A method, device and medium for identifying crop diseases and insect pests and following up on their pathology

Abstract

The present application provides a method, device and medium for identifying crop diseases and pests and for following up on their pathology, which relates to the technical field of identifying crop diseases and pests, and includes obtaining preliminary feature data of crop diseases and pests; encoding diagnostic hypotheses in text format into vector representations of fixed dimensions; configuring a comprehensive feature representation; generating soft labels from a teacher model through a multi-stage strategy, and using the soft labels as supervisory information to guide a student model; saving model checkpoints during training, and using a validation set to evaluate the performance of the student model; predicting the development condition parameters of various diseases; simulating the diffusion process of diseases in a spatial grid through convolution operations, and obtaining a propagation probability map within the next N days; and combining the disease progression prediction results generated under the disease propagation conditions to evaluate the development trend of crop diseases and pests. The present application realizes the detection and diagnosis of crop diseases and pests, and can also trace the propagation path of the diseases and pests and recommend root-causing blocking measures.

Description

Technical Field

[0001] The present application relates to the technical field of crop disease and insect pest identification, and in particular to a crop disease and insect pest identification and pathology follow-up method, equipment and medium. Background Art

[0002] The threat posed by crop diseases and insect pests to agricultural production cannot be underestimated. They severely impact crop yield and quality, hindering the sustainable development of the agricultural economy. In actual agricultural production, the growing environment of crops is complex and changeable, with frequent occurrences of overlapping leaves, reflections, and dew interference, posing a significant challenge to pest and disease identification.

[0003] In related technologies, the growing environment of crops is complex and changeable, and leaf overlap, reflection / dew interference, and other situations often occur, making it difficult for traditional pest and disease recognition models to accurately extract the characteristics of pests and diseases, resulting in low recognition accuracy. For example, overlapping leaves will block some pest and disease symptoms, and reflections and dew will affect the quality of the image, making it impossible for the model to obtain clear pest and disease information. There are differences in crop varieties, pest and disease types, and environmental conditions in different regions. Traditional pest and disease recognition models are often trained based on data from specific regions and lack the ability to generalize data from different regions. When the model is migrated to a new region, due to changes in data distribution, the model's performance will drop significantly, and it will not be able to accurately identify local pests and diseases. Summary of the Invention

[0004] The present invention provides a method for identifying crop diseases and pests and for following up on their pathology, which can realize the detection and diagnosis of crop diseases and pests, trace the transmission paths of the diseases and pests, and recommend root-cause blocking measures.

[0005] Methods include:

[0006] S101: Obtain preliminary characteristic data of crop diseases and pests;

[0007] S102: Generate a diagnostic hypothesis based on the preliminary feature data, and encode the diagnostic hypothesis in text format into a vector representation of fixed dimension;

[0008] S103: Using a gating mechanism to fuse preliminary feature data and vector representation to form a comprehensive feature representation;

[0009] S104: Use the teacher model to dynamically distill the original pest and disease recognition student model. Through a multi-stage strategy, the teacher model generates soft labels, which are used as supervision information to guide the student model.

[0010] S105: Train the student model, save model checkpoints during training, and evaluate the student model performance using the validation set;

[0011] S106: Construct a disease development prediction model based on the trained student model to predict the development condition parameters of various diseases;

[0012] S107: Simulate the disease diffusion process in the spatial grid through convolution operations, and adjust the spread rate based on environmental data to obtain a spread probability map within the next N days;

[0013] S108: Based on the student model generated by dynamic distillation, combined with the disease transmission condition and development condition parameters, by simulating the spread process of pests and diseases in the spatial grid, a propagation probability map within the next N days is generated, which is used to evaluate the development trend of crop diseases and pests.

[0014] It should be further explained that step S104 also includes:

[0015] Collect multi-source training sample sets, including composite feature inputs consisting of visible light images, multispectral data, and environmental sensing parameters;

[0016] The constructed learning model includes a feature extraction unit and a feature enhancement unit. The feature enhancement unit performs random masking on the extracted features.

[0017] The training samples are processed by the feature extraction unit and a primary representation is generated, so that the learning model generates a corrected representation with masking interference through the feature enhancement unit.

[0018] It is further necessary to explain that the difference between the corrected representation and the baseline representation generated by the comparison model is calculated to establish the representation alignment loss function;

[0019] Dynamically adjust model parameters based on representation alignment loss, and terminate the optimization process when the number of training iterations reaches a preset threshold or the loss value converges to a stable range;

[0020] The optimized learning model parameters are transferred to the pest and disease identification student model as the initialization benchmark for dynamic distillation.

[0021] It should be further explained that step S105 specifically includes:

[0022] Set the training objective and loss function, clearly define the training objective, and define the loss function to measure the difference between the student model's prediction results and the target features;

[0023] Dynamically adjust the knowledge distillation strategy according to the learning progress of the student model;

[0024] Based on the optimized loss function and distillation strategy, the student model is trained and the gradient descent algorithm is used to update the parameters.

[0025] It is further necessary to explain that the performance of the student model is evaluated through the validation set, and its adaptability in complex scenarios is monitored;

[0026] The model that performs best on the validation set is selected from the saved checkpoints as the candidate model. The candidate model is fine-tuned and the model structure is adjusted to enhance its ability to capture local features. The optimized model is finally determined as the final student model.

[0027] It should be further explained that step S107 specifically includes the following steps:

[0028] Construct a two-dimensional grid space to simulate the spatial relationship between farmlands;

[0029] Design an individual mobility algorithm based on the regeneration influence of dynamic location resources to simulate the spread of pests and diseases among farmlands;

[0030] Construct a resource regeneration module and set the resource regeneration algorithm based on the number of crops in the grid and the infection status of the crops;

[0031] A dynamic location resource regeneration influence model is established, in which resources affect the spread of pests and diseases by affecting the recovery rate of pests and diseases; a pest and disease transmission module is constructed, and the SIS model is applied to a two-dimensional grid space; the SIS model is constructed on the farmland network, and the changes in the number of uninfected crops and the number of infected crops in the field are described by ordinary differential equations.

[0032] It should be further explained that step S108 specifically includes: constructing a two-dimensional grid system including crop distribution, soil characteristics and environmental parameters, with each grid cell recording the basic infection rate of pests and diseases, coverage rate of disease-resistant crops and soil nutrient content;

[0033] Based on the planting area of ​​disease-resistant crops and crop rotation strategies within the grid unit, the regeneration amount of disease-resistant resources in each unit is calculated;

[0034] Integrate environmental parameters such as temperature, humidity, wind speed and planting density to construct a dynamic influencing factor set;

[0035] The SIS model is used to describe the spread of pests and diseases between grids, and the dynamic transformation equations between infected and disease-resistant crops are defined.

[0036] Nine transmission paths are defined based on grid neighborhood relationships, a probability migration matrix based on environmental influencing factors is designed, and the Monte Carlo method is used to simulate the disease spread trajectory over the next N days, generating a dynamic map containing transmission path probabilities and risk hotspots.

[0037] When the resource regeneration capacity of a preset grid reaches a threshold, the resource transfer protocol of the adjacent grid is automatically triggered. The resource allocation weight is dynamically adjusted according to the direction of disease spread, forming a spatially optimized allocation plan for disease-fighting resources.

[0038] According to another embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the crop disease and pest identification and pathology follow-up method when executing the program.

[0039] According to another embodiment of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for identifying crop diseases and pests and following up on pathology are performed.

[0040] It can be seen from the above technical solutions that the present invention has the following advantages:

[0041] The crop disease and pest identification and pathology follow-up method provided in this application uses multi-stage feature processing, such as generating diagnostic hypotheses from preliminary feature data, encoding them into vectors and fusing them with preliminary features, and using a teacher model to dynamically distill the student model, thereby improving the accuracy of disease and pest identification. The model performs well in cross-regional adaptability. During the dynamic distillation process, the teacher model's powerful knowledge reserve is used to guide the student model's learning, so that the student model can better adapt to the characteristics of disease and pests and environmental conditions in different regions, thereby expanding the scope of application of the model. A disease development prediction model is constructed based on the trained student model, and the diffusion process of the disease in the spatial grid is simulated through convolution operations. The propagation rate is adjusted in combination with environmental data, which can accurately predict the development condition parameters of the disease and the propagation probability map for the next N days. This helps plant protection personnel understand the development trend of the disease in advance, take timely prevention and control measures, and reduce the harm of diseases and pests to crops. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 Flowchart for crop pest and disease identification and pathology follow-up methods;

[0044] Figure 2 A flow chart of an embodiment of a method for identifying crop diseases and pests and following up on pathology;

[0045] Figure 3 Schematic diagram of an electronic device. DETAILED DESCRIPTION

[0046] In the following description, specific details such as specific system structures and technologies are provided for illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details.

[0047] It should be understood that when used in this specification, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their collections. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0048] The phrases "one embodiment" or "some embodiments" described in this application mean that the specific features, structures, or characteristics described in the embodiment are included in one or more embodiments of the application. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in other embodiments," etc. that appear in different places in this application do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized.

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] See also Figure 1 The figure is a flow chart of a method for identifying crop diseases and insect pests and following up on their pathology in a specific embodiment, the method comprising:

[0051] S101: Obtain preliminary characteristic data of crop diseases and pests.

[0052] In some implementations, crop images and environmental data are collected in real time using hyperspectral cameras, temperature and humidity sensors, and soil detectors deployed in farmland. Image data undergoes preprocessing, including denoising, illumination equalization, and leaf segmentation to extract visual features such as lesion shape and color distribution. Environmental data, including parameters such as temperature, humidity, wind speed, and soil pH, is converted into a uniformly dimensioned feature vector through standardization.

[0053] S102: Generate a diagnostic hypothesis based on the preliminary feature data, and encode the diagnostic hypothesis in text format into a vector representation of fixed dimension.

[0054] Specifically, based on preliminary feature data, a pre-trained pest and disease knowledge graph is used to generate diagnostic hypotheses (e.g., "wheat powdery mildew, moderate infection"). The textual hypotheses are encoded into 256-dimensional vectors using the Sentence-BERT model. Cosine similarity is used to calculate the match with the historical case database, and the top three similar cases are selected as diagnostic references. An attention mechanism is introduced into the encoding process to prioritize the semantic weight of pathogen-specific keywords.

[0055] S103: A gating mechanism is used to fuse the preliminary feature data and the vector representation to form a comprehensive feature representation.

[0056] In some embodiments, a two-layer fusion network consisting of GRU units and attention gates is constructed. The first GRU layer processes raw image features, while the second gating layer dynamically assigns fusion weights for image and environmental features based on the semantic importance of the diagnostic hypothesis vector. For example, in dew-induced interference scenarios, the weight coefficient of the environmental humidity feature is automatically increased to 0.7 to suppress the influence of image noise.

[0057] The teacher model is a pretrained pest and disease recognition model using the ResNet-50 deep learning model. It is used to dynamically distill the student model. The teacher model predicts training samples and generates soft labels in the form of a probability distribution. The teacher model dynamically adjusts its soft label generation method at different training stages to gradually guide the student model to learn more refined feature representations. The teacher model transfers knowledge to the student model through a dynamic distillation strategy, improving the student model's diagnostic capabilities and robustness.

[0058] The comprehensive feature representation is formed by fusing preliminary feature data with the embedded form of soft labels generated by the teacher model through a gating mechanism. This fused feature representation serves as supplementary information for the input data during subsequent student model training, indirectly improving the model's diagnostic capabilities.

[0059] S104: Use the teacher model to dynamically distill the original pest and disease identification student model. Through a multi-stage strategy, the teacher model generates soft labels, which are used as supervision information to guide the student model.

[0060] The learning model is a component of the dynamic distillation process, processing multi-source training sample sets to generate primary representations and modified representations with masking interference. The learning model and the teacher model are related in that the feature extraction unit and feature enhancement unit process the training samples to generate primary and modified representations. The teacher model uses the representations generated by the learning model to make predictions and generate soft labels, which serve as supervisory information to guide the student model. The feature enhancement unit of the learning model simulates data noise in real scenarios, helping the teacher model generate more robust soft labels.

[0061] In some embodiments, a progressive distillation strategy is employed: in the first 100 epochs, a teacher model generates hard labels. In the second stage, soft labels are used, with the temperature parameter T linearly decaying from 3.0 to 0.5 to smooth the probability distribution. In the third stage, adversarial training is introduced to enhance model robustness through gradient penalty. The teacher model is based on the ResNet-50 backbone network, and the student model uses a lightweight MobileNetV3 architecture.

[0062] S105: Train the student model, save model checkpoints during training, and evaluate the student model performance using the validation set.

[0063] Specifically, a transfer learning strategy was employed, with initial weights loaded from an ImageNet pre-trained model. Focal Loss was used as the loss function to mitigate class imbalance, and AdamW was used as the optimizer with β1=0.9 and β2=0.999. Checkpoints were saved every five epochs of training, and early stopping was triggered when the validation set loss did not decrease for three consecutive times.

[0064] S106: Construct a disease development prediction model based on the trained student model to predict the development condition parameters of various diseases.

[0065] In some embodiments, based on the improved Logistic growth model, a disease development equation is defined:

[0066] dI / dt = rI(1 - I / K) – αIP;

[0067] Where I is the percentage of infected area, r is the disease growth rate, K is the upper limit of the environmental load, and α is the control coefficient. Parameters are determined by fitting historical data, and the model input variables are modified using weather forecasts. Weather forecasts can include rainfall and diurnal temperature differences over the next seven days.

[0068] S107: Simulate the disease diffusion process in the spatial grid through convolution operations, and adjust the propagation rate based on environmental data to obtain a propagation probability map within the next N days.

[0069] In some embodiments, the crop planting area is divided into spatial grids. Using the current distribution of disease spots as the initial state, a two-dimensional convolution operation is used to simulate the spread of the disease within the spatial grid. Based on real-time environmental data such as temperature, humidity, and wind speed, the weights of the convolution kernel are adjusted to adjust the disease spread rate, thereby calculating a probability map of disease spread over the next N days.

[0070] S108: Based on comprehensive feature representation, student model performance optimization and development condition parameters, and combined with disease propagation conditions, disease progression prediction results are generated to evaluate the development trend of crop diseases and insect pests.

[0071] Specifically, a three-level early warning indicator system is established:

[0072] Level 1 (red): The infection area is expected to expand by more than 30% within 3 days;

[0073] Level 2 (orange): 15%-30%;

[0074] Level 3 (yellow): <15%;

[0075] A prevention and control plan is generated based on the results of the propagation simulation. For example, it is recommended to immediately spray the A5-B7 area of ​​the grid with a suspension concentrate of tebuconazole (concentration 0.3%), and to implement drone spraying in adjacent grids, and to complete the operation within 72 hours.

[0076] In some specific examples, an unidentified disease broke out in a major wheat-producing region. On May 10th, the monitoring system captured an image of an abnormal leaf. Sensor data indicated field humidity of 85%, a temperature of 22°C, and a southeasterly wind of force 3. The image feature extraction module identified spindle-shaped lesions, with a color histogram showing a predominantly grayish-white hue. The diagnostic module generated the hypothesis "early stage infection of wheat fusarium head blight," and the encoded vector matched the 2019 Jiangsu case with an 89% accuracy.

[0077] The teacher model (ResNet-50) generates soft labels based on historical data, with an initial temperature of T=2.5. The student model (MobileNetV3) gradually focuses on key features during training: by the 50th epoch, the gating network increased the weight of the humidity feature to 0.68. At the 120th epoch, adversarial training reduced the model's misclassification rate for reflective leaves by 18%. Validation set evaluation showed that the improved model's accuracy in overlapping leaf scenes increased from 82.3% to 94.1%.

[0078] The disease progression prediction module inputs the current infected area percentage of 8%, combined with the weather forecast (rainfall for the next five days), and calculates a disease growth rate of r = 0.22 / day. Spatial spread simulations indicate that the pathogen will spread along southeasterly winds, affecting three adjacent grids (approximately 45 mu) within 48 hours. The system generates a Level 3 alert, recommending the establishment of a bio-isolation zone (for garlic planting) at the grid intersection and the completion of foliar spraying with pyraclostrobin by May 13.

[0079] In actual prevention and control efforts, farmers deployed traps in key areas as recommended. Three days later, monitoring showed a 62% decrease in new cases. Based on this feedback, the system updated the recovery rate parameter μ to 0.18, recalculating the alert level to level 2. This case study demonstrated the effectiveness of the technical solution in complex farmland environments, reducing the time from disease discovery to spread control to just five days, a 40% reduction in the response time compared to traditional methods.

[0080] In some specific embodiments, such as Figure 2As shown in the figure, the crop disease and pest identification and pathology follow-up method uses Sentence-BERT to encode text hypotheses into 256-dimensional vectors and fuses original features and hypothetical features through a gating mechanism.

[0081] This example uses a large model database as a teacher model to dynamically distill the existing pest and disease recognition model. A multi-stage distillation strategy is employed to generate high-quality soft labels, which are then passed to the student model as supervisory information. During the distillation process, the temperature parameter is dynamically adjusted to control the smoothness of the soft labels. During training, model checkpoints are regularly saved, and the student model's performance is evaluated using a validation set. Dynamic distillation continuously optimizes the accuracy of the inference model, improving the inference hit rate in diverse and complex situations.

[0082] For pathology follow-up and decision optimization, disease progression prediction models must first integrate the development conditions of various diseases and make reasonable predictions about their progression. Here, we present a Python code snippet implementing a spatiotemporal propagation model. This type of model is typically used to predict the spread of diseases, information, or phenomena across time and space. Common models include diffusion models based on partial differential equations (PDEs), cellular automata (CAs), agent-based models (ABMs), and deep learning methods such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs).

[0083] In the context of agricultural pests and diseases, spatiotemporal transmission models need to consider how diseases spread among crops, such as airborne, soil-borne, and insect-borne, while also combining environmental factors (temperature, humidity, wind speed, etc.) and time factors (different stages of disease development).

[0084] The model uses convolution operations to simulate the spread of disease across a spatial grid, adjusting the propagation rate based on environmental data. This approach has the advantages of high computational efficiency and the ability to process gridded spatial data, making it suitable for large-scale farmland monitoring.

[0085] def disease_spread_prediction(current_state):

[0086] # Input: Current lesion distribution matrix (10x10 grid)

[0087] # Output: Propagation probability map for the next N days

[0088] # Propagation constraints based on physical mechanisms

[0089] kernel = np.array([[0,1,0],

[0090] [1,0,1],

[0091] [0,1,0]])# Four-neighborhood propagation

[0092] # DeepSeek environment factor correction

[0093] env_factor = deepseek.query("Current wind speed 2.3m / s, plant spacing 0.4m, calculate spore diffusion coefficient")

[0094] adjusted_kernel = kernel * env_factor['diffusion_coeff']

[0095] # Convolution prediction

[0096] return convolve2d(current_state, adjusted_kernel, mode='same').

[0097] After disease prediction and pathological follow-up, the above reasoning results are used to optimize the multi-dimensional decision-making method. At the same time, the NSGA-Ⅱ multi-objective genetic algorithm is used to perform Pareto frontier search.

[0098] In some embodiments, the original static image analysis and recognition process is abandoned, and a transparent "feature-mechanism-decision" chain is established. This enables the detection and diagnosis of crop pests and diseases to go beyond identifying symptoms and trace the transmission path of the pests and diseases and recommend root cause prevention measures.

[0099] The difference in recognition accuracy of complex scenes:

[0100] Leaf overlap scenario: 82.3% -> 94.1% (increase of 11.8%);

[0101] Reflection / dew interference scene: 68.7% -> 89.5% (increase of 20.8%);

[0102] Cross-regional adaptability:

[0103] The F1 value of the model migrated to the new region: 0.61->0.87 (an increase of 42.6%).

[0104] By integrating a large language model for interactive question-answering during decision-making, and utilizing multi-objective optimization and dynamic distillation knowledge graphs, we can achieve autonomous evolution of "practical feedback->model iteration->strategy optimization", thereby simultaneously improving the diagnostic success rate under various conditions and the decision-making friendliness for different situations.

[0105] In an exemplary embodiment, step S104 further includes the following steps:

[0106] S1041: Collect multi-source training sample sets, including composite feature inputs consisting of visible light images, multispectral data, and environmental sensing parameters: In order for the model to learn rich and comprehensive crop characteristics, data needs to be collected from multiple dimensions.

[0107] This embodiment uses a high-definition camera to capture visible light images of crops. These images can intuitively display visual characteristics such as color, shape, and texture. Multispectral sensors collect data across different spectral bands. Crops in different health states have different reflectance and absorption characteristics across these spectral bands, enabling the capture of pathological conditions that are difficult to detect with the naked eye.

[0108] S1042: Construct a heterogeneous model architecture, where the learning model includes a feature extraction unit and a feature enhancement unit. The feature enhancement unit performs random masking on the extracted features. The comparison model retains only the feature extraction unit structure and initialization parameters that are identical to the learning model. This involves constructing two structurally different models. The learning model includes a feature extraction unit and a feature enhancement unit. The feature extraction unit is responsible for extracting key features from multi-source training samples. The feature enhancement unit performs random masking on the extracted features to simulate the partial data loss or interference that may occur in actual applications, prompting the model to learn more robust and versatile feature representations.

[0109] S1043: The training sample is processed by the feature extraction unit to generate a primary representation, and the learning model additionally generates a modified representation with masking interference by the feature enhancement unit.

[0110] In this embodiment, multi-source training samples are input into the feature extraction units of the learning model and the comparison model. The feature extraction units analyze and process the sample data to generate primary representations, which contain basic feature information of crop diseases and pests.

[0111] For the learning model, after the feature extraction unit generates the primary representation, it will also be input into the feature enhancement unit. The feature enhancement unit will mask some features in the primary representation according to preset random rules, thereby generating a corrected representation with masking interference, simulating scenarios where data is missing or interfered with.

[0112] S1044: Calculate the difference between the corrected representation and the baseline representation generated by the comparison model, and establish a representation alignment loss function, whose mathematical form is: L = ||f_s (X)⊙M - f_t (X)||² . Here, M is a random masking matrix, and f_s and f_t represent the representation outputs of the learning model and the comparison model, respectively.

[0113] The modified representation obtained by the learning model after processing by the feature enhancement unit is compared with the baseline representation generated by the comparison model. The difference between the two is quantified using the representation alignment loss function.

[0114] S1045: Dynamically adjust model parameters based on the representation alignment loss, and terminate the optimization process when the number of training iterations reaches a preset threshold or the loss value converges to a stable range.

[0115] This embodiment uses the alignment loss value as a feedback signal and applies the stochastic gradient descent algorithm and its variants to dynamically adjust the parameters of the learning model.

[0116] In each training iteration, the algorithm adjusts model parameters based on the magnitude of the loss and the direction of the gradient, gradually reducing the difference between the corrected representation generated by the learning model and the baseline representation of the comparison model. This process continues until the number of training iterations reaches a pre-set threshold, or the loss converges to a stable range, indicating that the model has learned a relatively stable and effective feature representation. At this point, the optimization process terminates.

[0117] S1046: Migrate the optimized learning model parameters to the pest and disease identification student model as the initialization benchmark for dynamic distillation.

[0118] In this example, the learning model parameters optimized through the above steps are transferred to the pest and disease recognition student model, serving as initialization parameters for the student model during the dynamic distillation process. This parameter transfer allows the student model to converge to a better state more quickly, leveraging the effective feature representations and knowledge acquired by the learning model during early training. This improves the efficiency and effectiveness of dynamic distillation.

[0119] As can be seen, the multi-source data used in model training significantly enriches the feature dimensions, enabling the model to capture information about crop pests and diseases. The use of random masking and representation alignment loss functions enhances the model's robustness, enabling it to accurately identify pests and diseases even in complex situations such as missing data and noise interference. Furthermore, during the dynamic distillation process, the student model is able to better learn from the teacher model, further improving the accuracy of pest and disease identification.

[0120] Step S105 of this embodiment specifically includes: clarifying the training objectives of the student model. For example, if the goal is to identify specific types of crop pests and diseases, the training objectives may include accurately classifying the pest type and assessing the severity of the disease. One or more loss functions are defined to measure the difference between the student model's predictions and the actual target features. Common loss functions include cross-entropy loss (for classification tasks) and mean squared error (for regression tasks). This helps guide the model to learn the correct patterns and regularities.

[0121] This embodiment dynamically adjusts the knowledge distillation strategy based on the student model's learning progress. During the knowledge distillation process, the teacher model's knowledge is transferred to the student model via soft labels. As the student model's learning progresses, the specific knowledge distillation strategy can be dynamically adjusted, such as by changing the method or ratio of soft labels generated by the teacher model to adapt to the student model's learning state. This adaptive strategy helps accelerate the student model's learning process and improve its ultimate performance.

[0122] Based on the optimized loss function and distillation strategy, the student model is trained and its parameters are updated using gradient descent. The model is trained using the selected loss function and a knowledge distillation strategy adjusted based on the student model's learning progress. In each iteration, the loss between the current model output and the target value is calculated, and the model parameters are updated using gradient descent to minimize the loss function. This step is the core of model training and determines whether the model can learn effective feature representations from the data.

[0123] Evaluate the student model's performance on a validation set and monitor its adaptability in complex scenarios: To ensure that the model not only performs well on the training set but also maintains good generalization capabilities on unseen data, an independent validation set is typically used to evaluate the model's performance. Furthermore, a series of complex scenarios can be designed to test the model's adaptability and robustness, ensuring that the model can handle a variety of real-world scenarios.

[0124] The model with the best performance on the validation set is selected from the saved checkpoints as the candidate model, and the candidate model is fine-tuned to adjust the model structure to enhance its ability to capture local features. The optimized model is finally determined as the final student model: throughout the training process, the state of the model (i.e., checkpoints) is saved regularly. After the training is completed, the model version with the best performance on the validation set is selected as the basis for further optimization. In this embodiment, the model can be fine-tuned, especially its ability to capture local features, such as adjusting the convolution kernel size, adding an attention mechanism, etc., to further improve the model performance. The optimized model is finally determined as the completed student model.

[0125] In step S107 of this embodiment, a convolution operation is used to simulate the disease diffusion process in the spatial grid, and the propagation rate is adjusted in combination with environmental data to obtain a propagation probability map within the next N days.

[0126] Step S107 of this embodiment specifically involves the following steps:

[0127] S701: Construct a 2D grid space to simulate the spatial relationships between farmlands. A rectangular 2D grid network with side lengths W and H is created to simulate a real farmland area. The network consists of W × H grids. Each grid represents a small plot of farmland and records crop-related attributes, such as crop type, health status, and environmental parameters.

[0128] S702: Design an individual mobility algorithm based on the dynamic location resource regeneration influence to simulate the spread of pests and diseases among farmlands.

[0129] In this embodiment, pests can only move in nine simple ways between adjacent grids on a two-dimensional grid. Each pest spreads according to the probability of the dynamic location resource regeneration influence in the above nine directions.

[0130] S703: Construct a resource regeneration module and set a resource regeneration capacity algorithm based on the number of crops in the grid and the crop infection status. Considering that the uninfected crops in the field at each time point are capable of participating in photosynthesis and nutrient absorption, a farmland resource regeneration capacity algorithm based on the number of uninfected crops at a location is proposed to calculate the farmland resource regeneration capacity at each time step.

[0131] S704: When the amount of farmland resources is excessive, in order to effectively utilize the resources, the farmland can donate the resources to other farmland nearby. When the amount of farmland renewable resources exceeds a certain amount, a willingness to donate resources is generated, and the willingness to donate resources of the farmland is quantified.

[0132] S705: Build a dynamic site resource regeneration influence model. Site resource regeneration reflects the ability of a field to resist the spread of pests and diseases by generating resources. Field resource regeneration influences the decision-making behavior of organisms as it changes over time.

[0133] Resources influence the spread of pests and diseases by affecting their recovery rate. The stronger the field's resource regeneration capacity, the greater the amount of resources produced, and the easier it is to control the spread of pests and diseases.

[0134] S706: Construct a pest and disease transmission module and apply the SIS model to a two-dimensional grid space.

[0135] The SIS model is constructed on a farmland network. The changes in the number of uninfected and infected crops in a field can be described using ordinary differential equations. Considering that the infection rate of a field is affected by multiple factors in reality, such as soil fertility and irrigation conditions, it is negatively correlated with the size of the field. That is, the larger the field, the lower the corresponding infection rate.

[0136] It can be seen that this embodiment aims to enhance the function of the disease development prediction model based on S106, by introducing resource management and allocation mechanisms, and a grid-based spatial propagation model to simulate the spread of pests and diseases among farmlands and improve prediction accuracy.

[0137] In step S108 of this embodiment, a two-dimensional grid system is constructed, encompassing crop distribution, soil characteristics, and environmental parameters. Each grid cell records the baseline pest infection rate, disease-resistant crop coverage, and soil nutrient content. A mapping relationship is established between grid coordinates and actual geographic locations, forming a topological structure for the farmland ecological communication network.

[0138] This example calculates the regeneration of disease-resistant resources in each grid unit based on the planting area and crop rotation strategy of disease-resistant crops within the unit. A quantitative relationship model is established between resource regeneration and disease suppression effectiveness. When resource levels exceed a preset threshold, a cross-grid resource allocation mechanism is triggered. Environmental parameters such as temperature, humidity, wind speed, and planting density are integrated to construct a dynamic set of influencing factors. A weighted algorithm is used to determine the correction coefficient for each factor on the disease transmission rate, establishing a nonlinear relationship between the probability of transmission and environmental influencing factors.

[0139] This embodiment also uses the SIS model to describe the spread of pests and diseases between grids, defining dynamic transformation equations for infected and disease-resistant crops. Resource regeneration is introduced as a recovery rate correction parameter, establishing a set of differential equations encompassing spatial diffusion and resource allocation. Nine propagation paths are defined based on grid neighborhood relationships, and a probability migration matrix based on environmental influencing factors is designed. A Monte Carlo method is used to simulate the spread of pests over the next N days, generating a dynamic map containing propagation path probabilities and risk hotspots. During the simulation, meteorological forecast data is overlaid in real time to dynamically correct the diffusion rate parameters.

[0140] When monitoring indicates that resource regeneration in a specific grid reaches a threshold, a resource transfer protocol is automatically triggered between adjacent grids. Resource allocation weights are dynamically adjusted based on the direction of disease spread, forming a spatially optimized allocation plan for disease-fighting resources. A resource transportation cost model is established, and a path planning algorithm is used to select the optimal allocation path, ensuring a balance between resource timeliness and economic efficiency.

[0141] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0142] like Figure 3As shown, the present application also provides an electronic device, including a display module 103, a memory 102, a processor 101, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, the steps of the crop disease and pest identification and pathology follow-up method are implemented.

[0143] In the embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present application described and / or claimed herein.

[0144] In the embodiment of the present application, the processor 101 can be implemented by using at least one of an application-specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, and an electronic unit designed to perform the functions described herein. In some cases, such an embodiment can be implemented in a controller. For software implementation, an embodiment such as a process or function can be implemented with a separate software module that allows the execution of at least one function or operation. The software code can be implemented by a software application (or program) written in any appropriate programming language, and the software code can be stored in a memory and executed by a controller.

[0145] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc.

[0146] The memory 102 can be used to store software programs and various data. The memory 102 can include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0147] The present application also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for identifying crop diseases and pests and for following up on their pathology.

[0148] The storage medium can be any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0149] The above description of the disclosed embodiments will enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying crop diseases and insect pests and for following up on their pathology, characterized in that: Methods include: S101: Obtain preliminary characteristic data of crop diseases and pests; S102: Generate a diagnostic hypothesis based on the preliminary feature data, and encode the diagnostic hypothesis in text format into a vector representation of fixed dimension; S103: Using a gating mechanism to fuse preliminary feature data and vector representation to form a comprehensive feature representation; S104: Use the teacher model to dynamically distill the original pest and disease recognition student model. Through a multi-stage strategy, the teacher model generates soft labels, which are used as supervision information to guide the student model. S105: Train the student model, save model checkpoints during training, and evaluate the student model performance using the validation set; S106: Constructing a disease development prediction model based on the trained student model to predict the development condition parameters of various diseases; S107: Simulate the disease diffusion process in the spatial grid through convolution operations, and adjust the spread rate based on environmental data to obtain a spread probability map within the next N days; S108: Based on the student model generated by dynamic distillation, combined with the disease transmission condition and development condition parameters, by simulating the spread process of pests and diseases in the spatial grid, a propagation probability map within the next N days is generated, which is used to evaluate the development trend of crop diseases and pests.

2. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 1, characterized in that: Step S104 further includes: Collect multi-source training sample sets, including composite feature inputs consisting of visible light images, multispectral data, and environmental sensing parameters; The constructed learning model includes a feature extraction unit and a feature enhancement unit. The feature enhancement unit performs random masking on the extracted features. The training samples are processed by the feature extraction unit and a primary representation is generated, so that the learning model generates a corrected representation with masking interference through the feature enhancement unit.

3. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 2, characterized in that: Calculate the difference between the corrected representation and the baseline representation generated by the comparison model, and establish the representation alignment loss function; Dynamically adjust model parameters based on representation alignment loss, and terminate the optimization process when the number of training iterations reaches a preset threshold or the loss value converges to a stable range; The optimized learning model parameters are transferred to the pest and disease identification student model as the initialization benchmark for dynamic distillation.

4. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 1, characterized in that: Step S105 specifically includes: Set the training objective and loss function, clearly define the training objective, and define the loss function to measure the difference between the student model's prediction results and the target features; Dynamically adjust the knowledge distillation strategy according to the learning progress of the student model; Based on the optimized loss function and distillation strategy, the student model is trained and the gradient descent algorithm is used to update the parameters.

5. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 4, characterized in that: The performance of the student model is evaluated through the validation set, and its adaptability in complex scenarios is monitored. The model with the best performance on the validation set is selected from the saved checkpoints as the candidate model. The candidate model is fine-tuned and the model structure is adjusted to enhance its ability to capture local features. Finally, the optimized model is determined as the final student model.

6. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 1, characterized in that: Step S107 specifically includes the following steps: Construct a two-dimensional grid space to simulate the spatial relationship between farmlands; Design an individual mobility algorithm based on the regeneration influence of dynamic location resources to simulate the spread of pests and diseases among farmlands; Construct a resource regeneration module and set the resource regeneration algorithm based on the number of crops in the grid and the infection status of the crops; A dynamic location resource regeneration influence model is established, in which resources affect the spread of pests and diseases by affecting the recovery rate of pests and diseases; a pest and disease transmission module is constructed, and the SIS model is applied to a two-dimensional grid space; the SIS model is constructed on the farmland network, and the changes in the number of uninfected crops and the number of infected crops in the field are described by ordinary differential equations.

7. The method for identifying crop diseases and insect pests and following up on their pathology according to claim 1, characterized in that: Step S108 specifically includes: constructing a two-dimensional grid system including crop distribution, soil characteristics and environmental parameters, with each grid unit recording the basic infection rate of pests and diseases, coverage rate of disease-resistant crops and soil nutrient content; Based on the planting area of ​​disease-resistant crops and crop rotation strategies within the grid unit, the regeneration amount of disease-resistant resources in each unit is calculated; Integrate temperature, humidity, wind speed and planting density parameters to construct a dynamic influencing factor set; The SIS model is used to describe the spread of pests and diseases between grids, and the dynamic transformation equations between infected and disease-resistant crops are defined. Nine transmission paths are defined based on grid neighborhood relationships, a probability migration matrix based on environmental influencing factors is designed, and the Monte Carlo method is used to simulate the disease spread trajectory over the next N days, generating a dynamic map containing transmission path probabilities and risk hotspots. When it is monitored that the regeneration amount of preset grid resources reaches the threshold, the resource transfer protocol of the adjacent grid is automatically triggered, and the resource allocation weight is dynamically adjusted according to the direction of disease propagation to form a spatial optimization configuration plan for disease-resistant resources.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the crop disease and insect pest identification and pathology follow-up method according to any one of claims 1 to 7 are implemented.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the crop disease and insect pest identification and pathology follow-up method according to any one of claims 1 to 7 are implemented.

Images