A forest pest control method based on disease and insect pest monitoring

Abstract

The present invention discloses a forest pest and disease monitoring method, belonging to the field of forestry pest and disease control technology. The method includes: zoning the target forest according to vegetation type, collecting historical multi-source data and filling in gaps, analyzing time series and causal relationships to construct a time series causal diagram, simulating the spatiotemporal dynamics of pests and diseases in combination with current data, determining the prevention and control risk level and key inducing factors for each region, implementing prevention and control strategies accordingly, and then updating the causal diagram through reinforcement learning based on forest health response data. The method forms a closed loop from data processing to model construction, achieving precise prevention and control, improving efficiency and scientificity, and can dynamically optimize and adapt to changes in forest ecology to ensure effective prevention and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of forestry prevention and control, and in particular relates to a forest prevention and control method based on disease and insect pest monitoring. Background Art

[0002] In the field of forest pest control, with the improvement of ecological protection awareness and the growth of demand for forest resource management, existing technical means can no longer meet the goal of efficient and precise prevention and control.

[0003] At the forest management level, traditional forest pest control practices largely adopt an extensive management model. A common practice is to treat the entire forest as a single unit, implementing uniform pest control measures while completely ignoring the differences in vegetation types and distribution within the forest. For example, a mixed forest area contains both pine trees susceptible to pine wilt disease and poplar trees susceptible to infestation by poplar longhorn beetles. However, traditional control methods often involve uniformly spraying broad-spectrum insecticides across the entire forest. This method not only fails to precisely target the pest and disease needs of different vegetation types, but also severely damages the forest's beneficial organisms and ecological environment, resulting in high control costs and minimal effectiveness.

[0004] In terms of data collection and processing, existing technologies generally suffer from insufficient data utilization. Most forest pest and disease monitoring systems focus solely on collecting data on pest and disease occurrences, such as recording the location and scale of pest and disease outbreaks through manual inspections. However, they rarely simultaneously collect forest environmental data (including temperature, humidity, light intensity, soil fertility, etc.) and vegetation growth data (covering tree age, tree height, and diameter at breast height growth). However, in forest ecosystems, environmental changes can significantly affect vegetation growth, thereby altering vegetation's resistance to pests and diseases. These data are closely correlated. Due to the lack of multi-source data integration and comprehensive analysis, it is difficult to fully and deeply understand the inherent connection between forest ecosystems and the occurrence and development of pests and diseases, and it is impossible to provide sufficient data support and scientific basis for pest and disease prevention and control.

[0005] Existing technologies for studying the mechanisms of pest and disease outbreaks often rely on empirical judgment or simple statistical analysis, lacking in-depth understanding of the complex temporal and causal relationships between environmental change, vegetation growth, and pest and disease outbreaks. For example, observations have shown that the number of certain forest pests increases after rising temperatures in a certain region. However, existing technologies struggle to clearly define how rising temperatures affect physiological processes such as vegetation metabolism and nutrient absorption, ultimately leading to the outbreak of pests and diseases. The inability to establish a systematic logical relationship between these factors makes the formulation of prevention and control strategies lacking in precision and scientificity.

[0006] When it comes to pest and disease prediction and risk assessment, traditional methods primarily rely on simple trend forecasts based on historical pest and disease occurrences, failing to accurately simulate the dynamic temporal and spatial changes in pest and disease patterns. Due to the dynamic and complex nature of forest ecosystems, influenced by multiple factors such as climate change and human activities, the forest environment and vegetation growth conditions are constantly changing. Traditional prediction methods are unable to capture these real-time changes and accurately assess the control risks in different regions. This often results in missing optimal control opportunities or overinvesting resources in low-risk areas during pest and disease control.

[0007] When it comes to formulating and implementing prevention and control strategies, existing technologies often adopt a one-size-fits-all approach, implementing the same prevention and control measures regardless of regional risk or key pest-causing factors. This results in insufficient prevention and control efforts in high-risk areas, leading to the rapid spread of pests and diseases, while excessive prevention and control efforts in low-risk areas lead to a waste of human, material, and financial resources, seriously impacting the efficiency and effectiveness of forest pest and disease control. Summary of the Invention

[0008] In response to the above-mentioned deficiencies in the existing technology, the forest prevention and control method based on pest and disease monitoring provided by the present invention solves the problem in existing prevention and control methods that multi-source data is not fully utilized, and there is no in-depth analysis of the causal relationship between the occurrence and spread of pests and diseases and environmental changes and vegetation growth status, making it difficult to accurately achieve precise forest prevention and control through pest and disease monitoring.

[0009] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a forest pest control method based on disease and insect pest monitoring, comprising:

[0010] Divide the target forest area into several control zones based on vegetation type and distribution;

[0011] Collect historical multi-source data of the target forest area and fill in missing data to obtain complete data; the multi-source data includes forest environment data, vegetation growth data, and pest and disease occurrence data;

[0012] Conduct temporal and causal relationship analysis on complete data, and construct a temporal causal diagram that reflects the causal and temporal relationships among environmental changes, vegetation growth, and the occurrence of pests and diseases;

[0013] Based on the current forest environment data and vegetation growth data collected in each prevention and control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases along the causal path in the time series causal diagram is conducted to determine the prevention and control risk level and key inducing factors of each prevention and control area;

[0014] According to the prevention and control risk level and key inducing factors of each prevention and control area, the corresponding prevention and control strategy is implemented, and based on the forest health status response data after the implementation of the prevention and control strategy, the temporal causal diagram is updated through the reinforcement learning model.

[0015] Furthermore, missing data is filled in the collected historical multi-source data to obtain complete data, including:

[0016] Align various data samples in historical multi-source data in time series and organize them into independent two-dimensional data tables;

[0017] Processing each two-dimensional data table into a data table to be filled with a unified data structure and the same feature dimension and time length;

[0018] A column-level filling method based on a linear structure is used to fill missing data in the data table to be filled one by one to obtain complete data; in the missing data filling process, missing data is filled according to the left and / or right boundary values ​​of the missing data interval.

[0019] Furthermore, a temporal causal diagram is constructed to reflect the causal and temporal relationships among environmental changes, vegetation growth, and the occurrence of pests and diseases, including:

[0020] For complete environmental data, key variables that have a significant impact on the occurrence of pests and diseases in the target forest area are screened from forest environmental data and vegetation growth data; key variables that have a significant impact on vegetation growth are screened from pest and disease occurrence data;

[0021] Construct a VAR model based on the selected key variables and determine the lag order between the variables in the VAR model;

[0022] Constructing an initial graph structure based on the screened environmental variables, growth variables, and pest and disease variables; the initial graph structure includes a completely directed graph and / or a random graph;

[0023] The initial graph structure is tested for significance based on the VAR model coefficients. The effective edges of the initial graph structure are screened according to the significance of the VAR model coefficients. An improved Granger causality discovery method is introduced to construct the initial temporal causal graph.

[0024] Construct a scoring function that takes into account variable lags;

[0025] A sparse greedy iterative search is performed in the initial temporal causal graph, and the score values ​​after the operations performed during the search process are calculated. The operation that maximizes the score value is retained until the number of iterations is reached or the score value reaches a set threshold, thereby obtaining a temporal causal graph; the operations include adding edges, deleting edges, and reversing edge directions.

[0026] Furthermore, in the process of constructing the initial temporal causal graph by introducing the improved Granger causality discovery method, the causal relationships in the initial structure graph are screened by setting a dynamic lower bound threshold, and the screened causal relationships are verified by using a block perturbation method based on data characteristics.

[0027] Among them, the dynamic lower threshold for:

[0028]

[0029] Where, Indicates the front of the data Quantile, represents the mean score, represents the standard deviation, represents the coefficient for adjusting the standard deviation weight, A coefficient that represents the range of the control score.

[0030] Furthermore, considering the variable lag scoring function for:

[0031]

[0032] Where, Represents the initial temporal causal graph constructed, Represents a dataset containing the selected environmental variables, growth variables, and pest and disease variables. Indicates hysteresis Order pair variables The weight of Representing variables exist The value of the moment, Representing variables exist The set of parent nodes of the order lag, Representing variables exist The parent node set when represents the time-delay decay function, represents the regularization coefficient, represents the model parameters, represents the conditional probability function, Indicates the number of variables, represents the maximum delay order.

[0033] Furthermore, based on the current forest environment data and vegetation growth data collected in each prevention and control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases along the causal path in the time series causal diagram is conducted, including:

[0034] The current forest environment data and vegetation growth data collected in the prevention and control areas are integrated and standardized to form regional environmental characteristic vectors and growth characteristic vectors, which are used as the initial states of the environmental variables and growth variables in the time series causal diagram, respectively. The initial state of the pest and disease variable is set to 0.

[0035] In the current initial state, all directed causal paths from environmental variables and / or growth variables to pest and disease variables are extracted from the time series causal graph. The time lag order and VAR model coefficient of each edge in the directed causal path are recorded. The directed causal paths are ranked by significance according to the VAR model coefficient, and the path weight of each directed causal path is set according to the significance ranking result.

[0036] Based on the superposition effect of multi-path influence, a dynamic time-lag response model from environmental variables and / or growth variables to pest and disease variables is constructed; the dynamic time-lag response model is used to calculate the pest and disease response value of each pest and disease variable under the superposition of the mutual influence of all the directed causal paths associated with it;

[0037] Based on the constructed dynamic time-lag response model, all directed causal paths pointing to pest and disease variables are traversed, and the path weights of the directed causal paths are combined to simulate the spatiotemporal dynamics of pest and disease occurrence and spread, and predict the pest and disease response value;

[0038] The spatiotemporal dynamic simulation includes a time series simulation within the control area, and when the pest and disease response value of the time series simulation within the control area is greater than the pest and disease area diffusion threshold, the spatiotemporal dynamic simulation also includes a diffusion simulation between control areas.

[0039] Furthermore, the time series simulation within the control area includes:

[0040] Establishing a time-delay sequence of each variable in the directed causal path; the length of the time-delay sequence is the maximum time-delay order of the variable in the time-series causal graph;

[0041] The directed causal path whose total time delay order is greater than the preset simulation time step is taken as the trigger path;

[0042] In each simulation time step, for the same pest and disease variable, the dynamic time-delay response model is used to calculate and summarize the pest and disease response values ​​of all associated trigger paths, which are used as the current state of the pest and disease variable. The current states of the growth variable and environmental variable are updated synchronously, and the current pest and disease variable, growth variable, and environmental variable are saved in the corresponding time-delay sequence.

[0043] The time series simulation process is iterated according to the simulation time step until the preset simulation time step is reached or the current pest and disease response value in the time lag sequence exceeds the warning threshold.

[0044] Furthermore, the diffusion simulation between the control areas includes:

[0045] Construct a regional linkage network among the control areas within the target forest area;

[0046] Based on the regional association network, multi-factor weighted diffusion coefficients and time lag coefficients are set between each prevention and control area, and a diffusion simulation framework based on cellular automata is constructed to simulate the diffusion between prevention and control areas.

[0047] In the diffusion simulation framework, the state transition rule is set as follows: the node state of the control area is determined by superimposing the pest response value obtained by time series simulation in the control area and the pest response value obtained by diffusion from the associated control area;

[0048] When the superimposed pest and disease response value of the prevention and control area is greater than the set pest and disease area diffusion threshold, the prevention and control area will be regarded as a new diffusion source.

[0049] Furthermore, the risk level and key inducing factors of each prevention and control area are determined, including:

[0050] According to the control threshold range of the pest and disease response value, the control risk level is determined, and the environmental variables and / or growth variables in all directed causal paths where the pest and disease variables are located are traced back, and the environmental variables and / or growth variables that appear the most times are taken as key inducing factors.

[0051] Furthermore, based on the forest health status response data after the implementation of the prevention and control strategy, the reinforcement learning model updates the constructed temporal causal graph, including:

[0052] Collect data on forest health status after implementing prevention and control strategies, including pest and disease indicator data, vegetation health indicator data, and forest environment indicator data;

[0053] Mapping multi-dimensional forest health status response data into the state vector of the reinforcement learning model to construct the state space;

[0054] Based on an update operation on the temporal causal graph, an action space of the reinforcement learning model is constructed; the update operation includes structural adjustment and parameter optimization of the temporal causal graph, the structural adjustment includes deleting and / or adding causal edges, and the parameter optimization includes adjusting the causal edge weights and lag orders;

[0055] Constructing a reward function for the reinforcement learning model, wherein the reward items in the reward function include prediction error reward, causal logic consistency reward, and prevention and control strategy explainability reward;

[0056] Based on the constructed state space, action space and reward function, a reinforcement learning model is run to obtain an updated temporal causal graph; the reinforcement learning model adopts a network structure that combines a graph neural network and a deep Q network.

[0057] The beneficial effects of the present invention are:

[0058] (1) The present invention visualizes the complex abstract relationship between environmental changes, vegetation growth and the occurrence of pests and diseases in forest ecosystems by constructing a temporal causal diagram, and intuitively displays the causal chain and time sequence among the three. Based on the temporal causal diagram, the occurrence and spread of pests and diseases can be dynamically simulated in time and space along the causal path according to the current environmental data and vegetation growth data. This enables forest prevention and control work to be transformed from passive response to active defense, and targeted measures can be taken before pests and diseases break out on a large scale.

[0059] (2) In the method of the present invention, when performing spatiotemporal simulation of the occurrence and spread of pests and diseases based on the time series causal diagram, the time series simulation is first performed within the control area, and then the inter-regional spread simulation is performed after the response value exceeds the threshold. The design logic of this phased simulation strategy reflects the hierarchical path of the endogenous law of the occurrence of pests and diseases, which is "breeding within the region → spreading within the region → spreading across regions". Through this hierarchical simulation path, when achieving forest control, local lesions can be controlled first, and then cross-regional spread can be blocked.

[0060] (3) The phased spatiotemporal dynamic simulation strategy in the present invention starts from within the control area and performs inter-regional diffusion simulation only when the pest response value exceeds the threshold. This avoids the computing resource explosion caused by "full-scale simultaneous simulation" and allows computing resources to be concentrated in high-risk scenarios.

[0061] (4) In the method of the present invention, after determining the control risk level of each control area with the help of a time-series causal diagram, resources can be reasonably allocated according to the level of risk. For areas with high risk levels and complex causal relationships, more monitoring equipment, professional control personnel, and control materials are invested; while for areas with lower risks, resource investment is appropriately reduced and routine monitoring and prevention are carried out. This resource optimization allocation method can not only ensure the effective protection of forest resources, but also avoid resource waste, thereby improving the efficiency and economic benefits of overall control work.

[0062] (5) In the method of the present invention, precise policies are implemented based on the determined risk levels and key factors, resources are intelligently allocated, and combined strategies are used to enhance synergy, improve efficiency, and avoid waste of resources; at the same time, costs are reduced through hierarchical prevention and control, and benefits are guaranteed through quantitative evaluation, thus achieving precise, efficient, and economical pest and disease control.

[0063] (6) The method of the present invention is based on the forest health response data after the implementation of the prevention and control strategy, and uses the reinforcement learning model to update the temporal causal graph, which can dynamically capture the complex causal relationship between pests and diseases and the environment and tree growth in the forest ecosystem; according to the actual data feedback, the structure and parameters of the temporal causal graph are automatically adjusted to make the causal graph more in line with the real scene; compared with the traditional static causal analysis, this method has the ability of adaptive learning and can adapt to the dynamic changes of the forest ecosystem in real time, providing more accurate and timely decision-making basis for pest and disease prevention and control, and effectively improving the scientificity and effectiveness of the prevention and control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of the forest pest control method based on disease and insect pest monitoring provided by the present invention. DETAILED DESCRIPTION

[0065] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0066] The embodiment of the present invention provides a forest pest control method based on disease and insect pest monitoring.

[0067] See also Figure 1 , the figure shows that the method includes steps S100 to S500, including:

[0068] Step S100: Divide the target forest area into several control areas according to vegetation type and distribution;

[0069] Step S200: Collect historical multi-source data of the target forest area and fill in missing data to obtain complete data; the multi-source data includes forest environment data, vegetation growth data, and pest and disease occurrence data;

[0070] Step S300: Analyze the temporal relationship and causal relationship of the complete data, and construct a temporal causal graph reflecting the causal relationship and temporal relationship between environmental changes, vegetation growth and the occurrence of pests and diseases;

[0071] Step S400: Based on the current forest environment data and vegetation growth data collected from each control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases is performed along the causal path in the time series causal graph to determine the control risk level and key inducing factors of each control area;

[0072] Step S500: Execute corresponding prevention and control strategies according to the prevention and control risk level and key inducing factors of each prevention and control area, and update the temporal causal diagram through the reinforcement learning model based on the forest health status response data after the implementation of the prevention and control strategies.

[0073] In step S100 of the present embodiment, it should be noted that the vegetation types in the forest are interrelated and influence each other with other ecological factors. There are differences in the soil conditions, climate conditions, biodiversity, etc. in the areas where different vegetation types are located, and these factors will affect the occurrence and development of pests and diseases. Therefore, in the present invention, the control areas are first divided based on the vegetation types and distribution, which can comprehensively consider these ecological factors and adapt the control measures to the characteristics of the ecosystem. For example, the forest area near the wetland is suitable for the survival of certain moisture-loving pests and diseases due to the high humidity. After the area is divided, an eco-friendly control plan can be formulated in combination with the ecological characteristics of the wetland, such as by improving the regional ventilation conditions, adjusting the vegetation structure, etc., to destroy the living environment of the pests and diseases and achieve ecological control.

[0074] In a specific embodiment of the present invention, a method for dividing a target forest area into several control areas based on vegetation type and distribution is provided, including:

[0075] Step S101: collecting remote sensing data and ground survey data of the target forest area, performing pre-processing, and establishing a correlation database;

[0076] Specifically, for remote sensing data: Image data of target forest areas is acquired using high-resolution satellite remote sensing imagery (such as Landsat and Sentinel satellite data) and drone aerial photography. For ground survey data, professionals conduct field surveys to record information such as vegetation type, quantity, height, and diameter at breast height (DBH), while also annotating GPS coordinates. Satellite remote sensing data provides large-scale, periodic observations, suitable for macro-scale vegetation information acquisition. Drone aerial photography allows for high-resolution capture of key areas, providing more detailed vegetation information, including canopy structure and texture. Ground survey data can verify and supplement the interpretation of remote sensing data, improving the accuracy of vegetation type identification.

[0077] Furthermore, the preprocessing of the collected data includes performing radiation correction, geometric correction and other processing on the collected remote sensing images to eliminate errors caused by sensor characteristics, atmospheric conditions and other factors; spatially matching the ground survey data with the remote sensing images to establish an associated database.

[0078] Step S102: Identify and classify vegetation types in the target forest area based on the constructed association database;

[0079] Specifically, based on the unique spectral reflectance characteristics of different vegetation types, multispectral bands (such as red, near-infrared, and green bands) of remote sensing images are used to classify vegetation into different types, such as coniferous forests, broad-leaved forests, and mixed forests, through supervised or unsupervised classification methods (such as the K-means clustering algorithm). For example, supervised classification methods include maximum likelihood classification and support vector machine classification, while unsupervised classification methods include the K-means clustering algorithm.

[0080] Furthermore, on the basis of identifying vegetation types based on spectral characteristics, the texture information (such as roughness, contrast) and spatial distribution patterns (such as patch size and shape) of vegetation are used to further distinguish different vegetation types; for example, an object-oriented classification method can be used to segment the image into objects with similar characteristics, and then classify them according to the spectral, texture and spatial properties of the objects.

[0081] Step S103: performing regional division based on the identified vegetation type and distribution;

[0082] Specifically, the DBSCAN density clustering algorithm is used to cluster the target forest area according to the distribution of vegetation types; on the basis of clustering, the division results are adjusted by comprehensively considering ecological and geographical factors such as topography (such as mountains and rivers) and climatic conditions (such as precipitation and temperature gradient); the boundaries of each prevention and control area are determined and digitized to generate a prevention and control area distribution map. At the same time, each prevention and control area is numbered and attributed, and its main vegetation type, area, geographical location and other information are recorded.

[0083] In this embodiment, the DBSCAN density clustering algorithm is selected to aggregate spatially adjacent areas of similar vegetation types into one prevention and control area based on the spatial density and aggregation degree of vegetation types.

[0084] In this embodiment, the clustering results are adjusted by comprehensively considering the topography and climatic conditions so that the distribution of the control areas obtained by division is more consistent with the growth of vegetation and the spread of pests and diseases. For example, mountains and rivers may become natural barriers to the spread of pests and diseases, and they can be used as boundaries for dividing control areas. Regions with large differences in climatic conditions may also have different patterns and characteristics of pest and disease occurrence and should be divided into different control areas.

[0085] In step S200 of this embodiment, the integrity of multi-source data directly impacts the accuracy of the subsequent construction of the temporal causal diagram and the reliability of the results obtained from the spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases. For example, if three consecutive years of soil moisture data for a region are missing, the model may misjudge the drought-induced mechanism of pine wilt disease. Forest environmental data (such as temperature, humidity, and soil pH), vegetation growth data (diameter at breast height and tree height growth rates), and pest and disease data (onset time and severity of damage) must form a closed loop.

[0086] In step S200 of this embodiment, missing data is filled in the collected historical multi-source data to obtain complete data, including:

[0087] Step S201: Time-series alignment of various data samples in historical multi-source data, and arranging them into independent two-dimensional data tables;

[0088] Specifically, among the multi-source data for forest prevention and control, the collection time and frequency of forest environment data, vegetation growth data, and pest and disease occurrence data may differ. To facilitate subsequent analysis, these data need to be time-series aligned.

[0089] Based on the timestamp, resampling technology is used to organize various types of data for data with different collection frequencies. For example, meteorological data such as temperature and humidity in forest environmental data are usually collected hourly, while vegetation growth data (such as tree diameter at breast height measurement) may be collected monthly. For meteorological data, if it is to be aligned with vegetation growth data collected monthly, the hourly meteorological data can be aggregated monthly to calculate monthly averages, maximum values, minimum values ​​and other statistics. For pest and disease occurrence data with irregular collection times, their timestamps are aligned to a unified time series through linear interpolation or nearest neighbor interpolation.

[0090] The aligned data are organized into independent two-dimensional data tables. Taking forest environment data as an example, the columns of the two-dimensional table can include time (such as year-month), temperature, humidity, wind speed, precipitation and other fields, and each row represents the observation data of a time node. The vegetation growth data table can include time, tree species, diameter at breast height, tree height, crown width and other fields. The pest and disease occurrence data table includes time, pest and disease type, occurrence area, degree of damage and other fields.

[0091] Step S202: Processing each two-dimensional data table into a data table to be filled with a unified data structure and the same characteristic dimension and time length;

[0092] Specifically, to uniformly process different types of data, each two-dimensional data table needs to be processed into the same data structure. This includes standardizing data types (e.g., unifying all numeric data into floating-point data) and field naming conventions (e.g., naming the "temperature" field "Temperature" in all tables).

[0093] Furthermore, by adding dummy variables or truncating and padding data, each data table to be filled has the same feature dimensions and time length. For example, if a forest environment data table lacks the feature "wind speed," a column for "wind speed" can be added, and the missing values ​​can be temporarily marked with a special symbol (such as -999). If a pest and disease outbreak data table has a short time span, data padding can be used to fill in the missing time period in the time series, and the missing values ​​are also marked with a special symbol. The resulting data table to be filled will have each row representing a time node and each column representing a feature, and all tables will have the same number of columns (feature dimensions) and rows (time length).

[0094] Step S203: Using a column-level filling method based on a linear structure to fill missing data in the data table to be filled one by one to obtain complete data; in the missing data filling process, the missing data is filled according to the left and / or right boundary values ​​of the missing data interval;

[0095] Specifically, the column-level imputation method based on a linear structure uses the existing data in the column where the missing data is located to estimate the missing value through a linear relationship based on the left and / or right boundary values ​​of the missing data interval. This method can be divided into the following two cases:

[0096] Single Boundary Value Filling: When missing data is at the beginning or end of a data series, you can fill it in based on the boundary value on one side only. If the missing data is at the beginning, the first non-missing value on the right is used as the fill value; if the missing data is at the end, the last non-missing value on the left is used as the fill value.

[0097] Double boundary value linear interpolation: When the missing data is located in the middle of the data sequence, the missing value is calculated using the linear interpolation formula based on the left and right boundary values ​​of the missing data interval.

[0098] For example, assume that the left boundary value of the missing data interval is , the corresponding time is , the right boundary value is , the corresponding time is The time required to fill is , whose corresponding missing value is , then the linear interpolation formula is:

[0099]

[0100] In this embodiment, the above missing data filling method is used to systematically process historical multi-source data, and ultimately obtain complete and unified format data, providing a reliable data basis for the subsequent construction of a time series causal diagram and pest and disease monitoring and prevention analysis.

[0101] In step S300 of this embodiment, by constructing a temporal causal diagram that reflects the causal and temporal relationships among environmental changes, vegetation generation, and the occurrence of pests and diseases, the intricate abstract relationship among environmental changes, vegetation generation, and the occurrence of pests and diseases in the forest ecosystem is visualized, and the causal chain and temporal sequence among the three are intuitively displayed, so as to facilitate the subsequent spatiotemporal simulation of the occurrence and spread of pests and diseases, trace the key inducing factors of the occurrence of pests and diseases, and adopt targeted prevention and control strategies.

[0102] In step S300 of this embodiment, a temporal causal graph is constructed to reflect the causal relationship and temporal relationship between environmental changes, vegetation growth, and the occurrence of pests and diseases, including:

[0103] Step S301: For the complete environmental data, the key variables that have a significant impact on the occurrence of pests and diseases in the target forest area are screened from the forest environmental data and the vegetation growth data; and the key variables that have a significant impact on vegetation growth are screened from the pest and disease occurrence data;

[0104] In a specific embodiment of this embodiment, the method for screening key variables in the forest environment data, vegetation growth data, and pest and disease occurrence data in the above process is:

[0105] Forest environmental data (including temperature, humidity, light, soil nutrients, etc.), vegetation growth data (such as tree height, diameter at breast height, leaf area, biomass, etc.) and pest and disease occurrence data (such as pest and disease type, occurrence time, occurrence area, degree of damage, etc.) were standardized and converted into a standard normal distribution with a mean of 0 and a standard deviation of 1 to facilitate subsequent analysis.

[0106] Correlation analysis, random forest feature importance, and Lasso regression were used to screen variables related to pest and disease occurrence in forest environmental data and vegetation growth data. Variables with an absolute correlation coefficient greater than 0.5, a random forest feature importance ranking in the top 30%, and a non-zero Lasso regression coefficient were retained as candidate key variables. Similarly, variables related to vegetation growth in the pest and disease occurrence data were preliminarily screened.

[0107] Among them, correlation analysis is performed by calculating the Pearson correlation coefficient between variables to screen variables. The calculation formula of the Pearson correlation coefficient is:

[0108]

[0109] Where, and Represents the first observations, and represent the means of the two variables, represents the number of samples, when When there is a strong linear correlation between variables, for example, when analyzing forest environmental data, if the Pearson correlation coefficient between temperature and the number of occurrences of a certain disease or insect pest is found to be 0.8, then temperature may be an important variable affecting the occurrence of the disease or insect pest.

[0110] The candidate key variables initially screened were further screened using stepwise regression analysis or variance analysis. Stepwise regression was used to identify key variables that significantly affected pest and disease occurrence or vegetation growth. Variance analysis was used to verify the significance of differences between these variables across different groups. Ultimately, key variables in forest environmental data and vegetation growth data that significantly affected pest and disease occurrence in the target forest area were obtained, as well as key variables in pest and disease occurrence data that significantly affected vegetation growth.

[0111] Step S302: constructing a VAR model based on the selected key variables and determining the lag order between the variables in the VAR model;

[0112] Among them, the VAR model (Vector Autoregression Model) is a quantitative model used to analyze the dynamic interactive relationship between multiple variables. It regards each variable in the system as being determined by the lag order of itself and other variables. By capturing the lag effect between variables, it describes the dynamic association mechanism between variables.

[0113] Specifically, the key variables obtained by screening are organized into a unified data set according to time series, ensuring that the time span and sampling frequency of each key variable are consistent, and then the VAR model is constructed:

[0114]

[0115] Where, represents an endogenous variable vector, which includes the key variables that have been screened out and have a significant impact on the occurrence of pests and diseases and vegetation growth (such as temperature and humidity in forest environment variables, tree height and diameter at breast height in vegetation growth variables, and diseased area and insect population density in pest and disease occurrence variables); represents the parameter matrix, characterizing the influence coefficient of the lag terms of different variables, k represents the lag order, represents a random disturbance term, which follows a multivariate normal distribution with mean 0 and covariance matrix Σ.

[0116] Parameter estimation of the VAR model constructed above: Ordinary least squares (OLS) is used to solve the parameter matrix of the VAR model parameters by minimizing the residual sum of squares For the VAR model, multiple linear regression can be performed on each endogenous variable, that is, OLS estimation can be used for each equation in the model. For example, if Containing two variables, temperature and diseased area, we can establish a regression equation with temperature and diseased area as dependent variables and the lagged terms of other variables as independent variables, and then obtain the parameter estimates of the VAR model.

[0117] Determine the lag order between variables in the VAR model: Set a search range for the lag order (e.g., from 1 to 10), calculate the AIC, BIC, and HQIC values ​​of the VAR model under different lag orders, and select the lag order that minimizes these criteria as the optimal order. AIC, BIC, and HQIC are the Akaike Information Criterion, Bayesian Information Criterion, and Hannan-Quinn Information Criterion, respectively, and their calculation formulas are:

[0118]

[0119]

[0120]

[0121] Where, represents the number of parameters to be estimated in the model, represents the number of samples, Indicates the number of internal variables, Represents the estimated value of the covariance matrix; in the above formula, the smaller the AIC value, the better the trade-off between goodness of fit and complexity of the simulation. BIC, based on AIC, penalizes the complexity of the model more severely, and is suitable for avoiding overfitting of the model when the sample size is large.

[0122] Step S303: constructing an initial graph structure based on the screened environmental variables, growth variables, and pest and disease variables; the initial graph structure includes a completely directed graph and / or a random graph;

[0123] Specifically, a completely directed graph means that in the graph structure, there are bidirectional directed edges between any two variable nodes. When constructing a completely directed graph based on environmental variables, growth variables, and pest and disease variables, it is necessary to first assume that there may be a causal relationship between all variables, and then determine the actual causal edges through screening and testing.

[0124] A random graph randomly generates edges between nodes with a certain probability. By setting the probability parameters for edge generation, the sparsity of the graph structure can be controlled. This initial structure is more suitable when there is a certain fuzzy understanding of the relationship between variables. When we neither believe that all variables may be related (such as a completely directed graph) nor want to start from a completely unconnected state (such as an empty graph), it can balance the exploration scope and structural complexity to a certain extent.

[0125] Step S304: Perform a significance test on the initial graph structure based on the VAR model coefficients, screen the effective edges of the initial graph structure according to the significance of the VAR model coefficients, and introduce an improved Granger causality discovery method to construct an initial temporal causal graph.

[0126] In this embodiment, the VAR model coefficients represent the influence of the corresponding variable lag term on the current endogenous variable. The significance test of these coefficients can determine the validity of the relationship between variables. In this embodiment, Inspection is the main inspection method. The statistical calculation formula is:

[0127]

[0128] Where, represents the coefficient estimate to be tested in the VAR model, represents the standard error of the coefficient estimate, The larger the absolute value of the statistic, the more significant the coefficient is, and the more likely the relationship between the corresponding variables is to exist.

[0129] Specifically, in this embodiment, the method for screening valid edges based on the above t-statistic is:

[0130] Extract the coefficient matrix of each equation from the constructed VAR model and clarify the variable relationship corresponding to each coefficient. For example, in a VAR model containing four variables: temperature, humidity, tree height, and area of ​​pest and disease occurrence, the elements in the coefficient matrix may represent the influence coefficient of the first-order lag of temperature on the current value of humidity. For each coefficient in the coefficient matrix, according to Test formula calculation Statistics, and by looking up Distribution table or use statistical software (such as Python's scipy library, R language) to obtain the corresponding Value; In the initial graph structure, if the VAR model coefficient corresponding to an edge is not significant after testing (i.e. Value greater than ), then delete the edge; if the coefficient is significant ( Value is less than or equal to ), then retain the edge. For example, in a completely directed graph structure, if the VAR model coefficient corresponding to the edge of temperature pointing to the area of ​​pests and diseases If the value is greater than 0.05, it means that the effect of temperature on the area of ​​pests and diseases is not significant, and this edge is deleted from the graph structure; after completing a round of screening for all edges, the significance level can be appropriately adjusted according to actual needs. Or use other inspection methods to conduct re-inspection and screening to further optimize the graph structure.

[0131] In this embodiment, in the process of constructing the initial temporal causal graph by introducing the improved Granger causality discovery method, the causal relationship in the initial structure graph is screened by setting a dynamic lower bound threshold, and the screened causal relationship is verified by using a block perturbation method based on data characteristics;

[0132] Among them, the dynamic lower threshold for:

[0133]

[0134] Where, Indicates the front of the data Quantile, represents the mean score, represents the standard deviation, represents the coefficient for adjusting the standard deviation weight, A coefficient that represents the range of the control score.

[0135] Specifically, the method for determining the dynamic lower bound threshold is:

[0136] Target variable Attention score sequence , adjustment coefficient (control quantile range) and (Adjust standard deviation weight);

[0137] calculate Quantile , filter high score intervals: for example, The top 20% of high-scoring nodes are retained to ensure that weak signals are not missed and to determine the lower bound of the quantile. , that is, the score is not less than The minimum index position of .

[0138] Mean-standard deviation method: calculate the mean score and standard deviation , dynamically expand the threshold to ; Determine the lower bound of the standard deviation , that is, the score exceeds The minimum index position of .

[0139] Threshold fusion: take and and constrain its minimum value to 1 to avoid the risk of an empty set.

[0140] The above-mentioned dynamic setting of the lower bound threshold replaces the manually set fixed threshold, automatically adjusts according to the data distribution, and adapts to the characteristics of different data sets; the quantile and mean-standard deviation are jointly optimized, taking into account both position and dispersion information.

[0141] In the causal verification stage, a block perturbation optimization method based on data characteristics is proposed. The core idea is to preserve the macro structure of the data through block perturbation, while destroying local pseudo-correlations, improving the accuracy and interpretability of causal verification, and then constructing the initial time series causal graph.

[0142] Step S305: constructing a scoring function that takes into account variable time lag;

[0143] Specifically, the scoring function considering variable lags is for:

[0144]

[0145] Where, Used to evaluate a given initial temporal causal graph structure With the dataset The higher the score (or the lower it is, depending on the specific definition), the better the causal diagram structure can explain the data. represents a data set containing the screened environmental variables, growth variables, and pest and disease variables; Indicates hysteresis Order pair variables The weight of When the other variables have an influence on the variable the relative importance of the impact; Representing variables exist The value of the moment; Representing variables exist The parent node set of the order lag, that is, the variable The set of other variables that have causal effects (lag k) at time t; Representing variables exist The parent node set when ; represents the time-delay attenuation function, which is about the time-delay order The function is to adjust the penalty intensity of connections with different time delay orders. Generally, as k increases, The value of will decrease; Represents the regularization coefficient, which plays a role in balancing the fitting term and the penalty term; represents the model parameters, represents the conditional probability function, Indicates the number of variables, represents the maximum delay order.

[0146] Step S306: Perform a sparse greedy iterative search in the initial temporal causal graph, calculate the score value after the operations are performed during the search process, retain the operation that maximizes the score value, and obtain the temporal causal graph until the number of iterations is reached or the score value reaches a set threshold. The operations include adding edges, deleting edges, and reversing edge directions.

[0147] Specifically, sparse greedy iterative search uses the core idea of ​​"promoting global optimization from local optimization." Each iteration performs only one local operation on the current graph structure (adding an edge, removing an edge, or reversing the edge direction), selecting and retaining the operation that maximizes the scoring function, and gradually optimizing the graph structure. This strategy ensures search efficiency while effectively avoiding being trapped in a complex global search space, and is suitable for constructing time-series causal graphs for high-dimensional variables.

[0148] After each operation, the score for the entire graph structure must be recalculated. The scoring function constructed in this embodiment comprehensively considers factors such as the strength of causal relationships between variables, model complexity, and variable lag. By comparing the scores before and after the operation, we can determine whether the operation has optimized the graph structure.

[0149] When setting the iteration termination condition, you can use either of the following two conditions:

[0150] When the number of iterations reaches this upper limit, the search process stops, regardless of whether the score can be further improved. This approach effectively controls computing resource consumption and search time, and is suitable for scenarios requiring high computational efficiency. For example, when processing large-scale forest monitoring data, set N to 300. After 300 iterations, the optimal graph structure is output as a temporal causal diagram.

[0151] A scoring threshold is set. When the score after a certain iteration reaches or exceeds this threshold, the graph structure is considered sufficiently optimal and the search stops. The scoring threshold should be determined based on the specific problem and data characteristics, and can be determined through preliminary experiments or domain knowledge. For example, after multiple preliminary experiments, it was found that when the score reaches 85 (assuming the score range is 0-100), the graph structure accurately depicts the causal relationships between variables. Therefore, 85 is used as the stopping threshold. Once the score after a certain iteration is greater than or equal to 85, the current real-time causal graph is output.

[0152] In step S400 of this embodiment, based on the current forest environment data and vegetation growth data collected in each prevention and control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases is performed along the causal path in the temporal causal graph, including:

[0153] Step S401: Integrate the current forest environment data and vegetation growth data collected in the prevention and control area and perform normalization processing to form regional environmental feature vectors and growth feature vectors, which are used as the initial states of the environmental variables and growth variables in the time series causal graph, respectively, and the initial state of the pest and disease variable is set to 0;

[0154] Specifically, the current environmental data (such as temperature, humidity, precipitation, soil nutrients, etc.) and tree growth data (such as diameter at breast height, tree height, leaf area index, chlorophyll content, etc.) of each prevention and control area are collected to form a regional feature vector; variables of different dimensions are normalized to avoid the impact of numerical differences on simulation accuracy.

[0155] Step S402: In the current initial state, all directed causal paths from environmental variables and / or growth variables to pest and disease variables are extracted from the temporal causal graph. The time lag order and VAR model coefficient of each edge in the directed causal path are recorded. The directed causal paths are ranked by significance according to the VAR model coefficient, and the path weight of each directed causal path is set based on the significance ranking result.

[0156] For example, the extracted directed causal path is temperature → tree resistance → pest and disease incidence;

[0157] Step S403: Based on the multi-path influence superposition effect, a dynamic time-lag response model from environmental variables and / or growth variables to pest and disease variables is constructed; the dynamic time-lag response model is used to calculate the pest and disease response value of each pest and disease variable under the superposition of the mutual influence of all the directed causal paths associated with it;

[0158] Specifically, for the same pest and disease variable, when there are multiple directed causal paths (e.g., both temperature and humidity affect pest and disease variable B), the weighted superposition method is used to construct a dynamic time-lag response model:

[0159]

[0160] Where, represents the comprehensive impact on the pest and disease variable B after d time steps, represents the response function of the lth directed causal path, represents the model parameters, represents the variable on the lth causal directed path, represents the weight of the lth directed causal path, and L represents the number of causal paths.

[0161] Step S404: Based on the constructed dynamic time-lag response model, all directed causal paths pointing to the pest and disease variables are traversed, and the path weights of the directed causal paths are combined to perform spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases, and predict the pest and disease response value;

[0162] Specifically, starting from the current time t, the state of the next T time steps is iteratively calculated according to the following steps:

[0163] For each pest variable , traverse all pointing , calculate the directed causal paths of each path in The impact value after time steps;

[0164] According to the weighted superposition of the influence values, we can get The predicted value of

[0165] Update vegetation growth variables (e.g., vegetation resistance declines due to pests and diseases).

[0166] Follow the above steps to perform time-step simulation step by step to obtain the pest and disease response value; at the same time, use a queue to store the historical state of each variable (the length is the maximum time lag order) to ensure that the required lag data can be accessed during the simulation.

[0167] In this embodiment, the spatiotemporal dynamic simulation includes a time series simulation within the control area, and when the pest and disease response value of the time series simulation within the control area is greater than the pest and disease area diffusion threshold, the spatiotemporal dynamic simulation also includes a diffusion simulation between control areas.

[0168] Specifically, in the process of implementing forest control, a time-series simulation is first conducted within the control area. Then, after the pest response value exceeds a threshold, an inter-regional diffusion simulation is conducted. Pest and disease occurrence typically follows a hierarchical path: "intra-regional initiation → regional spread → inter-regional transmission." This phased, spatiotemporal dynamic simulation strategy reflects the endogenous patterns of pest and disease occurrence. For example, pine wilt disease first colonizes within a single pine tree and spreads through the tree's vascular system (microscopic diffusion within the region). Only when the number of infected trees exceeds a certain proportion (e.g., a regional infection rate >10%) does it spread to surrounding areas through the flight activities of vector insects (such as Monochamus alternatus) (macroscopic diffusion between regions).

[0169] In this example, time series simulation within a control area simulates the dynamic development of pests and diseases over time within a single control area, based on a causal graph and current environmental growth data. This process achieves prediction from the "current state" to the "future trajectory" through time-delayed response calculations and iterative state updates along the causal path. Specifically, it includes:

[0170] Step S404-1: Establish a time-delay sequence for each variable in the directed causal path; the length of the time-delay sequence is the maximum time-delay order of the variable in the time-series causal graph;

[0171] For example, the maximum time lag of temperature on pests and diseases is 3 periods, so the temperature queue needs to save the data of the current and previous 3 periods;

[0172] Step S404-2: taking the directed causal path whose total time delay order is greater than the preset simulation time step as the triggering path;

[0173] Specifically, for the current simulation time step, all directed causal paths are traversed to check whether the time delay triggering conditions are met:

[0174] If the total time lag order D of the path is greater than the preset simulation time step T, the response calculation of the directed causal path is triggered (for example, if the total time lag order of the path is 3, its impact calculation begins at t=3); at the same time, the required historical data is retrieved from the time lag sequence. For example, the path "temperature → tree resistance → pests and diseases" requires the retrieval of temperature data from the TD period at time T.

[0175] For example, trigger path 1: temperature Tree resistance Incidence of pests and diseases; where the lag orders are 2 and 1 respectively; trigger path 2: humidity Pest and disease density; the lag order is 3.

[0176] Step S404-3: Within each simulation time step, for the same pest and disease variable, use the dynamic time-delay response model to calculate and summarize the pest and disease response values ​​of all associated trigger paths, use this as the current state of the pest and disease variable, synchronously update the current states of the growth variable and environmental variable, and save the current pest and disease variable, growth variable, and environmental variable to the corresponding time-delay sequence;

[0177] Step S404-4: iterating the time series simulation process according to the simulation time step until a preset simulation time step is reached or the current pest response value in the time lag sequence exceeds the warning threshold;

[0178] Specifically, the simulation process ends when the iteration reaches the preset simulation time step, or when the pest and disease response value exceeds the warning threshold, the warning is triggered and the simulation process is terminated early; further, in each simulation time step, the current state of each variable (such as temperature, vegetation resistance or pest and disease density) and the real-time impact contribution of the key trigger path (such as the current contribution of the temperature path is 35%, and the contribution of the humidity path is 20%) are output to generate a time series curve (such as the changing trend of pest and disease density over time).

[0179] In this embodiment, based on the time series simulation within the control area, when the pest and disease response value of the time series simulation within the control area is greater than the pest and disease area diffusion threshold, the diffusion simulation between the control areas is performed. In this embodiment, the diffusion simulation is performed in combination with spatial correlation, environmental similarity and time lag effect, forming a linkage with the previous dynamic time lag response model.

[0180] In this embodiment, the diffusion simulation between control areas includes:

[0181] Step S404-5: constructing a regional association network among the control areas within the target forest area;

[0182] Specifically, in the regional association network:

[0183] Each control area is considered a network node, and the node attributes include regional environmental characteristic vectors (such as temperature, humidity, vegetation type, etc.), growth characteristic vectors (such as tree density, tree age structure, etc.), and pest and disease response values ​​(based on time series simulation results within the area);

[0184] Connecting edges represent potential diffusion pathways between regions, and their weights are determined by the following factors: geographic distance (adjacent regions are directly connected, and the strength of association between non-adjacent regions is calculated using a distance decay coefficient (e.g., Gaussian function)), transportation network (transportation channels such as roads and railways enhance the probability of diffusion), and ecological corridors (natural channels such as rivers and mountains promote or block the migration of pests and diseases).

[0185] Furthermore, chordal similarity or Euclidean distance is used to quantify the differences in environmental feature vectors between any two regions and generate a similarity matrix. The higher the similarity, the stronger the adaptability of pests and diseases between the control areas and the higher the probability of spread.

[0186] Step S404-6: Based on the regional association network, set the multi-factor weighted diffusion coefficient and time lag coefficient between each control area, build a diffusion simulation framework based on cellular automation, and perform diffusion simulation between control areas;

[0187] Specifically, the diffusion coefficient from region p to region q is Defined as:

[0188]

[0189] Where, represents the regional geographical distance from region p to region q, represents the distance decay effect, represents the traffic impact factor, Indicates the environmental similarity, 、 and They are 、 and The weight coefficient of .

[0190] Combined with the time delay order in the previous dynamic time delay response model , the time lag coefficient of inter-regional diffusion is defined as:

[0191] When the pest and disease status of region p reaches the threshold at time t, region q will be Affected, Related to the length of the diffusion path and the speed of vector transmission (e.g., the transmission delay of insects is related to the flight cycle).

[0192] Step S404-7: In the diffusion simulation framework, a state transition rule is set as follows: the node state of the control area is determined by superimposing the pest response value obtained through time series simulation within the control area and the pest response value obtained by diffusion from the associated control area; when the superimposed pest response value of the control area is greater than a set pest area diffusion threshold, the control area is used as a new diffusion source;

[0193] Specifically, the diffusion impact of adjacent control areas for:

[0194]

[0195] Where, Indicates all control areas connected to the region. Indicates the region p lag Time pest and disease status, Indicates the node status of region q;

[0196] when When it is greater than the pest and disease diffusion threshold, region q enters the pest and disease outbreak state and serves as a new diffusion source.

[0197] The diffusion process proposed in this embodiment integrates geographical, environmental, and human factors, and is more consistent with the actual diffusion mechanism of pests and diseases. It shares parameters with the previous dynamic time-lag response model, ensuring the logical unity of simulation within and between control areas, and is applicable to control areas of different scales.

[0198] In step S400 of this embodiment, the control risk level and key inducing factors of each control area are determined, including:

[0199] According to the control threshold range of the pest and disease response value, the control risk level is determined, and the environmental variables and / or growth variables in all directed causal paths where the pest and disease variables are located are traced back, and the environmental variables and / or growth variables that appear the most times are taken as key inducing factors.

[0200] Specifically, based on the biological characteristics of pests and diseases, historical outbreak data and prevention and control objectives, the response values ​​of pests and diseases (such as incidence rate and insect population density) are divided into three intervals of low risk, medium risk and high risk, and the threshold intervals are dynamically adjusted by introducing a dynamic adjustment threshold based on the environmental correction coefficient; for example, the response value data of pests and diseases in the past 5-10 years are counted, with the 30th percentile as the upper limit of low risk, the 70th percentile as the upper limit of medium risk, and above the 70th percentile as high risk; when the average temperature in the area exceeds the suitable survival temperature of pests and diseases by 10%, the risk threshold is lowered (such as the high risk threshold is lowered from 20% to 15%); if the regional density is lower than the standard value, the risk threshold is raised (because good ventilation reduces the risk of transmission).

[0201] In the process of tracing the environmental variables and / or growth variables in all directed causal paths where the pest and disease variables are located, the depth-first search (DFS) algorithm is used to extract all directed paths from the environmental and / or growth variables to the pest and disease variables, and the number of occurrences of each environmental and / or growth variable is recorded and weighted summed up; if the weighted occurrence number of a variable exceeds 50% of the total weighted occurrence number, it is directly identified as a key inducing factor. When the contribution of a single variable is insufficient, the combination of variables with the top 3 occurrences is extracted as the key inducing factor (such as the combined effect of "temperature + humidity + tree density").

[0202] In step S500 of this embodiment, corresponding prevention and control strategies are executed according to the prevention and control risk level and key inducing factors of each prevention and control area, including:

[0203] For low-risk prevention and control areas, the goal of the prevention and control strategy is to prevent the occurrence of pests and diseases and enhance the resistance of forest systems. The main prevention and control strategy adopts ecological regulation, while strengthening monitoring.

[0204] For example, when the key inducing factor is "poor soil": implement organic fertilizer improvement (apply 200kg of decomposed wood chips + 50kg of nitrogen fertilizer per mu) to enhance the tree's resistance to stress; when the key factor is "high canopy density": carry out selective thinning (felling intensity 15%-20%) to improve ventilation and light conditions in the forest.

[0205] For medium-risk prevention and control areas, the goal of the prevention and control strategy is to locally control the spread of pests and diseases and block the transmission of the causal chain. The main prevention and control strategy adopted is a combination of targeted intervention and physical barrier.

[0206] For example, if the key factor is rising temperature, shade nets (50%-60% light transmittance) should be erected in the forest to reduce the local temperature by 1-2°C. If the key factor is high aphid density, natural predators such as ladybugs should be released at a 1:100 ratio or a biological pesticide (1.2% bitter tobacco emulsifiable concentrate, 1000 times diluted) should be sprayed. At the same time, isolation zones (50-100 meters wide) should be established at the edges of the pest-infested areas, and diseased plants should be felled and dead branches and leaves cleared to block the spread of the disease.

[0207] For high-risk prevention and control areas, the goal of the prevention and control strategy is to urgently curb the outbreak of pests and diseases and prevent cross-regional spread. It mainly adopts chemical and biological coordinated prevention and control, and implements regional blockade and emergency response prevention and control strategies.

[0208] For example, when the key factor is pine wood nematodes, a combination of punch injection and insect vector control is employed (inject 5-10 ml of 40% oxydemeton-methyl emulsifiable concentrate per tree, and simultaneously spray 1500 times diluted 25% thiazolinone wettable powder). When the key factor is fungal diseases, a triazole fungicide (such as tebuconazole) is applied foliarly (at a concentration of 0.1%-0.2%), combined with a release of Trichoderma fungicide (2 kg per mu) for biological control. Furthermore, the boundaries of the epidemic area are demarcated, quarantine checkpoints are established, and the export of timber and seedlings is prohibited. Emergency plans are activated, and drone swarms are deployed for large-scale pesticide spraying.

[0209] In step S500 of this embodiment, based on the forest health status response data after the implementation of the prevention and control strategy, the temporal causal graph constructed by the reinforcement learning model is updated, including:

[0210] Step S501: collecting forest health status response data after implementing the prevention and control strategy, including pest and disease indicator data, vegetation health indicator data, and forest environment indicator data;

[0211] For example, pest and disease indicators include the rate of change in incidence, the decrease in insect population density, and the attenuation of disease severity; vegetation health indicators include the growth of new shoots and the recovery value of leaf chlorophyll content; forest environment indicators include the temperature and humidity fluctuation range and changes in soil nutrient utilization rate.

[0212] Step S502: Mapping the multi-dimensional forest health status response data into a state vector of the reinforcement learning model to construct a state space;

[0213] Specifically, the forest health status response data is mapped into a state vector for reinforcement learning, and combined with the current parameters of the temporal causal graph (such as edge weights) to form a joint state, thereby constructing a state space.

[0214] Step S503: constructing an action space of a reinforcement learning model based on an update operation on the temporal causal graph; the update operation includes structural adjustment and parameter optimization of the temporal causal graph;

[0215] Among them, structural adjustment includes deleting and / or adding causal edges, for example, adding new causal edges based on the relationship between environmental humidity and disease spread; parameter optimization includes adjusting the causal edge weights and time lag order, for example, increasing the temperature rise The impact coefficient of increased insect population density corrects the time difference in the response of pests and diseases to environmental changes.

[0216] Step S504: Constructing a reward function for the reinforcement learning model, wherein the reward items in the reward function include a prediction error reward, a causal logic consistency reward, and a prevention and control strategy explainability reward;

[0217] Specifically, the prediction error reward is the error reward between the pest and disease response value predicted based on the time-series causal graph and the actual observed value, the causal logic consistency reward is the confidence reward of each causal edge in the causal time-series graph, and the prevention and control strategy interpretability reward is the interpretability reward of the effect of the time-series causal graph on the actual implementation of the prevention and control strategy.

[0218] Step S505: Based on the constructed state space, action space, and reward function, a reinforcement learning model is run to obtain an updated temporal causal graph; the reinforcement learning model adopts a network structure that combines a graph neural network and a deep Q network;

[0219] Among them, the graph neural network processes the topological structure of the temporal causal graph, the neural network CNN processes the temporal data of the temporal causal graph, and uses the Q-value function to predict the expected reward of the state-action pair.

[0220] In a specific embodiment of the present invention, taking the spraying control strategy as an example, the process of updating the temporal causal graph is as follows:

[0221] The initial state definitions include: the cause-and-effect diagram before spraying (e.g., "temperature and humidity → insect population density" with a weight of 0.6), the density change rate one month after spraying (-40%), and the temperature and humidity fluctuation values ​​(+5°C, +10% humidity).

[0222] Reinforcement learning model output action: Increase the weight of "temperature and humidity → insect population density" from 0.6 to 0.8, and add a causal edge of "tree age → pesticide sensitivity" (weight 0.5);

[0223] Reward calculation: If the updated temporal causal graph's prediction error for a 15% rebound density decreases from 25% to 10%, the prediction error reward is +15. If the "tree age → drug sensitivity" correlation is verified by data with p < 0.05, the logical consistency reward is +20, for a total reward of 25.

[0224] Updated time-series causal diagram: New causal paths added: Temperature and humidity ↑ → insect population density ↑ (weighted), tree age → pesticide effect (explains why older trees recover more slowly), making the causal diagram more consistent with actual responses after prevention and control measures.

[0225] In this embodiment, in the above-mentioned temporal causal graph update process based on the reinforcement learning model, the reinforcement learning model can serve as a bridge between "data-causal graph", so that the temporal causal graph can continuously evolve with the prevention and control practice, thereby improving the accuracy and dynamic adaptability of forest pest and disease prevention and control.

[0226] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

[0227] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A forest pest control method based on disease and insect pest monitoring, characterized in that: include: Divide the target forest area into several control zones based on vegetation type and distribution; Collect historical multi-source data of the target forest area and fill in missing data to obtain complete data; the multi-source data includes forest environment data, vegetation growth data, and pest and disease occurrence data; Conduct temporal and causal relationship analysis on complete data, and construct a temporal causal diagram that reflects the causal and temporal relationships among environmental changes, vegetation growth, and the occurrence of pests and diseases; Based on the current forest environment data and vegetation growth data collected in each prevention and control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases along the causal path in the time series causal diagram is conducted to determine the prevention and control risk level and key inducing factors of each prevention and control area; According to the prevention and control risk level and key inducing factors of each prevention and control area, the corresponding prevention and control strategy is implemented, and based on the forest health status response data after the implementation of the prevention and control strategy, the temporal causal diagram is updated through the reinforcement learning model.

2. The forest pest control method based on disease and insect pest monitoring according to claim 1, characterized in that: Fill missing data in the collected historical multi-source data to obtain complete data, including: Align various data samples in historical multi-source data in time series and organize them into independent two-dimensional data tables; Processing each two-dimensional data table into a data table to be filled with a unified data structure and the same feature dimension and time length; A column-level filling method based on a linear structure is used to fill missing data in the data table to be filled one by one to obtain complete data; in the missing data filling process, missing data is filled according to the left and / or right boundary values ​​of the missing data interval.

3. The forest pest control method based on disease and insect pest monitoring according to claim 1, characterized in that: Construct a temporal causal diagram that reflects the causal and temporal relationships among environmental changes, vegetation growth, and the occurrence of pests and diseases, including: For complete environmental data, key variables that have a significant impact on the occurrence of pests and diseases in the target forest area are screened from forest environmental data and vegetation growth data; key variables that have a significant impact on vegetation growth are screened from pest and disease occurrence data; Construct a VAR model based on the selected key variables and determine the lag order between the variables in the VAR model; Constructing an initial graph structure based on the screened environmental variables, growth variables, and pest and disease variables; the initial graph structure includes a completely directed graph and / or a random graph; The initial graph structure is tested for significance based on the VAR model coefficients. The effective edges of the initial graph structure are screened according to the significance of the VAR model coefficients. An improved Granger causality discovery method is introduced to construct the initial temporal causal graph. Construct a scoring function that takes into account variable lags; A sparse greedy iterative search is performed in the initial temporal causal graph, and the score values ​​after the operations performed during the search process are calculated. The operation that maximizes the score value is retained until the number of iterations is reached or the score value reaches a set threshold, thereby obtaining a temporal causal graph; the operations include adding edges, deleting edges, and reversing edge directions.

4. The forest pest control method based on disease and insect pest monitoring according to claim 3, characterized in that: In the process of constructing the initial temporal causal graph by introducing the improved Granger causality discovery method, the causal relationships in the initial structure graph are screened by setting a dynamic lower bound threshold, and the screened causal relationships are verified by using a block perturbation method based on data characteristics. Among them, the dynamic lower threshold for: Where, Indicates the front of the data Quantile, represents the mean score, represents the standard deviation, represents the coefficient for adjusting the standard deviation weight, A coefficient that represents the range of the control score.

5. The forest pest control method based on disease and insect pest monitoring according to claim 3, characterized in that: Scoring function considering variable lag for: Where, Represents the initial temporal causal graph constructed, Represents a dataset containing the selected environmental variables, growth variables, and pest and disease variables. Indicates hysteresis Order pair variables The weight of Representing variables exist The value of the moment, Representing variables exist The set of parent nodes of the order lag, Representing variables exist The parent node set when represents the time-delay decay function, represents the regularization coefficient, represents the model parameters, represents the conditional probability function, Indicates the number of variables, represents the maximum delay order.

6. The forest pest control method based on pest monitoring according to claim 1, characterized in that: Based on the current forest environment data and vegetation growth data collected in each prevention and control area, a spatiotemporal dynamic simulation of the occurrence and spread of pests and diseases is conducted along the causal path in the time series causal diagram, including: The current forest environment data and vegetation growth data collected in the prevention and control areas are integrated and standardized to form regional environmental characteristic vectors and growth characteristic vectors, which are used as the initial states of the environmental variables and growth variables in the time series causal diagram, respectively. The initial state of the pest and disease variable is set to 0. In the current initial state, all directed causal paths from environmental variables and / or growth variables to pest and disease variables are extracted from the time series causal graph. The time lag order and VAR model coefficient of each edge in the directed causal path are recorded. The directed causal paths are ranked by significance according to the VAR model coefficient, and the path weight of each directed causal path is set according to the significance ranking result. Based on the superposition effect of multi-path influence, a dynamic time-lag response model from environmental variables and / or growth variables to pest and disease variables is constructed; the dynamic time-lag response model is used to calculate the pest and disease response value of each pest and disease variable under the superposition of the mutual influence of all the directed causal paths associated with it; Based on the constructed dynamic time-lag response model, all directed causal paths pointing to pest and disease variables are traversed, and the path weights of the directed causal paths are combined to simulate the spatiotemporal dynamics of pest and disease occurrence and spread, and predict the pest and disease response value; The spatiotemporal dynamic simulation includes a time series simulation within the control area, and when the pest and disease response value of the time series simulation within the control area is greater than the pest and disease area diffusion threshold, the spatiotemporal dynamic simulation also includes a diffusion simulation between control areas.

7. The forest pest control method based on disease and insect pest monitoring according to claim 6, characterized in that: The time series simulation within the control area includes: Establishing a time-delay sequence of each variable in the directed causal path; the length of the time-delay sequence is the maximum time-delay order of the variable in the time-series causal graph; The directed causal path whose total time delay order is greater than the preset simulation time step is taken as the trigger path; In each simulation time step, for the same pest and disease variable, the dynamic time-delay response model is used to calculate and summarize the pest and disease response values ​​of all associated trigger paths, which are used as the current state of the pest and disease variable. The current states of the growth variable and environmental variable are updated synchronously, and the current pest and disease variable, growth variable, and environmental variable are saved in the corresponding time-delay sequence. The time series simulation process is iterated according to the simulation time step until the preset simulation time step is reached or the current pest and disease response value in the time lag sequence exceeds the warning threshold.

8. The forest pest control method based on disease and insect pest monitoring according to claim 6, characterized in that: The diffusion simulation between the control areas includes: Construct a regional linkage network among the control areas within the target forest area; Based on the regional association network, multi-factor weighted diffusion coefficients and time lag coefficients are set between each prevention and control area, and a diffusion simulation framework based on cellular automata is constructed to simulate the diffusion between prevention and control areas. In the diffusion simulation framework, the state transition rule is set as follows: the node state of the control area is determined by superimposing the pest response value obtained by time series simulation in the control area and the pest response value obtained by diffusion from the associated control area; When the superimposed pest and disease response value of the prevention and control area is greater than the set pest and disease area diffusion threshold, the prevention and control area will be regarded as a new diffusion source.

9. The forest pest control method based on disease and insect pest monitoring according to claim 6, characterized in that: Determine the prevention and control risk level and key inducing factors for each prevention and control area, including: According to the control threshold range of the pest and disease response value, the control risk level is determined, and the environmental variables and / or growth variables in all directed causal paths where the pest and disease variables are located are traced back, and the environmental variables and / or growth variables that appear the most times are taken as key inducing factors.

10. The forest pest control method based on disease and insect pest monitoring according to claim 9, characterized in that: Based on the forest health status response data after the implementation of the prevention and control strategy, the reinforcement learning model updates the constructed temporal causal graph, including: Collect data on forest health status after implementing prevention and control strategies, including pest and disease indicator data, vegetation health indicator data, and forest environment indicator data; Mapping multi-dimensional forest health status response data into the state vector of the reinforcement learning model to construct the state space; Based on an update operation on the temporal causal graph, an action space of the reinforcement learning model is constructed; the update operation includes structural adjustment and parameter optimization of the temporal causal graph, the structural adjustment includes deleting and / or adding causal edges, and the parameter optimization includes adjusting the causal edge weights and lag orders; Constructing a reward function for the reinforcement learning model, wherein the reward items in the reward function include prediction error reward, causal logic consistency reward, and prevention and control strategy explainability reward; Based on the constructed state space, action space and reward function, a reinforcement learning model is run to obtain an updated temporal causal graph; the reinforcement learning model adopts a network structure that combines a graph neural network and a deep Q network.

Images