the system automatically compares the two options of "covering and keeping warm" (high cost but significant improvement in ΔH) and "spraying antifreeze" (low cost but limited improvement in ΔH), and selects the strategy with the highest net benefit (such as giving priority to covering high-value variety plots) based on preset weights (such as α=0.6, β=0.3, γ=0.1).

[0131] Flexibility of dynamic weight adjustment: Farmers can manually adjust α / β / γ based on factors such as season and market price (such as increasing the β weight during the production season), or automatically optimize the weights through reinforcement learning to adapt to diverse needs.

[0132] Localized Data Processing: Edge devices on each farm only upload policy gradient update data (not raw planting data), ensuring that sensitive information such as farmers' soil characteristics and variety genetics remain local, complying with agricultural data privacy requirements. Scenario: When multiple farms jointly train a cold-resistance model, federated learning allows the model to incorporate different regional cold response patterns (e.g., extreme cold data from northern farms versus late spring cold weather data from southern farms), while also avoiding the risk of "variety leakage" due to data sharing.

[0133] Cross-farm collaborative evolution: Through the FedAvg mechanism, the cloud collects the policy optimization experience of each edge node, generates a global cold resistance strategy, and then distributes it to each farm for iterative updates, forming a virtuous cycle of "data does not leave the farm, and models evolve together."

[0134] Online learning and feedback loop: Leveraging the efficient update capabilities of the Proximal Policy Optimization (PPO) algorithm, the system dynamically adjusts cold tolerance measures based on real-time monitoring data (such as sudden nighttime temperature drops and soil moisture changes), achieving a minute-by-minute response through a "monitor-assess-decide-execute" process. In contrast, traditional solutions rely on preset thresholds (e.g., heating is activated when the temperature falls below 0°C) and are unable to adapt to the varying cold tolerance thresholds of different varieties. Federated reinforcement learning, however, uses historical data to learn the personalized critical temperatures for each variety (e.g., variety A begins to suffer frost damage at -2°C, while variety B only shows damage at -4°C), enabling precise intervention.

[0135] Long-term strategy optimization: Through a cumulative reward mechanism, the model can learn cross-seasonal cold resistance patterns (such as the relationship between the duration of low temperatures in the previous winter and the cold resistance during the greening period in the following year) and optimize long-term planting strategies (such as applying potassium fertilizer in advance in autumn to improve winter cold resistance).

[0136] Automated decision-making replaces empiricism: This avoids misjudgments caused by farmers' lack of expertise (such as over-reliance on a single cold-hardiness measure, resulting in costly waste). Automated decision-making recommends optimal solutions through quantitative analysis using algorithms. This approach is particularly suitable for large-scale management by emerging business entities, such as family farms and cooperatives. For example, a cooperative managing 1,000 mu of mixed alfalfa plots used a system that automatically generated differentiated cold-hardiness plans based on soil temperature and variety distribution across plots. This reduced operational costs by 30% compared to manual decision-making and increased cold-hardiness compliance by 25%.

[0137] Explainable policy output: Decisions generated by reinforcement learning can be linked to specific reward items (e.g., "This recommendation is to cover the ground with film. Due to a 4-point increase in ΔH, the expected yield increase is 5%, and the cost increases by 200 yuan per mu"), making it easier for farmers to understand and trust the algorithm's recommendations.

[0138] The visual interface is implemented by the following steps:

[0139] Real-time rendering of cold resistance heat map, the color mapping function is:

[0140] ;

[0141] in, and It is the cold tolerance classification threshold calibrated based on experimental data.

[0142] The color-coded heat map dynamically presents the distribution of cold resistance of mixed alfalfa in the field. Farmers can quickly identify high-risk areas (such as red warning plots) and accurately associate the location of plots with geographic information, assisting farmers in implementing differentiated cold-resistant measures for areas with different risk levels (red / yellow / green) to avoid blind operations across the entire region. The three-level color classification (red / yellow / green) based on the threshold clearly defines the risk level, guides resource allocation according to priority, and improves decision-making efficiency. The visual interface lowers the threshold for technology use, and non-professionals can understand the assessment results based on color alone, reducing the cost of interpreting complex data.

[0143] Furthermore, 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 number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0144] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction 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, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0145] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A deep learning-based cold tolerance evaluation system for alfalfa, characterized in that: include: A data generation module is used to generate phenotypic images and corresponding physiological data of alfalfa under low temperature stress by embedding a diffusion model with physical constraints on plant low temperature response; The physical constraint diffusion model includes physical constraint terms in the inverse denoising process; An evaluation model module that uses a causal-driven two-channel adaptive network to extract cold tolerance features from image and physiological data; The decision-making module includes a hierarchical model distillation unit and a federated reinforcement learning unit. These units implement 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. The model generation process of the physical constraint diffusion is as follows: Among them, x t is the noise image of step t, α t is the noise scheduling coefficient, ∈ is the random noise that obeys the standard normal distribution; The objective function of the inverse denoising process contains physical constraints: in, is the input temperature gradient, S is the generated leaf morphology, k is the cell membrane permeability constant measured by the laboratory, f(T) is the chlorophyll-temperature calibration function, λ1 and λ2 are weight coefficients, SPAD gen is the generated chlorophyll content SPAD value, is the partial derivative of leaf morphology over time; The evaluation model module is used to adopt a causal driven dual channel and includes: Image processing channel: Extract local features of leaves based on Vision Transformer and calculate self-attention weights: Where Q, K and V are query, key and value matrices, and d is the dimension of the feature vector; Physiological data channel: A dynamic weight allocation network adjusts weights based on sensor confidence and optimizes learnable parameters through a back-propagation algorithm; The output features of the image processing channel and the physiological data channel are aligned in dimension and then spliced ​​into a feature matrix F; The hierarchical model distillation unit of the decision module performs a hierarchical model distillation process including: The loss function of the cloud teacher model is: Where, is the cross entropy loss, is the knowledge distillation loss, γ: the weight coefficient for balancing the two types of losses The marginal student model optimizes the feature matching loss via gradient descent: Where, F tea (x) and F stu (x) is the feature output of the teacher model and the student model; β is the weight coefficient of task loss; The reward function of the federated reinforcement learning unit is defined as: r t =αΔH+βY-γC; Among them, ΔH is the change in cold resistance score, Y is the yield prediction value, C is the operating cost, α, β and γ are weight coefficients, and the strategy update adopts the proximal strategy optimization algorithm.

2. The deep learning-based cold tolerance evaluation system for alfalfa according to claim 1, characterized in that: The chlorophyll-temperature calibration function f(T) is constructed by the following steps: The SPAD values ​​at different temperatures T were measured in the laboratory, and a quadratic polynomial was fitted using the least squares method; The fitting function f(T) = aT 2 +bT+c is embedded in the diffusion model to constrain the physiological indicators of the generated data to conform to biological laws; a, b, and c are the coefficients of the quadratic polynomial fitted by the least squares method.

3. The deep learning-based cold tolerance evaluation system for alfalfa according to claim 1, characterized in that: The causal-driven two-channel adaptive network integrates features in and out through the following process: Use the ICECI causal discovery algorithm to construct a causal graph G from historical data to identify the causal relationship between environmental temperature, genetic markers and phenotypic characteristics; The cross-channel causal attention output is: Where F is the input matrix after the image features and physiological features are concatenated; W Q 、W K With W V is the projection matrix of query, key, and value; d is the feature dimension; M is a binary mask matrix generated based on the causal graph G.

4. The deep learning-based cold tolerance evaluation system for alfalfa according to claim 1, characterized in that: The visual interface is implemented by the following steps: Real-time rendering of cold resistance heat map, the color mapping function is: Among them, H low With H high It is the cold tolerance classification threshold calibrated based on experimental data.

Images