Agricultural pest and disease prediction method and device based on big data and medium

Abstract

The invention discloses an agricultural pest prediction method and device based on big data, and a medium, and relates to the technical field of big data analysis, and the method comprises the steps: building a propagation mode prototype through carrying out the multi-scale spatial-temporal feature coding and clustering of historical pest propagation cases, and carrying out the prediction of the propagation mode prototype; and performing feature comparison on the current observation sample and the propagation mode under different scales to obtain a feature trend reflecting propagation similarity. And performing multi-scale supervised learning on a graph neural network based on the feature trend, and performing adaptive calibration on model parameters in combination with a self-supervised learning process of a current observation sample, thereby realizing modeling and prediction of a disease and pest propagation relationship. Meanwhile, a prevention event flow is introduced as a propagation inhibition factor, a prediction result is intervened, and a pest and disease damage prediction result considering prevention and control influences is generated. The accuracy and adaptability of pest and disease prediction can be improved, and support is provided for agricultural prevention and control decisions.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method, equipment, and medium for predicting agricultural pests and diseases based on big data. Background Technology

[0002] Agricultural pests and diseases are among the main factors affecting crop yield and quality. Their occurrence and spread are influenced by a variety of factors, including meteorological conditions, soil environment, crop growth stage, and human agricultural activities, exhibiting significant spatiotemporal heterogeneity and transmission characteristics. With the development of agricultural informatization and sensing technology, multi-source data from meteorological monitoring, pest monitoring, disease monitoring, and agricultural management are continuously accumulating, providing a data foundation for pest and disease prediction.

[0003] Existing methods for predicting agricultural pests and diseases mostly employ statistical analysis or traditional machine learning models to fit historical monitoring data in order to predict the probability of pest and disease occurrence at a certain time or in a certain area. However, these methods typically treat each monitoring point or area as an independent prediction object, focusing on single-point or local time series characteristics, making it difficult to effectively characterize the spatial spread path of pests and diseases and their evolution over time. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a big data-based method for predicting agricultural pests and diseases, which solves the problems that existing pest and disease prediction methods are unable to depict the transmission process, adapt to real-time observation changes, and reflect the intervention effect of control measures.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting agricultural pests and diseases based on big data, which includes acquiring multi-source data of the area to be predicted, wherein the multi-source data includes meteorological data, soil environmental data, crop phenological data, agricultural pest and disease control data, and monitoring data of pest and disease conditions. The multi-source data is spatiotemporally aligned to form a sample sequence indexed by grid cells and time windows of the region to be predicted; The observed cases in the database and the sample sequence are compared in terms of features to obtain the feature trend of the sample sequence. By using the graph neural network constrained by the aforementioned features to learn the sample sequence, the tendency of pest and disease events to occur in each unit is predicted; and based on the prevention and control event flow, the results of the tendency prediction are intervened to obtain the prediction results of agricultural pests and diseases.

[0007] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the spatiotemporal alignment includes resampling and synchronizing the multi-source data according to a preset time window, and mapping the spatial data to the grid cell; After removing outlier data, missing data is imputed, and sample vectors are generated for each grid cell in each time window to form the sample sequence.

[0008] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the observation cases include obtaining spatiotemporal coding features at different observation scales by performing feature clustering processing on sample sequences corresponding to historical pest and disease transmission cases; the specific steps are as follows: When extracting features from sample sequences corresponding to historical pest and disease transmission cases, a spatiotemporal coupled pyramid coding network is used for spatiotemporal joint coding to obtain spatiotemporal coding features at each scale. In each case, the similarity of the spatiotemporal coding features at each scale is calculated, and the similarity is used as a distance metric. For all cases at the same scale, the spatiotemporal coding features are clustered, and the spatiotemporal coding feature that minimizes the sum of intra-cluster distances in each cluster is selected as the cluster center. Each cluster center is labeled with its corresponding case and scale, and its spatiotemporal coding features are marked to obtain each observation case. The spatiotemporal coupled pyramid coding network includes a spatiotemporal joint embedding module, which constructs a spatiotemporal joint input tensor from sample sequences indexed by grid cells and time windows, and performs three-dimensional convolutional embedding to obtain basic features; The spatiotemporal coupled feature extraction module includes at least one spatiotemporal residual block, which contains two three-dimensional convolutional layers and achieves joint spatiotemporal domain feature extraction through residual connections; based on the basic features, the spatiotemporal correlation in the process of pest and disease transmission is jointly modeled, and spatiotemporal coding features reflecting the evolutionary characteristics of the pest and disease transmission process are extracted. The spatiotemporal joint downsampling module uses 3D convolution with a preset stride to simultaneously downsample basic features in the temporal and spatial dimensions to form multi-scale pyramid features. At different levels of the spatiotemporal coupled feature extraction module, the resolution of the temporal and spatial dimensions is simultaneously reduced to form multi-level spatiotemporal feature representations with different observation scales, thereby constructing a spatiotemporal feature pyramid. The multi-scale spatiotemporal coding output module outputs spatiotemporal coding features at each level of the spatiotemporal feature pyramid, corresponding to the observation scale, as spatiotemporal coding features at different observation scales.

[0009] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the feature comparison includes obtaining the original sample sequences corresponding to all cluster centers; The scale information in the cluster center labels is used as the scale for calculation between the corresponding original sample sequence and the currently observed sample sequence. For each cluster center at any scale i, calculate the similarity at label scale i: A random truncation length is extracted from the original sample sequence for use by the spatiotemporally coupled pyramid coding network to extract spatiotemporal coding features at scale i. The extracted spatiotemporal coding features are denoted as: F i,j ;F i,j Let i represent the spatiotemporal coding features of the j-th cluster center at scale i; Simultaneously, based on the aforementioned spatiotemporal coupled pyramid coding network, spatiotemporal coding features are extracted at scale i for the currently observed sample sequence. The extracted spatiotemporal coding features are denoted as: f i,0 ;f i,0 This represents the spatiotemporal coding features of the currently observed sample sequence at scale i; For F i,j and f i,0 Calculate the similarity S, and select the maximum value of S as the similarity between the currently observed sample sequence and the j-th cluster center at scale i by randomly truncating the sequence length. In this process, based on the feature comparison at each scale, the similarity with each cluster center is obtained, which is used to represent the feature trend of the current observed sample sequence at each scale.

[0010] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the graph neural network using the feature trend constraint includes constructing a basic graph structure based on the grid cells of the region to be predicted, wherein each grid cell serves as a node in the graph, and the spatial adjacency relationship between grid cells serves as a basic propagation relationship. The similarity between the current observed sample sequence and each cluster center is used as a propagation relationship influencing factor to constrain the propagation relationship between nodes in the basic graph structure; The influencing factors are used as the attention of the original sample sequences corresponding to the cluster centers. The original sample sequences corresponding to each cluster center are used as supervised learning samples. The graph neural network is subjected to multi-scale supervised learning according to the observation scale corresponding to the cluster centers, thereby enhancing the sensitivity of the graph neural network to different propagation features at different scales.

[0011] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the learning of the sample sequence is a self-supervised learning process based on the current observed sample sequence, and the parameters of the graph neural network are adaptively calibrated by utilizing the spatiotemporal consistency constraints of the current observed sample sequence itself. After completing the self-supervised learning process, the graph neural network calibrated by self-supervision is used to predict the tendency of pest and disease events to occur in each grid cell, so as to obtain the probability of pest and disease events in each grid cell at each prediction time node.

[0012] As a preferred embodiment of the big data-based agricultural pest and disease prediction method of the present invention, the prevention and control event flow includes a sequence of prevention and control events performed by each grid unit; By mapping the corresponding pest and disease transmission inhibition factors to each prevention and control event, and applying them to a self-supervised calibrated graph neural network, the final agricultural pest and disease prediction results are generated.

[0013] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the big data-based agricultural pest and disease prediction method described in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the big data-based agricultural pest and disease prediction method described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: By performing multi-scale spatiotemporal joint modeling of historical pest and disease transmission processes, this invention constructs a transmission pattern prototype and compares the features of current observation samples with the transmission pattern. This effectively characterizes the transmission trends of pests and diseases at different temporal and spatial scales, avoiding dependence on origin and single-point features. Based on this, multi-scale supervised learning of the graph neural network is performed using the described feature trends, guiding the model to learn the transmission relationships of pests and diseases rather than simply predicting the probability of occurrence, significantly improving its ability to express the transmission mechanism. Simultaneously, by introducing a self-supervised learning process based on current observation samples, the model parameters are adaptively calibrated, improving the model's generalization ability in different regions, years, and under different control conditions. Furthermore, this invention introduces control event flows as transmission inhibition factors into the prediction process, dynamically intervening in the prediction results to ensure that the prediction results truly reflect the impact of control measures, thereby improving the accuracy, interpretability, and practical value of the prediction results for agricultural pest and disease control decision-making. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a big data-based method for predicting agricultural pests and diseases. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Reference Figure 1 This is one embodiment of the present invention, which provides a method for predicting agricultural pests and diseases based on big data, including the following steps: S1: Obtain multi-source data for the area to be predicted. The multi-source data includes meteorological data, soil environmental data, crop phenological data, agricultural pest and disease control data, and pest and disease monitoring data.

[0022] The occurrence and spread of pests and diseases are not only influenced by meteorological conditions, but also closely related to soil environment, crop growth stage, and agricultural control practices. Furthermore, pest and disease monitoring data directly reflect the actual development status of pests and diseases. A single data source cannot fully depict the driving factors and evolutionary patterns of pest and disease transmission, easily leading to biased or distorted prediction results.

[0023] By comprehensively collecting meteorological data, soil environmental data, crop phenological data, agricultural control data, and pest and disease monitoring data, the model in S3 can simultaneously perceive the external environmental conditions, crop susceptibility, and human intervention behaviors related to pest and disease transmission. This provides multi-dimensional input support for subsequent spatiotemporal alignment, transmission feature extraction, and transmission relationship modeling. Building a multi-source data foundation improves the completeness of subsequent pest and disease transmission modeling and the reliability of prediction results, thereby meeting the needs of actual agricultural production for the accuracy and applicability of pest and disease predictions.

[0024] S2: Perform spatiotemporal alignment on the multi-source data to form a sample sequence indexed by the grid cells of the region to be predicted and the time window.

[0025] Furthermore, the spatiotemporal alignment includes resampling and synchronizing the multi-source data according to a preset time window, and mapping the spatial data to the grid cells. After removing outlier data, missing data is imputed, and sample vectors are generated for each grid cell in each time window to form the sample sequence.

[0026] It is worth noting that by performing unified spatiotemporal alignment processing on multi-source data, heterogeneous agricultural data with different sources, sampling frequencies, and spatial representations are transformed into structurally consistent sample sequences that can be jointly modeled, laying the foundation for subsequent extraction and modeling of pest and disease transmission characteristics. Because meteorological data, soil environmental data, crop phenological data, and pest and disease monitoring data differ significantly in terms of collection period, spatial resolution, and data completeness, direct modeling can easily introduce temporal misalignment and spatial mismatch, affecting the accurate depiction of pest and disease transmission processes.

[0027] By resampling and synchronizing multi-source data according to preset time windows, the temporal scale differences between different data sources are eliminated. Spatial data is uniformly mapped to grid cells of the region to be predicted, achieving a consistent spatial representation. Based on this, outlier data is removed and missing data is imputed, reducing the interference of data noise and incompleteness on model learning, thereby generating a structured sample vector for each grid cell within each time window. Through spatiotemporal alignment processing, a sample sequence indexed by grid cells and time windows is constructed, enabling the learning of pest and disease transmission patterns within a unified spatiotemporal reference frame, thus improving the modeling ability of spatiotemporal evolution characteristics and the stability of prediction results.

[0028] S3: Compare the features of the observed cases in the database with the sample sequence to obtain the feature trend of the sample sequence.

[0029] The observation cases include obtaining spatiotemporal coding features at different observation scales by performing feature clustering processing on sample sequences corresponding to historical pest and disease transmission cases. The specific steps are as follows: When extracting features from sample sequences corresponding to historical pest and disease transmission cases, a spatiotemporal coupled pyramid coding network is used for spatiotemporal joint coding to obtain spatiotemporal coding features at each scale.

[0030] In each case, the similarity of the spatiotemporal coding features at each scale is calculated and the similarity is used as a distance metric. For all cases at the same scale, the spatiotemporal coding features are clustered, and the spatiotemporal coding feature that minimizes the sum of intra-cluster distances is selected as the cluster center in each cluster.

[0031] Each cluster center is labeled with its corresponding case and scale, and its spatiotemporal coding features are marked to obtain each observation case.

[0032] Because historical pest and disease transmission processes vary significantly in duration, spread range, and pace, directly comparing all historical samples with current samples is not only computationally complex but also susceptible to interference from individual abnormal transmission events, making it difficult to reflect typical transmission patterns. This paper addresses this issue by extracting features from sample sequences corresponding to historical pest and disease transmission cases and employing a spatiotemporally coupled pyramid coding network to jointly encode the transmission process. This allows each historical transmission process to generate corresponding feature representations at different temporal and spatial scales, reflecting the changing characteristics of the transmission process at different observation scales. Based on this, similarity calculations and clustering are performed on the features of each historical transmission process at the same scale. Cases with similar transmission processes are grouped into one category, and samples representing the transmission characteristics of that category are selected as cluster centers. By labeling the cases corresponding to the cluster centers and their observation scales, the resulting observation cases can serve as representatives of typical transmission patterns, used to describe different types and scales of pest and disease transmission processes throughout history. The advantage of this setup is that, by simply comparing the current observed samples with these typical transmission patterns, it is possible to determine which historical scenario the current pest and disease transmission is closer to, thereby reducing redundant calculations, avoiding noise interference, and improving the relevance and reliability of pest and disease transmission trend analysis.

[0033] It's important to note that in pest and disease transmission analysis, historical transmission cases are typically numerous, and different cases often exhibit repetition and similarities in transmission range, duration, and rate of change. Directly using all historical samples for comparative analysis would not only significantly increase computational costs but also easily lead to the influence of numerous similar or redundant samples, thus weakening the ability to identify typical transmission patterns. The features corresponding to cluster centers best represent the common characteristics of the transmission process within that category, effectively reflecting the typical evolutionary pattern of pest and disease transmission. Compared to using all samples, using cluster centers for analysis is more conducive to highlighting the dominant trends of transmission characteristics and avoiding interference from individual anomalies or marginal samples. Furthermore, by selecting only cluster centers as representative samples, the most representative feature information from historical transmission data can be retained while significantly reducing data size and computational complexity, thereby improving the overall system's operational efficiency and practicality while ensuring analytical effectiveness.

[0034] The spatiotemporal coupled pyramid coding network includes: a spatiotemporal joint embedding module, which constructs a spatiotemporal joint input tensor from sample sequences indexed by grid cells and time windows, and performs 3D convolutional embedding to obtain basic features; a spatiotemporal coupled feature extraction module, which includes at least one spatiotemporal residual block, wherein the spatiotemporal residual block contains two 3D convolutional layers and achieves joint spatiotemporal domain feature extraction through residual connections; based on the basic features, it jointly models the spatiotemporal correlation in the disease and pest transmission process and extracts spatiotemporal coding features reflecting the evolutionary characteristics of the disease and pest transmission process; a spatiotemporal joint downsampling module, which simultaneously downsamples the basic features in the temporal and spatial dimensions through 3D convolution with a preset stride to form multi-scale pyramid features; at different levels of the spatiotemporal coupled feature extraction module, the resolution of the temporal and spatial dimensions is simultaneously reduced to form multi-level spatiotemporal feature representations with different observation scales, thereby constructing a spatiotemporal feature pyramid; and a multi-scale spatiotemporal coding output module, which outputs the spatiotemporal coding features at the corresponding observation scale at each level of the spatiotemporal feature pyramid, serving as spatiotemporal coding features at different observation scales.

[0035] It is worth noting that the spread of pests and diseases exhibits significant temporal continuity and spatial diffusion characteristics, which are interdependent and difficult to separate and model. The spatiotemporally coupled pyramid coding network constructs a spatiotemporally joint input tensor from sample sequences indexed by grid cells and time windows, and uses 3D convolution for feature embedding. This allows the temporal and spatial dimensions to be modeled simultaneously in the same feature extraction process, which is beneficial for accurately reflecting the spread behavior of pests and diseases in different regions over time. By introducing a residual structure into the spatiotemporally coupled feature extraction module, the network can focus on learning the changing characteristics during the spread process while maintaining the original spread state information, thus more stably depicting the evolutionary trend of pest and disease spread and avoiding feature degradation or information loss due to the increase in network layers. Furthermore, in practice, pest and disease spread can manifest as localized, short-term concentrated events or as cross-regional, long-term diffusion processes. By simultaneously downsampling in the temporal and spatial dimensions to construct a multi-level spatiotemporal feature pyramid, the network can extract spread features at different observation scales, thus taking into account both local details and overall trends. The spatiotemporal coding features output at each level can be directly used for subsequent feature comparison and propagation relationship modeling, avoiding information bias caused by scale inconsistency and improving the coherence and effectiveness of the overall modeling process.

[0036] Furthermore, the feature comparison includes obtaining the original sample sequences corresponding to all cluster centers. Simultaneously, based on the scale information in the cluster center labels, a scale is used to calculate the relationship between the corresponding original sample sequences and the currently observed sample sequences.

[0037] It's important to note that during feature comparison, each cluster center corresponds to a specific observation scale during its formation. This scale reflects the typical time span and spatial impact range of the historical pest and disease transmission process. Therefore, when comparing the original sample sequences corresponding to the cluster centers with the currently observed sample sequences, a uniform or arbitrarily set scale is not used. Instead, the scale information contained in the cluster center labels is directly used as the scale basis for feature calculation. This ensures that the currently observed samples are always compared within a time and spatial range that matches the historical transmission pattern, avoiding distortion in similarity calculations due to scale inconsistencies.

[0038] For each cluster center at any scale i, calculate the similarity at label scale i: A random truncation length is extracted from the original sample sequence for use by the spatiotemporally coupled pyramid coding network to extract spatiotemporal coding features at scale i. The extracted spatiotemporal coding features are denoted as: F i,j ;F i,j Let i represent the spatiotemporal coding features of the j-th cluster center at scale i.

[0039] Simultaneously, based on the aforementioned spatiotemporal coupled pyramid coding network, spatiotemporal coding features are extracted at scale i for the currently observed sample sequence. The extracted spatiotemporal coding features are denoted as: f i,0 ;f i,0 This represents the spatiotemporal coding features of the currently observed sample sequence at scale i.

[0040] For F i,j and f i,0 Calculate the similarity S, and select the maximum value of S by randomly truncating the sequence length as the similarity between the currently observed sample sequence and the j-th cluster center at scale i.

[0041] In this process, based on the feature comparison at each scale, the similarity with each cluster center is obtained, which is used to represent the feature trend of the current observed sample sequence at each scale.

[0042] It's important to note that in reality, the pests and diseases currently observed are often at a certain stage of transmission, but it's difficult to determine in advance which specific point in the historical transmission process they correspond to. Directly comparing them with historical data of fixed locations or lengths can easily lead to an underestimation of true similarity due to stage misalignment. Therefore, this method uses random truncation of the original sample sequences corresponding to the cluster centers, essentially continuously searching for possible transmission stages that correspond to the current observation state during the historical transmission process, and then extracting spatiotemporal coding features at the corresponding scale for comparison.

[0043] By extracting spatiotemporal coding features from historical sample fragments and current observation samples at the same scale, consistency in time span and spatial scope between the compared samples can be ensured, making similarity calculation more reasonable. Selecting the maximum similarity score from multiple random cuts as the final result is equivalent to determining the stage position of the current observation sample that most closely resembles this type of historical propagation pattern at that scale, thus avoiding misjudgments caused by differences in the speed of propagation or different start times. The similarity obtained in this way better reflects the degree of similarity within the propagation process itself, rather than a simple time alignment relationship.

[0044] S4: Using the graph neural network constrained by the aforementioned features, the sample sequence is learned, and the tendency of pest and disease events to occur in each unit is predicted; and based on the prevention and control event flow, the result of the tendency prediction is intervened to obtain the prediction result of agricultural pests and diseases.

[0045] Furthermore, the graph neural network utilizing the aforementioned characteristic trend constraints includes constructing a basic graph structure based on grid cells of the region to be predicted, wherein each grid cell serves as a node in the graph, and the spatial adjacency relationship between grid cells serves as a basic propagation relationship.

[0046] The similarity between the current observed sample sequence and each cluster center is used as a propagation relationship influencing factor to constrain the propagation relationship between nodes in the basic graph structure.

[0047] The influencing factors are used as the attention of the original sample sequences corresponding to the cluster centers. The original sample sequences corresponding to each cluster center are used as supervised learning samples. The graph neural network is subjected to multi-scale supervised learning according to the observation scale corresponding to the cluster centers, thereby enhancing the sensitivity of the graph neural network to different propagation features at different scales.

[0048] It should be noted that in actual agricultural scenarios, pests and diseases often do not appear in isolation, but rather spread gradually along spatially adjacent areas. Therefore, by first using grid cells as nodes and spatial adjacency relationships as the basic propagation relationships, the model can have the basic structure to describe spatial propagation paths.

[0049] Building upon this, the similarity between the current observed samples and each cluster center is incorporated into the graph structure. Its purpose is not simply to adjust the numerical values, but rather to determine which propagation relationships are more likely to hold true in the current context and which should be weakened. In other words, different historical propagation patterns have varying reference value for the current propagation process; propagation patterns with high similarity should participate more in model learning, while patterns with low similarity have less guiding significance for current predictions.

[0050] Furthermore, the similarity is used as the attention level of the original sample sequence corresponding to the cluster center, enabling the graph neural network to automatically focus on historical propagation samples that are closer to the current propagation trend during training. By performing multi-scale supervised learning according to the observation scale corresponding to the cluster center, the model can learn different types of propagation characteristics such as short-term local propagation and long-term regional diffusion, thereby avoiding the bias caused by single-scale modeling.

[0051] Multi-scale supervised training is achieved by using historical pest and disease transmission processes at different observation scales as training references. Specifically, each cluster center corresponds to a specific observation scale at its formation, reflecting the typical time span and spatial impact range of that type of transmission process. When training the graph neural network, instead of mixing all historical samples together for unified training, the historical transmission samples are divided into multiple scale levels according to the scale corresponding to the cluster centers, and each level participates in model training.

[0052] During training, the graph neural network learns in batches from supervised samples at different scales, based on the same fundamental graph structure. For training at a specific scale, the network focuses only on the propagation characteristics corresponding to that scale, using the original sample sequence of cluster centers at that scale as a supervised reference, enabling the model to learn the changing patterns of pest and disease propagation relationships within that time and space. In this way, the network can learn the characteristics of rapid local propagation at short scales and the characteristics of regional diffusion or long-term evolution at larger scales.

[0053] Meanwhile, the similarity between the current observed samples and each cluster center is incorporated into the training process to adjust the influence of samples from different scales and cluster centers on training. Samples from cluster centers with higher similarity play a stronger guiding role in training, while samples with lower similarity have a relatively weaker impact on model updates, thus ensuring that the model's learning focus remains consistent with the current propagation trend.

[0054] By repeatedly performing supervised training at different scales and gradually updating the same model parameters, the graph neural network can simultaneously perceive multiple propagation scales without biasing towards a single scale. Ultimately, when faced with new observation samples, the model can naturally invoke matching scale information for prediction based on the current propagation characteristics, thereby improving its modeling ability and prediction accuracy for complex pest and disease propagation processes.

[0055] Learning from the sample sequence is a self-supervised learning process based on the current observed sample sequence. It utilizes the spatiotemporal consistency constraints of the current observed sample sequence itself to adaptively calibrate the parameters of the graph neural network.

[0056] After completing the self-supervised learning process, the graph neural network calibrated by self-supervision is used to predict the tendency of pest and disease events to occur in each grid cell, so as to obtain the probability of pest and disease events in each grid cell at each prediction time node.

[0057] Through this self-supervised calibration process, the model can automatically correct biases caused by environmental changes or differences in data distribution, ensuring that subsequent predictions are based on conditions more closely aligned with current realities. After completing self-supervised learning, the calibrated graph neural network is then used to predict pest and disease event tendencies in each grid cell. This yields the probability distribution of pest and disease occurrence in each grid cell at each prediction time point, thereby improving the reliability and practicality of the prediction results and providing a more accurate reference for pest and disease control.

[0058] The prevention and control event flow includes a sequence of prevention and control events performed by each grid cell. Each prevention and control event is mapped to a corresponding pest and disease transmission inhibition factor, which is then applied to a self-supervised calibrated graph neural network to generate the final agricultural pest and disease prediction result.

[0059] It should be noted that the prevention and control event flow includes a sequence of prevention and control events performed by each grid cell. By mapping each prevention and control event to a corresponding pest and disease transmission inhibition factor, and applying this factor to a self-supervised calibrated graph neural network, the final agricultural pest and disease prediction result is generated.

[0060] The control events in each grid cell are arranged chronologically to form a control event sequence, and different types of control events are mapped to corresponding transmission inhibition factors to describe the degree to which the control measures weaken the ability of pests to spread. After self-supervised calibration, the transmission inhibition factors are applied to the graph neural network, enabling the model to proactively reduce the transmission impact on the control area and its adjacent areas when making predictions, thereby simulating the inhibitory effect of control measures on transmission paths and intensity.

[0061] In this way, the prediction results are no longer based solely on natural transmission trends, but comprehensively consider the intervention effects of human control, making the final agricultural pest and disease prediction results closer to the actual production situation. This helps to evaluate the effectiveness of existing control measures and provides more valuable information for subsequent control decisions.

[0062] In this embodiment, the spread of pests and diseases is expressed through the information transmission between grid cells in a graph neural network. When a grid cell “transmits risk” to surrounding cells, it corresponds to the possibility of the pest or disease spreading.

[0063] When control measures are implemented in a grid cell, the survival rate and transmission capacity of pests and diseases within that area will significantly decrease, and their potential for outward spread will also weaken. Therefore, the control event is mapped to a transmission inhibition factor d, which describes the degree to which the transmission capacity of the grid cell is weakened over a certain period of time. The transmission inhibition factor d can be understood as an "attenuation coefficient" acting on the transmission process, and its magnitude reflects the strength of the control measures.

[0064] In graph neural networks, the propagation inhibition factor 'd' acts on the propagation process of grid cells. This weakens the propagation influence transmitted from a grid cell to its neighboring grid cells when the grid cell is in a preventative state, or restricts its participation in propagation calculations within a certain time window. Thus, when the model performs propagation inference, it naturally reduces the propagation risk contribution from that grid cell, thereby reflecting the blocking or mitigation effect of preventative measures on the propagation path.

[0065] For example, in a certain area, grid cell A is adjacent to surrounding grid cells B and C. Without preventative measures, the pest and disease risk in A will affect B and C through transmission, causing the predicted risk of B and C to increase over time.

[0066] When pesticide spraying is implemented within grid cell A, the corresponding control event is mapped to a transmission inhibition factor d. If the control measure is effective, d takes a smaller value. In this case, during the model's transmission process, the transmission impact from A is proportionally weakened, meaning the risk of A spreading to B and C is significantly reduced, or even no longer participates in the transmission calculation during the effective control period. As a result, the predicted risk growth rate of B and C slows down significantly, reflecting the actual inhibitory effect of pesticide spraying on the spread of pests and diseases.

[0067] This embodiment also provides a computer device applicable to the big data-based agricultural pest and disease prediction method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the big data-based agricultural pest and disease prediction method proposed in the above embodiment.

[0068] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0069] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the big data-based agricultural pest and disease prediction method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0070] In summary, this invention achieves the following: First, it unifies the spatiotemporal alignment of multi-source agricultural data to construct a sample sequence indexed by grid cells and time windows. Second, it extracts and clusters multi-scale spatiotemporal features of historical pest and disease transmission processes to form representative transmission pattern prototypes. Third, it compares the features of current observed samples with these transmission patterns at different scales to obtain feature information reflecting the current pest and disease transmission trends. Based on this, it utilizes these feature trends to perform multi-scale supervised learning on a graph neural network and adaptively calibrates the model parameters using a self-supervised learning process based on current observed samples. This enables the model to more accurately depict the transmission relationships and evolutionary processes of pests and diseases. Simultaneously, it transforms control events into transmission-inhibiting factors that influence the transmission process, dynamically intervening in the prediction results. This allows for the generation of pest and disease prediction results that better reflect actual agricultural production conditions, providing reliable decision support for pest and disease control, while fully considering the impact of control measures.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting agricultural pests and diseases based on big data, characterized in that: This includes acquiring multi-source data for the area to be predicted, including meteorological data, soil environmental data, crop phenological data, agricultural pest and disease control data, and pest and disease monitoring data. The multi-source data is spatiotemporally aligned to form a sample sequence indexed by grid cells and time windows of the region to be predicted; The observed cases in the database and the sample sequence are compared in terms of features to obtain the feature trend of the sample sequence. After learning the sample sequence using the graph neural network constrained by the aforementioned features, the tendency of pest and disease events to occur in each unit is predicted. Based on the prevention and control event flow, the predicted trend results are intervened to obtain agricultural pest and disease prediction results.

2. The method for predicting agricultural pests and diseases based on big data as described in claim 1, characterized in that: The spatiotemporal alignment includes resampling and synchronizing the multi-source data according to a preset time window, and mapping the spatial data to the grid cell; After removing outlier data, missing data is imputed, and sample vectors are generated for each grid cell in each time window to form the sample sequence.

3. The method for predicting agricultural pests and diseases based on big data as described in claim 2, characterized in that: The observation cases include obtaining spatiotemporal coding features at different observation scales by performing feature clustering on sample sequences corresponding to historical pest and disease transmission cases. The specific steps are as follows: When extracting features from sample sequences corresponding to historical pest and disease transmission cases, a spatiotemporal coupled pyramid coding network is used for spatiotemporal joint coding to obtain spatiotemporal coding features at each scale. In each case, the similarity of the spatiotemporal coding features at each scale is calculated, and the similarity is used as a distance metric. For all cases at the same scale, the spatiotemporal coding features are clustered, and the spatiotemporal coding feature that minimizes the sum of intra-cluster distances in each cluster is selected as the cluster center. Each cluster center is labeled with its corresponding case and scale, and its spatiotemporal coding features are marked to obtain each observation case. The spatiotemporal coupled pyramid coding network includes a spatiotemporal joint embedding module, which constructs a spatiotemporal joint input tensor from sample sequences indexed by grid cells and time windows, and performs three-dimensional convolutional embedding to obtain basic features; The spatiotemporal coupled feature extraction module includes at least one spatiotemporal residual block, which contains two three-dimensional convolutional layers and achieves joint spatiotemporal domain feature extraction through residual connections; based on the basic features, the spatiotemporal correlation in the process of pest and disease transmission is jointly modeled, and spatiotemporal coding features reflecting the evolutionary characteristics of the pest and disease transmission process are extracted. The spatiotemporal joint downsampling module uses 3D convolution with a preset stride to simultaneously downsample basic features in the temporal and spatial dimensions to form multi-scale pyramid features. At different levels of the spatiotemporal coupled feature extraction module, the resolution of the temporal and spatial dimensions is simultaneously reduced to form multi-level spatiotemporal feature representations with different observation scales, thereby constructing a spatiotemporal feature pyramid. The multi-scale spatiotemporal coding output module outputs spatiotemporal coding features at each level of the spatiotemporal feature pyramid, corresponding to the observation scale, as spatiotemporal coding features at different observation scales.

4. The method for predicting agricultural pests and diseases based on big data as described in claim 3, characterized in that: The feature comparison includes obtaining the original sample sequences corresponding to all cluster centers; The scale information in the cluster center labels is used as the scale for calculation between the corresponding original sample sequence and the currently observed sample sequence. For each cluster center at any scale i, calculate the similarity at label scale i: A random truncation length is extracted from the original sample sequence for use by the spatiotemporally coupled pyramid coding network to extract spatiotemporal coding features at scale i. The extracted spatiotemporal coding features are denoted as: F i,j ;F i,j Let i represent the spatiotemporal coding features of the j-th cluster center at scale i; Simultaneously, based on the aforementioned spatiotemporal coupled pyramid coding network, spatiotemporal coding features are extracted at scale i for the currently observed sample sequence. The extracted spatiotemporal coding features are denoted as: f i,0 ;f i,0 This represents the spatiotemporal coding features of the currently observed sample sequence at scale i; right F i,j and f i,0 Calculate the similarity S, and select the maximum value of S as the similarity between the currently observed sample sequence and the j-th cluster center at scale i by randomly truncating the sequence length. In this process, based on the feature comparison at each scale, the similarity with each cluster center is obtained, which is used to represent the feature trend of the current observed sample sequence at each scale.

5. The method for predicting agricultural pests and diseases based on big data as described in claim 4, characterized in that: The graph neural network utilizing the aforementioned characteristic trend constraints includes constructing a basic graph structure based on grid cells of the region to be predicted, wherein each grid cell serves as a node in the graph, and the spatial adjacency relationship between grid cells serves as a basic propagation relationship; The similarity between the current observed sample sequence and each cluster center is used as a propagation relationship influencing factor to constrain the propagation relationship between nodes in the basic graph structure; The influencing factors are used as the attention of the original sample sequences corresponding to the cluster centers. The original sample sequences corresponding to each cluster center are used as supervised learning samples. The graph neural network is subjected to multi-scale supervised learning according to the observation scale corresponding to the cluster centers, thereby enhancing the sensitivity of the graph neural network to different propagation features at different scales.

6. The method for predicting agricultural pests and diseases based on big data as described in claim 5, characterized in that: Learning the sample sequence is a self-supervised learning process based on the current observed sample sequence. The parameters of the graph neural network are adaptively calibrated by utilizing the spatiotemporal consistency constraints of the current observed sample sequence itself. After completing the self-supervised learning process, the graph neural network calibrated by self-supervision is used to predict the tendency of pest and disease events to occur in each grid cell, so as to obtain the probability of pest and disease events in each grid cell at each prediction time node.

7. The method for predicting agricultural pests and diseases based on big data as described in claim 6, characterized in that: The prevention and control event flow includes a sequence of prevention and control events performed by each grid cell; By mapping the corresponding pest and disease transmission inhibition factors to each prevention and control event, and applying them to a self-supervised calibrated graph neural network, the final agricultural pest and disease prediction results are generated.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the big data-based agricultural pest and disease prediction method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the big data-based agricultural pest and disease prediction method as described in any one of claims 1 to 7.

Images