Artificial intelligence-based crop disease and pest control scheme recommendation method and system

Abstract

The application relates to the technical field of disease and pest control, and discloses a crop disease and pest control scheme recommendation method and system based on artificial intelligence, which collects multi-modal data of a target farmland, pre-processes the multi-modal data, inputs the pre-processed multi-modal data into a deep multi-modal fusion network for disease and pest identification, and outputs disease and pest types; the identified diseases and pests are subjected to causal reasoning analysis, key driving factors and key intervention nodes of the diseases and pests are identified; based on the causal reasoning result, a digital twin model is constructed, the digital twin model is taken as an environment simulator, a PPO algorithm is used to optimize a control strategy, an optimal control strategy is obtained, and the optimal control strategy is issued; and the application realizes accurate, dynamic and personalized recommendation of a disease and pest control scheme.

Description

Technical Field

[0001] This invention relates to the field of pest and disease control technology, specifically to a method and system for recommending crop pest and disease control programs based on artificial intelligence. Background Technology

[0002] Crop diseases and pests are major biological stressors affecting agricultural production and food security, characterized by their diversity, complex transmission routes, and high outbreak frequency. Traditional pest and disease control mainly relies on manual inspections and experience-based judgment, which suffers from problems such as delayed diagnosis, lack of personalized control plans, and imprecise pesticide application. Existing methods are mostly based on statistical associations for pest and disease identification and control recommendations, failing to understand the causal mechanisms of pest and disease occurrence, resulting in a significant decrease in recommendation accuracy when facing new pests and diseases or drastic changes in environmental conditions. Existing methods are mostly static recommendations, failing to fully consider the temporal and spatial evolution of pest and disease transmission, as well as the differentiated responses of crops at different growth stages. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing an artificial intelligence-based method and system for recommending crop pest and disease control programs.

[0004] The first aspect of this invention provides a method for recommending crop pest and disease control programs based on artificial intelligence, the method comprising the following steps: Multimodal data of the target farmland is collected and preprocessed, wherein the multimodal data includes crop image data, environmental data, soil data, meteorological data and agricultural data; The preprocessed multimodal data is input into a deep multimodal fusion network for pest and disease identification, and the pest and disease type is output. Conduct causal reasoning analysis on the identified pests and diseases to identify the key driving factors and key intervention points for the occurrence of pests and diseases. Based on the causal reasoning results, a digital twin model is constructed. The digital twin model is used as an environment simulator. The PPO algorithm is used to optimize the prevention and control strategy to obtain the optimal prevention and control strategy, and then the optimal prevention and control strategy is issued.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of collecting multimodal data of the target farmland and performing preprocessing includes: Interpolation alignment is used to resample multi-source data to a unified frequency reference according to their respective timestamps. Local outlier factors are used to detect outliers in continuous data and mean interpolation is performed. Linear minimum-maximum scaling is used to map all numerical features to the same order of magnitude. The image data is sequentially subjected to geometric correction, uniform cropping normalization, random rotation, flipping, scaling, and color dithering enhancement operations. A sliding time window is used to segment continuous environmental, meteorological, and agricultural time-series data into input sequence segments of fixed historical length.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of inputting the preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and outputting the pest and disease type includes: The preprocessed image is divided into a fixed-size sequence of two-dimensional image blocks. Each image block is flattened and mapped to the latent space dimension through a linear projection layer. Learnable positional codes are then superimposed on each vector. The image block vector sequence after overlay position encoding is input into the Vision Transformer encoder, and a multi-head self-attention operation is performed to generate query, key, and value triples. Attention weights are obtained by scaling dot product and normalization exponential function processing. The weighted aggregated value vector is then transformed by a multilayer perceptron and output as a visual feature tensor.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and outputting the pest and disease type further includes: The environment time series sequence is input into the temporal Transformer encoder. Relative position encoding is superimposed on each time step, and the dependencies are calculated in time order through a masked multi-head self-attention mechanism to output the environment time series feature tensor. The visual feature tensor is used as the query input, and the environmental temporal feature tensor is used as the key and value input. The correlation weights between image features and temporal features are calculated through the cross-attention module to generate a fused feature tensor. The fused feature tensor is aggregated into a single global feature vector through global average pooling, and then subjected to nonlinear transformation through a multi-layer fully connected network. Finally, the probability distribution of each category is calculated by a normalized exponential function classifier, and the category corresponding to the maximum value is taken as the output pest and disease type.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing causal reasoning analysis on the identified pests and diseases to identify key driving factors and key intervention nodes for the occurrence of pests and diseases includes: Construct a causal diagram with environmental factors, soil factors, pest and disease status, crop growth status, and control actions as nodes; Perform maximum likelihood estimation on the conditional probability distribution of each node in the causal graph under the given conditions of its parent node; For each candidate environmental factor node in the causal graph, perform the do operator intervention operation and recalculate the probability distribution of the target pest node state after intervention; The average causal effect value is calculated by comparing the differences in the probability distribution of the occurrence of target pests and diseases before and after the intervention of each candidate factor. Key driving factors are screened by sorting them according to the effect intensity, and key intervention nodes are identified by searching along the action path in combination with the feasibility constraints of prevention and control.

[0009] Optionally, in the fifth implementation of the first aspect of the present invention, the step of constructing a digital twin model based on the causal reasoning result, using the digital twin model as an environment simulator, optimizing the prevention and control strategy using the PPO algorithm to obtain the optimal prevention and control strategy, and issuing the optimal prevention and control strategy includes: The digital twin model is encapsulated into an interactive reinforcement learning environment, and the positive benefits of prevention and control, negative benefits of costs, and negative ecological impacts are weighted and combined into a single-step reward function. In the digital twin environment, the probability distribution of actions is output based on the current state and the actions are sampled and executed. After receiving the actions, the environment returns the next state, reward value and round termination flag. Repeated interactions generate multi-step experience trajectories and store them in the experience cache pool. Batch trajectory data are sampled from the experience buffer pool, and the generalized dominance estimation algorithm is used to calculate the dominance function value at each time step; Calculate the probability ratio of the current policy relative to the old policy for sampling actions as the policy ratio, and update the network based on the policy ratio. The training convergence is determined by repeatedly collecting and updating data until the cumulative reward value remains unchanged. The current real-time status data of the farmland is then input into the converged Actor network, and the optimal prevention and control strategy is output through forward propagation. The optimal prevention and control strategy, including the type of pesticide, dosage, and spraying area, is then sent to the execution terminal.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, a near-end policy optimization objective function is used to limit the policy ratio within a preset range to calculate the sheared policy loss and update the Actor network, while the mean square error of the state value prediction and cumulative reward is calculated to update the Critic network.

[0011] A second aspect of this invention provides an artificial intelligence-based system for recommending crop pest and disease control solutions, the system comprising: The multimodal data acquisition module is used to collect multimodal data of the target farmland and perform preprocessing, wherein the multimodal data includes crop image data, environmental data, soil data, meteorological data and agricultural data; The multimodal fusion identification module is used to input preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and output the pest and disease type. The causal reasoning module is used to perform causal reasoning analysis on the identified pests and diseases, and to identify the key driving factors and key intervention points for the occurrence of pests and diseases. The recommendation generation module is used to construct a digital twin model based on the causal reasoning results, use the digital twin model as an environment simulator, optimize the prevention and control strategy using the PPO algorithm, obtain the optimal prevention and control strategy, and then issue the optimal prevention and control strategy.

[0012] A third aspect of the present invention provides an artificial intelligence-based crop pest and disease control scheme recommendation device, the artificial intelligence-based crop pest and disease control scheme recommendation device comprising a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the artificial intelligence-based crop pest and disease control scheme recommendation device to perform the various steps of the artificial intelligence-based crop pest and disease control scheme recommendation method as described in any of the preceding claims.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the method for recommending crop pest and disease control schemes based on artificial intelligence as described in any of the preceding claims.

[0014] The technical solution provided by this invention involves collecting multimodal data from target farmland and preprocessing it. The preprocessed multimodal data is then input into a deep multimodal fusion network for pest and disease identification, outputting pest and disease types. Causal reasoning analysis is performed on the identified pests and diseases to identify key driving factors and key intervention nodes. Based on the causal reasoning results, a digital twin model is constructed. This digital twin model is used as an environmental simulator, and the PPO algorithm is employed to optimize the control strategy, obtaining the optimal control strategy, which is then distributed. This invention achieves precise, dynamic, personalized, and interpretable recommendations for pest and disease control schemes. The deep multimodal fusion network integrates multi-source heterogeneous data such as images, environment, soil, and meteorology, improving the accuracy of pest and disease identification. It can understand the causal mechanisms of pest and disease occurrence and maintain good recommendation accuracy even under new pests and diseases or environmental changes. The digital twin model is used for forward-looking simulation evaluation of control schemes, and deep reinforcement learning is combined to achieve dynamic optimization of the strategy, enabling the control scheme to adapt to the spatiotemporal evolution of pests and diseases and environmental changes. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A flowchart illustrating the method for recommending crop pest and disease control solutions based on artificial intelligence, as provided in this embodiment of the invention. Figure 2A schematic diagram of the structure of an artificial intelligence-based crop pest and disease control solution recommendation system provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of the recommended equipment for crop pest and disease control based on artificial intelligence provided in the embodiments of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 The flowchart of the method for recommending crop pest and disease control programs based on artificial intelligence provided in this embodiment of the invention includes the following steps: Step 101: Collect multimodal data of the target farmland and perform preprocessing; In this embodiment, multimodal data includes crop image data, environmental data, soil data, meteorological data, and agricultural data. Based on the timestamp field of each acquisition device, such as IoT sensors, drones, satellites, and ground cameras, an interpolation alignment algorithm is used to resample all data to a unified time-frequency reference, ensuring that different modal data have a consistent correspondence in the time dimension during subsequent fusion analysis. For continuous environmental, soil, and meteorological data, a detection algorithm based on local outlier factors or interquartile range criteria is used to identify and mark abnormal records that deviate from the normal distribution range. Outliers are imputed using mean or median to fill missing data segments. All numerical features are mapped to the same order of magnitude using a linear minimum-maximum scaling algorithm or robust normalization algorithm to eliminate the impact of different dimensions on subsequent neural network training. Image data is then subjected to geometric correction, uniform cropping and normalization, random rotation, flipping, scaling, and color dithering enhancement operations in sequence. Geometric correction involves first performing distortion correction on original images acquired from different acquisition platforms based on the calibration parameters of each acquisition device to eliminate radial and tangential distortion caused by lens optical characteristics. Simultaneously, using preset ground control points or geometric feature points of image edges, the image projection is corrected to a unified geospatial reference coordinate system through affine or perspective transformation to ensure that images from different sources have accurate pixel-level correspondences in spatial location. Uniform cropping and normalization involves cropping the corrected image to a uniform size according to a preset region of interest, and then scaling the cropped sub-image to a fixed pixel size using a bicubic interpolation algorithm. The pixel values ​​are normalized, linearly mapping each channel's pixel values ​​from the original bit depth range to a standardized range with zero mean and unit variance, thus eliminating the inconsistency in pixel value distribution caused by differences in exposure and sensitivity between different acquisition devices. The random rotation, flipping, scaling, and color jitter enhancement operations are performed during the online data preprocessing phase of the model training stage. Specifically, for the cropped and normalized image, the image is rotated in-plane by randomly selecting rotation angles of 0°, 90°, 180°, or 270° with uniform probability. Whether to perform horizontal or vertical flipping on the image is determined with equal probability. The scaling factor is uniformly sampled according to a preset scaling factor range, and the image is randomly scaled before being cropped back to a fixed size. Random perturbations following a uniform distribution are added to the brightness, contrast, saturation, and hue channels of the image in the color space to generate diverse training samples with different spatial transformation and color change characteristics. For time-series data such as environmental, meteorological, and agricultural records, a sliding time window algorithm is used to cut the continuous time series into input sequence segments with fixed historical lengths.

[0019] Step 102: Input the preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification, and output the pest and disease type; In this embodiment, the preprocessed image is divided into a fixed-size two-dimensional image patch sequence. Each image patch is flattened into a one-dimensional vector and mapped to the latent space dimension through a trainable linear projection layer. Subsequently, a learnable positional encoding vector is superimposed on each image patch vector to preserve its spatial position information. The sequence of image patch vectors with superimposed positional encoding is then input into the Vision Transformer encoder. Inside the encoder, multi-head self-attention operations are performed sequentially: a query, key, and value triple is generated for each input vector; the dot product attention score of the query and all keys is calculated; the attention weights are obtained after scaling and normalizing the exponential function; then, the weighted aggregated value vector is weighted and nonlinearly transformed through a multilayer perceptron to output a visual feature tensor. The temporal sequence of the environment is input into the temporal Transformer encoder. First, relative positional encoding is superimposed on each time step in the sequence. Then, through a masked multi-head self-attention mechanism, the dependencies between each time step are calculated sequentially in chronological order, outputting an environmental temporal feature tensor containing historical dependencies. The visual feature tensor is used as the query input, and the environmental temporal feature tensor is used as the key and value input, and then fed into the cross-modal cross-attention module. This module calculates the correlation weight between image features and each temporal feature, and dynamically weights and fuses environmental information into the visual representation accordingly, generating a fusion feature tensor for pests and diseases with context awareness. The fusion feature tensor is aggregated into a single global feature vector through global average pooling, and then nonlinearly transformed through a multi-layer fully connected network. Finally, a normalized exponential function classifier calculates the probability distribution of the sample belonging to each preset pest and disease category, and takes the category corresponding to the maximum probability as the output pest and disease type.

[0020] Step 103: Conduct causal reasoning analysis on the identified pests and diseases to identify the key driving factors and key intervention points for the occurrence of pests and diseases; In this embodiment, based on prior expert knowledge in the field of agricultural pests and diseases, an initial causal directed acyclic graph (DAG) framework is constructed with pest and disease-related variables as nodes. Nodes cover environmental factors, soil factors, pest and disease status, crop growth status, and control actions. Directed edges between nodes represent known potential causal relationships. Using collected historical multimodal big data, maximum likelihood estimation is performed on the conditional probability distribution of each node in the causal graph under given conditions of its parent node. This involves statistically analyzing the frequency of each state of the child node under each combination of parent nodes, thus constructing a complete set of structured causal model parameters. For each candidate environmental factor node in the causal graph, a do-operator intervention operation is performed to fix the node at a specific intervention value, cutting off all input edges from other nodes to that node. The probability distribution change of the target pest and disease node state after intervention is recalculated in the forward propagation algorithm. For each candidate factor, the difference in the probability distribution of the target pest and disease occurrence before and after intervention is compared, and its average causal effect value is calculated to quantify the marginal contribution of the factor change to the degree of impact on the outbreak of pests and diseases. All calculated causal effect values ​​are sorted, and environmental factors with the highest effect intensity are selected as key driving factors. Subsequently, in conjunction with the feasibility constraints of prevention and control, the search along the action paths of these driving factors on the causal graph was conducted to identify intermediate nodes that can be adjusted through prevention and control measures as key intervention nodes.

[0021] Step 104: Based on the causal reasoning results, construct a digital twin model, use the digital twin model as an environment simulator, use the PPO algorithm to optimize the prevention and control strategy, obtain the optimal prevention and control strategy, and issue the optimal prevention and control strategy.

[0022] In this embodiment, real-time data of the current farmland is collected as the initial boundary conditions, such as crop variety, sowing date, current growth stage, and current soil and environmental conditions. Pre-set crop variety genetic parameters and local soil hydraulic parameters are loaded to complete the initial state setting of the digital twin model. Within the digital twin, the crop growth sub-model is driven to run iteratively on a daily basis. Based on daily meteorological input data, this sub-model calculates the crop's photosynthetic rate, respiration consumption, dry matter accumulation and distribution, leaf area index dynamics, and phenological progression, outputting daily updated crop biomass and canopy structure status. Based on the identified initial distribution and severity of current pests and diseases, the pest and disease transmission sub-model is driven. This sub-model employs a spatial grid discretization method, combined with local meteorological conditions, to calculate the disease infection rate and insect population density changes in each spatial grid unit according to the propagation dynamics rules, and to predict the spatial spread of diseases and pests over multiple time steps in the future. Based on a pre-set library of control measures, multiple candidate control schemes are generated. Each candidate scheme is sequentially loaded into the digital twin model as an external intervention input, driving the control measure effect sub-model to simulate changes in disease and pest mortality rates and crop growth inhibition or promotion effects under each scheme. For the result dataset generated after the simulation of each candidate scheme, three dimensions of quantitative evaluation values ​​are calculated: first, the control effect dimension, the final reduction rate of diseases and pests; second, the economic dimension, the net benefit calculated from pesticide costs, labor costs, and estimated yield loss; and third, the ecological dimension, the pesticide residue index and the impact index on beneficial organisms. Each candidate scheme is considered a solution point in the three-dimensional evaluation space. Using the Pareto dominance comparison algorithm, all candidate schemes are traversed, and solutions that are not simultaneously dominated by any other scheme in all three dimensions are selected, forming the Pareto optimal frontier solution set, which serves as the candidate action reference set for reinforcement learning.

[0023] In this embodiment, the converged digital twin model is encapsulated into an interactive reinforcement learning environment. Within this environment, a state space is explicitly defined, including crop growth status, spatial distribution of pests and diseases, future weather forecast sequences, and the time interval since the last treatment. An action space is also defined, including candidate pesticide types, application dosages, and application area grid selection. A multi-objective reward function is designed, weighting and combining the positive benefits of pest and disease reduction, positive benefits of yield maintenance, negative benefits of cost, and negative benefits of ecological impact into a single-step reward scalar value. An actor network and a critic network are constructed, with their network weights set using a random initialization method. An experience cache pool is also initialized to store interaction samples. The digital twin environment is reset to the current initial state of the farmland. The actor network outputs an action probability distribution based on the current state, samples and executes actions according to this distribution, and the environment receives actions, performs internal deduction, and returns the next state, reward value, and round termination flag. Repeat this interaction process to generate an experience trajectory containing a multi-step state-action-reward sequence and store it in the experience cache pool; sample batch trajectory data from the experience cache pool, and for each time step, use the generalized advantage estimation algorithm to calculate the advantage function value of that step, which measures the degree of superiority or inferiority of the actual action performed in that step compared to the average level of the current policy. Based on the sampled data, the policy ratio of the actor network is calculated, which is the probability ratio of the current policy relative to the old policy on the sampled action. Using the pruning objective function of Proximal Policy Optimization (PPO), the policy ratio is restricted to a preset range, truncating excessively large or small update amplitudes. The pruned policy loss gradient is calculated, and the actor network parameters are updated via backpropagation. Based on the sampled data, the mean squared error between the predicted state value and the actual cumulative discounted reward is calculated and used as the loss function of the critic network. The gradient is calculated, and the critic network parameters are updated via backpropagation. This collection and update process is repeated until the cumulative reward value obtained by the action policy output by the policy network in the digital twin environment tends to stabilize and no longer increases significantly, indicating training convergence. The current actor network parameters are saved as the optimal prevention and control strategy model. The latest farmland state data collected in real time is input into the converged optimal prevention and control strategy network. After forward propagation, the network outputs the current optimal action vector, which contains specific prevention and control instructions including the optimal pesticide type, dosage, and spraying area. This instruction is formatted, packaged, and sent to the farmer's mobile terminal or smart application equipment terminal for execution.

[0024] Please see Figure 2 A schematic diagram of the structure of an artificial intelligence-based crop pest and disease control program recommendation system provided in this embodiment of the invention. The system includes: The multimodal data acquisition module is used to collect multimodal data of the target farmland and perform preprocessing. The multimodal data includes crop image data, environmental data, soil data, meteorological data, and agricultural data. The multimodal fusion identification module is used to input preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and output the pest and disease type. The causal reasoning module is used to perform causal reasoning analysis on the identified pests and diseases, and to identify the key driving factors and key intervention points for the occurrence of pests and diseases. The recommendation generation module is used to construct a digital twin model based on the causal reasoning results. The digital twin model is used as an environment simulator. The PPO algorithm is used to optimize the prevention and control strategy to obtain the optimal prevention and control strategy and then distribute the optimal prevention and control strategy.

[0025] Figure 3 This is a schematic diagram of the structure of an AI-based crop pest and disease control solution recommendation device 300 provided in an embodiment of the present invention. This AI-based crop pest and disease control solution recommendation device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the AI-based crop pest and disease control solution recommendation device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the AI-based crop pest and disease control solution recommendation device 300 to implement the method provided in the above embodiment.

[0026] The AI-based crop pest and disease control solution recommendation device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The recommended equipment structure for the AI-based crop pest and disease control scheme shown does not constitute a limitation on the computer equipment provided by this invention. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0027] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform each step of the method for recommending crop pest and disease control schemes based on artificial intelligence provided in the above embodiments.

[0028] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described equipment or apparatus / unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0029] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A method for recommending crop pest and disease control programs based on artificial intelligence, characterized in that, The method includes the following steps: Multimodal data of the target farmland is collected and preprocessed, wherein the multimodal data includes crop image data, environmental data, soil data, meteorological data and agricultural data; The preprocessed multimodal data is input into a deep multimodal fusion network for pest and disease identification, and the pest and disease type is output. Conduct causal reasoning analysis on the identified pests and diseases to identify the key driving factors and key intervention points for the occurrence of pests and diseases. Based on the causal reasoning results, a digital twin model is constructed. The digital twin model is used as an environment simulator. The PPO algorithm is used to optimize the prevention and control strategy to obtain the optimal prevention and control strategy, and then the optimal prevention and control strategy is issued.

2. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 1, characterized in that, The process of collecting multimodal data from the target farmland and preprocessing it includes: Interpolation alignment is used to resample multi-source data to a unified frequency reference according to their respective timestamps. Local outlier factors are used to detect outliers in continuous data and mean interpolation is performed. Linear minimum-maximum scaling is used to map all numerical features to the same order of magnitude. The image data is sequentially subjected to geometric correction, uniform cropping normalization, random rotation, flipping, scaling, and color dithering enhancement operations. A sliding time window is used to segment continuous environmental, meteorological, and agricultural time-series data into input sequence segments of fixed historical length.

3. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 1, characterized in that, The process of inputting preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and outputting pest and disease types includes: The preprocessed image is divided into a fixed-size sequence of two-dimensional image blocks. Each image block is flattened and mapped to the latent space dimension through a linear projection layer. Learnable positional codes are then superimposed on each vector. The image block vector sequence after overlay position encoding is input into the Vision Transformer encoder, and a multi-head self-attention operation is performed to generate query, key, and value triples. Attention weights are obtained by scaling dot product and normalization exponential function processing. The weighted aggregated value vector is then transformed by a multilayer perceptron and output as a visual feature tensor.

4. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 3, characterized in that, The step of inputting the preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and outputting pest and disease types also includes: The environment time series sequence is input into the temporal Transformer encoder. Relative position encoding is superimposed on each time step, and the dependencies are calculated in time order through a masked multi-head self-attention mechanism to output the environment time series feature tensor. The visual feature tensor is used as the query input, and the environmental temporal feature tensor is used as the key and value input. The correlation weights between image features and temporal features are calculated through the cross-attention module to generate a fused feature tensor. The fused feature tensor is aggregated into a single global feature vector through global average pooling, and then subjected to nonlinear transformation through a multi-layer fully connected network. Finally, the probability distribution of each category is calculated by a normalized exponential function classifier, and the category corresponding to the maximum value is taken as the output pest and disease type.

5. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 1, characterized in that, The causal reasoning analysis of the identified pests and diseases identifies key driving factors and key intervention points for their occurrence, including: Construct a causal diagram with environmental factors, soil factors, pest and disease status, crop growth status, and control actions as nodes; Perform maximum likelihood estimation on the conditional probability distribution of each node in the causal graph under the given conditions of its parent node; For each candidate environmental factor node in the causal graph, perform the do operator intervention operation and recalculate the probability distribution of the target pest node state after intervention; The average causal effect value is calculated by comparing the differences in the probability distribution of the occurrence of target pests and diseases before and after the intervention of each candidate factor. Key driving factors are screened by sorting them according to the effect intensity, and key intervention nodes are identified by searching along the action path in combination with the feasibility constraints of prevention and control.

6. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 1, characterized in that, The process involves constructing a digital twin model based on causal reasoning results, using this digital twin model as an environmental simulator, optimizing the prevention and control strategy using the PPO algorithm to obtain the optimal prevention and control strategy, and then distributing the optimal prevention and control strategy. This includes: The digital twin model is encapsulated into an interactive reinforcement learning environment, and the positive benefits of prevention and control, negative benefits of costs, and negative ecological impacts are weighted and combined into a single-step reward function. In the digital twin environment, the probability distribution of actions is output based on the current state and the actions are sampled and executed. After receiving the actions, the environment returns the next state, reward value and round termination flag. Repeated interactions generate multi-step experience trajectories and store them in the experience cache pool. Batch trajectory data are sampled from the experience buffer pool, and the generalized dominance estimation algorithm is used to calculate the dominance function value at each time step; Calculate the probability ratio of the current policy relative to the old policy for sampling actions as the policy ratio, and update the network based on the policy ratio. The training convergence is determined by repeatedly collecting and updating data until the cumulative reward value remains unchanged. The current real-time status data of the farmland is then input into the converged Actor network, and the optimal prevention and control strategy is output through forward propagation. The optimal prevention and control strategy, including the type of pesticide, dosage, and spraying area, is then sent to the execution terminal.

7. The method for recommending crop pest and disease control programs based on artificial intelligence as described in claim 6, characterized in that, The objective function of near-end policy optimization is used to limit the policy ratio within a preset range to calculate the policy loss after shearing and update the Actor network. At the same time, the mean square error of state value prediction and cumulative reward is calculated to update the Critic network.

8. An artificial intelligence-based crop pest and disease control program recommendation system, characterized in that, The system includes: The multimodal data acquisition module is used to collect multimodal data of the target farmland and perform preprocessing, wherein the multimodal data includes crop image data, environmental data, soil data, meteorological data and agricultural data; The multimodal fusion identification module is used to input preprocessed multimodal data into a deep multimodal fusion network for pest and disease identification and output the pest and disease type. The causal reasoning module is used to perform causal reasoning analysis on the identified pests and diseases, and to identify the key driving factors and key intervention points for the occurrence of pests and diseases. The recommendation generation module is used to construct a digital twin model based on the causal reasoning results, use the digital twin model as an environment simulator, optimize the prevention and control strategy using the PPO algorithm, obtain the optimal prevention and control strategy, and then issue the optimal prevention and control strategy.

9. An artificial intelligence-based crop pest and disease control program recommendation device, characterized in that: The AI-based crop pest and disease control scheme recommendation device includes a memory and at least one processor. The memory stores instructions. The at least one processor invokes the instructions in the memory to cause the AI-based crop pest and disease control scheme recommendation device to perform each step of the AI-based crop pest and disease control scheme recommendation method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement each step of the method for recommending crop pest and disease control schemes based on artificial intelligence as described in any one of claims 1-7.

Images