A multi-source data assimilation rolling prediction method for chilo irridans in hangjiahu rice area

Abstract

The application discloses a kind of Chilo suppressalis multi-source data assimilation rolling prediction methods suitable for Hangjiahu rice area. In view of the defects, such as poor model migration ability, lack of full system dynamic calibration, early warning index and field disaster disconnection and insufficient system robustness in the prior art, the application solves the problem by the following scheme: a population process mechanism model library is established, including a nonlinear development rate model of a divided insect state and a full-generation population process model; data assimilation algorithm is used to dynamically globally calibrate the model parameters of each insect state and the population state variables; a multi-dimensional field risk index of integrated insect quantity stress, lack of natural enemy control and pesticide resistance stress is calculated and graded warning is carried out. Through multiple repeated field verification in 2023-2025, the MAE of the key insect state peak prediction of the application is 0.43 days, the RMSE is 0.61 days, the relative control efficiency reaches 92.9%, the heavy warning hit rate is 95%, the false alarm rate is 5%, the mRI and the rate of withered stem R are 0.91, and the accurate rolling prediction of Chilo suppressalis is realized, which is mechanism sound, robust and can be engineered.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest prediction technology, and in particular to a rolling prediction method for rice stem borer using multi-source data assimilation suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area. Background Technology

[0002] The rice stem borer is a major borer of single-season late rice in the Hangzhou-Jiaxing-Huzhou Plain. Accurate forecasting is crucial for scientific control and reducing yield losses. Existing forecasting technologies mainly include: field observation of larvae, pupae, and egg masses to assess development progress; historical forecasting based on effective accumulated temperature and phenological patterns; physical and chemical trapping using monitoring lights and pheromone traps; and intelligent forecasting methods (BDRW) based on machine learning, Bayesian updates, or the Internet of Things that have emerged in recent years.

[0003] However, under the specific single-season late rice cultivation system and subtropical monsoon microclimate conditions in the Hangzhou-Jiaxing-Huzhou region, the aforementioned existing methods have the following core technical defects:

[0004] The model has weak cross-regional and cross-generational migration capabilities: existing technologies generally use fixed empirical parameters and do not fully consider the biological differences between the overwintering generation and the first generation of rice stem borer, the differences in insect resistance among the main cultivated varieties in Hangzhou, Jiaxing and Huzhou, and the coupled effects of field microclimates, which leads to systematic prediction biases in the model during years of abnormal climate or when the planting structure is adjusted.

[0005] The data utilization level is shallow and lacks the ability to dynamically calibrate the entire system: existing Bayesian update methods only use a single development node as the update target, and core parameters such as egg production, development rate, and larval survival rate are not calibrated synchronously. Once there is a deviation in the prior information, the model lacks the ability to make global corrections.

[0006] The early warning index is disconnected from the actual disasters in the field: The existing risk index does not explicitly integrate multi-dimensional stress factors such as natural enemy control, variety susceptibility to insects, rice growth period, larval survival rate after borer entry, and farmers' specific control habits, resulting in low reliability of the early warning.

[0007] The system lacks robustness and is difficult to implement in engineering: existing solutions lack quality control and fault tolerance mechanisms for AI recognition errors and data gaps, and complex models cannot run stably on basic edge devices.

[0008] Therefore, there is an urgent need for a precise rolling prediction method for rice stem borer that is driven by both mechanism and data, has cross-generational adaptive calibration, deeply integrates multi-source information, and has strong robustness. Summary of the Invention

[0009] The purpose of this invention is to provide a rolling prediction method for multi-source data assimilation of rice stem borers suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area, so as to solve the above-mentioned defects of the prior art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A rolling prediction method for multi-source data assimilation of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area includes:

[0012] S1. Construct a multi-source data acquisition and quality control network.

[0013] S2. Establish a generational population process mechanism model library, including nonlinear development rate models of different stages and full-generation population process models, and complete the initialization of the model parameter set.

[0014] S3. Using a data assimilation algorithm, the population state variables and model parameters are used as assimilation vectors. Multi-source real-world observation data that has undergone quality control is fused together to perform dynamic global calibration on the model parameters and population state variables. Based on the calibration results, the model is driven to predict the future and outputs a daily rolling prediction sequence.

[0015] S4. Based on the calibrated population status, calculate a multidimensional field risk index that integrates insect population stress, lack of natural enemy control, and pesticide resistance stress, and implement graded early warning based on thresholds calibrated from historical disaster damage data.

[0016] Preferably, step S2: the generational mechanistic model library includes

[0017] 1. Model of generational development rate of larval stages

[0018] This invention establishes developmental rate models for the three stages of the insect: egg (e), larva (l), and pupa (p), according to the overwintering generation (G=1) and the first generation (G=2), with temperature and relative humidity as independent variables. The expressions are as follows:

[0019]

[0020] in This is a humidity correction function, applied when the ambient relative humidity is... When humidity falls below the minimum required for development, the development rate is reduced; when humidity is sufficient... Humidity is not a limiting factor. The specific form of the humidity correction function and the values ​​of its parameters are given in detail in the specific implementation method.

[0021] The model parameters (Developmental rate scaling factor) (Developmental onset temperature) (Lethal temperature of high temperature) The (high-temperature boundary layer width) is a generation- and insect-stage-specific parameter. It is determined by fitting historical monitoring data of the target area during the model initialization stage and dynamically optimized and updated by the data assimilation algorithm in step S3 during system operation.

[0022] 2. Full-generation population process model and larval survival function

[0023] The probability of larvae surviving to the voracious stage (after the 3rd instar) Determined by a multifactor function:

[0024]

[0025] The biological meaning and mathematical form of each sub-function are as follows:

[0026] Temperature stress function The survival probability decreases symmetrically with the optimal temperature as the center.

[0027] Variety-Growth Period Interaction Function Characterizing the main cultivated varieties in the Hangzhou-Jiaxing-Huzhou region. (e.g., "Xiushui 134" and "Jia 58") at different developmental stages of rice The success rate of larvae boring into the plant (most susceptible during the tillering stage) increases with... It exhibits a linear change;

[0028] Enemy suppression function : The density of natural enemies such as rice stem borer Trichogramma wasps As the independent variable, it exhibits an exponential decay form, quantifying the negative correlation between the population density of natural enemies and the survival probability of pests;

[0029] Drug resistance effect function Based on the frequency of drug resistance in the population The independent variable is linearly decaying, which clearly quantifies the impact of drug resistance on larval survival.

[0030] 3. Treatment of multiple insect stages and generation overlap

[0031] To address the highly uneven development and severe generational overlap of the overwintering generation of the rice stem borer in the Hangzhou-Jiaxing-Huzhou region, the following processing mechanism was constructed: Developmental progress was discretized into multiple intervals based on physiological age (preferably 20 equal-width intervals); a developmental progress distribution vector for day t was constructed, incorporating the developmental progress distribution of daily oviposition batches within a preset time window (preferably 14 days); the variation in individual developmental rate within the same oviposition batch was assigned a probability distribution (preferably log-normal distribution), with the dispersion of developmental progress variation in the overwintering generation being greater than that in the first generation, accurately characterizing the uneven developmental biological characteristics of the overwintering generation. The statistical time windows for each developmental stage were set based on multi-year field observations in the Hangzhou-Jiaxing-Huzhou region.

[0032] Preferably, step S3: Data assimilation dynamic calibration

[0033] The data assimilation algorithm of this invention integrates all key variables requiring dynamic updates into a state vector, including at least the current effective egg quantity, development rate deviation term, larval survival rate deviation term, and larval equivalent during the voracious feeding phase. The assimilation algorithm employs a recursive filtering method that includes a prediction step and an analysis step.

[0034] Forecasting step: Based on the global calibration results and meteorological driving data of the previous moment, the current population state background field and its error covariance are recursively inferred using the mechanistic model constructed in step S2.

[0035] Analysis Step: Once valid observation data after quality control is obtained, the population state variables are weighted and corrected according to the following form by calculating gain coefficients that are appropriate for observation errors and background errors:

[0036]

[0037] In the formula To analyze the optimal estimation of the field, For background field prediction, For the observation vector, For observation operators (mapping the model state space to the actual field observation space). This is the adaptive gain matrix.

[0038] Under low computing power conditions at the edge, the gain matrix It is automatically simplified to a form that only uses diagonal element operations. Each diagonal gain coefficient is determined by the ratio of the background error variance to the observation error variance of the corresponding state. No matrix inversion operation is required, and millisecond-level operations can be achieved on embedded devices in the field.

[0039] Preferred, data quality control

[0040] Quality control includes a dual criterion based on the confidence level of the observed data itself and the degree of deviation from the background field:

[0041] Confidence-based weight reduction processing: When the confidence of the observation data source (such as AI image recognition of intelligent insect monitoring lamp) is lower than the first threshold, the observation data is assigned a weight reduction processing lower than the standard weight, so that it participates in subsequent calculations with reduced weight; when the confidence further decreases to a lower second threshold, it is marked as questionable data and manual review is triggered.

[0042] Outlier detection based on background field deviation: When the deviation of an observed value from its predicted background field value exceeds a preset multiple of the standard deviation, it is identified as an outlier and automatically removed; the gaps generated after removal are filled by an interpolation method based on the nearest neighbor of the time series.

[0043] The specific values ​​of the above thresholds and the parameter settings of the interpolation method can be adjusted according to the monitoring area, equipment accuracy, and data characteristics during implementation.

[0044] In this solution, data processing adopts the following deployment method: cloud-edge collaborative deployment.

[0045] The cloud-edge collaborative architecture adopts a layered computing model: the edge is responsible for local data collection, real-time quality control, simplified data assimilation calculations, and short-term early warning generation; the cloud is responsible for global assimilation calculations based on multi-region aggregated data, batch optimization of model parameters, and long-term historical data storage and management. When the network connection is normal, data synchronization and parameter exchange are performed according to a preset time period; when the network is interrupted, the edge independently runs a local simplified model, maintaining basic prediction services with locally stored historical data and the most recently synchronized parameters, and automatically completes data retransmission and parameter synchronization after the network is restored.

[0046] Preferably, step S4: Multidimensional Risk Index (mRI) and Tiered Early Warning

[0047] mRI calculation formula

[0048]

[0049] The biological logic behind each term in the formula is explained below:

[0050] The ratio of the larval equivalent during the voracious feeding stage to the dynamic economic threshold shows that the higher the larval quantity, the higher the risk, thus contributing positively.

[0051] The degree of absence of natural enemy density relative to the normal baseline density. The more abundant the natural enemies, the closer this item is to zero. The more scarce the natural enemies, the larger this item is. It accurately reflects the positive increment of risk caused by insufficient natural enemy control.

[0052] The increase in the frequency of drug resistance in a population relative to the baseline frequency indicates that the stronger the resistance, the worse the control effect and the higher the risk, thus making a positive contribution.

[0053] The weight coefficients of each stress factor are objectively determined by analyzing historical disaster data of the target area using structural equation modeling or hierarchical analysis methods, and the sum of the weights is 1. The weight values ​​used in a preferred embodiment are given in the specific implementation details.

[0054] Objective calibration of tiered early warning thresholds: Collect multi-year field disaster event records for the target area. Use a preset crop damage rate threshold (e.g., a disaster event where the sheath blight rate exceeds a specific percentage) as a positive sample. Plot the receiver operating characteristic (ROC) curve of the mRI value on this sample set. Select the tiered threshold based on a comprehensive criterion that maximizes the correct early warning rate and minimizes the false alarm rate (e.g., the maximum value of the Youden index). Divide the risk index value space into multiple early warning levels (e.g., no warning, moderate warning, and severe warning). A preferred embodiment of the threshold scheme is given in the specific implementation details.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] 1. Significantly improved prediction accuracy: Verified by three years of fixed-point repeated experiments in townships under the jurisdiction of Jiaxing City from 2023 to 2025, the prediction of the peak of key insect stages of this invention has significantly improved the error compared with the national standard effective accumulated temperature method; the prediction error of the peak hatching of overwintering eggs has been greatly reduced compared with the traditional method.

[0057] 2. The early warning system is highly consistent with field disasters: the sign direction of each stress factor in the mRI risk index strictly follows biological logic. After multiple rice stem borer outbreaks, the hit rate of the serious warning reached 95%, and the false alarm rate was only 5%. Compared with the traditional method, the hit rate was significantly improved and the false alarm rate was significantly reduced. The mRI and the field measured sheath dead rate are also significantly better than the national standard method (accumulated temperature method).

[0058] 3. Strong adaptability to extreme climates: Under three extreme climate scenarios, namely, years with low temperature and low sunshine, years with continuous high temperature and heat damage in summer, and years with periodic drought stress in the field, the prediction error of this invention does not exceed 1.5 days. Compared with the large error of the traditional accumulated temperature method, the prediction ability has been significantly improved and enhanced.

[0059] 4. Significant control benefits: Guided by this invention, the number of times pesticides are used can be reduced compared to traditional methods, but the insect control effect and target achievement rate are increased, achieving the green control goal of "reducing pesticides and increasing efficiency".

[0060] 5. Extremely high engineering robustness: The dual-criteria quality control mechanism and the degradation strategy in the cloud-edge collaborative architecture work together to maintain the prediction error basically stable even under extreme conditions of limited data loss, and the system remains available, fully verifying the field engineering implementation capability of this invention. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall architecture of the two-generation prediction system for rice stem borers according to the present invention.

[0062] Figure 2 The temperature development rate response curve for the diagenetic stage Logan is shown.

[0063] Figure 3 This is a logic diagram for lightweight data assimilation and iterative computation.

[0064] Figure 4 The curves show the comparison between the measured values ​​of insect pests in the field and the predicted values ​​of the model over three years.

[0065] Figure 5 This is a schematic diagram illustrating the threshold division for the multidimensional risk index (mRI).

[0066] Figure 6 A bar chart comparing the prediction error indices of different control models.

[0067] Figure 7 The figure shows the test results of robustness to perturbation of model parameters.

[0068] Figures 8a-8c A comparative chart showing the predicted population occurrence of the rice stem borer under extreme climatic conditions. Detailed Implementation

[0069] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0071] Before proceeding with the embodiments, the core mathematical model and algorithm of this invention are first formally defined to ensure that those skilled in the art can reproduce all technical solutions. The specific parameter values ​​below are preferred embodiments calibrated based on historical monitoring data from the Jiaxing Plant Protection Station and do not constitute a limitation on the scope of protection of the claims.

[0072] I. Model of Generational Development Rate of Different Stages

[0073] In one specific implementation, developmental rate models were established for the three stages of the rice stem borer: egg (e), larva (l), and pupa (p), based on the overwintering generation (G=1) and the first generation (G=2). These models used environmental temperature as a reference. and relative humidity Development rate is the independent variable. The dependent variable is expressed as:

[0074] Among them, humidity correction function The following segmentation format can be used:

[0075] In a preferred parameter setting, the minimum humidity required for development is... Take 70%, humidity response coefficient The value is set to 0.1. When the relative humidity is below 70%, the development rate decreases linearly; when the humidity reaches or exceeds 70%, humidity does not constitute a limiting factor. .

[0076] The optimal values ​​for the generational parameters of each insect developmental stage in the above model were calibrated using indoor constant-temperature rearing experimental data from the Jiaxing Plant Protection Station from 2018 to 2022. The experiment included 15 constant-temperature gradients (12.0℃ to 37.0℃, with a step size of approximately 2℃), with 200 insects tested at each gradient. The Levenberg-Marquardt nonlinear least squares algorithm was used for parameter fitting, and the goodness-of-fit for each generational stage was determined. The residuals range from 0.86 to 0.94, and the residuals conform to a normal distribution according to the Shapiro-Wilk test (P>0.05). The coefficient of variation (CV) using the Jackknife leave-one-out cross-validation method is <15%, meeting the stability requirements. Optimal parameter values ​​are shown in the table below:

[0077] 0.042 / 0.048 0.018 / 0.022 0.035 / 0.040 11.8 / 12.0 12.9 / 13.2 12.7 / 12.9 36.0 / 37.0 38.0 / 39.0 37.0 / 38.0 2.5 / 2.5 3.2 / 3.2 3.0 / 3.0

[0078] During system operation, the above model parameters (Developmental rate scaling factor) (Developmental onset temperature) (Lethal temperature of high temperature) The (high-temperature boundary layer width) is treated as part of the state variables and dynamically optimized and updated by subsequent data assimilation algorithms, thereby enabling the model to have adaptive calibration capabilities.

[0079] II. Full-generation population process model and larval survival function

[0080] In one specific implementation, this technical solution constructs a full-generation population process model encompassing the entire process of tandem egg hatching, larval development at each instar, transfer to a new strain, pupation, and emergence. The probability of larvae surviving to the voracious feeding stage (i.e., developing to the 3rd instar and beyond) is included. It is determined by the following multifactor function:

[0081] The specific expressions and coefficients of each sub-function are as follows:

[0082] Temperature stress function:

[0083] In a preferred parameter setting, the temperature response coefficient Take 0.15, optimal temperature The temperature was set to 28℃. This function symmetrically reduces the larval survival probability when the ambient temperature deviates from 28℃. The above coefficients were determined by fitting experimental data on larval survival rates under eight isothermal gradients within the range of 15℃ to 35℃.

[0084] For the interaction function of variety-growth period The DVS can be expressed as a linear change with the rice's growth stage:

[0085]

[0086] in This is the code for the rice development stage, specifically the peak tillering stage ( =0.3 to 0.5) is the most susceptible period for rice stem borer larvae. The coefficient for the main cultivated varieties in the Hangzhou-Jiaxing-Huzhou region was determined through a field inoculation trial conducted in Jiaxing City from 2020 to 2021, involving four varieties at three transplanting periods (each treatment inoculated with 200 newly hatched larvae). The coefficient for the variety "Xiushui 134" was taken as... , ; variety "Jia 58" , .

[0087] For the enemy suppression function It can be expressed in an exponential decay form:

[0088]

[0089] in The field parasitism rate (%) of natural enemies, represented by the rice stem borer Trichogramma wasp, on the rice stem borer egg masses can be obtained by surveying every 3 days; the natural enemy inhibition coefficient d is taken as 0.02, which is determined by regression analysis of field parasitism rate and larval survival rate from 2019 to 2021.

[0090] For the influence function of drug resistance It can be implemented using a linear decay form:

[0091]

[0092] in The resistance frequency of rice stem borer to chlorantraniliprole (based on a diagnostic dose of 1 μg / mg insect weight), i.e., the proportion of surviving individuals, can be obtained by uniformly measuring overwintering larvae annually; resistance effect coefficient. The value was set to 0.8, which was derived from the dose-response curve of resistance monitoring data and field efficacy from 2019 to 2022.

[0093] III. Treatment of Multiple Insect Stages and Generation Overlap

[0094] In response to the highly uneven development and severe generational overlap of the overwintering generation of the rice stem borer in the Hangzhou-Jiaxing-Huzhou region, this technical solution introduces a multi-generational overlap treatment mechanism into the population process model.

[0095] In one specific implementation, the developmental progress is discretized into n=20 equally wide intervals based on physiological age, with each interval corresponding to approximately 5% of the developmental progress. A developmental progress distribution vector for day t is then constructed. Each element represents the number of worms whose physiological age falls within the corresponding range. Daily data is included from the past. The developmental progress of each oviposition batch within a day is analyzed by superimposing the contributions of each batch to obtain the distribution of multiple insect life stages coexisting in the field on that day.

[0096] The variation in developmental rate within the same spawning batch can be described using a log-normal distribution. Specifically, the standard deviation of the developmental progress variation of the overwintering generation... Take 0.25 as the standard deviation of the first generation. We set the value to 0.12. For the overwintering generation, we used a larger dispersion value to accurately characterize its uneven developmental biological characteristics.

[0097] The statistical time windows for each insect developmental stage can be set as follows: egg quantity is contributed by all batches of unhatched developmental stages within the past 8 days; young larvae quantity is contributed by egg-hatched developmental stages within the past 20 days; voracious larvae quantity is contributed by developmental stages within the past 35 days; and pupae quantity is contributed by pupating stages within the past 50 days. The specific number of days in the above time windows is determined based on long-term field observation data in the Hangzhou-Jiaxing-Huzhou area and can be appropriately adjusted according to actual monitoring conditions.

[0098] IV. Data Assimilation Algorithms

[0099] In one specific implementation, a recursive filtering method including a prediction step and an analysis step is used for data assimilation. All key variables requiring dynamic updates are integrated into a state vector, defined as follows:

[0100]

[0101] in, The effective number of eggs in the current period (eggs / acre). This is the global deviation term for developmental rate (dimensionless, initial value 1.0). This is the larval survival rate deviation term (dimensionless, initial value 1.0). The equivalent number of voracious larvae (heads / acre).

[0102] During the algorithm initialization phase, the following parameters can be set: initial value of the state vector. Initial error covariance matrix ,in It is a 4×4 identity matrix; the process noise covariance matrix ; Observation noise covariance matrix , where m is the number of valid observed variables for the day, which is dynamically adjusted according to the valid observed data.

[0103] Observation Operator H is an m×4 matrix used to map a 4-dimensional state space to an m-dimensional observation space. Taking the simultaneous acquisition of two observations (m=2) – the number of female moths attracted by lights and the density of larvae found in the field – as an example, one construction form of the H matrix is:

[0104]

[0105] Among them, elements The mapping coefficient between the number of female moths under the light and the effective egg quantity is given. This value can be calibrated through a label-release-recovery experiment, with a 95% confidence interval of [0.62, 0.78]. The field larval density is used to measure the larval equivalent of the voracious larvae. This value can be determined by comparing it with the full larval density, and its 95% confidence interval is [0.78, 0.92].

[0106] In the forecast step, based on the global calibration results and meteorological driving data from the previous moment, the current population state background field and its error covariance are recursively derived using a mechanistic model:

[0107]

[0108] in For the 24-hour recursive operator of the whole-generation population process model, This is hourly weather data from the previous day. for exist The 4×4 Jacobian matrix at a given location can be calculated daily using the numerical difference method.

[0109] In the analysis step, once quality-controlled valid observation data is obtained, the population state variables are weighted and corrected by calculating gain coefficients that are appropriate for observation and background errors. Under cloud-based or sufficiently powerful computing conditions, a complete set approach can be used.

[0110]

[0111]

[0112] Under low computing power conditions at the edge, the above analysis steps can be automatically simplified to a form using only diagonal element operations. Specifically, for the i-th element of the state vector, its gain coefficient and update formula are:

[0113]

[0114] This simplified form requires only four scalar division and multiplication operations. Actual testing on an ARM Cortex-A72 processor (1.5GHz) showed that the computation time for a single diagonal simplified assimilation step is approximately 0.3 milliseconds.

[0115] V. Data Quality Control

[0116] In one specific implementation, data quality control includes a dual criterion based on both the confidence level of the observed data itself and the degree of deviation from the background field.

[0117] For confidence-based weighting, the rules can be set as follows: when the classification confidence of the intelligent insect pest monitoring lamp's AI image recognition... When the confidence level reaches or exceeds 0.80, the observation data is used in the calculation with a standard weight of 1.0; when the confidence level... When the confidence level is between 0.60 and 0.80, the observation data is assigned a reduction weight of 0.3; when the confidence level is... If the value is below 0.60, the observed data is marked as questionable, its weight is set to zero, and a manual review process is triggered. The specific values ​​of the above thresholds and weight coefficients can be adjusted according to the monitoring area, equipment accuracy, and actual data quality.

[0118] For outlier detection and imputation based on background field deviation, the rules can be set as follows: For each observation... Calculate the deviation between the predicted value and the background field value. ,in If the standard deviation of the observations at the same time over the past 7 days is used, the observation is considered an outlier and is removed. Data gaps resulting from removal are filled using an interpolation method based on the nearest neighbor in the time series. In a preferred implementation, the mean of valid observations at the same time three days before and after can be calculated as the filler value (i.e., K-nearest neighbor time interpolation, K=3). The value of K can be adaptively adjusted within the range of 2 to 5 based on the monitoring point density.

[0119] VI. Multidimensional Risk Index

[0120] In one specific implementation, the field real risk index The formula for calculation is:

[0121]

[0122] In the formula, This is the standard normal cumulative distribution function, used to map the comprehensive stress score to the interval between 0 and 1; The predicted number of voracious larvae (heads / acre) on day d. The optimal value for the dynamic economic threshold during the peak tillering period is 120 heads / mu; This is the ratio of the current predator density to the baseline density in a normal year; This is the ratio of the current population's drug resistance frequency to the baseline frequency.

[0123] Weighting coefficients of each stress factor The weights can be objectively determined by analyzing historical disaster data of the target area using structural equation modeling or analytic hierarchy process (AHP), with the sum of weights being 1. In a preferred embodiment, AHP is used to determine the weights. Specifically, a third-order judgment matrix is ​​constructed (using the 1-9 scale, with the geometric mean obtained after independent scoring by three plant protection experts), the eigenvectors are calculated and normalized to obtain the weight values: (Weight of damage caused by larvae in the voracious feeding stage) (Weight of control over natural enemies) (Weight of drug resistance control failure). The consistency test result is: maximum eigenvalue. Consistency indicators Random consistency ratio If the test threshold is less than 0.10, the consistency test is passed.

[0124] VII. Tiered Early Warning Thresholds

[0125] The grading of early warning thresholds can be achieved by collecting field disaster event records for the target area over many years and distinguishing between positive and negative samples using a preset crop damage rate threshold. In a preferred embodiment, events with a sheath blight rate exceeding 3% are considered positive samples. Receiver operating characteristic (ROC) curves of the mRI values ​​on this sample set are plotted, and the grading threshold is selected based on a comprehensive criterion that maximizes the correct warning rate and minimizes the false alarm rate (such as the maximum value of the Youden index).

[0126] Based on ROC analysis of 96 rice stem borer occurrence events (52 positive and 44 negative) in Jiaxing from 2018 to 2022, the preferred grading scheme is as follows: Set to "No Alarm" level; it is recommended to refrain from administering medication for the time being. Set to "Medium Alert" level; it is recommended to closely monitor changes in dead sheaths in the field within 3 days. The alert level is set to "Severe Alert," and immediate application of pesticides is recommended. The severe alert threshold of 0.50 corresponds to a maximum Youden's Index of 0.86.

[0127] VIII. Summary of Optimized Parameters

[0128] To facilitate understanding and reproduction, the following list presents a set of preferred parameter values ​​for each of the core algorithms described above. These parameters together constitute a complete set of preferred implementation methods.

[0129] (1) Developmental rate model parameters:

[0130] Egg / Larva / Pupa Egg / Larva / Pupa Developmental rate scaling factor 0.042 / 0.018 / 0.035 0.048 / 0.022 / 0.040 Developmental start temperature ℃ 11.8 / 12.9 / 12.7 12.0 / 13.2 / 12.9 High temperature that can kill ℃ 36.0 / 38.0 / 37.0 37.0 / 39.0 / 38.0 High-temperature boundary layer width ℃ 2.5 / 3.2 / 3.0 2.5 / 3.2 / 3.0

[0131] (2) Humidity correction function parameters

[0132] Minimum humidity required for development 70 % Common to all insect states / generations Humidity response coefficient 0.1 — The value can be adjusted between 0.08 and 0.12 depending on the ecological zone.

[0133] (3) Larval survival function parameters

[0134] 15~35℃ constant temperature feeding experiment Field parasitism rate regression analysis Dose-response curve derivation

[0135] (4) Parameters of the interaction function between variety and growth period

[0136] Xiushui134 0.12 0.35 0.45 0.40 Jia58 0.10 0.40 0.48 0.42

[0137] (5) Generation overlap processing parameters

[0138] Number of discrete intervals for development progress n 20 5% development progress per interval Include in time window 14 days Daily summation of spawning batches from the past 14 days Developmental dispersion of overwintering generation 0.25 Log-normal distribution standard deviation First-generation developmental dispersion 0.12 Log-normal distribution standard deviation

[0139] (6) Data assimilation parameters

[0140] Initial error covariance 4×4 unit array Process noise covariance Reflecting the uncertainty of model recursion Observation noise covariance m is the number of effective observed variables. Mapping coefficient of female moth under lamp 0.7 Larval exfoliation efficiency coefficient 0.85

[0141] (7) Quality control parameters

[0142] AI confidence high threshold 0.80 Q≥0.80 Normal weight (1.0) AI confidence threshold 0.60 Q < 0.60 is flagged as questionable (w=0) weighting coefficient 0.3 Enabled when Q < 0.80. Outlier detection factor Standard deviation of the observed values ​​for the same period over the past 7 days K-nearest neighbor K value 3 Adaptive adjustment (2~5)

[0143] (8) mRI risk index parameters

[0144] Weight 0.55 Damage weight of voracious larvae Weight 0.25 The lack of control over natural enemies is a significant factor. Weight 0.20 Weight of drug resistance control failure Weighted Consistency Ratio (CR) 0.016 <0.10, passed the test. Leveling threshold: No alarm / Medium alarm boundary 0.20 The corresponding Youden Index is 0.83. Classification threshold: Medium alert / Severe alert boundary 0.50 This corresponds to the maximum value of the Youden Index, 0.86. Economic threshold (Peak tillering period) 120 head / mu

[0145] (9) Cloud-Edge Collaboration Parameters

[0146] Edge processor clock speed ≥1.5GHz ARM architecture Edge memory ≥4GB RAM Data synchronization cycle Daily 00:00 24 hours / time Future rolling forecast days 7 days Output step size 1 day Confidence interval 95% 1.96 × standard deviation

[0147] Example

[0148] The following specific field example further verifies and illustrates the technical effects of the present invention.

[0149] I. Basic Experimental Setup

[0150] 1.1 Experimental Area and Duration

[0151] The technical solution of this invention involves a three-year (2023, 2024, and 2025) fixed-point repeated experiment in the main rice-producing areas under the jurisdiction of Jiaxing City (Xincheng Town, Xiuzhou District, Jiaxing City; Xujiazhuang Village, Haining City; and Shuangfeng Village, Yuanhua Town, Haining City, Zhejiang Province). The experimental period fully covers the overwintering generation and the entire first generation of the rice stem borer. Regional characteristics: Located in the hinterland of the Hangjiahu Plain, at an altitude of 3-5m, with a high groundwater level, a dense river network, and a typical subtropical monsoon climate.

[0152] 1.2 Rice Varieties and Cultivation Management

[0153] type conventional late-season japonica rice conventional late-season japonica rice Throughout the reproductive period 152 days 150 days Resistance to rice stem borer Infected (moderate to high risk) Infected (moderate) Transplanting period May 18-20 May 18-20 Row spacing 30cm×16cm 30cm×16cm Community area 66.7㎡ (8.33m×8.0m) 66.7㎡ Community protection 1.0m wide, same variety of rice Tongzuo

[0154] II. Experimental Grouping and Evaluation System

[0155] 2.1 Multiple parallel control group setup

[0156] The experiment set up multiple parallel control systems, and standardized the collection of field micro-meteorological data, insect trapping observation data, rice growth period survey data, and field damage survey data. Three field biological replicates were set up simultaneously for each experimental plot.

[0157] The designs for each group are as follows:

[0158] The experimental group of this invention fully utilizes the generational mechanism model library, data assimilation dynamic calibration system, and multidimensional risk index mRI described in this invention to guide drug use.

[0159] Existing technology control group 1 (1-accumulated temperature method): The effective accumulated temperature period method in the national standard is used to predict and guide the application of pesticides.

[0160] Existing technology control group 2 (p. 2-BDRW): A prediction scheme based on Bayesian dynamic updates was adopted, with fixed parameters and updates only on the peak day of egg hatching to guide drug administration.

[0161] The existing technology control group 3 (pair 3 - single temperature driven population model): uses a single temperature driven (without humidity correction) population process model to predict and guide pesticide application, without performing multi-generational overlap and superposition calculations, and without performing data assimilation dynamic calibration.

[0162] The control group (CK) was used to observe the dynamics of the natural population and establish a disaster baseline, without any rice stem borer control measures during the entire growth period.

[0163] 2.2 Performance Evaluation and Quality Inspection System

[0164] Mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) were used as core quantitative evaluation indicators. Analysis of variance (ANOVA) was used to test the significance of differences between groups, and the overall model fit was verified by combining residual distribution analysis and residual normality test (Shapiro-Wilk test).

[0165] Robustness tests include: small-range perturbation test of core parameters (random perturbation is applied to multiple core parameters of the development rate model within ±15% of the calibration value), random noise superposition test of input data (Gaussian white noise with different signal-to-noise ratios is superimposed on three types of observation data respectively), and local data missing simulation test (different proportions of observation data points are randomly deleted to test the system's interpolation and degradation strategy recovery capabilities).

[0166] III. Basic Observation Data for Each Year and Each Day

[0167] Table 1-1. Daily baseline observation data of overwintering generation in Xiushui 134 blank group, Xincheng Town, Xiuzhou District, 2023

[0168] June 5 3 18 5 0.2 15 23.5 78 June 8 8 68 18 0.5 38 24.2 82 June 12 22 127 55 1.4 62 25.8 80 June 15 (peak hatching rate of 50%) 50 210 102 2.8 85 26.3 72 June 18 73 285 156 4.1 44 27.0 76 June 22 91 342 228 5.6 18 27.5 70 June 26 98 365 290 7.2 7 28.1 74

[0169] Table 1-2. Daily baseline observation data of the overwintering generation in the Xujiazhuang 58 blank group in Haining City in 2024.

[0170] June 4 2 15 4 0.2 12 23.8 80 June 8 7 60 16 0.4 35 24.5 83 June 12 18 110 48 1.2 55 26.1 78 June 15 45 190 88 2.5 72 26.8 73 June 18 (peak hatching rate of 50%) 68 260 135 3.7 40 27.3 75 June 22 88 310 205 5.1 16 27.9 71 June 26 96 332 262 6.8 6 28.3 74

[0171] Table 1-3. Daily baseline observation data of the overwintering generation in Xiushui 134 blank group, Shuangfeng Village, Yuanhua Town, Haining City, in 2025.

[0172] June 6 3 20 6 0.3 18 24.0 76 June 9 9 75 20 0.6 42 24.8 81 June 12 24 135 58 1.5 68 25.5 79 June 16 (peak hatching rate of 50%) 55 225 110 3.0 90 26.6 72 June 19 76 300 168 4.3 47 27.1 75 June 23 93 360 245 6.0 20 27.8 68 June 27 99 388 305 7.8 8 28.2 73

[0173] Table 2. Summary of complete data for each generation and variety of the three-year blank group

[0174] 2023-overwintering generation Xiushui134 1 / 2 / 3 June 15 / 15 / 14 365 / 378 / 352 290 / 305 / 275 8.2 / 8.5 / 7.9 12 / 10 / 14 0.35 2023 - Generation Xiushui134 1 / 2 / 3 July 28 / 29 / 28 420 / 445 / 398 350 / 372 / 330 9.1 / 9.8 / 8.7 8 / 7 / 9 0.38 2024-overwintering generation Jia58 1 / 2 / 3 June 13 / 13 / 14 340 / 358 / 332 268 / 285 / 260 7.5 / 7.8 / 7.3 15 / 13 / 16 0.32 2024 - Generation Jia58 1 / 2 / 3 July 26 / 27 / 26 385 / 410 / 370 315 / 340 / 300 8.4 / 8.9 / 8.1 10 / 9 / 11 0.36 2025-overwintering generation Xiushui134 1 / 2 / 3 June 17 / 17 / 16 388 / 402 / 375 305 / 325 / 298 8.8 / 9.2 / 8.5 11 / 10 / 12 0.40 2025 - Generation Xiushui134 1 / 2 / 3 July 30 / 31 / 29 450 / 478 / 432 378 / 402 / 360 10.1 / 10.5 / 9.6 6 / 5 / 7 0.42

[0175] It should be noted that the natural enemy parasitism rate is calculated based on the natural parasitism rate of rice stem borer Trichogramma on rice stem borer egg masses in the field; the resistance frequency Rm is expressed as the proportion of surviving individuals at the diagnostic dose of chlorantraniliprole (1 μg / mg insect weight).

[0176] IV. Model Parameter Calibration and Sensitivity Analysis

[0177] 4.1 Parameter Calibration Process

[0178] Taking the calibration of developmental rate model parameters as an example, the complete process is as follows:

[0179] Data on indoor constant-temperature rearing of insects were collected from Jiaxing Plant Protection Station from 2018 to 2022 (15 constant-temperature gradients, 200 test insects per gradient).

[0180] Calculate the development rate (reciprocal of developmental duration) of each insect stage at each temperature.

[0181] The Levenberg-Marquardt algorithm was used to fit the model parameters, and the initial values ​​were set based on empirical values ​​from the literature.

[0182] The goodness of fit was evaluated using the coefficient of determination R² and the root mean square error RMSE.

[0183] Cross-validation was performed using the Jackknife leave-one-out method (each time a temperature gradient data point was removed and refitted, the coefficient of variation of the prediction error was calculated to be <15% to indicate that the stability was acceptable).

[0184] 4.2 Parameter Sensitivity Analysis

[0185] The Sobol global sensitivity analysis method was used to analyze multiple core parameters of the developmental rate models for each insect stage. First-order and total-effect sensitivity analyses were performed, with developmental rate r as the output variable and a sample size of N=10000 (Saltelli sampling method). The results are as follows:

[0186] 0.38 / 0.45 0.32 / 0.40 0.35 / 0.42 0.28 / 0.33 0.25 / 0.30 0.27 / 0.32 0.12 / 0.15 0.18 / 0.22 0.15 / 0.18 0.08 / 0.10 0.10 / 0.13 0.09 / 0.11 (Humidity coefficient) 0.05 / 0.07 0.06 / 0.08 0.05 / 0.07

[0187] Analysis shows that: developmental rate scaling factor The developmental threshold temperature (#imgpt167#) is the two parameter with the greatest impact on the output (total first-order sensitivity > 0.60). The estimation accuracy of these two parameters should be a key focus during model parameter initialization and data assimilation. The humidity coefficient (#imgpt168#) has a first-order sensitivity < 0.07, indicating that the model has low sensitivity to humidity correction within the current parameter value range, demonstrating good robustness.

[0188] V. Detailed Experimental Results and Statistical Tests

[0189] 5.1 Comparison of Three-Year Average Forecast Accuracy Indicators

[0190] For 1 (accumulated temperature method) 2.89a 3.62a 42.1a 0.032* (non-normal) For 2 (BDRW) 1.56b 2.18b 22.8b 0.089 (normal) For 3 (single temperature driven) 1.72b 2.45b 25.3b 0.078 (normal) This invention (experimental) 0.43c 0.61c 8.2c 0.215 (normal)

[0191] In this paper, #imgpt170# < 0.05 indicates that the residuals do not follow a normal distribution, and the model may have systematic bias. Different letters indicate significant differences in the Tukey HSD test (#imgpt171# < 0.05). The residuals of this invention follow a normal distribution, indicating that the model has fully captured the systematic variation of population dynamics, and the remaining residuals are random errors.

[0192] 5.2 Comparison of prediction errors for peak stages of key insect life forms (days, mean ± standard deviation)

[0193] For 1 +2.1±1.5a +3.5±2.8a +2.8±2.0a +4.1±3.3a 2 +0.8±1.2b +1.9±2.1b +1.5±1.6b +2.5±2.7b 3 +1.1±1.4b +2.3±2.3b +1.8±1.8b +2.9±2.9b This invention +0.2±0.5c +0.6±0.8c +0.4±0.7c +0.9±1.1c

[0194] 5.3 Comparison of Early Warning Effect and Prevention Effect

[0195] CK 0 8.5±0.7a — — — For 1 2.0±0 2.5±0.4b 70.6b 65 35 2 1.7±0.5 1.5±0.3c 82.4c 80 20 3 1.8±0.4 1.7±0.3c 80.0c 75 25 This invention 1.3±0.5 0.6±0.2d 92.9d 95 5

[0196] 5.4 Correlation between mRI risk index and sheath dryness rate

[0197] For 1 y = 0.45x + 1.12 0.38 1.85 <0.05 2 y = 0.72x + 0.58 0.62 1.12 <0.01 3 y = 0.55x + 0.95 0.48 1.52 <0.05 This invention y = 1.01x + 0.05 0.91 0.42 <0.001

[0198] 5.5 Robustness Test Results

[0199] (1) Core parameter disturbance test

[0200] Random perturbations of ±5%, ±10%, and ±15% were applied to multiple core parameters of the developmental rate model, with each perturbation level repeated 100 times in Monte Carlo simulations.

[0201] ±5% 0.3±0.5 0.2±0.4 2.1±1.5 ±10% 0.5±0.8 0.3±0.6 4.5±2.8 ±15% 0.9±1.2 0.5±0.9 8.2±4.5

[0202] (2) Input data noise superposition test

[0203] Egg hatching peak error (days) 0.3±0.5 0.6±0.8 1.2±1.5 Equivalent error (%) of larvae in the voracious feeding stage 5.2±3.1 12.8±7.5 28.5±15.2

[0204] (3) Simulation of missing local data

[0205] Peak error (days) 0.2±0.5 0.4±0.6 0.5±0.8 1.8±2.1 System status Normal operation Normal operation Automatic degradation operation Degraded operation

[0206] 5.6 Comparison of the effects of generational overlap processing mechanisms (data from 2023-2025)

[0207] This invention (discrete interval n=20, log-normal variation) 0.31 0.52 ±0.7 days Control: Fixed timeframe overlay (no developmental progress distribution) 1.23 1.86 ±2.4 days Control: Only mean developmental progress (no variation) was used. 0.98 1.45 ±1.9 days

[0208] VI. Verification under extreme climate scenarios

[0209] Based on three years of conventional climate field trials, in order to fully verify the regional adaptability and stability under extreme conditions of the model of this invention, a multi-climate year group comparison verification was carried out.

[0210] 6.1 Classification of Extreme Climate Years

[0211] Based on historical meteorological monitoring data of Jiaxing area over the past ten years (2014-2023), the following three representative extreme climate types are objectively classified:

[0212] Low temperature and low light delayed year 2019 The average temperature in April and May was 2.1℃ lower than normal, and the sunshine duration decreased by 35%. The overwintering generation experiences delayed development, with the peak hatching period postponed by 5-8 days. Years of persistent high temperatures and heat damage in summer 2022 The number of days with temperatures ≥35℃ in July and August reached 42, 18 more than the average for this time of year. First-generation larval development is inhibited, and the mortality rate due to high temperatures increases. Field-stage drought stress year 2018 Rainfall in June and July was 52% lower than normal, and relative humidity in the fields remained below 60%. The egg hatching rate decreased, and the larvae's success rate of borer entry decreased.

[0213] 6.2 Validation Results of Extreme Climate Scenarios

[0214] Low temperature and low light delayed year Egg hatching peak error (days) +7.2 +2.5 +3.1 +0.9 Occurrence prediction R² 0.35 0.58 0.48 0.84 High temperature heat damage year Peak larval feeding period error (days) +5.8 +3.2 +2.8 +1.2 Occurrence prediction R² 0.28 0.52 0.45 0.81 Periodic drought years Egg hatching peak error (days) +3.5 +1.8 +4.2 +1.5 Occurrence prediction R² 0.42 0.61 0.38 0.78

[0215] VII. Engineering Implementation Architecture

[0216] 7.1 Hardware Deployment Plan

[0217] Perception layer Intelligent insect monitoring lamp IMS-100 (365nm wavelength, effective radius 50m) 1 unit / village Center of the field Perception layer Automatic counting sex trap Contains rice stem borer pheromone lure (effective for 30 days) 3 per monitoring point equilateral triangle layout Perception layer Field microclimate station Temperature and humidity accuracy ±0.3℃ / ±3%, data collected every 10 minutes. 1 set / village Field ridge (1.5m high) Edge computing layer Embedded ARM processor Cortex-A72 1.5GHz, 4GB RAM, 64GB eMMC 1 unit / village Village monitoring station cabinet Communication layer 4G / 5G communication module CAT-4 LTE module 1 / edge device Integrated into edge processor cloud server 8-core CPU 2.5GHz, 32GB RAM, GPU acceleration 1 unit County Plant Protection Station / Cloud Data Center

[0218] 7.2 Software Module Architecture

[0219] Data acquisition and quality control module edge end Image acquisition, AI inference, data cleaning, outlier removal, missing value imputation Python 3.9, ONNXRuntime, OpenCV 4.8 Mechanism Model Library Module Edge + Cloud Models of insect developmental rates, population processes, and superposition of multiple insect stages. NumPy 1.24, SciPy 1.10, Cython acceleration Data Assimilation Dynamic Calibration Module Edge (diagonally simplified) + Cloud (complete set) State forecasting, observation data assimilation analysis, error covariance recursion Self-developed KalmanLib (implemented in C++17) Risk Decision Module Edge + Cloud mRI calculation, ROC threshold determination, and tiered early warning generation and push. Python 3.9, Scikit-learn 1.3 Cloud-Edge Collaboration Module Edge + Cloud Daily 00:00 data synchronization, parameter update package distribution, downgrade switching management, interruption recovery and data retransmission MQTT 5.0, SQLite 3.42 (edge), PostgreSQL 15 (cloud)

[0220] 7.3 Edge Performance Benchmark

[0221] Performance benchmark tests (100 repetitions) of the complete algorithm chain of this invention were performed on the ARM Cortex-A72 1.5GHz platform:

[0222] AI Image Reasoning (Single Image) 85 120 Data quality control and filling 5 8 Model prediction step (24-hour recursion) 12 15 Diagonal Simplification Assimilation Analysis Step 0.3 2 Multi-state superposition calculation 8 10 7-day rolling forecast 45 18 mRI Calculation and Early Warning 2 3 Total (excluding AI inference) ≈72 56

[0223] 7.4 Inter-module timing linkage and data flow

[0224] The mechanism model library module, data assimilation dynamic calibration module, risk decision-making module, data acquisition and quality control module, and cloud-edge collaboration module of this invention operate in tandem according to a fixed daily time-series scheduling rule, forming a complete data closed loop. The coupling relationship, data flow path, and timing logic of each module are as follows:

[0225] Daily 08:00 – Data Acquisition and Preprocessing: The data acquisition and quality control module automatically aggregates image data from intelligent insect monitoring lamps, automatic pheromone counting data, and hourly temperature and humidity data from field microclimate stations from 19:00 of the previous day to 07:00 of the current day. It then performs quality control operations such as AI image recognition, confidence level determination, outlier detection, and missing value imputation. The standardized dataset after quality control is sent in a structured message format to the mechanism model library module and the data assimilation dynamic calibration module.

[0226] Daily 08:05 – Mechanism Model State Recursion: The mechanism model library module receives the quality-controlled meteorological driving data (hourly temperature and relative humidity), calls the insect stage and generation development rate model to calculate the hourly development rate of each insect stage; simultaneously, it calls the full-generation population process model to recursively calculate the density distribution of each insect stage on the current day based on the analysis field of the previous day, including the multi-stage coexistence distribution vector output by the generation overlap processing submodule. The recursive results (background field state vector and error covariance) are sent to the data assimilation dynamic calibration module.

[0227] Daily 08:10 — Data Assimilation and State Update: The data assimilation dynamic calibration module receives the background field output from the mechanism model library module and the effective observation vector y_t output from the data acquisition and quality control module. If there is at least one effective observation data point for the day, the analysis step is executed: calculating the observation innovation and adaptive gain coefficient, weighting and correcting the state vector, and outputting the optimal estimate of the analysis field x^a_t and the updated error covariance P^a_t; if there is no effective observation data for the day, the analysis field degenerates into the background field. The updated global state vector is synchronously fed back to the mechanism model library module to update the development rate bias term r_bias and the larval survival rate bias term S_larv,bias, forming a parameter dynamic calibration closed loop.

[0228] Daily 08:15 – 7-Day Rolling Forecast: The data assimilation dynamic calibration module uses the daily analysis field x^a_t as the initial condition, calls the 7-day weather forecast data (synchronously acquired from the cloud at 00:00 daily), and drives the mechanism model library module to recursively generate a 7-day daily egg hatching prediction sequence, a daily voracious larval equivalent prediction sequence, and a suitable window for control. Each predicted value comes with a 95% confidence interval (calculated from the error covariance through recursive propagation). The prediction results are sent to the risk decision module.

[0229] Daily 08:20 – Risk Index Calculation and Tiered Early Warning: The risk decision-making module receives rolling forecasts for the next 7 days and, combined with field natural enemy density survey data, population herbicide resistance monitoring data, and dynamic economic thresholds, calculates the risk index sequence for the current day and the next few days using the mRI formula. The mRI value is compared with the tiered early warning threshold to determine the warning level (no warning / moderate warning / severe warning). Warning results are pushed to farmers' mobile apps, village-level LED bulletin boards, and cloud monitoring platforms via the MQTT protocol, and simultaneously fed back to the data acquisition and quality control module for adjusting subsequent monitoring frequency (monitoring frequency is increased to twice daily in severe warning situations).

[0230] Daily 00:00 — Cloud-Edge Data Synchronization: The cloud-edge collaboration module executes a scheduled synchronization task, uploading the daily quality control data and calibration results, and downloading the cloud-based global parameter update package (including model parameters and mRI weight coefficients optimized based on multi-region data fusion). If synchronization fails due to network interruption, the edge device continues to run the local simplified model independently, and will automatically re-upload and synchronize upon recovery.

[0231] The above-mentioned inter-module coupling and linkage mechanism corresponds completely to the four-step method: step S1 corresponds to the data acquisition and quality control module; step S2 corresponds to the mechanism model library module; step S3 corresponds to the data assimilation dynamic calibration and rolling prediction module; step S4 corresponds to the risk decision module; and the cloud-edge collaboration module supports the distributed deployment and fault-tolerant operation of the overall system.

[0232] It should be noted that the technical solution of this invention can be directly applied to various practical scenarios, such as grassroots plant protection service stations, large-scale grain planting bases, and grid-based contiguous planting households. The specific process includes:

[0233] Step 1: Grid-based deployment of front-end monitoring equipment. Within the target rice planting area, grid-based deployment is completed according to plot distribution, water system conditions, and planting varieties. Field micro meteorological monitoring stations, intelligent insect monitoring lamps, and automatic monitoring equipment for rice stem borer pheromone traps are deployed in a standardized manner. At the same time, edge computing terminals are installed, network adaptation and debugging are completed, and the time of all devices in the area is synchronized.

[0234] Step Two: Multi-Source Data Aggregation and Intelligent Quality Control. Raw data collected by various monitoring devices is transmitted to the edge computing terminal in real time, automatically running preset quality control logic to sequentially complete abnormal value removal, intelligent filling of missing data, and confidence correction of pest image recognition.

[0235] Step 3: Model Time Series Calculation and Graded Risk Warning. The system automatically calls the insect development rate model, the full-generation population process model, and the data assimilation module daily to complete iterative calculations. Combining field data on differences in rice varieties, crop growth stages, and regional pest resistance baselines, the system calculates the multidimensional pest risk index mRI in real time and outputs predictions of the rice stem borer occurrence period, dynamic forecasts of insect population density, regional damage levels, and the optimal window for control.

[0236] Step 4: Implement precise prevention and control measures based on the early warning results. Grassroots plant protection personnel will implement targeted prevention and control measures by region, batch, and level, based on the regional early warning level, key prevention and control areas, and optimal application time suggested by the system.

[0237] Step 5: Field Inspection and Follow-up Visits and Model Closed-Loop Optimization. After the pesticide control operation is completed, regular field inspections are conducted. The actual control effect and pest occurrence are recorded back into the system for local fine-tuning of regional parameters and continuous model verification and iteration, forming a long-term closed-loop application model of "real-time monitoring → intelligent prediction → precise control → effect follow-up visit → model optimization".

[0238] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data assimilation rolling prediction method for Chilo suppressalis in Hangjiahu rice region, characterized in that: Includes the following steps S1. Construct a multi-source data acquisition and quality control network to obtain insect infestation data, pheromone trap data, field survey data, and microclimate monitoring data; S2. Establish a multigenerational population process mechanism model library, including a nonlinear development rate model coupled with the interaction of temperature and humidity, and a full-generation population process model that connects the developmental processes of each insect stage. The probability function of larvae surviving to the voracious stage integrates the effects of temperature, rice variety and growth period, natural enemy density and population resistance frequency. The population process model includes a multi-stage generation overlap treatment mechanism. By superimposing the developmental progress distribution of different oviposition batches on the time axis, the coexistence distribution of multiple insect stages in the field is obtained. S3. Using a data assimilation algorithm, the population state variables and model parameters are used as assimilation vectors. Multi-source real-time observation data after quality control are fused together to perform dynamic global calibration on the model parameters and population state variables. Based on the calibration results, the model is driven to predict the future and outputs a daily rolling prediction sequence. S4. Based on the calibrated population status, calculate a multidimensional field risk index that integrates insect population stress, lack of natural enemy control, and pesticide resistance stress, and conduct graded early warning based on thresholds calibrated from historical disaster damage data.

2. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The nonlinear development rate model described in step S2 is based on insect stages. and generations The nonlinear functions established separately are expressed in terms of temperature. and relative humidity Let be the independent variable, and its expression is: In the formula: For the first Generations of insects The development rate; Here is the humidity correction function, when It takes effect when the humidity is below the minimum required for development; otherwise, the value is 1. As a developmental rate scaling factor; This is the temperature at which development begins; This is the temperature at which high temperatures can cause death. This refers to the width of the high-temperature boundary layer. The model parameters represent the positive part function; , , , All are set as state variables that can be dynamically updated by the data assimilation algorithm described in step S3.

3. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The probability function for the larvae to survive to the gluttonous stage in step S2 is: In the formula: Basic survival rate; It is a temperature stress function that exhibits symmetrical logical decay around the optimum temperature; This is a variety-growth period interaction function, which varies with the rice development stage. It exhibits a linear change; predator density The resulting survival decay function exhibits an exponential decay form; For the frequency of drug resistance in the population The resulting change in baseline mortality rate; The generation overlap processing mechanism is implemented by discretizing the developmental progress into multiple intervals and assigning a probability distribution to the variation of the individual development rate within each spawning batch, wherein the developmental progress variation dispersion of the overwintering generation is greater than that of the first generation.

4. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The data assimilation algorithm in step S3 employs a recursive filtering method that includes a prediction step and an analysis step, wherein... The forecast step is based on the global calibration results and meteorological driving data of the previous moment, and uses a mechanistic model to recursively deduce the background field of the population state and its error covariance at the current moment; In the analysis step, when valid observation data is obtained, the population state variables are weighted and corrected in the following form by calculating gain coefficients that are appropriate for observation errors and background errors. In the formula To analyze the optimal estimation of the field, For background field prediction, For the observation vector, For the observation operator, This is an adaptive gain matrix; under low computing power conditions at the edge, the gain matrix... It is automatically simplified to a form that uses only diagonal element operations, and each diagonal gain coefficient is determined by the ratio of the background error variance to the observation error variance of the corresponding state.

5. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The quality control described in step S3 includes a dual criterion based on the confidence level of the observed data itself and the degree of deviation from the background field. When the confidence level of the source of the observation data is lower than the first threshold, the observation data is given a weight reduction process with a weight lower than the standard weight, so that it participates in the subsequent calculation with a reduced weight; When the deviation between the observed data and its predicted background field value exceeds the preset multiple standard deviation range, it is judged as an outlier and removed. The gaps generated after removal are filled by an interpolation method based on the nearest neighbor of the time series.

6. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The multidimensional field risk index mentioned in step S4 The calculation formula is In the formula: The standard normal cumulative distribution function maps the comprehensive stress score to the interval between 0 and 1. For the first Daily larval equivalent during voracious feeding; For dynamic economic thresholds; The degree of absence of natural enemy density relative to the normal baseline density is included in the risk calculation as a positive contribution; The increase in the frequency of antibiotic resistance in the population relative to the baseline frequency is included in the risk calculation as a positive contribution; The weight coefficients for each stress factor are objectively determined through analysis of historical disaster data, and the sum of the weights is 1.

7. A rolling prediction method for multi-source data assimilation of rice stem borers suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area, as described in claim 6, is characterized in that: The method for determining the graded early warning threshold is as follows: collect field disaster event records of the target area over many years, distinguish positive and negative samples by a preset crop damage rate threshold, plot the receiver operating characteristic curve of the risk index on the sample set, select the graded threshold by a comprehensive criterion of maximizing the correct early warning rate and minimizing the false alarm rate, and divide the risk index value space into multiple early warning levels.

8. The method for multi-source data assimilation and rolling prediction of rice stem borer suitable for the Hangzhou-Jiaxing-Huzhou rice-growing area according to claim 1, characterized in that: The daily rolling prediction sequence output in step S3 covers the daily population dynamics information within a preset number of days in the future, including at least the daily predicted egg hatching rate and the predicted larval equivalent during the voracious feeding period, and includes the confidence intervals of each prediction value calculated based on the error covariance recursive propagation.

Images