Soybean pod borer predicting and forecasting method based on machine learning

Abstract

The invention discloses a soybean pod borer predicting and forecasting method based on machine learning, and relates to the technical field of insect situation prediction.The soybean pod borer predicting and forecasting method comprises the steps that trapping counting and state information are collected, the total trapping amount of the day is summarized, and an observation credibility index is calculated through S-shaped attenuation punishment; determining a key occurrence window for annual phase alignment of total trapping amount historical data, and calculating a phenological continuous index of a future prediction window based on a sowing date and day-by-day air temperature; establishing a mechanism constraint machine learning model based on a feedforward full-connection neural network, and outputting a continuous time trapping intensity predicted value and change trend information; carrying out numerical integration on the predicted value to obtain a future insect situation cumulant, and calculating an insect situation risk amount by combining the environmental suitability and the phenological suitability; and generating a maintenance instruction according to the observation credibility index, and outputting an early warning and patrol frequency adjustment instruction in a graded manner according to the insect condition risk quantity. According to the method, quantitative suppression of low-quality trapping data, constraint stabilization of a trapping intensity prediction mechanism and comparable and reviewable risk early warning are realized.

Description

Technical Field

[0001] This invention relates to the field of insect infestation prediction technology, and in particular to a method for predicting and forecasting soybean pod borers based on machine learning. Background Technology

[0002] Forecasting and prediction of soybean pod borer usually relies on field monitoring and meteorological data. Sex traps are deployed in the target field to obtain daily trap counts. Combined with environmental data such as sowing date, crop growth period, and daily temperature, methods such as accumulated temperature method, empirical discrimination of occurrence period, or statistical regression are used to determine the occurrence period and infer the trend of insect population. In recent years, some studies have also organized multi-source time series data into training samples and used machine learning to predict insect population and output early warning levels to support field inspection arrangements and control decision-making.

[0003] However, existing methods still have two limitations: First, crop phenological progression and insect infestation changes often have significant intra-year phase differences. Without intra-year alignment of historical insect infestation sequences and focus on key occurrence windows, model training is easily affected by yearly differences, resulting in insufficient cross-year generalization stability. Second, early warnings often rely on single-point insect populations or empirical thresholds, making it difficult to consistently constrain the cumulative changes in insect infestations within future windows with environmental and phenological conditions. This leads to insufficient verifiability of risk classification, treatment timing, and threshold sources. Summary of the Invention

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

[0005] Therefore, this invention provides a machine learning-based method for predicting and forecasting soybean pod borers, which solves the problems of insufficient model training and prediction stability caused by cross-year phase differences and the difficulty in quantifying the magnitude of future pod infestation increases in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a machine learning-based method for predicting and forecasting soybean pod borers. The method includes: collecting trap counts, recording status information, summarizing the total daily trap count, and calculating the daily observation reliability index using an S-shaped decay penalty; performing phase alignment within the year based on historical data of the total trap count to determine the key occurrence window for the field, and calculating phenological continuity indices for future prediction windows based on sowing dates and daily temperature data; establishing a mechanistic constraint machine learning model based on a feedforward fully connected neural network, constructing a training input sequence using the total trap count and phenological continuity indices, inputting the training input sequence into the mechanistic constraint machine learning model to obtain the predicted trap intensity and its changing trend information; performing numerical integration on the predicted trap intensity to obtain the cumulative pest population for future prediction windows, determining environmental suitability and phenological suitability, and calculating the pest risk; making a judgment based on the observation reliability index, generating maintenance instructions, and performing multi-level risk assessment based on the pest risk, generating patrol frequency adjustment instructions.

[0007] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the following steps are taken: collecting trap counts, recording status information, summarizing the total number of traps caught that day, and calculating the observation reliability index for that day using an S-shaped decay penalty: recording the status flag and the number of days the lure core is used for each trap, summarizing the total number of traps caught that day, and determining the effective trap set for that day; performing global normalization by the ratio of the number of effective traps to the total number of traps deployed, and applying an exponential penalty to the status flag of each trap; and applying a multiplicative correction to the number of days the lure core is used relative to the maximum number of days the lure core can be used, and calculating the observation reliability index.

[0008] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the determination of the key occurrence window of the field includes: starting from the monitoring start date and searching backwards to find the earliest date on which the total number of traps is greater than zero in three consecutive inspections, and taking the earliest date as the stable detection start point; shifting the entire trapping sequence along the annual day sequence to ensure that the stable detection start points of all years fall at the same day sequence position, thus completing the phase alignment within the year; taking the median of the total number of traps in the same day sequence for multiple years to form the annual baseline sequence, finding the main peak day sequence on the annual baseline sequence, and using the drop to half the height of the main peak on both sides as the window boundary to determine the key occurrence window of the field.

[0009] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the calculation of the phenological continuous indicators for the future prediction window includes: collecting the daily average temperature of the field, accumulating the positive difference between the daily average temperature and the base temperature day by day to obtain the effective accumulated temperature accumulation; calculating the effective accumulated temperature accumulation corresponding to the entry into the grain-filling stage and the late grain-filling stage respectively, obtaining the accumulated temperature benchmark point, normalizing and converting the daily effective accumulated temperature accumulation and the accumulated temperature benchmark point to obtain the phenological indicators; reading the daily average temperature forecast value, and recursively obtaining the phenological continuous indicator sequence for the future prediction window.

[0010] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the establishment of the mechanism-constrained machine learning model includes: constructing a data fitting term by minimizing the integral of the squared difference between the predicted value of continuous-time trapping intensity and the continuous-time observation sequence within the critical occurrence window; constructing a mechanism constraint term by constraining the time variation of the predicted value of continuous-time trapping intensity to satisfy the saturation growth inhibition mechanism modulated by the environmental modulation term; and establishing a mechanism-constrained machine learning model by using the joint training objective of the data fitting term and the mechanism constraint term.

[0011] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the step of obtaining the predicted value of trapping intensity and the information on its changing trend includes: establishing a daily data package; retrieving historical daily data packages of the same field within the critical occurrence window to form a historical sample sequence; constructing a training input sequence; concatenating the training input sequence into a feature vector in chronological order; inputting the feature vector into a feedforward fully connected neural network; and outputting the predicted value of trapping intensity over continuous time; and calculating the information on the changing trend based on the predicted value of trapping intensity over continuous time at two adjacent consecutive time points.

[0012] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the steps of obtaining the cumulative insect population in the future prediction window by numerically integrating the predicted values ​​of trapping intensity and determining environmental suitability and phenological suitability are as follows: numerically integrating the sequence of predicted values ​​of trapping intensity over a continuous time period within the future prediction window, obtaining the future cumulative population using the trapezoidal method and calculating the historical cumulative population to obtain the cumulative insect population in the future prediction window; determining environmental suitability based on the deviation from the optimal interval, and determining phenological suitability based on the key growth stages.

[0013] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the calculation of the pest risk includes: normalizing the cumulative pest amount in the future prediction window by ratio, and performing gating correction through environmental suitability and phenological suitability to calculate the pest risk.

[0014] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the step of determining the reliability index and generating maintenance instructions by means of observation is as follows: comparing the daily observation reliability index with the reliability threshold; if the reliability index is lower than the reliability threshold, it is determined to be low reliability and the maintenance instructions are set to the highest priority; if the reliability index is not lower than the reliability threshold, it is determined to be usable and the daily data is allowed to be included in the incremental update of the mechanism-constrained machine learning model; when it is determined to be low reliability, according to the trap status record, drainage and sticky plate replacement instructions are generated for the water inlet record, repair or replacement instructions are generated for the damaged record, and lure core failure instructions are generated for the lure core failure record.

[0015] As a preferred embodiment of the machine learning-based soybean pod borer prediction and forecasting method of the present invention, the step of performing multi-level risk assessment based on pest risk quantity and generating patrol frequency adjustment instructions specifically involves: comparing the daily risk quantity with the key risk threshold to output a risk warning level; when a high-risk warning is output, determining the peak time point of the continuous-time trapping intensity prediction value and combining it with trend information to determine the disposal time window; and generating patrol frequency adjustment instructions for the next day based on the warning level; the recent risk baseline is obtained from the geometric mean of the risk quantity sequence of the most recent fixed number of days, and the fixed multiple is determined by the median of the risk quantity ratio sequence of two adjacent days within the historical key occurrence window.

[0016] The beneficial effects of this invention are as follows: by observing the credibility index and credibility threshold, it is possible to quantitatively identify low-quality trapping data and suppress its interference with forecast results; by using the joint training objective of the machine learning model constrained by the mechanism and the low-credibility freeze incremental update, it is possible to make the continuous time trapping intensity prediction more in line with the growth suppression law; by calculating the risk amount accumulated in the future and in the historical background, gating the environmental suitability and phenological suitability, and adaptively determining the key risk threshold, it is possible to make the risk warning comparable and verifiable. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a machine learning-based method for predicting and forecasting soybean pod borers.

[0019] Figure 2 This is a flowchart for calculating the reliability index of observations for the day.

[0020] Figure 3 This is a flowchart for obtaining continuous phenological indicators.

[0021] Figure 4 A flowchart for generating early warning instructions.

[0022] Figure 5 This is a comparative data graph showing the predicted trapping intensity curves. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides a machine learning-based method for predicting and forecasting soybean pod borers, comprising the following steps: S1. Collect trap counts, record status information, summarize the total number of traps for the day, and calculate the daily observation reliability index using S-shaped decay penalty.

[0027] Select the boundary of the target field and evenly deploy multiple sex traps inside the field. The traps are arranged in a triangular pattern. Set the installation height of each trap to a fixed height above the crop canopy (e.g., 20–30 cm) and adjust the height as the crop grows so that the trap entrance is always kept within a fixed height range above the canopy.

[0028] A daily inspection is conducted at a fixed time (e.g., 08:00–10:00). The collection box of each trap is opened, and the number of male insects captured that day is manually counted to obtain the trap count. On the same day, three types of status indicators are recorded for each trap, including a missing test indicator (if the data cannot be read or the count fails), a water ingress indicator (if there is water accumulation in the collection box or the rubber plate has fallen off), and a damage indicator (if the trap's outer shell is cracked, the inlet is deformed, or the fixing device is malfunctioning). On the same day, the number of days the lure has been used is recorded, the date of the most recent replacement of the lure is read, and the number of days the lure has been used is calculated.

[0029] The total number of traps caught each day is summarized by field. The set of effective traps for the day is determined if and only if the traps are not missing, not flooded, not damaged, and the number of days the lure has been used has not exceeded the limit (e.g., 45 days). Otherwise, the traps are considered invalid.

[0030] The field observation reliability index is calculated by globally normalizing the ratio of the effective number of traps to the total number of traps deployed, and applying exponential penalties to the water ingress and damage indicators of each trap. An S-shaped decay penalty is applied to correct the ratio of the number of days the lure was used to its maximum usage days. The expression is as follows: ; in, Indicates field exist Daily observation reliability index Indicates the discrete day number after the monitoring start date. Indicates field exist The number of decoys determined to be effective per day. For fields The total number of traps deployed inside. Indicates field Inner Taiwan trap The water inlet indicator is set to 1 if water is inlet and 0 otherwise. This indicates a damaged item; a value of 1 indicates damage, otherwise a value of 0. This indicates the number of days the decoy core has been used. This represents a constant indicating the maximum number of days a decoy can be used. This indicates the steepness coefficient of the aging of the inducing core.

[0031] It should be noted that, The method involves using the same batch of lures in the same field on different activation dates, continuously recording the daily trap count for each lure's corresponding trap, and using a drop in the trap count to half of the average daily trap count for the first three days after the lure's activation, with the decline occurring for three consecutive days, as the failure criterion. The number of days the lure is used when the failure criterion is first met is determined as the lure's maximum usable days constant. The reason for using three consecutive days as the failure criterion is that the daily trap count is affected by short-term weather fluctuations, field operation disturbances, and insect activity rhythms, resulting in random fluctuations that temporarily drop and then recover. If only a single day or two consecutive days of low values ​​are used as the failure criterion, it is easy to misjudge the above short-term fluctuations as lure failure, thus leading to a systematic underestimation of the lure's maximum usable days. Fixing the criterion window to three days can cover common short-term fluctuation cycles and impose time consistency constraints on random fluctuations. This was achieved by establishing a sample set of lure usage days and daily trap counts for the same batch of lures within the same field, and determining the lure failure criterion. As the inflection point, the normalized curve of the decrease in trap count with the number of days of use is fitted with an S-shaped decay function, and the slope parameter obtained by fitting is used as the aging steepness coefficient of the trap core.

[0032] When the observation credibility index is greater than or equal to the credibility threshold, the observation is considered credible.

[0033] It should be noted that the credibility threshold is calculated by dividing historical monitoring days into two categories: normal days (on which all traps are recorded as having no missing measurements, no water ingress, no damage, and the lure core is not determined to be faulty) and abnormal days (on which at least one trap is recorded as having any of the following: missing measurements, water ingress, damage, and lure core failure). The distribution of the observation credibility index values ​​corresponding to the two types of days is calculated for each category, and the dividing point with the greatest distinction between the two distributions is used as the credibility threshold.

[0034] S2. Based on historical data of total trapping, perform phase alignment within the year to determine the key occurrence window of the field, and calculate the phenological continuity index of the future prediction window based on sowing date and daily temperature data.

[0035] After collecting daily total trapping records for several years in the same field, the trapping records for each year are uniformly converted into a timeline of daily occurrences within the year (i.e., converting each observation date in a year into the current year's date starting from January 1st), so that data from different years are on the same time scale that can be directly compared. The stable detection date at the start of the occurrence is used as the phase anchor point. Specifically, for each year, starting from the monitoring start date of that year, the earliest date on which the total trapping amount is greater than zero for three consecutive patrols is found, and the earliest date is regarded as the stable detection start point for that year. The entire trapping sequence of that year is shifted along the daily sequence of the year, so that the stable detection start point for all years falls on the same daily sequence position, completing the phase alignment within the year. After phase alignment, the median of the total trapping amount for the same daily sequence over multiple years is taken to form the baseline sequence for the year. Finally, the main peak daily sequence (the maximum point of the baseline) is found on the baseline sequence for the year, and the window boundaries are defined by the point on both sides of the main peak falling back to half the height of the main peak. The key occurrence window of the field is then determined.

[0036] It should be noted that the reason for three consecutive inspections is that the total number of traps may experience short-term fluctuations in the early stages of the event, such as occasional individuals entering the trap, counting errors, or short-term environmental disturbances, resulting in a single non-zero value followed by a return to zero. If a single non-zero value or two consecutive non-zero values ​​are taken as the starting point, it is easy to misjudge occasional fluctuations as stable occurrences, thereby leading to instability of the phase anchor point within the year and amplifying the cross-year alignment error. Therefore, stable detection is defined as the total number of traps being greater than zero in all three consecutive inspection cycles.

[0037] During the current season, the sowing date of the field is read; the daily average temperature of the field is obtained from the sowing date, and the difference between the daily average temperature and the base temperature is accumulated if it is positive, otherwise it is recorded as zero and not included in the accumulation. The effective accumulated temperature from the sowing date to the current day is accumulated daily. The base temperature is determined by statistically analyzing the lowest effective temperature limit from emergence to flowering in the soybean growth records of the same region over many years. In the soybean growth period survey records of previous years, the dates of entering the grain-filling stage and entering the late grain-filling stage are found respectively, and the effective accumulated temperature corresponding to the grain-filling stage and entering the late grain-filling stage is calculated. The multi-year average is taken as the two phenological accumulated temperature benchmarks for the field. The daily effective accumulated temperature and the two phenological accumulated temperature benchmarks are normalized to obtain a continuously changing phenological index. Specifically, when the phenological index is less than zero, it means that the grain-filling stage has not yet begun. When the phenological index is between zero and one, it means that the grain-filling stage is between the late grain-filling stage and the late grain-filling stage. When the phenological index is greater than one, it means that the late grain-filling stage has been exceeded.

[0038] During daily updates, the daily average temperature forecast values ​​for the next few days are obtained from meteorological data sources. Following the same effective accumulated temperature accumulation rules as the current day, the forecast temperatures for each future day are added to the accumulated amount, and the effective accumulated temperature accumulation for each future day is recursively obtained to obtain the phenological continuous index sequence for the future forecast window (e.g., the next 7 days).

[0039] S3. Establish a mechanism-constrained machine learning model based on a feedforward fully connected neural network. Construct a training input sequence using the total amount of traps and continuous phenological indicators. Input the training input sequence into the mechanism-constrained machine learning model to obtain the predicted value of trapping intensity and information on its changing trends.

[0040] A daily data package is established, which includes the total number of traps, the observation reliability index, the continuous phenological index, the temperature, and the relative humidity of the day. Historical daily data packages of the same field within the critical occurrence window are retrieved to form a historical sample sequence. Dates are converted into continuous time identifiers (e.g., the continuous time is accumulated in days with the monitoring start date as zero point), so that each day's data corresponds to a continuous time point. A training input sequence is constructed.

[0041] It should be noted that relative humidity represents the maximum amount of water vapor that air can hold relative to the same temperature.

[0042] For the continuous time domain within the critical occurrence window, a mechanistic-constrained machine learning model is established using a joint training objective of data fitting and mechanistic constraints. The data fitting term ensures that the continuous-time trapping intensity prediction function is consistent with the total number of traps at the observation time point, while the mechanistic constraint term ensures that the continuous-time trapping intensity prediction function conforms to growth and suppression mechanisms over continuous time. The expression for the joint training objective is: ; ; ; ; in, Indicates the training target value. Represents a continuous-time variable. This represents the predicted value of trapping intensity over continuous time. Represents a continuous-time observation sequence. Represents the mechanism residual function, Indicates the mechanism constraint strength coefficient. Indicates the saturation level parameter. Indicates the environmental modulation term. Indicates temperature, Indicates relative humidity. Indicates time Continuous phenological indicators at the location Represents the benchmark coefficient. Indicates the temperature sensitivity coefficient. This represents the relative humidity sensitivity coefficient. Indicates the phenological sensitivity coefficient. Indicates the benchmark growth rate. Represents the residual scaling constant. Indicates the daily patrol interval.

[0043] It should be noted that continuous time variables The baseline growth rate is measured in days. The dimensions are , , , All are dimensionless features after normalization.

[0044] It should be noted that the training process of the mechanism-constrained machine learning model is as follows: Using fields as the object, training input samples are obtained; the training input sequence is concatenated into a feature vector according to time order, and this feature vector is input into a feedforward fully connected neural network (multilayer perceptron, MLP), outputting the predicted trapping intensity value at the current time point, thus obtaining a prediction curve that continuously changes over time; a joint training objective is constructed using data fitting terms and mechanism constraint terms, and the gradient of the joint training objective with respect to network parameters is calculated using backpropagation. In each iteration, an adaptive moment estimation optimization algorithm is used to update the network parameters, and a stagnation-triggered decay rule is used to update the learning rate; two types of constraints are applied simultaneously during training: the first is a data constraint, where the mechanism-constrained machine learning model outputs data at each observed date point. The first type of constraint is to closely approximate the total number of insects trapped on that day. The second type is mechanistic constraint, which requires that the direction and magnitude of the change in the output curve of the mechanistic constraint machine learning model between adjacent days conform to the basic law of population growth saturation (for example, an increase is allowed during the low insect population stage, and the increase slows down after the insect population increases, avoiding unwarranted jumps). Furthermore, the effects of temperature, humidity, and phenology on the increase or inhibition are incorporated into the basic law as modulators. The training uses data within the historical critical occurrence window for repeated iterations until the observation fitting error and the degree of mechanistic deviation both decrease to a stable level. When running on a given day, if the confidence level of the observation on that day is lower than the confidence level threshold, the data on that day is only used to generate predictions and not to update parameters, keeping the model parameters of the previous day unchanged, in order to avoid the drift of the mechanistic constraint machine learning model caused by abnormal trapping data.

[0045] It should be noted that, This refers to a discrete observation of the total number of traps per day, using linear interpolation to make a straight transition between two adjacent days, so that the change is smooth within the whole day and that calculable values ​​are filled in between adjacent days; , , as well as It is learned by minimizing the training target values ​​of the data fitting term and the mechanism constraint term, together with the weight parameters of the feedforward fully connected neural network, as optimization variables for the joint training objective; It was obtained by taking the peak value and median of the total daily trapping volume within key occurrence windows over multiple historical years; The method involves replaying and calibrating the model within the historical critical occurrence window of the same field, sequentially training the model with multiple sets of different mechanistic constraint strength coefficients, calculating the mean absolute error between the predicted and measured total trapping amounts within the critical window, and calculating the number of anomalous jumps in the prediction curve without observational support (defined as the difference between predictions of two adjacent days exceeding three times the median absolute value of the difference within the critical window). The mechanistic constraint strength coefficient set that minimizes both the mean absolute error and the number of anomalous jumps without observational support is selected as the model. .

[0046] Starting from the current day, the future forecast window is divided into several consecutive time points (e.g., every hour). At each time point, the corresponding meteorological and phenological data are input to obtain a sequence of continuous time-based trapping intensity prediction values ​​for that time point. Based on the continuous time-based trapping intensity prediction values ​​of two adjacent consecutive time points, the trend information is calculated. Specifically, when the continuous time-based trapping intensity prediction value of the later time point is greater than that of the previous time point, the time interval is considered to be in an upward trend. When the continuous time-based trapping intensity prediction value of the later time point is less than that of the previous time point, the time interval is considered to be in a downward trend. The duration of the upward trend, the duration of the downward trend, and the time point corresponding to the maximum continuous time-based trapping intensity prediction value are recorded as the prediction peak period to obtain trend information.

[0047] S4. By numerically integrating the predicted value of trapping intensity, the cumulative amount of insect infestation in the future prediction window is obtained, and the environmental suitability and phenological suitability are determined to calculate the insect infestation risk.

[0048] To calculate the cumulative insect infestation amount in the future prediction window, specifically, the cumulative amount in the future is obtained by numerically integrating the sequence of predicted trapping intensity values ​​over a continuous time period within the future prediction window and using the trapezoidal method. The cumulative amount in the future is also calculated by reading the sequence of predicted trapping intensity values ​​over a continuous time period output in the historical window and using the same trapezoidal method.

[0049] Environmental suitability is determined by the deviation from the optimal range. Specifically, the optimal temperature center value and optimal temperature half-width of the insect species are obtained. When the average temperature of the future prediction window deviates from the center value by more than the optimal temperature half-width, the environmental suitability is set to 0. When the average temperature deviates from the center value by less than the optimal temperature half-width, the environmental suitability is calculated by linearly decreasing the value, so that the closer to the center value, the higher the suitability.

[0050] Phenological suitability is determined based on key reproductive stages. Specifically, the average value of continuous phenological indicators within the future forecast window is taken as the window phenological level. The median of the continuous phenological indicators corresponding to the grain-filling stage is taken as the starting point of the key stage, and the median of the continuous phenological indicators corresponding to the late grain-filling stage is taken as the ending point of the key stage. To form a smooth transition zone, the daily increment sequence of continuous phenological indicators from the grain-filling stage to the late grain-filling stage in historical data is read, and the median is taken as the typical daily advance. The typical daily advance amplitude corresponding to the typical daily advance is taken as the width of the transition zone. When the window phenological level is less than the difference between the starting point of the key stage and the width of the transition zone, phenological suitability is considered... The phenological suitability is set to 0. When the window phenological level is between the difference between the critical stage start point and the transition zone width and the critical stage start point, the phenological suitability increases linearly from 0 to 1. When the window phenological level is between the critical stage start point and the critical stage end point, the phenological suitability is set to 1. When the window phenological level is between the critical stage end point and the sum of the critical stage end point and the transition zone width, the phenological suitability decreases linearly from 1 to 0. When the window phenological level is greater than the sum of the critical stage end point and the transition zone width, the phenological suitability is set to 0. This ensures that the risk is amplified only in the critical phenological stage and suppressed in non-critical stages, reducing false alarms.

[0051] It should be noted that the optimal temperature center value and optimal temperature half-width of the insect species are determined by identifying the date when the total number of insects trapped reaches its peak each year within the historical critical occurrence window of the same field. Then, a peak neighborhood temperature sample set is formed by taking the daily average temperature samples of several days before and after the peak date (e.g., 3 days each). The arithmetic mean of the peak neighborhood temperature sample set is taken as the optimal temperature center value, and half of the difference between the maximum and minimum temperatures in the peak neighborhood temperature sample set is taken as the optimal temperature half-width.

[0052] The magnitude of risk escalation is characterized by future cumulative and historical cumulative approaches. Furthermore, risk is gating using environmental and phenological suitability to ensure consistency between the risk level and the necessary conditions for pest outbreaks. The pest risk level is calculated using the following expression: ; in, Indicates the first The corresponding insect infestation risk level per day Indicates the first Phenological suitability of the day Indicates environmental suitability.

[0053] Save the critical occurrence windows of multiple historical years, calculate the risk amount daily within the critical occurrence window for each historical year to form a set of historical risk amounts; sort the set of historical risk amounts from smallest to largest, calculate the difference sequence between two adjacent items, find the adjacent position with the largest difference, sort the risk amounts of adjacent positions from smallest to largest, and take the risk amount at the top of the list as the critical risk threshold.

[0054] In this embodiment, to verify the impact of low-confidence freeze incremental updates on the stability of continuous-time prediction of trapping intensity and the conformity of growth inhibition patterns, trapping sequences from the same monitoring season were selected, and predictions were compared using a control group employing a smoothing prediction method based on traditional experience and the complete scheme of this invention. Figure 5 As shown, within the critical occurrence window, the observation confidence index is calculated and compared with the confidence threshold. When the observation confidence is low on the day, the low confidence freeze incremental update is triggered, so that low-quality trap data does not enter the incremental learning, thereby suppressing the pull and jitter of the prediction curve caused by abnormal surges and missing measurements. Combining the overview and local zoom-in, it can be seen that the complete scheme constrains the machine learning model through mechanism to make the curve conform to the growth suppression law and reduce unreasonable jumps.

[0055] S5. Make judgments based on observation reliability indicators and generate maintenance instructions. Make multi-level risk judgments based on insect infestation risk and generate patrol frequency adjustment instructions.

[0056] The daily observation reliability index is compared with the reliability threshold. If the daily observation reliability index is lower than the reliability threshold, it is judged as low reliability, and the maintenance command priority is set to the highest. If the daily observation reliability index is not lower than the reliability threshold, it is judged as "available", and the daily data can be included in the incremental update of the mechanism constraint machine learning model. When it is judged as low reliability, maintenance commands are generated according to the trap status record. For traps with water ingress records, drainage and sticky plate replacement commands are generated. For traps with damage records, repair or replacement commands are generated. For traps with lure failure records, lure core replacement commands are generated.

[0057] The system compares the daily risk level with the critical risk threshold and outputs the risk level. Specifically, when the daily risk level is less than the critical risk threshold, a low-risk warning is output; when the daily risk level reaches or exceeds the critical risk threshold, a high-risk warning is output. The system takes the risk level sequence of the most recent fixed number of days (e.g., 3 days) and calculates the geometric mean as the recent risk baseline. When the daily risk level is less than the critical risk threshold and the daily risk level is not less than a fixed multiple of the recent risk baseline, a medium-risk warning is output.

[0058] It should be noted that the fixed multiple is calculated by calculating the risk amount daily within the historical critical occurrence window, forming a sequence of risk amount ratios between adjacent days, and taking the median of the ratio sequence as the fixed multiple.

[0059] When a high-risk warning is issued, the peak time point of the continuous-time trapping intensity prediction value is searched within the future prediction window. The disposal time window is determined based on the peak time point and trend information. Specifically, if the continuous-time trapping intensity prediction value sequence before the peak time point shows a continuous upward trend, the disposal window is determined to be a number of time periods before the peak time point; if the peak time point has already appeared and the continuous-time trapping intensity prediction value sequence shows a downward trend, the disposal window is determined to be the earliest executable time period between the current day and the peak time point, so as to avoid missing the best disposal opportunity.

[0060] Based on the warning level, an instruction to adjust the patrol frequency for the next day is generated. Specifically, when under a low-risk warning, the regular patrol frequency is maintained; when under a medium-risk warning, the patrol frequency is increased to once a day and additional recording of the trap status information is required; when under a high-risk warning, the patrol frequency is increased to once a day and an additional verification patrol is required before and after the disposal window to verify the risk changes and disposal effectiveness.

[0061] In summary, this invention achieves quantitative identification of low-quality trapping data and suppresses its interference with forecast results by observing credibility indicators and credibility thresholds; it makes continuous-time trapping intensity prediction more consistent with the growth suppression law by using mechanistic constraints on the joint training objectives of machine learning models and low-credibility freeze incremental updates; and it achieves comparability and verifiability of risk warnings by calculating the risk amount accumulated in the future and in the historical background, gating environmental suitability and phenological suitability, and adaptively determining key risk thresholds.

[0062] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based method for predicting and forecasting soybean pod borers, characterized in that: include, Collect trap counts, record status information, summarize the total number of traps for the day, and use S-shaped decay penalty to calculate the daily observation reliability index; Based on historical data of total trapping, phase alignment within the year is performed to determine the key occurrence window of the field, and phenological continuity indicators for future prediction windows are calculated based on sowing date and daily temperature data. A mechanism-constrained machine learning model is established based on a feedforward fully connected neural network. A training input sequence is constructed by using the total number of traps and continuous phenological indicators. The training input sequence is then input into the mechanism-constrained machine learning model to obtain the predicted value of trapping intensity and information on its changing trends. By numerically integrating the predicted value of trapping intensity, the cumulative amount of insect infestation in the future prediction window is obtained, and the environmental suitability and phenological suitability are determined to calculate the insect infestation risk. The system determines the reliability index and generates maintenance instructions. It also determines the multi-level risk based on the insect infestation risk level and generates instructions to adjust the patrol frequency.

2. The soybean pod borer prediction and forecasting method based on machine learning as described in claim 1, characterized in that: The specific steps for collecting and counting the trapped animals, recording status information, summarizing the total number of animals trapped that day, and calculating the daily observation reliability index using an S-shaped decay penalty are as follows: Record the status flags and the number of days the lure was used for each trap, summarize the total number of traps caught that day, and determine the set of effective traps for that day. The ratio of the number of effective traps to the total number of traps deployed is used for global normalization, and an exponential penalty is applied to the status flag of each trap. The observation reliability index is calculated by multiplying the number of days the decoy was used relative to the maximum number of days the decoy was used, using an S-shaped attenuation penalty.

3. The machine learning-based prediction and forecasting method for soybean pod borers as described in claim 2, characterized in that: The key occurrence window for determining the field includes, Starting from the monitoring start date, search backwards to find the earliest date when the total number of traps is greater than zero in three consecutive patrols. The earliest date is regarded as the stable detection start point. The entire trap sequence is shifted along the day sequence within the year, so that the stable detection starting point of all years falls at the same day sequence position, thus completing the phase alignment within the year; The median total number of traps for the same date over multiple years is used to form the annual baseline sequence. The main peak date is then identified on the annual baseline sequence. The critical occurrence window for the field is determined by using the drop to half the height of the main peak on both sides as the window boundary.

4. The machine learning-based soybean pod borer prediction and forecasting method as described in claim 3, characterized in that: The phenological continuity indexes used to calculate the future prediction window include, Collect the daily average temperature of the field, and accumulate the positive difference between the daily average temperature and the base temperature day by day to obtain the effective accumulated temperature. Calculate the effective accumulated temperature for entering the grain-filling stage and the late grain-filling stage respectively, obtain the accumulated temperature benchmark, and normalize the daily effective accumulated temperature and the accumulated temperature benchmark to obtain phenological indicators. By reading the daily average temperature forecast value, the phenological continuous index sequence for the future forecast window is obtained recursively.

5. The machine learning-based soybean pod borer prediction and forecasting method as described in claim 4, characterized in that: The establishment of the mechanism-constrained machine learning model includes constructing a data fitting term by minimizing the integral of the squared difference between the predicted value of continuous-time trapping intensity and the continuous-time observation sequence within the critical occurrence window; By limiting the temporal variation of the predicted trapping intensity over a continuous period to satisfy the saturation growth inhibition mechanism modulated by the environmental modulation term, a mechanism constraint term is constructed. A mechanism-constrained machine learning model is established by using a joint training objective of data fitting terms and mechanism constraint terms.

6. The machine learning-based soybean pod borer prediction and forecasting method as described in claim 5, characterized in that: The process of obtaining the predicted value and trend information of trapping intensity includes: establishing a daily data packet, retrieving historical daily data packets of the same field within the critical occurrence window, forming a historical sample sequence, and constructing a training input sequence; The training input sequence is concatenated into a feature vector in chronological order. The feature vector is then input into a feedforward fully connected neural network, which outputs a continuous-time trapping intensity prediction value. The trend information is calculated based on the predicted value of the trapping intensity at two consecutive time points.

7. The machine learning-based prediction and forecasting method for soybean pod borers as described in claim 6, characterized in that: The specific steps for obtaining the cumulative insect infestation amount in the future prediction window by numerically integrating the predicted value of the trapping intensity, and determining the environmental suitability and phenological suitability are as follows: Numerical integration is performed on the sequence of predicted trapping intensity values ​​within the future prediction window. The trapezoidal method is used to obtain the future cumulative amount and calculate the historical cumulative amount to obtain the insect infestation cumulative amount within the future prediction window. Environmental suitability is determined by deviation from the optimal range, and phenological suitability is determined by the key reproductive stages.

8. The machine learning-based prediction and forecasting method for soybean pod borers as described in claim 7, characterized in that: The calculation of insect infestation risk includes normalizing the cumulative insect infestation amount in the future prediction window by ratio, and performing gating corrections through environmental suitability and phenological suitability to calculate the insect infestation risk.

9. The machine learning-based prediction and forecasting method for soybean pod borers as described in claim 8, characterized in that: The process of determining reliability indicators and generating maintenance instructions involves the following steps: The daily observation credibility index is compared with the credibility threshold. If it is lower than the credibility threshold, it is judged as low credibility and the maintenance instruction priority is set to the highest. If it is not lower than the credibility threshold, it is judged as usable and the daily data is allowed to be included in the mechanism constraint machine learning model for incremental updates. When the status is determined to be low, the system generates drainage and sticky plate replacement instructions for water ingress records, repair or replacement instructions for damage records, and lure core replacement instructions for lure core failure records, based on the lure status records.

10. The machine learning-based prediction and forecasting method for soybean pod borers as described in claim 9, characterized in that: The process of determining multi-level risks based on insect infestation risk levels and generating patrol frequency adjustment instructions involves the following steps: The risk level is output by comparing the daily risk level with the key risk threshold. When a high-risk warning is issued, the peak time point of the continuous time-based trapping intensity prediction value is determined and the disposal time window is determined in combination with trend information. An instruction to adjust the patrol frequency for the next day is generated according to the warning level. The recent risk baseline is obtained from the geometric mean of the most recent fixed-day risk quantity series, and the fixed multiple is determined by the median of the risk quantity ratio series of two adjacent days within the historical critical occurrence window.

Images