Integrated ecological prevention and control method for citrus psylla

Abstract

The application relates to the technical field of citrus pest prevention and control, and discloses a comprehensive ecological prevention and control method for citrus psylla. The method obtains current pest state information based on real-time monitoring data of a citrus planting area; extracts biological characteristics of the citrus psylla and related ecological environment characteristics through an ecological characteristic library to form an initial ecological characteristic set; fuses the real-time monitoring data and the initial ecological characteristic set to generate a dynamic ecological characteristic set; calculates the priority scores of various prevention and control strategies according to the spatial and temporal distribution density, migration trend intensity, natural enemy inhibition coefficient, host plant resistance index and environmental stress factor of each characteristic in the dynamic set; generates a strategy ranking table in descending order of the scores, selects the to-be-executed strategies in sequence, and judges whether the target pest density reduction amount is completed within a preset period. The method combines real-time data and ecological characteristics, dynamically adjusts the strategies, and improves the accuracy and effectiveness of the prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of citrus pest control, in particular to a comprehensive ecological control method for citrus psylla. BACKGROUND

[0002] Citrus is an important economic crop and is widely planted worldwide. Its industry development has a significant impact on regional economy and agricultural structure. However, as one of the main pests in the citrus industry, citrus psylla not only directly sucks the sap of citrus tender shoots, causing the plant to grow weak and the new shoots to be deformed, but more importantly, it is the main vector of citrus Huanglongbing disease. As a devastating disease, there is currently no effective cure, and once it occurs, it will cause great losses to the citrus industry.

[0003] Currently, the main method for controlling citrus psylla is chemical control, which uses insecticides to suppress the population density. However, long-term reliance on chemical control can cause a series of problems. Citrus psylla is prone to developing resistance to insecticides, leading to an increase in the amount of pesticide used and an increase in control costs. At the same time, it also exacerbates pollution of the ecological environment. Chemicals not only kill citrus psylla, but also harm its natural enemies, disrupting the natural balance of the ecosystem and making the citrus psylla population more likely to rebound, creating a vicious cycle.

[0004] Although ecological control methods such as biological control and agricultural control have the characteristics of environmental friendliness, their current applications often have limitations. In biological control, the introduction and use of natural enemies are greatly influenced by environmental conditions and climate factors, making it difficult to achieve stable and effective control. Agricultural control measures such as proper pruning and cleaning orchards can reduce the source of pests to some extent, but lack precise combination with pest occurrence dynamics, resulting in less than ideal control effect. In addition, most existing control methods rely on empirical operations and lack effective integration and analysis of real-time monitoring data in citrus planting areas, making it impossible to adjust control strategies in a timely manner based on the dynamic changes of pests, resulting in insufficient pertinence and timeliness of control. SUMMARY

[0005] The present application aims to provide a comprehensive ecological control method for citrus psylla to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a comprehensive ecological control method for citrus psylla, which comprises:

[0007] Based on real-time monitoring data of citrus planting areas, current pest status information is obtained;

[0008] The biological characteristics of citrus psylla and the associated ecological environment characteristics are extracted from the ecological feature library to form an initial ecological feature set;

[0009] The real-time monitoring data is fused with the initial ecological characteristic set to generate a dynamic ecological characteristic set;

[0010] According to the spatiotemporal distribution density, migration trend intensity, natural enemy suppression coefficient, host plant resistance index and environmental stress factor of each characteristic in the dynamic ecological characteristic set, the priority score of each ecological prevention and control strategy is calculated.

[0011] The prevention and control strategy ranking table is generated in descending order of the priority score, and the current to-be-executed strategy is selected in sequence and it is judged whether it can complete the target pest density reduction amount within a preset period.

[0012] Preferably, the biological characteristics of the citrus psylla and the associated ecological environment characteristics extracted from the ecological characteristic library include:

[0013] The individual behavior characteristics, population diffusion characteristics, natural enemy restriction characteristics, plant volatile characteristics and microclimate response characteristics of the citrus psylla are extracted from the ecological characteristic library in hierarchical levels;

[0014] The individual behavior characteristics and population diffusion characteristics of the same hierarchical level are subjected to mean value processing to obtain primary fusion characteristics;

[0015] The primary fusion characteristics and the microclimate response characteristics are subjected to weighted superposition to generate secondary fusion characteristics;

[0016] The hierarchical characteristic fusion is iteratively executed until the highest level characteristic processing is completed, and a dynamic ecological characteristic set containing multi-level correlation is output.

[0017] Preferably, the calculation of the priority score of each ecological prevention and control strategy includes:

[0018] The spatiotemporal distribution density proportion value, migration trend intensity proportion value, natural enemy suppression coefficient proportion value, host plant resistance index proportion value and environmental stress factor proportion value corresponding to each strategy are obtained;

[0019] The migration trend intensity proportion value is multiplied by a first weight coefficient, the natural enemy suppression coefficient proportion value is multiplied by a second weight coefficient, the host plant resistance index proportion value is multiplied by a third weight coefficient, the environmental stress factor proportion value is multiplied by a fourth weight coefficient, and the spatiotemporal distribution density proportion value is multiplied by a fifth weight coefficient;

[0020] The sum of the five product results is obtained and the resource consumption cost proportion value is deducted to obtain the priority score.

[0021] Preferably, the selection of the current to-be-executed strategy in sequence and the judgment of whether it can complete the target pest density reduction amount within a preset period include:

[0022] The pest density reduction amount that can be achieved by a single strategy within a preset period is calculated;

[0023] Comparing the achievable pest density reduction amount with the target pest density reduction amount;

[0024] If the achievable amount is greater than or equal to the target amount, it is determined that a single strategy can be completed;

[0025] If the achievable amount is less than the target amount, the number of strategies that need to be supplemented and executed is calculated.

[0026] Preferably, the calculation of the number of strategies that need to be supplemented and executed includes:

[0027] Divide the target pest density reduction amount by the single-strategy achievable amount, and round up to get the number of strategy instances;

[0028] Detect the total amount of available ecological resources according to the number of strategy instances;

[0029] When the available resources are sufficient, start the corresponding number of strategy instances;

[0030] When the available resources are insufficient, select instances with lower rankings than the current strategy from the started strategies to terminate and release resources until the resource requirements are met or there are no instances to terminate.

[0031] Preferably, the method further includes:

[0032] Based on the geographic information system topology, analyze the pest spread path;

[0033] Determine the sink point and migration corridor of the pest source through gradient field calculation;

[0034] Extract key spread characteristics that meet the pest density threshold and the duration of continuous infestation;

[0035] Input the key spread characteristics into the pre-trained ecological model to generate a reinforced prevention and control field.

[0036] Preferably, the input of the key spread characteristics into the pre-trained ecological model to generate a reinforced prevention and control field includes:

[0037] Introduce topological structure consistency constraints in the loss function of the ecological model;

[0038] Optimize the matching degree of the reinforced prevention and control field and the real ecological gradient field through adversarial training;

[0039] Spatially couple the reinforced prevention and control field with the basic environmental factors to generate a composite prevention and control field containing biological and non-biological factors.

[0040] Preferably, the method further includes:

[0041] Perform spatiotemporal evolution simulation on the composite prevention and control field;

[0042] Separate the biological action field and the environmental action field after evolution;

[0043] The biological action field is reconstructed with real-time monitoring data, and an optimized prevention and control instruction set is output.

[0044] Preferably, the method further comprises:

[0045] Continuously monitoring the implementation state of the executed strategy;

[0046] Waking up the dormant strategy instance when new monitoring data triggers a change in insect situation event;

[0047] If the strategy instance is not matched after waking up, it enters a dormant waiting state;

[0048] If a valid instruction is matched, the implementation state is modified to be in execution, and the execution result data is fed back to the dynamic ecological feature set.

[0049] Preferably, the output of the optimized prevention and control instruction set further comprises:

[0050] Verifying the actual insect population density change curve after the execution of the instruction set;

[0051] When the actual reduction amount deviates from the predicted value by more than a threshold value, the key diffusion features are re-extracted;

[0052] Iteratively updating the reinforcement prevention and control field parameters until the deviation value falls within an acceptable range.

[0053] Compared with the prior art, the beneficial effects of the present application are:

[0054] By obtaining current pest status information based on real-time monitoring data of citrus planting areas, the prevention and control measures can closely match the actual pest occurrence situation, avoiding improper measures due to information lag in traditional prevention and control. The use of real-time monitoring data enables more accurate understanding of the occurrence dynamics and population changes of citrus psylla, providing reliable real basis for subsequent prevention and control strategy formulation.

[0055] By extracting the biological characteristics of citrus psylla and the associated ecological environment characteristics from the ecological feature library, an initial ecological feature set is formed, combining the biological characteristics of citrus psylla with the ecological environment factors, comprehensively considering various factors affecting pest occurrence. This multi-dimensional feature extraction breaks through the limitations of single factor consideration in traditional prevention and control, enabling more systematic analysis of the root causes and rules of pest occurrence, laying a comprehensive foundation for subsequent feature fusion and strategy formulation.

[0056] The real-time monitoring data is processed by feature fusion with the initial ecological feature set to generate a dynamic ecological feature set, and the organic combination of real-time information and basic feature information is realized. The dynamic ecological feature set can change with the update of the monitoring data, reflecting the dynamic characteristics of the insect pests under different time and space conditions, so that the understanding of the insect pests is more timely and comprehensive, and the dynamic adjustment of the prevention and control strategy is possible.

[0057] According to the spatiotemporal distribution density, migration trend intensity, natural enemy suppression coefficient, host plant resistance index and environmental stress factor of each feature in the dynamic ecological feature set, the priority score of each ecological prevention and control strategy is calculated, so that the selection of the prevention and control strategy is more scientific and reasonable. By comprehensively considering these various feature factors, the applicability and potential effect of different prevention and control strategies under the current situation can be accurately evaluated, the blindness of the selection of the strategy in the traditional prevention and control is avoided, and the pertinence of the prevention and control measures is improved.

[0058] The prevention and control strategy sorting table is generated in descending order of the priority score, and the current to-be-executed strategy is selected and judged whether it can complete the target insect population density reduction amount within a preset period, so that the orderly implementation and effect evaluation of the prevention and control measures are ensured. The mechanism of the orderly execution and effect judgment can timely find the problems that may occur in the prevention and control process, if a strategy fails to achieve the expected effect, the subsequent strategy can be adjusted in time, the flexibility and effectiveness of the prevention and control process are ensured, and the continuous optimization of the prevention and control cycle is formed, so that the ecological prevention and control of the citrus psylla is better achieved. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The timing diagram of the comprehensive ecological prevention and control method for the citrus psylla;

[0060] Figure 2 The flowchart of the ecological feature extraction and fusion;

[0061] Figure 3 The flowchart of the strategy execution judgment;

[0062] Figure 4 The flowchart of the insect pest diffusion analysis and the generation of the reinforced prevention and control field;

[0063] Figure 5 The flowchart of the strategy dynamic monitoring and awakening. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0065] Please refer to Figure 1 The present application provides a comprehensive ecological prevention and control method for the citrus psylla, which comprises:

[0066] The method is realized by integrating real-time monitoring data, ecological feature analysis and strategy optimization execution through digital means. The method first obtains current pest status information based on real-time monitoring data of citrus planting areas. These monitoring data come from sensor networks installed in planting areas, including infrared cameras, soil humidity sensors, anemometers, etc., with a sampling frequency of every 30 minutes, covering pest density, temperature, humidity, etc. The system receives sensor data through a pre-set data interface, removes noise and outliers, and outputs structured pest status information. Then, the system accesses the pre-constructed ecological feature library to extract biological features of citrus psylla and related ecological environment features, forming an initial ecological feature set. The ecological feature library is stored in a distributed database, containing historical data sets and pre-defined ecological model parameters. The extracted features include pest behavior data, plant volatile concentration, microclimate factors, etc. The system performs feature fusion processing on real-time monitoring data and the initial ecological feature set, aligns and normalizes based on spatio-temporal coordinates using multi-source data fusion algorithms, and generates a dynamic ecological feature set. This set contains updated feature values, such as corrected population distribution positions. Next, the system quantitatively calculates the spatio-temporal distribution density, migration trend intensity, natural enemy suppression coefficient, host plant resistance index, and environmental stress factors in the dynamic ecological feature set. After these factors are numerized, they are input into the priority calculation module. This module calculates the priority score of each ecological control strategy according to pre-set scoring rules. The priority score is used to sort strategies and generate a control strategy ranking table, ranking the strategy list in descending order. After sorting, the system selects the current strategy to be executed, and judges whether it can complete the target pest density reduction amount within the pre-set period through the prediction model. The pre-set period is set to 7 days, and the target reduction amount threshold is set to reduce the pest density by 50%. If the strategy can be completed, the system executes immediately; otherwise, it goes to the supplementary strategy processing flow. The whole process is managed by a central control unit and implemented with cloud computing resource scheduling. This automatic circulation mechanism ensures dynamic response to changes in pests.

[0067] Example 1: see Figure 2In the ecological feature extraction stage, the distributed database stores multi-level structure data. The database adopts B+ tree index mechanism, and the basic layer stores individual behavior feature data set, including the spatial coordinate sequence and time stamp of foraging trajectory, the time sequence statistical value of mating frequency and the static duration distribution histogram. The middle layer contains population diffusion feature data group, which is specifically composed of migration distance probability distribution model, contour map coordinate point array of distribution range and matrix expression of population age structure. The high-level layer integrates the natural enemy constraint feature vector (natural enemy species number and predation rate correlation curve), plant volatile concentration gradient field (with volatile organic compound chromatographic analysis value as core) and microclimate response feature matrix (tensor expression of temperature and humidity correlation coefficient).

[0068] After the feature fusion operation is started, the system loads data in hierarchical order. In the basic level processing unit, the algorithm automatically pairs the spatiotemporal tags of individual behavior features and population diffusion features. For each paired coordinate point, the system performs data alignment: the foraging trajectory coordinate point and the migration distance data are matched through spatial interpolation, and the mating frequency time series value and the distribution range data are aligned through time window sliding average. After alignment, the mean value calculation engine is implemented: the arithmetic mean value is calculated for the matched points, for example, the foraging intensity value (0.7) and the migration intensity value (0.5) of a certain spatial point generate the primary fusion feature value (0.6). This calculation covers all paired points, forming the initial feature plane.

[0069] Microclimate response features are input through an independent channel. The system collects real-time environmental sensor data stream, and updates the temperature and humidity response coefficient table every five minutes. The coefficient table establishes a spatial mapping relationship with the primary fusion feature plane. When the weighted superposition module is activated, the environmental monitoring unit dynamically generates weight factors: when the temperature data rises continuously for three hours, the microclimate weight is automatically increased to 0.65; when the relative humidity is lower than the critical value of 40%, the humidity related weight is set to 0.35. The weighting process performs point-by-point matrix multiplication: the primary fusion feature value of each spatial point is multiplied by the corresponding microclimate weight, and then superimposed with the original microclimate response value. For example, for the point with coordinates (X, Y), the calculation formula is (0.6 x 0.65) + 0.8 (microclimate value) = 1.19, generating the secondary fusion feature data layer.

[0070] The hierarchical iteration process adopts a loop control mechanism. During intermediate layer processing, the secondary fused feature data serves as the input source for channel fusion with the high-level layer features. The fusion operation platform creates a dual-buffer area: the front buffer stores the secondary fused data, and the rear buffer loads the natural enemy constraint feature vector. The system establishes an index mapping through a feature matching algorithm, which specifically resamples the predation rate curve to the same spatial resolution. The same algorithm kernel is used for weighted superposition, but the weight parameter is switched to the hierarchical weight coefficient table. When processing is complete, the new feature plane is looped into the next level. In the final level processing, the high-level layer plant volatile feature generates a three-dimensional concentration field through a gas chromatography simulator, which is spatially fused with the input features. The entire iteration process performs integrity checks: each layer output performs a tensor dimension check, and if the dimensions do not match, a data reconstruction process is triggered.

[0071] After the feature fusion is completed to form a dynamic ecological feature set, the priority scoring module is immediately started. The scoring parameters are obtained through five independent collectors: the spatio-temporal distribution density collector scans the monitoring area grid, divides the current insect density by the five-year historical peak value; the migration trend analyzer processes the motion vector field, calculates the motion direction consistency strength value; the natural enemy biodiversity counter counts the number of natural enemy species per unit area; the plant resistance detector measures the leaf antibody protein concentration; the environmental stress monitoring station integrates soil salinity and heavy metal content data. Each collector outputs a normalized proportion value (0-1 range).

[0072] The weight allocation unit presets a fixed coefficient group: migration trend 0.20, natural enemy suppression 0.25, plant resistance 0.15, environmental stress 0.30, spatio-temporal density 0.10. The calculation core performs five-way parallel multiplication: migration trend strength proportion value; natural enemy suppression coefficient proportion value; plant resistance index proportion value; environmental stress factor proportion value; spatio-temporal distribution density proportion value.

[0073] The product result is input into the addition accumulator: 0.17 + 0.175 + 0.09 + 0.135 + 0.09 = 0.66. The resource cost accounting module operates synchronously: reads the strategy resource registration table, extracts the labor consumption hours, pesticide consumption, and equipment occupation time, and converts them into total cost proportion according to the resource cost conversion table. Finally, a subtraction operation is performed to generate the priority score result. This score is automatically associated with the strategy number and stored in the priority queue for sorting and calling.

[0074] All computing processes are completed on dedicated hardware accelerators: tensor processors are used in the feature extraction stage to fuse three-dimensional feature fields; priority calculation is implemented using a five-channel floating-point operation unit for parallel computing. The error handling mechanism includes three layers of verification: feature value range verification (resampling if it exceeds the 0-1 range), weight and verification (the total weight coefficient is always equal to 1.0), and calculation result verification (trigger manual review if the score value is out of limit). The entire process execution time is allocated by the central controller to ensure that all processing tasks are completed under the time constraint.

[0075] Example 2: see Figure 3 In the policy execution evaluation stage, the system calls preset parameters and prediction models to judge the effectiveness of the policy. The preset period is defined by a fixed time window, and the standard setting is a seven-day period; the target reduction amount is initialized by the configuration module and stored in the system core database, and the typical setting is to reduce the original insect population density by fifty percent. The policy effect prediction model accesses the data stream of the dynamic ecological feature set, continuously receives the population size change rate, the natural enemy activity frequency matrix, and the host plant response curve as dynamic input parameters. The model architecture is built on a time series prediction algorithm, which contains three layers of recurrent neural network processing time dependence, and performs full-amount prediction operation every sixty minutes.

[0076] When starting the single-policy effect prediction process, the simulation engine loads the feature parameter cluster of the target policy. For example, when evaluating the "release predatory natural enemies" policy, the natural enemy species number, unit area release amount, and release point geographic coordinate set need to be input. The model initializes the baseline environment scenario: input the insect population density distribution field, temperature gradient map, and wind speed vector field at the current time into the simulation space. During the operation process, the time stepping mechanism is executed, and the state calculation is advanced every hour: first, update the biological interaction sub-model, calculate the natural enemy predation amount based on the Lotka-Volterra equation; simultaneously update the environmental factor influence sub-module, and apply the temperature and humidity coupling algorithm to correct the insect activity rate. At the end of each time step, record the insect population density change value, and after continuously advancing one hundred and sixty-eight steps (corresponding to a seven-day period), terminate the simulation. The final output is a quantitative result value, typically presented as the insect population density reduction percentage at the end of the prediction period, such as reducing from five hundred units to three hundred units, which indicates that the achievable reduction amount is forty percent.

[0077] The system then activates the reduction amount comparator. The comparator core algorithm performs a numerical comparison: the achievable reduction amount extracted from the simulation output is compared with the target reduction amount stored in the database. The comparison logic uses an exact threshold decision criterion: a Boolean true value is generated when the achievable amount is greater than or equal to the target amount; otherwise, a Boolean false value is generated. The decision result triggers the corresponding operation instruction. If a true value is returned, an activation instruction is sent to the strategy executor, which contains the strategy execution code, resource allocation list, and spatial coordinate set. The executor calls the hardware interface to control the drone swarm to release natural enemies or start the spray device array according to the coordinate sequence. The entire implementation process records operation logs, including operation timestamps and mechanical state codes.

[0078] When the decision returns a false value, the system switches to the strategy supplement mode. This mode includes a double-computation and resource coordination mechanism. First, the strategy quantity demand is calculated: the target reduction amount is divided by the single-strategy achievable amount to obtain the theoretical multiple. The mathematical processor uses the ceiling function to process this multiple, and the ceil(1.25) function in the standard function library outputs an integer value of 2. This result is stored as the strategy implementation quantity parameter.

[0079] The resource scheduling module then starts the collaborative operation. The resource database maintains a multi-dimensional resource list in real time: the biological resource pool records the available natural enemy species inventory; the chemical resource stores the remaining amount and expiration date of the pesticide; the equipment resource table records the available number of drones and battery status; the human resource pool counts the standby working hours of technical personnel. The resource demand estimator calculates the total consumption based on the strategy implementation quantity parameter: for example, if a single strategy requires 50 units of natural enemies, the total number requires 100 units; if a single strategy consumes 20 liters of pesticide, the total number requires 40 liters; if a single strategy occupies six working hours, the total requires twelve working hours. The resource sufficiency analysis engine performs parallel detection: it compares whether the available biological resource amount is greater than the total demand, whether the available pesticide amount is greater than the calculated total demand, and whether the idle equipment working hours meet the total occupancy demand. If all three detections pass, a resource confirmation signal is generated.

[0080] Under the condition of resource sufficiency, the execution controller calls the strategy cloning function: based on the strategy implementation quantity parameter, multiple parallel execution instances are created. The instance startup uses a distributed architecture, and each instance is assigned an independent process ID and resource slot. For example, if the implementation quantity is two, two instance identifiers S001A and S001B are generated, which are bound to the execution resources of the north and south regions respectively. The implementation process maintains state monitoring, and each instance reports progress data to the central coordinator every thirty minutes.

[0081] In resource-scarce scenarios, a resource reorganization process is triggered. The priority queue scanner retrieves subsequent items of the current policy from the sorting table: if the priority score of the policy to be started is 85, it retrieves running policies with scores between 70 and 80. The instance status analyzer filters out instances with a running status of "active" and which are not core protection policies. The instance terminator sends a termination command to the selected instances: first, it saves intermediate execution data to persistent storage; then, it removes the resource occupancy mark and populates the released resource value back into the resource pool. The amount of released resources is updated in real-time in the resource database; for example, when 10 liters of reagent are released, the chemical resource table is updated immediately. The resource reorganization cycle continues: after terminating an instance, a resource sufficiency check is re-executed; if the requirement is met, a supplementary policy is started; if not, the next level policy instance is terminated until the requirement is met or there are no more instances to terminate. When there are no more instances to terminate, a resource shortage alert code "ERR-507" is generated, and the current policy start request is suspended and placed in the standby queue. The standby queue is configured with a periodic wake-up mechanism, retrying the resource coordination process every six hours until resource conditions are met or administrator intervention terminates the process.

[0082] The resource reorganization process employs atomic transaction management to ensure data consistency: instance termination and resource release operations are encapsulated within transactions, and any failure in any step automatically rolls back to the state before the operation. Concurrency conflict resolution uses a timestamp sorting protocol to avoid resource calculation errors caused by multiple terminations. All operation details are written to the audit log, including the terminated instance ID, released resource type, operation time, and changes in remaining resource quantity. Log entries are transmitted via an encrypted channel to a remote backup server for persistent storage.

[0083] Example 3: See Figure 4 During the spatial analysis phase, the geographic information platform loads a multidimensional environmental dataset. The elevation raster data uses a digital terrain model, with a resolution set so that each pixel represents five square meters of actual terrain. The vegetation distribution polygon layer includes tree species classification codes and canopy coverage attribute values. The road network dataset consists of line features, with each road recording its width and pavement material code. The topology modeling engine converts monitoring point coordinates into network nodes, and the distance between adjacent nodes is automatically calculated using the Euclidean distance algorithm. Path connectivity weights are determined by three parameters: the vegetation coverage influence coefficient is a normalized value of 0-1 (0.3 for broadleaf forests and 0.7 for coniferous forests), the slope resistance coefficient is calculated using an inverse proportional function (the weight decreases by 0.1 for every five degrees increase in slope angle), and the road guidance factor is fixed at 0.9. The system constructs a weighted adjacency matrix to store the migration resistance between nodes, with the matrix dimension consistent with the number of monitoring points.

[0084] The migration path simulation is based on an improved Dijkstra algorithm. The algorithm initializes all nodes with infinite distance and selects the pest source core area as the starting point set. During the iteration process, the relaxation operation is performed: when the shortest path of the current node is updated, its adjacent nodes are traversed and the temporary distance is calculated. The weight parameter dynamic adjustment mechanism is activated: when the temperature sensor returns a value greater than 32°C, the path weight is temporarily multiplied by the temperature correction factor 1.2; when the rainfall exceeds 10mm, the humidity correction factor 0.8 is triggered. Finally, a set of minimum resistance paths is generated, each path is labeled with the cumulative resistance value and node sequence.

[0085] The gradient field calculation introduces a potential energy field model. The input basic field includes the spatial distribution matrix of the pest density (resolution 20m x 20m) and the wind speed vector field (0.5 second level update). The physical modeling core solves the partial differential equation set:

[0086]

[0087] where: represents the migration potential energy at coordinate (x, y) at time t, is the pest density gradient operator (two-dimensional spatial partial derivative), is the wind speed vector (including direction and speed components), is the temperature time change rate. The coefficient takes 0.75 to correspond to the density weight, takes 0.55 to control the influence strength of the wind field, takes -0.35 to represent that the temperature rise reduces the migration potential energy (the temperature parameter is derived from the meteorological station network data). The solver uses the finite difference method for discretization processing, the grid step matches the spatial resolution, and the time step interval is ten minutes. After twenty continuous iterations, the stable potential energy field is output, the local minimum value point is marked as the pest source and sink point (output in GeoJSON format coordinates), and the zonal area with a gradient of potential energy falling below the threshold is identified as the migration corridor.

[0088] The key diffusion feature extraction unit is configured with a double condition filter. The spatial density condition detection module scans each grid cell: when the pest density sensor returns a value continuously higher than the preset threshold of 100 per square meter, a candidate area is generated. The time persistence verifier interfaces with the historical database to check whether the location has met the density condition for three consecutive days. The intersection area of the two conditions is marked as a red warning area, and the system extracts its spatial feature parameters: core coordinate point set, area expansion rate (square meters per day), boundary curvature radius list. At the same time, the associated environmental parameter snapshot is collected: the temperature mean value, humidity fluctuation range, and vegetation chlorophyll index value corresponding to the red warning area.

[0089] The ecological model adopts a three-dimensional convolutional neural network architecture. The input layer receives the key diffusion feature tensor with dimensions of 64x64x18 (the first two dimensions correspond to the spatial grid, and the third dimension contains thirteen biological parameters and five environmental parameters). The convolution kernel size is designed as 5x5x5, and the features are extracted by three-dimensional sliding on the feature tensor . The loss function improvement contains a topological constraint term:

[0090]

[0091] wherein is the mean square error of the predicted field and the measured field, takes 0.5 as the constraint strength coefficient, represents the output feature map of the i-th layer of the model, corresponds to the down-sampled version of the real topological feature map at the same scale. Adversarial training operates under the generator-discriminator framework: the generator inputs the key features to generate the predicted prevention and control field, and the discriminator inputs the actual ecological gradient field data (collected from the sensor network). Training iterations are performed for fifteen rounds, each round containing three generator optimizations and two discriminator updates. The optimization algorithm uses adaptive moment estimation, with an initial learning rate of 0.001 decaying by 20% every five rounds. After convergence, the enhanced prevention and control field data is generated, with a format of three-channel grid: the first channel stores the biological inhibition intensity value (0-1 normalized), the second channel records the environmental adjustment parameter, and the third channel is the spatial confidence index.

[0092] The spatial coupling processor performs field data integration operations. The basic environmental factor database outputs temperature grid, humidity distribution map, and soil type classification map. The registration engine uses an affine transformation algorithm to align the enhanced prevention and control field with the basic environmental map at the pixel level. The coupling operation uses pixel-by-pixel weight superposition: for a pixel point with coordinates (x, y), the composite value

[0093]

[0094] wherein: is the biological inhibition intensity value, is the environmental adjustment parameter, represents the environmental factor value. The weight coefficients = 0.6 (biological weight), = 0.25 (adjustment coefficient), = 0.15 (environmental basis). Sigmoid activation function is used to limit the output range. A composite prevention and control field data structure is finally generated, each storage unit contains thirty feature dimensions, and the spatial coverage range completely matches the monitoring area. The timestamp is marked as the UTC time when it is generated. The data compression module uses the Zstandard algorithm for compression storage, and the compression level is set to 12 to balance efficiency and resource consumption. The composite field is stored by a distributed storage system, and each piece of data is attached with a check code for transmission integrity verification.

[0095] Example 4: The system processes the composite prevention and control field data, performs spatio-temporal evolution simulation, field separation, feature reconstruction and iterative verification operations. The entire process relies on a cloud computing platform for operation, and the central processing unit coordinates the interaction of each module. Take a specific example: select a typical citrus planting area as the implementation object. The area is divided into 10,000 grid units (each unit is 10 meters x 10 meters), and the simulation period is set to seven days (168 hours) with a fixed time step of one hour. The initial input data is the composite prevention and control field data, which is a multi-dimensional grid array: each unit stores biological inhibition intensity value (normalized 0-1), environmental regulation parameters (temperature, humidity comprehensive index), and associated confidence value, totaling 30 feature dimensions. The system loads the composite prevention and control field through the data interface, and the format uses HDF5 binary files with a capacity of about 5GB, and transmission is completed through high-speed optical fiber network.

[0096] The spatio-temporal evolution simulation stage starts the model calculation engine. The simulation core is based on the cellular automata architecture: the model initializes the starting state data of the composite prevention and control field (time point T0), and the parameter configuration includes biological action coefficient and environmental factor parameters. The simulation rule is defined as a neighborhood interaction mechanism: the state change of each unit depends on the biological and environmental parameters of its eight adjacent units. For example, when calculating the evolution of grid unit (50, 60), the system scans the neighborhood unit data: obtains the temperature mean value, natural enemy density value, and insect population density baseline value; then applies the rule set to judge the migration behavior, such as the migration probability decreases by 20% when the temperature is above 30°C, and the predation rate increases by 15% when the natural enemy density is high. The model performs a time step loop: the state is updated every hour, and 168 steps are calculated. The system records the state data of each unit at key time points (such as every hour), including the predicted insect population density change value (in percentage form). The simulation outputs the evolved composite field data, including the complete spatio-temporal sequence, and the format is converted to GeoJSON and NetCDF hybrid format for subsequent processing. The simulation process takes about 45 minutes, and GPU cluster is used for accelerated calculation.

[0097] Next, the post-evolution field separation operation is performed. The system inputs the completed composite field data to the separation module. The separation uses a principal component analysis algorithm: the input data is first standardized, reducing the multi-dimensional features to the core dimension. Specifically for the 30 dimensions of the composite field, the covariance matrix is calculated and the eigenvectors are solved; the first three principal components are selected according to the eigenvalue sorting: the first principal component explains biological action, the second principal component explains environmental action, and the third principal component represents noise. The system outputs two independent data fields: the biological action field contains the reduced biological feature vector (core parameters such as natural enemy distribution density); the environmental action field contains the reduced environmental factor vector (core parameters such as heat stress index). The data format is converted to a raster layer, with each cell associated with geographic coordinates and a timestamp. The numerical snapshot of part of the cells after separation (time point is the 24th hour of simulation), see Table 1.

[0098] Table 1: Separated field numerical snapshot table.

[0099] Grid coordinates (x, y) Biological action field intensity Environmental action field value Change in prediction of insect density (%) (50,60) 0.82 0.58 -12.5 (51,60) 0.75 0.63 -11.8 (50,61) 0.78 0.61 -12.0 (51,61) 0.80 0.59 -12.3

[0100] This table is generated by database query and stored in the memory cache area for real-time access.

[0101] The feature reconstruction phase integrates the separated biological action field with real-time monitoring data. Real-time monitoring data comes from the sensor network within the region, with a sampling frequency of once per minute, including real-time insect population density values, temperature and humidity readings. The reconstruction algorithm calls the Kalman filter: input the predicted value of the biological action field and the real-time sensor stream; perform time series alignment, apply the data fusion function after matching to the same space-time coordinates. For example, at the point (50, 60), integrate the biological action value 0.82, the real-time insect population density measurement value 550 per square meter, and the temperature value 28°C; output the reconstructed feature set, including the updated biological activity curve and the environmental impact matrix. After reconstruction, an optimized prevention and control instruction set is generated: the instruction generator analyzes the feature set and maps it to a specific operation command sequence. The instruction format is a JSON array: each instruction contains the operation type, target location (coordinate list), execution time window, and resource requirements. Output example: the instruction set contains 20 commands, which are stored in the central queue after priority sorting and issued to the execution device through the API interface.

[0102] The verification process continuously monitors the results of the instruction set execution. After the system starts the executor, it collects real-time data on the actual change in insect population density: the sensor uploads the density value every 10 minutes, and a time series curve is constructed in the database. The execution cycle covers seven days (168 hours), and after it ends, the actual reduction curve is calculated: the difference between the starting and ending insect population densities is compared. The system has a preset deviation threshold of 10% (adjustable in the configuration file), and uses the absolute deviation calculation formula: the absolute value of the actual reduction minus the predicted reduction divided by the predicted value. When the deviation exceeds 10%, the system triggers the iterative update process: the key diffusion feature extraction module is called again, and an updated feature set is generated based on the new sensor data. Then, the ecological model is input, and parameter adjustment is performed: the loss coefficient or weight matrix is modified. The reinforced prevention and control field is regenerated and fed back to the starting point of the embodiment, and the spatio-temporal evolution simulation, field separation, and reconstruction are repeated. The upper limit of iteration is set to five times: parameter change logs are recorded each time the cycle is repeated; the process is terminated when the deviation value falls within 10%, and the final instruction set is locked and stored permanently. The cycle is managed by event triggers: 20 minutes are required for each round of verification detection; and an alarm code is output when resources are exceeded.

[0103] The resource consumption throughout the process is recorded by monitoring tools: the CPU usage peaks at 85%, and the memory consumption is about 20 GB; data is stored in a distributed system to ensure redundant backup. The system performs atomic transaction processing: any iteration failure is automatically rolled back to the previous stable state. The final output of the optimized instruction set is delivered to the execution unit, forming a closed-loop control system.

[0104] Embodiment 5: see Figure 5 During the operation of the system, a strategy state monitoring and management process is implemented. The monitoring module is configured with a periodic polling mechanism, with a preset of automatically scanning the state data of all active strategy instances every 60 minutes. The state data types include execution progress percentage, resource consumption record, and operation interruption code. These data are transmitted through a structured protocol, formatted as a sequence of key-value pairs, and stored in a log database. The database uses a distributed architecture, with partitioned storage of strategy instance identifiers and timestamps, supporting fast indexed queries. For example, after the instance state of identifier "STRAT-001" is updated, the record contains the progress value, resource usage value, and UTC time stamp. The monitoring process is designed based on an event-driven model: the core event trigger listens to the sensor network stream in real time, including the insect population density detector and the environmental parameter sampler. When new monitoring data triggers an insect situation change event, the trigger starts the state change engine. The event is defined as a change rate of a key monitoring indicator exceeding a preset threshold, typically set at a daily fluctuation amplitude of 10% of the original data value. The detection algorithm uses a difference calculation method: real-time reading of sensor stream data (refreshed every 5 minutes), calculation of the absolute difference between the current value and the value of the previous hour divided by the baseline mean value; if the calculated value is greater than 0.10, an event code "EVENT-507" is generated and sent to the dormant strategy processing queue.

[0105] The hibernation policy instance management adopts a memory resident mechanism. The hibernation state instance is stored in a dedicated memory pool area, and the pool structure is a priority queue, which is sorted according to the policy priority score. Each instance is associated with a set of metadata: policy type code, resource reservation flag, and wake-up waiting counter. When the event trigger outputs the pest situation change event, the system performs the instance wake-up operation. The wake-up function calls the kernel-level restart program: the first step is to unfreeze the hibernation instance state, extract the instance data from the memory pool, and load it to the active process; the second step is to reset the instance internal timer to the initial value; the third step is to modify the instance state flag to "waking up". For example, after the policy instance "SLP-002" is loaded, the original configuration parameters such as the target coordinate set and the operation time window are restored. After the wake-up operation is completed, the system immediately checks the instruction matching conditions.

[0106] The instruction matching check is performed by the buffer pool retrieval module. The instruction buffer pool is built on the message queue system and stores the latest generated optimized prevention and control instruction set. The buffer pool structure is aligned with the policy type code: each policy type corresponds to an independent sub-queue, and the instruction format is a JSON object, which includes an operation command string, a geographic coordinate array, and a validity timestamp. The matching process performs double verification: the first verification checks whether the instruction validity timestamp covers the current time; the second verification compares the consistency of the policy type code and the instruction command type. If it fails any verification, the system determines that no valid instruction is matched. At this time, the instance state automatically switches to the hibernation waiting state: the waiting state manager sets the periodic detection period to poll the buffer pool every thirty minutes; at the same time, the instance resource occupation is frozen; the waiting timeout counter accumulates a value, and the upper limit is set to twenty-four hours. The timeout triggers the automatic re-entry of the hibernation program, and the original state data is preserved.

[0107] When the matching is successful, the system activates the policy execution process. First, modify the instance state flag to "executing", update the log record and send the state broadcast message to the coordinator. The instruction execution unit then receives the parsed command parameters: extract the operation command string, coordinate point set, and resource requirement list. The executor calls the underlying hardware interface: release the resource reservation flag to bind the physical device, and start the operation sequence. For example, execute the coordinate array to drive the drone to fly along the path and drop a specified amount of biological agent at each point. During the execution process, the system collects feedback data every fifteen minutes: the execution result data unit includes the actual completion progress, resource consumption increment, and effect indicators. The data collection adopts a transaction locking mechanism to ensure the integrity of the write.

[0108] The execution result data is fed back to the dynamic ecological characteristic set in real time. The feedback channel is realized through a data stream pipeline: the input end of the pipeline is connected to the execution result unit, and the output end is connected to the characteristic fusion processing module. The data transmission protocol adopts binary encoding, and about one thousand records are processed per second. After receiving the data, the fusion module executes the characteristic update algorithm: compares the new result data with the original set value, and replaces the old parameters by applying the differential update function. Specifically, it includes updating the natural enemy suppression coefficient (based on actual natural enemy activity data) and correcting the host plant resistance index (according to the plant response sensor value). After the update is completed, the dynamic ecological characteristic set triggers the overall system re-entry mechanism: when the characteristic change rate exceeds the limit, the prevention and control strategy evaluation cycle is automatically restarted. The whole process forms a closed-loop control: the monitoring state change drives the awakening and feedback, the feedback data promotes the characteristic iteration, and the system runs adaptively. The resource management adopts a rollback mechanism: when any state is abnormal, it is restored to the previous stable snapshot to ensure the integrity of continuous operation. During the implementation, the CPU load control is within the 70% threshold, and the network bandwidth occupancy is preset to 50Mbps.

[0109] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

Claims

1. A comprehensive ecological control method for citrus psyllids, characterized in that, Includes the following steps: Information on the current pest status is obtained based on real-time monitoring data of citrus planting areas; Biological characteristics and associated ecological environment characteristics of citrus psyllids were extracted from an ecological characteristic database to form an initial set of ecological characteristics; Real-time monitoring data is fused with an initial set of ecological features to generate a dynamic set of ecological features. Based on the spatiotemporal distribution density, migration trend intensity, natural enemy inhibition coefficient, host plant resistance index, and environmental stress factors of each feature in the dynamic ecological feature set, the priority score of each ecological control strategy is calculated. The prevention and control strategies are sorted in descending order of priority scores to generate a sorting table. The current strategies to be executed are selected in turn and it is determined whether they can achieve the target insect population density reduction within the preset period. The extraction of biological characteristics and associated ecological environment characteristics of citrus psyllids from an ecological feature database includes: Individual behavioral characteristics, population dispersal characteristics, natural enemy control characteristics, plant volatile characteristics, and microclimate response characteristics of citrus psyllids were extracted hierarchically from the ecological characteristic database. The individual behavioral characteristics and population diffusion characteristics at the same level are averaged to obtain the primary fusion characteristics; The primary fusion features and microclimate response features are weighted and superimposed to generate secondary fusion features; Iteratively perform hierarchical feature fusion until the highest level feature processing is completed, and output a dynamic ecological feature set containing multi-level relationships; The calculation of the priority score for each ecological control strategy includes: Obtain the proportion of spatiotemporal distribution density, migration trend intensity, natural enemy inhibition coefficient, host plant resistance index, and environmental stress factor corresponding to each strategy; Multiply the migration trend intensity ratio by the first weight coefficient, the natural enemy inhibition coefficient ratio by the second weight coefficient, the host plant resistance index ratio by the third weight coefficient, the environmental stress factor ratio by the fourth weight coefficient, and the spatiotemporal distribution density ratio by the fifth weight coefficient. The priority score is calculated by summing the results of the five products and subtracting the proportion of resource consumption costs. The step of sequentially selecting the current strategy to be executed and determining whether it can achieve the target insect population density reduction within a preset period includes: Calculate the insect population reduction that a single strategy can achieve within a preset period; Compare the achievable insect population reduction with the target insect population reduction; If the achievable quantity is greater than or equal to the target quantity, it is determined that a single strategy can complete the task. If the achievable quantity is less than the target quantity, then calculate the number of additional strategies that need to be executed.

2. The method according to claim 1, characterized in that, The calculation of the number of strategies that need to be additionally executed includes: Divide the target insect population density reduction by the amount achievable by a single strategy, and round up to get the number of strategies to be implemented. The total amount of available ecological resources is determined based on the number of measures implemented according to the strategy. Start the corresponding number of policy instances when sufficient resources are available; When available resources are insufficient, select instances from the started policies that are ranked lower than the current policy to terminate and release resources until the resource requirements are met or there are no more instances to terminate.

3. The method according to claim 1, characterized in that, Also includes: Analysis of pest spread paths based on geographic information system topology; The source areas and migration corridors of insects were determined by gradient field calculations. Extract key diffusion features that satisfy the insect population density threshold and duration of continuous infection; Key diffusion characteristics are input into a pre-trained ecological model to generate a reinforced control field.

4. The method according to claim 3, characterized in that, The step of inputting key diffusion features into a pre-trained ecological model to generate an enhanced control field includes: Introduce topological consistency constraints into the loss function of the ecological model; The matching degree between the control field and the real ecological gradient field is optimized and enhanced through adversarial training; The control field will be spatially coupled with basic environmental factors to generate a composite control field that includes both biological and abiotic factors.

5. The method according to claim 4, characterized in that, Also includes: Spatiotemporal evolution simulation of the composite defense field; The biological action field and the environmental action field after separation and evolution; By reconstructing the characteristics of biological action fields and real-time monitoring data, an optimized set of prevention and control instructions is output.

6. The method according to claim 1, characterized in that, Also includes: Continuously monitor the implementation status of executed strategies; The dormant strategy instance is awakened when new monitoring data triggers a pest change event. If no valid command is found after the policy instance is woken up, it will enter a sleep waiting state; If a valid instruction is matched, the implementation status is changed to "in execution", and the execution result data is fed back to the dynamic ecological feature set.

7. The method according to claim 5, characterized in that, The optimized control instruction set also includes: Verify the actual insect population density change curve after the execution of the instruction set; When the actual reduction amount deviates from the predicted value by more than a threshold, key diffusion features are re-extracted. Iteratively update and strengthen the field parameters for prevention and control until the deviation value falls within an acceptable range.

Images