A grey eulia amurensis forecasting method and system based on population dynamic model

Abstract

The present application relates to the technical field of agricultural pest monitoring and early warning and intelligent prediction, more specifically, the present application relates to a grey tea geometrid moth prediction method based on population dynamic model, the method faces the tea garden test site, obtains the trap counting sequence and the insect situation lamp counting sequence, and constructs the unified weekly observation record sequence; according to the key field integrity, the collection time consistency and the cross-source deviation, the input data level uncertainty information is generated; the adult population dynamic model of the source separation observation structure is constructed, and the model parameter level uncertainty information is generated; the future adult occurrence trend is prior evolution and assimilation correction, and the assimilation correction level uncertainty information is generated; further, the three kinds of uncertainty are positively propagated, the quantitative uncertainty interval of occurrence index, outbreak probability, risk level and early warning window is obtained, and based on the interval width and its change trend, the decision risk stratification is executed, and the fallback prediction result with low confidence label is generated in the abnormal situation.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest monitoring, early warning and intelligent prediction technology, and more specifically, to a method and system for predicting the gray tea geometrid moth based on a population dynamics model. Background Technology

[0002] The tea geometrid moth is one of the more common pests in tea garden production. The occurrence of adult moths is strongly correlated with subsequent population growth, early warning systems, and the timing of control measures. Therefore, continuous monitoring and prediction of the adult moth occurrence trend in tea gardens is of great significance for improving the timeliness of early warnings, reducing the burden of manual inspections, and assisting in the formulation of control strategies.

[0003] In existing technologies, pest monitoring schemes for tea gardens typically rely on a single collection method. For example, they may use only the counting results from traps for statistical analysis or only the collection results from insect lamps for trend judgment. While such schemes can reflect pest changes at a certain point in time, the differences in the observation mechanisms, response intensity, and degree of influence from the field environment among different collection methods mean that data from a single source is easily affected by local fluctuations, changes in equipment status, or collection anomalies. This results in insufficient stability of the prediction results and makes it difficult to accurately reflect the potential scale of adult tea geometrid moths in tea gardens.

[0004] Furthermore, while some existing pest prediction schemes can combine multi-source observation data for comprehensive analysis, they typically only involve simple overlaying, averaging, or empirical weighting of data from multiple sources. They lack a modeling mechanism that establishes independent observation channels around different sources and constrains potential population sizes. Especially in scenarios where traps and insect monitoring lights are deployed simultaneously, directly and roughly merging the two types of observation results can easily mask the differences between different sources, making it difficult for the system to distinguish between "true pest changes" and "single-source response bias," thereby reducing prediction accuracy.

[0005] Furthermore, existing technologies in practical applications typically focus more on the predicted central value itself, while rarely quantifying the uncertainties in the prediction chain. This is especially true in applications deployed locally in tea gardens and sampled weekly, where observational data may suffer from issues such as missing key fields, inconsistent collection periods, large cross-source bias, and missing measurements at local nodes. If the system only outputs a single predicted value or a single risk level without simultaneously providing the corresponding uncertainty range, actual users will find it difficult to determine whether the current prediction result is suitable as a direct warning basis, and will also struggle to distinguish between "high-confidence predictions" and "low-confidence predictions."

[0006] Therefore, we propose a prediction method and system for the gray tea geometrid moth based on a population dynamics model to solve the above problems. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for predicting the gray tea geometrid moth based on a population dynamics model, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the gray tea geometrid moth based on a population dynamics model, comprising: Acquire adult insect observation data collected at each monitoring point in the target tea garden experimental site according to a preset weekly sampling cycle. The adult insect observation data includes at least the trap counting sequence and the insect lamp counting sequence, and construct a unified weekly observation record sequence based on the trap counting sequence and the insect lamp counting sequence. Based on the completeness of key fields in the unified weekly observation record sequence, the consistency of collection time, and the cross-source deviation between the trap counting sequence and the insect lamp counting sequence, input data-level uncertainty information is generated. The input data-level uncertainty information includes at least the observation confidence coefficient and the corresponding observation confidence interval. Based on the trap counting sequence, insect lamp counting sequence and observation confidence coefficient, an adult insect population dynamic model is constructed. The adult insect population dynamic model includes at least the adult insect population size state quantity, the trapping response state quantity and the phototactic response state quantity. The model parameter level uncertainty information corresponding to the state transition parameters, trapping response parameters and phototactic response parameters in the adult insect population dynamic model is determined. The adult population dynamics model is used to perform prior evolution on the adult occurrence trend during the future target prediction period to obtain the prior prediction sequence; Based on the adult insect observation data and observation confidence coefficient within the current weekly sampling period, the prior prediction sequence is assimilated and corrected to obtain the corrected prediction sequence, and assimilation and correction level uncertainty information is generated. The uncertainty information at the input data level, the uncertainty information at the model parameter level, and the uncertainty information at the assimilation correction level are forward propagated through the error propagation formula or the Monte Carlo sampling method to obtain the quantitative uncertainty interval of the occurrence index, outbreak probability, risk level, and warning window corresponding to the correction prediction sequence; Based on the width and trend of the quantitative uncertainty interval, the prediction results are stratified into decision risks. Specifically, when the width of the quantitative uncertainty interval is less than a first preset threshold, it is determined to be a high-confidence prediction and an early warning window is directly output. When the width of the quantitative uncertainty interval is between the first and second preset thresholds, it is determined to be a medium-confidence prediction, and supplementary monitoring suggestions are output along with the early warning window. When the width of the quantitative uncertainty interval is greater than the second preset threshold, it is determined to be a low-confidence prediction, and a rollback mode or manual intervention request is triggered. When at least one of the following situations is detected: the cross-source deviation between the trap counting sequence and the insect lamp counting sequence exceeds a preset deviation threshold, the completeness of key fields is lower than a preset completeness threshold, or there are missing valid observations in consecutive weekly sampling periods, a regression prediction result is generated using the corrected prediction sequence corresponding to the previous valid weekly sampling period and the valid adult insect observation data corresponding to the current weekly sampling period. After attaching a low confidence marker to the regression prediction result, an early warning prompt is output.

[0009] In a preferred embodiment, constructing a unified weekly observation record sequence includes: Using a one-week rolling sampling period, the trap counting sequence and the insect lamp counting sequence are merged according to a unified time benchmark, and a unified weekly observation record table is generated according to the monitoring point number, collection date, collection time period, adult insect count value and source identifier, so that adult insect observation data from different sources can form a unified observation input that can be compared within the same weekly sampling period.

[0010] In a preferred embodiment, the input data level uncertainty information is determined by at least three items: key field completeness, collection time consistency, and cross-source bias. The cross-source bias is used to characterize the degree of difference between the trap counting sequence and the insect lamp counting sequence in terms of adult occurrence trends at the same monitoring point and within the same weekly sampling period. The observation reliability coefficient is generated by weighting the key field completeness, collection time consistency, and cross-source bias.

[0011] In a preferred embodiment, the adult population dynamic model is constructed using a source-separated observation structure. The adult population size state quantity is used to characterize the potential population size of the adult moths corresponding to the target monitoring point, the trapping response state quantity is used to characterize the trapping response intensity of the trappers to the adults, the phototactic response state quantity is used to characterize the phototactic response intensity of the insect lamps to the adults, and the trapping count sequence and the insect lamp count sequence are used as different observation channels to constrain the adult population size state quantity.

[0012] In a preferred embodiment, the model parameter level uncertainty information includes at least the parameter confidence intervals corresponding to the state transition parameters, trapping response parameters, and phototaxis response parameters. The parameter confidence intervals are obtained by estimating the parameters of the adult population dynamic model based on the effective adult observation data within the current weekly sampling period, and are used to constrain the fluctuation range of the prior prediction sequence.

[0013] In a preferred embodiment, assimilation correction of the prior prediction sequence includes: The trap counting sequence and the insect lamp counting sequence are assigned observation update weights according to the observation confidence coefficient, and the adult population size and state of the prior prediction sequence are corrected according to the observation update weight to obtain the corrected prediction sequence; among them, the higher the observation confidence coefficient, the greater the influence of the corresponding observation channel on the update of the adult population size and state.

[0014] In a preferred embodiment, the input data level uncertainty information, model parameter level uncertainty information, and assimilation correction level uncertainty information are forward propagated, including: The quantitative uncertainty intervals corresponding to the occurrence index, outbreak probability, risk level, and warning window can be calculated using the error propagation formula, or the interval distribution results corresponding to the occurrence index, outbreak probability, risk level, and warning window can be obtained by performing Monte Carlo sampling on the observation confidence coefficient, parameter confidence interval, and assimilation correction error.

[0015] In a preferred embodiment, decision risk stratification includes: Prediction results with an interval width less than the first preset threshold are marked as high-confidence predictions; Prediction results with an interval width greater than or equal to the first preset threshold and less than or equal to the second preset threshold are marked as medium confidence predictions, and an auxiliary prompt is output suggesting that the number of supplementary observations be increased in the current weekly sampling period. Prediction results with an interval width greater than the second preset threshold are marked as low-confidence predictions, and a prompt for manual review or a control command to switch to rollback mode is output.

[0016] In a preferred embodiment, a prediction system for the gray tea geometrid moth based on a population dynamics model includes: The data access module is used to acquire the trap counting sequence and insect lamp counting sequence collected at each monitoring point in the target tea garden experimental site according to a preset weekly sampling cycle; A unified observation construction module is used to construct a unified weekly observation record sequence based on the trap counting sequence and the insect lamp counting sequence. The input uncertainty assessment module is used to generate input data-level uncertainty information based on the completeness of key fields, consistency of acquisition time, and cross-source deviation. The dynamic model building module is used to construct a dynamic model of adult insect populations, including the population size state, the trapping response state, and the phototactic response state, and to determine the uncertainty information of the model parameters. The prediction and assimilation module is used to generate prior prediction sequences, perform assimilation corrections, and generate assimilation correction level uncertainty information. The uncertainty propagation module is used to forward propagate the uncertainty information at the input data level, the uncertainty information at the model parameter level, and the uncertainty information at the assimilation correction level to obtain the quantitative uncertainty range of the occurrence index, outbreak probability, risk level, and warning window. The risk stratification output module is used to perform decision risk stratification based on quantitative uncertainty intervals and output early warning prompts. The rollback processing module is used to generate rollback prediction results with low confidence flags when preset abnormal conditions are met.

[0017] The technical effects and advantages of this invention are as follows: 1. By simultaneously accessing the trap counting sequence and the insect lamp counting sequence, and constructing a unified weekly observation record sequence, this invention can form a more complete adult insect observation input in the tea garden setting. Compared with the scheme that relies on observation from only a single source, it is beneficial to improve the comprehensiveness and stability of the identification of the occurrence trend of adult tea geometrid moths.

[0018] 2. This invention constructs a dynamic model of adult insect populations with a source-separated observation structure. By using the trapping response state and phototactic response state as different observation channels to constrain the adult insect population size state, it can distinguish the difference between potential changes in adult insect size and single-source response deviation. Compared with the scheme of simply superimposing or roughly fusing multi-source data, it is beneficial to improve the accuracy and interpretability of prediction results.

[0019] 3. This invention generates input data-level uncertainty information and incorporates the completeness of key fields, consistency of acquisition time, and cross-source bias into the construction process of observation reliability coefficient and observation confidence interval. This enables a quantitative description of the reliability of observation data within the current weekly sampling period. Compared with schemes that only output a single observation value, this invention is beneficial to improving the pertinence of subsequent prediction and correction.

[0020] 4. This invention generates model parameter level uncertainty information by using state transition parameters, trapping response parameters, and phototactic response parameters, and uses parameter confidence intervals to constrain the fluctuation range of the prior prediction sequence. This can avoid excessive amplification of future prediction nodes due to abnormal parameter fluctuations. Compared with schemes with fixed parameters or no interval constraints, this invention is beneficial to improving the stability and controllability of the prediction process.

[0021] 5. This invention forward propagates uncertainty information at the input data level, uncertainty information at the model parameter level, and uncertainty information at the assimilation correction level, so that the occurrence index, outbreak probability, risk level, and warning window all correspond to quantitative uncertainty intervals. Compared with schemes that only output the center value or static level, this invention can more realistically reflect the reliability of the prediction results and provide quantifiable evidence for tea garden early warning.

[0022] 6. This invention performs decision risk stratification by performing high-confidence prediction, medium-confidence prediction, and low-confidence prediction based on the interval width and changing trend of the quantitative uncertainty interval. It can directly output early warning windows in high-confidence scenarios, simultaneously output supplementary monitoring suggestions in medium-uncertainty scenarios, and trigger manual intervention requests or rollback modes in low-confidence scenarios. Compared with a single early warning conclusion output method, it is beneficial to improve the decision adaptability of field applications. Attached Figure Description

[0023] Figure 1 This is a system module framework diagram of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0025] Reference Figure 1 A method for predicting the growth of the tea geometrid moth based on a population dynamics model, comprising: To obtain adult insect observation data collected at each monitoring point in the target tea garden experimental site according to a preset weekly sampling cycle, the following method can be used: First, multiple monitoring points were set up within the target tea garden experimental site according to a pre-defined layout rule. Each monitoring point was equipped with at least one set of traps and one set of insect-spreading lights, and each monitoring point was assigned a unique monitoring point number. After installation, each set of traps and insect-spreading lights was registered in the local prediction terminal, establishing a link between the device number and the corresponding monitoring point number. The local prediction terminal pre-stored the monitoring point number, device number, installation location, activation time, and sampling cycle configuration to ensure that subsequent adult insect observation data could be accurately mapped to the corresponding monitoring point.

[0026] The acquisition of trap counting sequences can be achieved through the following method: During each weekly sampling period, the adult insect count records generated by the traps at each monitoring point are periodically read. Adult insect count records can be directly output by the trap's own counting module, or they can be entered into the local prediction terminal by on-site personnel when changing lures, collecting insect samples, or inspecting the traps. Regardless of whether automatic output or manual entry is used, at least the monitoring point number, device number, collection date, collection time period, adult insect count value, and data source identifier must be retained. If the trap generates multiple records within a week, the original timestamp for each record is saved separately; if the trap is only counted on weekends or fixed dates, the corresponding statistical results are recorded as valid trap observations within that weekly sampling period. For automatically output trap data, the local prediction terminal can periodically retrieve it via wired interface, short-range wireless interface, or local file import; for manually entered trap data, the local prediction terminal provides an entry interface and automatically appends the entry time and entry personnel identifier upon saving for subsequent verification.

[0027] The insect lamp counting sequence can be obtained in the following way: During each weekly sampling period, the insect lamp operates within a preset working time period and generates corresponding adult insect collection records. These records can directly use the adult insect count values ​​output by the insect lamp's counting module, or they can use image records collected by the imaging device accompanying the insect lamp. The local prediction terminal then performs image recognition to generate the adult insect count values. If image recognition is used, after acquiring the insect lamp image, the local prediction terminal first performs cropping, noise reduction, brightness equalization, and background suppression processing on the image. Then, it uses connected component extraction, contour detection, template matching, or target recognition models to identify and count the insect targets in the image to obtain the adult insect count values ​​for the corresponding collection date and time period. For multiple observation records generated by the insect lamp within a week, the original timestamp, monitoring point number, device number, adult insect count value, and source identifier are saved for each record. The source identifier is used at least to distinguish whether the data comes from the trap or the insect lamp, so that subsequent merging and comparison can be performed according to the source.

[0028] After obtaining the trap counting sequence and insect lamp counting sequence, the local prediction terminal first performs preprocessing on the raw records. Preprocessing includes at least field integrity checks, time format standardization, duplicate record removal, outlier screening, and monitoring point mapping verification. Field integrity checks identify records with missing monitoring point numbers, missing collection dates, empty adult count values, or missing source identifiers; records with missing fields that cannot be filled in are directly marked as invalid and not included in the weekly unified observation record construction. Time format standardization converts dates and times generated by different devices or different input methods into a standard time format preset by the local prediction terminal. Duplicate record removal identifies duplicate upload records with the same monitoring point number, same device number, same collection date, same collection time period, and completely identical adult count values; only the earliest written record or the one with higher credibility is retained among the identified duplicate records. Outlier screening identifies abnormal records that significantly exceed the reasonable counting range of the equipment. This can be achieved through thresholding, moving average deviation, or adjacent record variation constraint methods. When a record's adult insect count exceeds a preset upper limit, its variation exceeds a preset proportion twice consecutively, or its deviation from other records at the same monitoring point in the same week is too large, the record is marked as a record to be verified and will not be directly used as input for the unified weekly observation until verification is completed. Monitoring point mapping verification checks the correctness of the binding relationship between equipment numbers and monitoring point numbers. If an equipment number is found to be unregistered, incorrectly bound, or the same equipment is repeatedly bound to multiple monitoring points, the relevant record is marked as a mapping anomaly record.

[0029] After preprocessing, the local prediction terminal constructs a unified weekly observation record sequence. Specifically, using a week as the rolling sampling period, trap records and insect light records belonging to the same monitoring point within the current processing window are merged according to a unified time benchmark. The unified time benchmark can be the start and end time of a natural week or a preset business week; for example, a weekly sampling period can be from Monday 00:00:00 to Sunday 23:59:59. For all valid records within this weekly sampling period, a unified weekly observation record table is generated according to the monitoring point number, collection date, collection time period, adult insect count value, and source identifier. The "collection time period" is not simply the original timestamp; instead, it can be mapped to preset time period labels as needed, such as early morning, daytime, evening, and nighttime, or mapped to several fixed time windows within a day, so that data from different sources can form comparable inputs with the same granularity within the same weekly sampling period.

[0030] During the merging process, if multiple trap records exist for the same monitoring point, the same collection date, and the same collection time period, the final count result for that source in that time period can be generated using summation, mean, maximum value, or last value retention methods. The same rules apply to insect lamp records. Preferably, if the adult insect count reflects the cumulative capture amount within that time period, the summation method is used; if the adult insect count reflects the instantaneous identification result, the mean or maximum value method is used. Once the merging rules are selected, they remain consistent throughout all implementations of the same application document to ensure the repeatability and auditability of subsequent predictive inputs.

[0031] When both trap records and insect lamp records exist at the same monitoring point within the same weekly sampling period, the local prediction terminal does not directly merge them into a single value. Instead, it retains them as parallel observation inputs from different sources and distinguishes them by source identifiers. In other words, in the unified weekly observation record table, the same monitoring point, the same collection date, and the same collection time period can correspond to two records, representing trap observations and insect lamp observations respectively; or they can correspond to the same composite record, which includes at least the trap adult count field and the insect lamp adult count field. This processing method preserves the independent characterization capabilities of different collection methods for adult occurrence trends during subsequent prediction processes, thus providing fundamental data for cross-source bias calculation, observation reliability generation, and assimilation correction.

[0032] For monitoring points where only a single valid source exists within a weekly sampling period, the local prediction terminal retains the count value corresponding to that valid source in the unified weekly observation record table and adds a missing marker to the missing source. For monitoring points where both sources are missing within a weekly sampling period, an empty observation placeholder record is generated and marked with a missing marker, so that subsequent modules can identify whether the monitoring point has valid input for that week. This processing method ensures that the unified weekly observation record table maintains a fixed field structure and does not change its record format due to incomplete actual data sources, thus facilitating direct reading and processing by subsequent modules.

[0033] After the unified weekly observation record table is generated, the local prediction terminal can further output a unified weekly observation record sequence. The unified weekly observation record sequence arranges the unified weekly observation records of each monitoring point in the order of the weekly sampling cycle and can be indexed and stored by monitoring point number. Each unified weekly observation record includes at least: monitoring point number, week number, collection date, collection time period, adult insect count value from the trap, adult insect count value from the insect lamp, source validity marker, missing record marker, and original record index. The purpose of retaining the original record index is that when subsequent prediction results are abnormal or the data source needs to be traced during the review stage, the unified weekly observation record can be used to trace back to the original observation record, thereby proving that the method of this application has a clear data link between data collection, data merging, and result formation.

[0034] After generating a unified weekly observation record sequence, the local prediction terminal also calculates basic statistics for the adult insect counts from traps and insect lamps at each monitoring point within the current weekly sampling period. These basic statistics include at least the weekly cumulative value, weekly average, weekly maximum value, and inter-source difference. These basic statistics are then used as auxiliary input for subsequent prediction modules. By first constructing a unified weekly observation record sequence and then extracting basic statistics, the input bias caused by inconsistencies in the original output formats, collection times, and recording granularities of different devices being directly input into the prediction module can be avoided, thereby improving the stability and repeatability of the adult insect prediction results.

[0035] Input data-level uncertainty information is generated based on the completeness of key fields in the unified weekly observation record sequence, the consistency of collection time, and the cross-source deviation between the trap counting sequence and the insect lamp counting sequence. This can be performed as follows: First, the local prediction terminal reads all valid records of the target monitoring point within the current weekly sampling period from the aforementioned unified weekly observation record table, and groups them according to the monitoring point number and week sequence number. For each group of records, at least the monitoring point number, collection date, collection time period, adult insect count value of the trap, adult insect count value of the insect lamp, source identifier, original timestamp, missing data marker, and original record index are extracted. If the unified weekly observation record table adopts a parallel source record format, the local prediction terminal first aligns the trap records and insect lamp records under the same monitoring point, the same collection date, and the same collection time period; if the unified weekly observation record table adopts a composite field format, the trap adult insect count value and insect lamp adult insect count value are directly read from the corresponding fields. After this processing, the input evaluation dataset corresponding to the current monitoring point and the current weekly sampling period is formed.

[0036] After obtaining the input evaluation dataset, the local prediction terminal calculates the key field completeness, collection time consistency, and cross-source bias, respectively.

[0037] The process of obtaining the completeness of key fields may include the following steps: First, a list of key fields is pre-configured in the local prediction terminal. This list includes at least the monitoring point number, collection date, collection time period, source identifier, and adult insect count value under the corresponding source. For implementations requiring traceability, the original timestamp and device number can also be included in the key field list. Then, the unified weekly observation records for the current monitoring point within the current weekly sampling period are scanned one by one. For each record, it is determined whether the key fields contain null values, illegal values, incorrect format values, or outliers exceeding the preset value range. Fields that exist and pass format validation are recorded as valid fields; fields that are missing, whose content cannot be parsed, do not match the source identifier, or whose content exceeds the preset reasonable range are recorded as invalid fields. After completing the scan, the completeness result of the key fields for the monitoring point within the current weekly sampling period is generated according to the ratio of the number of valid key fields to the number of required key fields. To avoid a single abnormal record having an excessive impact on the overall weekly evaluation, the local prediction terminal can first calculate the local completeness for each collection date or each collection period, and then perform mean processing or weighted average processing on the local completeness over the week to obtain the final key field completeness value. For trap-source and insect lamp-source, the local prediction terminal can either calculate the completeness within the source separately and then synthesize it, or directly calculate the overall completeness at the unified weekly observation record table level; preferably, the completeness within the source is calculated separately first, and then synthesized according to the number of valid records from the source, in order to preserve the quality differences between different sources.

[0038] The process of obtaining consistency in data collection time may include the following steps: The local prediction terminal pre-stores the weekly sampling plan table corresponding to the current monitoring point. The weekly sampling plan table includes at least the expected collection dates, expected collection periods, and planned collection requirements for each source on the corresponding dates and periods within a week. Then, the original timestamps in the unified weekly observation records are matched with the weekly sampling plan table to determine the planned collection window to which each record belongs. Records falling within the corresponding planned collection window are recorded as time-matched records; records exceeding the planned collection window but still within a preset tolerance range are recorded as time-offset records; and records exceeding the preset tolerance range are recorded as time-abnormal records. After matching, the local prediction terminal can generate a collection time consistency result based on these three types of results. In one embodiment, collection time consistency can be generated jointly by the "planned window hit rate" and the "time offset degree": first, the planned window hit rate within the current weekly sampling period is calculated, i.e., the proportion of time-matched records to the total number of planned collection windows; then, the time offset degree is calculated, i.e., the average or weighted average offset of all time-offset records relative to the center time of their respective planned collection windows; finally, the planned window hit rate and the time offset degree are normalized, and a collection time consistency value is generated. If there are instances within the current week's sampling period where no data is collected for an entire day, no data is collected for multiple consecutive planned periods, or the difference between the collection times of two sources exceeds a preset upper limit, the local prediction terminal can deduct the data collection time consistency value to reflect the stability of the week's observation input in the time dimension.

[0039] The process of obtaining cross-source bias may include the following steps: The local prediction terminal extracts the trap count sequence and insect lamp count sequence from the unified weekly observation records of the current monitoring point and the current weekly sampling period, and pairs them according to the same collection date and the same collection time. For successfully paired records, the pairing results are retained; for records with only single-source data, a missing data marker is added and they are not directly used for pairing bias calculation. Subsequently, the successfully paired trap count values ​​and insect lamp count values ​​are processed at the same scale. The same-scale processing can be performed using range normalization, mean-variance standardization, proportionalization relative to the cumulative value within the week, or proportionalization relative to the maximum value within the source, as long as it allows for comparison of differences between the two sources at the same scale. After completing the same-scale processing, the local prediction terminal calculates the cross-source bias from two aspects: numerical bias and trend bias. The numerical bias can be obtained by summing, averaging, or weighting the absolute values ​​of the differences between the two sources at the same collection date and the same collection time; the trend bias can be obtained by comparing whether the upward and downward directions of the two sources are consistent and whether the magnitude of change is synchronized between adjacent collection dates within a week, or by calculating the correlation coefficient, direction consistency rate, and slope bias of the weekly change sequences of the two sources. Preferably, a time-period difference sequence and a weekly trend direction sequence are first generated for each paired record within a week. Then, the mean result corresponding to the difference sequence and the consistency result corresponding to the trend direction sequence are jointly mapped to a cross-source deviation value. This reflects both the difference in single-point counts between the two sources and the degree of consistency in the two sources' judgments on adult emergence trends.

[0040] The generation of cross-source bias specifically includes: For all successfully paired records within the current weekly sampling period, calculate the absolute value of the difference between the trap count and the insect lamp count for each pairing time period, and obtain a time period difference set; calculate the mean of the time period difference set to obtain the first deviation; calculate the direction of change between adjacent records for the trap count sequence and the insect lamp count sequence in chronological order, count the number of times the direction of change between the two sources is inconsistent, and obtain the second deviation based on the proportion of inconsistent times to the total number of comparisons; normalize the first deviation and the second deviation and then weight them to obtain the cross-source deviation result of the current monitoring point within the current weekly sampling period.

[0041] This approach avoids misjudgment due to relying solely on single count differences, and also avoids ignoring absolute numerical differences due to relying solely on trend consistency.

[0042] After obtaining the key field completeness, data collection time consistency, and cross-source bias, the local prediction terminal generates an observation confidence coefficient. Specifically, weight combination rules can be pre-configured in the local prediction terminal, where key field completeness corresponds to the first weight, data collection time consistency corresponds to the second weight, and cross-source bias corresponds to the third weight. Since a larger cross-source bias value indicates a more significant difference between the two sources, the cross-source bias is converted into a cross-source consistency result before being included in the weighted synthesis; alternatively, a value comparable to the first two items can be obtained by directly subtracting the cross-source bias from the preset upper limit and normalizing it. After unifying the direction, the local prediction terminal performs a weighted summation of key field completeness, data collection time consistency, and cross-source consistency according to the preset weights to generate the observation confidence coefficient. Preferably, if the current business objective emphasizes prediction accuracy, the weight corresponding to cross-source consistency can be appropriately increased; if the current business objective emphasizes data collection standardization, the weights corresponding to key field completeness and data collection time consistency can be appropriately increased. For the scenario of predicting adult insects by weekly sampling in tea garden experimental sites, the weight of cross-source consistency can be set higher than the weight of sampling time consistency to highlight the mutual verification role of traps and insect lamps on the occurrence trend of the same adult insect.

[0043] In one implementation, after the observation confidence coefficient is generated, the local prediction terminal continues to generate the observation confidence interval. Specifically, the local prediction terminal first determines the baseline observation value for the current weekly sampling period. The baseline observation value can be a weighted fusion value of the adult insect count from the trap and the adult insect count from the insect lamp, or it can be the observation result obtained by fusing the cumulative count, mean count, or trend feature value of the two sources within a week. Preferably, the trap count and insect lamp count are weighted and fused using the same weighting rule as the observation confidence coefficient to generate the baseline observation value. Then, based on the missing error caused by insufficient completeness of key fields, the temporal error caused by inconsistent collection time, and the source difference error caused by cross-source bias, three types of error components are calculated respectively, and the three types of error components are synthesized to generate the observation interval width corresponding to the current weekly sampling period. Finally, the observation interval width is expanded upwards and downwards with the baseline observation value as the center to form the observation confidence interval.

[0044] The three error components can be calculated as follows: For missing errors caused by insufficient completeness of key fields, the first error component is generated according to the rule that "the lower the completeness, the larger the error"; for time-series errors caused by inconsistent acquisition times, the second error component is generated according to the rule that "the larger the time offset, the lower the hit rate of the planned window, and the larger the error"; for source difference errors caused by cross-source deviations, the third error component is generated according to the rule that "the larger the paired difference, the more inconsistent the trend, and the larger the error". Then, the first, second, and third error components are summed, weighted, or the sum of squares and square rooted to obtain the total error, and the observation interval width is generated based on the total error. If a weighted summation method is used, the weights corresponding to the three error components can be the same as those used in the aforementioned observation confidence coefficient generation stage, or they can be configured individually depending on whether more emphasis is placed on source consistency, time standardization, or field completeness.

[0045] Observation confidence intervals can also be generated through resampling. Specifically, the local prediction terminal resamples the trap count records and insect lamp count records within the current weekly sampling period multiple times. Each resampling assigns different sampling weights to the original records based on the evaluation results corresponding to the completeness of key fields, consistency of collection time, and cross-source bias. This ensures that higher-quality records have a higher probability of being selected, while lower-quality records have a lower probability of being selected. A fused observation value is generated from each resampling result. After multiple resampling iterations, multiple fused observation value samples are obtained. Then, the observation confidence interval is generated based on the distribution range of the multiple fused observation value samples. This method can directly reflect the differences in record quality in the width of the observation confidence interval without changing the original collection link, thus making the observation confidence interval closer to the uncertainty under actual collection conditions.

[0046] To ensure that subsequent prediction modules can directly access input data-level uncertainty information, the local prediction terminal outputs an input data-level uncertainty result table after completing the calculation. Each result includes at least: monitoring point number, week number, key field completeness value, acquisition time consistency value, cross-source deviation value, observation confidence coefficient, baseline observation value, lower bound of the observation confidence interval, upper bound of the observation confidence interval, and result generation time. For cases with missing data from a single source, simultaneous anomalies from two sources, or the number of effective pairings falling below a preset lower limit, an anomaly marker field can be added. Subsequent adult population dynamic models can directly read the observation confidence coefficient and observation confidence interval from the input data-level uncertainty result table as input for prior prediction correction and uncertainty propagation.

[0047] A dynamic model of adult insect population is constructed based on the trap counting sequence, insect lamp counting sequence, and observation confidence coefficient. The uncertainty information of the model parameter level corresponding to the state transition parameters, trap response parameters, and phototaxis response parameters in the dynamic model of adult insect population is determined. Specifically, it can be performed as follows.

[0048] First, the local prediction terminal reads the modeling input data for the current monitoring point within the current weekly sampling period from the aforementioned unified weekly observation record sequence and input data level uncertainty result table. The modeling input data includes at least: monitoring point number, week number, collection date, collection time period, adult insect count from traps, adult insect count from insect lamps, source validity marker, missing data marker, observation confidence coefficient, lower bound of the observation confidence interval, and upper bound of the observation confidence interval. For the same monitoring point with multiple collection time periods within a week, the local prediction terminal first sorts all records chronologically according to the collection date and collection time period, then removes or downweights invalid records, duplicate records, and marked abnormal records to generate the valid modeling record set corresponding to the current monitoring point and the current weekly sampling period.

[0049] After obtaining a valid set of modeling records, the local prediction terminal constructs a modeling input matrix with a source-separated observation structure. Specifically, instead of directly merging the trap counts and insect lamp counts into a single observation, independent observation channels are established for each source, forming parallel inputs on the same time axis. For each collection date and collection period, at least one corresponding period record is generated; the period record includes at least: the adult insect count of the traps in the current period, the adult insect count of the insect lamps in the current period, the observation confidence coefficient of the current period, the observation confidence interval of the current period, and the source validity marker. If valid records exist for both sources in the same period, it is recorded as a dual-channel valid period; if only a single source has a valid record, it is recorded as a single-channel valid period; if neither source has a valid record, it is recorded as a missing period. In this way, the subsequent model receives independent observation constraints from the traps and insect lamps in the same period, instead of pre-mixing the two and losing source difference information.

[0050] Before modeling, the local prediction terminal performs scaling and weight preparation on the effective modeling record set. Specifically, for the trap counting sequence and the insect lamp counting sequence, the cumulative value, mean, maximum value, and change between adjacent time periods are calculated, and observation weights are generated for each time period based on the observation confidence coefficient. Preferably, the observation confidence coefficient is directly mapped to the time period weight value, so that time periods with high confidence have a higher proportion in subsequent state updates and parameter estimation; at the same time, the observation confidence interval width is converted into an interval penalty, so that the wider the interval, the lower the constraint strength of the record. After this processing, a dual-channel time series input table for model construction is formed.

[0051] Subsequently, the local prediction terminal constructs a dynamic model of the adult insect population. This model is built using a rolling, time-period-based approach and includes at least three state variables: adult population size, trapping response, and phototactic response. Specifically, the adult population size corresponds to the estimated potential adult size at the target monitoring point for each collection period; the trapping response corresponds to the intensity of the trap's response to the current adult size during the same collection period; and the phototactic response corresponds to the intensity of the insect monitoring lamp's response to the current adult size during the same collection period. To ensure model execution, the local prediction terminal establishes a time-period-based recursive state record table for each of the three state variables. Each time period stores the current state value, the state value of the previous period, the predicted value before the state update, the corrected value after the state update, and the associated observation residuals.

[0052] When constructing the adult insect population dynamics model, the initial state values ​​for the first time period are determined. For the adult insect population size state quantity, the local prediction terminal preferably generates initial values ​​based on the adult count values ​​from traps, insect lamps, and the observation confidence coefficient during the first effective collection period. If both sources are valid during the first effective collection period, the count results from the two sources are weighted and synthesized, and the synthesized result is used as the initial estimate of the adult insect population size state quantity. If only a single source is valid during the first effective collection period, the count value from that single source is used in conjunction with a preset conversion coefficient to obtain the initial estimate, while widening the initial uncertainty range. For the trap response state quantity and the phototactic response state quantity, the local prediction terminal can use preset initial response values ​​or generate initial estimates based on the source count ratios of the previous few effective time periods within the current weekly sampling cycle. For example, when the trap count is consistently higher than the insect lamp count in the previous few time periods, the initial value of the trap response state quantity can be increased accordingly; conversely, when the insect lamp count is consistently higher than the trap count, the initial value of the phototactic response state quantity can be increased accordingly. In this way, even in the absence of long-term historical samples, the model can still be initialized based on valid observations within the current weekly sampling period.

[0053] After initialization, the local prediction terminal recursively updates three state variables according to the collection time period. For the adult population size state variable, the local prediction terminal first generates a predicted value based on the adult population size state value of the previous time period and the state transition parameters, and then corrects the predicted value using the current time period's trap count and insect lamp count. The state transition parameters are used to characterize the continuity, decay, or rate of change of the potential adult population size between adjacent collection time periods. For the trap response state variable, the local prediction terminal updates the current time period's trap response state value based on the deviation between the previous time period's trap response state value, the current trap count, and the current predicted adult population size. For the phototactic response state variable, the local prediction terminal updates the current time period's phototactic response state value based on the deviation between the previous time period's phototactic response state value, the current insect lamp count, and the current predicted adult population size. Since the three state variables are updated in tandem within the same time period, when a mutation occurs in the observation from one source while the observation from another source does not change synchronously, the model will not immediately interpret the mutation as a change in the adult population size. Instead, it will prioritize absorbing the local bias of the corresponding source by adjusting the response state variable of that source, thereby improving the model's ability to explain the differences between the two sources.

[0054] In one embodiment, the implementation of using the trap counting sequence and the insect lamp counting sequence as different observation channels to constrain the adult population size state quantity may include the following process: First, generate the predicted trap count value for the current time period based on the predicted adult population size and the predicted trapping response state value for the current time period; then, generate the predicted insect lamp count value for the current time period based on the predicted adult population size and the predicted phototactic response state value for the current time period; next, compare the predicted trap count value for the current time period with the actual trap count value to obtain the observation residual for the trapping channel; simultaneously, compare the predicted insect lamp count value for the current time period with the actual insect lamp count value to obtain the observation residual for the phototactic channel; finally, based on the magnitude, direction, and corresponding observation confidence weight of the two observation residuals, jointly correct the adult population size state quantity, the trapping response state quantity, and the phototactic response state quantity.

[0055] This processing method ensures that the two sources do not simply compete for the same output value, but rather impose constraints on the internal state of the model in different directions and with different intensities through their respective channel residuals.

[0056] To improve the stability of parameter estimation results, the local prediction terminal further determines state transition parameters, trapping response parameters, and phototactic response parameters. Specifically, the local prediction terminal extracts state recursion results and channel residual results from all valid time period records within the current weekly sampling cycle to construct a parameter estimation sample set. The parameter estimation sample set includes at least: adult population size state values, trapping response state values, phototactic response state values, current trap count values, current insect lamp count values, current observation confidence weight, and current observation residuals for adjacent time periods. Then, a constrained weighted estimation method is used to solve for the parameters. Preferably, a weighted least squares iterative solution method can be used: aiming to minimize the weighted sum of squares of the dual-channel observation residuals, the time periods with high confidence contribute more weight in the parameter solution, while setting allowable value ranges for each parameter to prevent abnormal results that are neither physically nor operationally meaningful. The state transition parameters obtained in this way are used to describe the recursive relationship between the potential adult size at different time periods within the week, the trapping response parameters are used to describe the coupling relationship between the change in trap count and the change in potential adult size, and the phototactic response parameters are used to describe the coupling relationship between the change in insect lamp count and the change in potential adult size.

[0057] When the number of valid time periods within the current weekly sampling period is small, or the number of valid time periods for both channels is below a preset lower limit, the local prediction terminal does not directly abandon parameter solving. Instead, it determines the parameters using a method of "preset initial parameter values ​​+ current week recursive correction." Specifically, it first calls the default state transition parameters, default trapping response parameters, and default phototactic response parameters preset during the device deployment phase as initial values. Then, it uses the limited number of valid records within the current weekly sampling period to perform small-scale corrections on the above parameters, while simultaneously expanding the parameter uncertainty range. If the current monitoring point has parameter results corresponding to the previous valid weekly sampling period, it prioritizes using the parameter results corresponding to the previous valid weekly sampling period as the initial parameters for the current week to enhance the continuity of model operation in the local deployment scenario.

[0058] After the parameters are solved, the local prediction terminal generates model parameter level uncertainty information. Specifically, the parameter fitting error is first calculated based on the final parameter results and the dual-channel observation residuals over all valid time periods, and then the interval range corresponding to each parameter is generated based on the parameter fitting error. Preferably, this can be achieved using any one or a combination of the following two methods: The first method is residual propagation. Specifically, the local prediction terminal calculates the residual variance in the parameter estimation process based on the dispersion of the residuals of the trapping channel and the phototactic channel in each time period. Then, combined with the number of effective samples in the current weekly sampling period, the proportion of effective time periods for both channels, and the observation confidence weight distribution, it generates the parameter interval widths corresponding to the state transition parameters, trapping response parameters, and phototactic response parameters. Subsequently, using the final estimated value of each parameter as the center, the parameter interval widths are expanded upwards and downwards to form the corresponding lower and upper bounds of the parameters.

[0059] The second method is resampling. Specifically, the local prediction terminal performs multiple resampling operations on the valid time period records within the current weekly sampling cycle. Each resampling is performed according to the observation confidence weight, making it easier to select time periods with high confidence and reducing the probability of selecting time periods with low confidence. Parameter estimation is performed on each resampling result to obtain a set of state transition parameters, trapping response parameters, and phototactic response parameters. After multiple repetitions, a sample distribution of each parameter is formed, and the corresponding model parameter level uncertainty interval is generated based on the upper and lower quantile values ​​of the sample distribution. This method can directly reflect the impact of time period quality differences, dual-channel differences, and local anomaly observations on parameter stability in the width of the parameter interval.

[0060] The local prediction terminal simultaneously generates parameter reliability labels. Specifically, based on the parameter interval width, the number of effective time periods in the dual channels, the coverage of time periods within the week, and whether the dual-channel residuals are consistently large, state transition parameters, trapping response parameters, and phototactic response parameters are assigned high-reliability, medium-reliability, or low-reliability labels, respectively. When a parameter is labeled as low-reliability, the local prediction terminal limits the update magnitude of that parameter in subsequent prior prediction stages, or prioritizes using the parameter value corresponding to the previous effective week's sampling period, to avoid excessive amplification of the impact of abnormal observations in the current week on subsequent model predictions.

[0061] To facilitate direct use by subsequent modules, after completing the above modeling and parameter estimation, the local prediction terminal outputs a result table of the adult population dynamics model and a result table of model parameter level uncertainty. The result table of the adult population dynamics model includes at least: monitoring point number, week number, collection date, collection period, predicted adult population size and state, corrected adult population size and state, predicted trapping response state, corrected trapping response state, predicted phototactic response state, corrected phototactic response state, observation residuals of the trapping channel, and observation residuals of the phototactic channel. The result table of model parameter level uncertainty includes at least: monitoring point number, week number, estimated state transition parameters, lower bound of state transition parameters, upper bound of state transition parameters, estimated trapping response parameters, lower bound of trapping response parameters, upper bound of trapping response parameters, estimated phototactic response parameters, lower bound of phototactic response parameters, upper bound of phototactic response parameters, parameter confidence markers, and result generation time. The subsequent prior prediction module can directly read the state variables, parameter values, and parameter ranges in the results table as input for the rolling evolution and uncertainty propagation within the future target prediction period.

[0062] The adult population dynamics model is used to perform prior evolution on the adult occurrence trend during the future target prediction period to obtain a prior prediction sequence. Based on the adult observation data and observation confidence coefficient in the current weekly sampling period, the prior prediction sequence is assimilated and corrected to obtain a corrected prediction sequence and generate assimilation correction level uncertainty information. Specifically, it can be performed as follows.

[0063] First, the local prediction terminal reads the aforementioned adult population dynamics model result table and model parameter level uncertainty result table to determine the latest state input corresponding to the current monitoring point at the end of the current weekly sampling period. The latest state input includes at least: adult population size state correction value, trapping response state correction value, phototactic response state correction value, state transition parameter estimate, trapping response parameter estimate, phototactic response parameter estimate, lower and upper bounds of each parameter, observation residuals of the trapping channel, observation residuals of the phototactic channel, and parameter confidence markers. The local prediction terminal also reads the observation confidence coefficients, observation confidence intervals, source validity markers, and anomaly markers corresponding to each collection date and collection time period within the current weekly sampling period from the input data level uncertainty result table, and completes the association according to the monitoring point number and week sequence number. After this processing, the prediction starting dataset corresponding to the current monitoring point is formed.

[0064] After obtaining the initial dataset for prediction, the local prediction terminal determines the future target prediction period. The future target prediction period can be divided using the same time granularity as the current weekly sampling period, such as by day, by a preset collection period, or by the rolling period of the following week, as long as it maintains the same temporal granularity as the aforementioned unified weekly observation record sequence. Preferably, the future target prediction period generates multiple future prediction nodes according to the same time period division method as the current weekly sampling period, allowing the prior prediction process to continuously iterate using the existing model structure and parameter structure. For each future prediction node, the local prediction terminal assigns at least one future time period number, a future date identifier, and a future time period identifier, so that subsequent output of the prior prediction sequence and the correction prediction sequence can correspond to each node.

[0065] Subsequently, the local prediction terminal performs prior evolution. Specifically, it uses the adult population size state correction value at the end of the current weekly sampling period as the starting point for future prediction, and the trapping response state correction value and phototactic response state correction value at the end of the current weekly sampling period as the response starting points for the two observation channels. It then recursively extrapolates to future nodes based on state transition parameters, trapping response parameters, and phototactic response parameters. For each future prediction node, it first generates the predicted adult population size value for the current future node based on the adult population size state value and state transition parameters of the previous node; then it generates the predicted trapping response value for the current future node based on the trapping response state value and trapping response parameters of the previous node; simultaneously, it generates the predicted phototactic response value for the current future node based on the phototactic response state value and phototactic response parameters of the previous node. After completing the recursion of the three state variables, it further generates the predicted count values ​​for the trapping channel and the phototactic channel based on the predicted adult population size value and the corresponding channel response prediction value for the current future node. In this way, prior prediction results are generated in the order of future nodes without the introduction of observation corrections for the current weekly sampling period.

[0066] The prior evolution is executed in a node-by-node rolling manner. That is, after the state prediction of the current future node is completed, the prediction result of the current future node is used as the input of the next future node, until the recursion of all future target prediction periods is completed. This ensures the temporal continuity of the prior prediction sequence and allows parameter changes, residual state effects, and channel response changes to be gradually propagated across multiple future nodes. To avoid the excessive amplification of abnormal recursion results of a certain future node in subsequent nodes, the local prediction terminal can set reasonable range constraints for the predicted values ​​of adult population size, trap response, and phototaxis response of each future node. When a predicted value exceeds the preset upper limit or falls below the preset lower limit, the excess portion is truncated, pulled back, or attenuated according to a preset ratio, and the processing result is written to the prior prediction log for subsequent traceability.

[0067] After completing the above recursion, the local prediction terminal generates a priori prediction sequence. The priori prediction sequence includes at least: monitoring point number, future node number, future date identifier, future time period identifier, priori prediction value of adult population size, priori prediction value of trapping response, priori prediction value of phototaxis response, priori count value of trapping channels, priori count value of phototaxis channels, and corresponding parameter version identifiers. If model parameter-level uncertainty information also participates in the priori evolution, prior intervals for adult population size, trapping channels, and phototaxis channels can also be added to the priori prediction sequence.

[0068] After generating the prior prediction sequence, the local prediction terminal performs assimilation correction on it. The data used for assimilation correction comes from the adult insect observation data acquired and preprocessed within the current weekly sampling period. Specifically, this includes: the trap count sequence, insect lamp count sequence, observation confidence coefficients for each time period, observation confidence intervals for each time period, source validity markers, missing data markers, and the completeness of key fields, collection time consistency, and cross-source bias results corresponding to the input data level uncertainty result table. To ensure that the assimilation correction corresponds to the prior prediction on the time axis, the local prediction terminal first maps the observation records within the current weekly sampling period to time nodes consistent with the prior prediction sequence according to the collection date and collection time, forming a correction observation table. For observation records that can correspond one-to-one with the prior prediction nodes, effective correction pairings are directly formed; for multiple observation records from the same source at the same node, the final observation value of that source under that node is generated by summation, mean, maximum value, or last value retention; for observation records that cannot correspond to the prior prediction nodes, they are marked as records not participating in the correction.

[0069] In one embodiment, the process of assigning observation update weights to the trap counting sequence and the insect lamp counting sequence according to the observation confidence coefficient may include the following steps.

[0070] First, for each valid calibration pairing node, read the observation values ​​of the trap, the insect lamp, the prior count value of the trapping channel, and the prior count value of the phototaxis channel under that node.

[0071] Then, the observation confidence coefficient corresponding to the node is read, and update weights are generated for different sources in combination with the source validity label. If both sources are valid, the first update weight and the second update weight are generated for the trap channel and the insect lamp channel, respectively; if only a single source is valid, update weights are generated only for the valid source, and the update weights corresponding to the missing sources are reset to zero or a preset extremely low value.

[0072] Subsequently, the update weight of the corresponding source is increased according to the principle that the higher the observation confidence coefficient, the narrower the observation confidence interval, and the better the source validity; the update weight of the corresponding source is decreased according to the principle that the lower the observation confidence coefficient, the wider the observation confidence interval, and the presence of anomaly markers in the source.

[0073] This approach allows the impact of different channels on the prior prediction sequence at different nodes to change dynamically based on the current observation quality, rather than using a fixed update ratio.

[0074] Preferably, when generating update weights, the local prediction terminal also considers the relative quality difference between the two sources. Specifically, it first generates basic source weights based on the observation confidence coefficients of the two sources at the same node. Then, it normalizes the basic source weights by combining the observation interval widths of the two sources, source anomaly markers, and whether there is cross-source bias amplification, so that the sum of the update weights of the two sources at the same node does not exceed the preset total update weight upper limit. This avoids applying excessive corrections to the adult population size and state when both sources have large biases.

[0075] After obtaining the update weights for each node, the local prediction terminal corrects the adult population size state in the prior prediction sequence based on the observation update weights. Specifically, for each valid correction pair node, the observation residuals of the trapping channel and the phototactic channel are first calculated. The observation residual of the trapping channel is obtained by comparing the actual observation value of the trapper at that node with the prior count value of the trapping channel, and the observation residual of the phototactic channel is obtained by comparing the actual observation value of the insect lamp at that node with the prior count value of the phototactic channel. Then, the two channel residuals are weighted according to the first update weight and the second update weight to form the node weighted residual. Subsequently, the prior prediction value of the adult population size of that node is corrected according to the node weighted residual to obtain the corrected value of the adult population size of that node. After the correction is completed, the trapping response state value and the phototactic response state value of that node are updated in reverse according to the corrected value of the adult population size to form the channel response correction value corresponding to that node. In this way, the correction of the adult population size status is not performed in isolation, but is updated in line with the response status of both channels, making the corrected results more consistent with the actual dual-channel observations.

[0076] When the observed residuals from both channels are in the same direction and have large absolute values, the local prediction terminal prioritizes increasing the correction magnitude for the adult population size state quantity to reflect changes in the potential adult size indicated by both sources. When the observed residuals from the two channels are in different directions, or when only one channel's residual is significantly larger, the local prediction terminal prioritizes using the response state quantity of the corresponding source to absorb the local bias of that source and reduces the direct correction ratio for the adult population size state quantity. This avoids misinterpreting anomalous disturbances from a single source as true jumps in the potential adult size, thereby improving the stability and reliability of the corrected prediction sequence.

[0077] To ensure the continuity of the assimilation correction process, the local prediction terminal adopts a node-by-node recursive correction method. That is, after correcting the adult population size and state of the current node, the corrected values ​​for the adult population size, trapping response state, and phototactic response state of that node are used as the starting point for the correction of the next node. The same observation residual comparison and weighted correction are then performed on the prior prediction value of the next node until all nodes requiring correction have been processed. For nodes lacking effective observations within the current weekly sampling period, the corrected state of the previous node is directly inherited as the intermediate state of the current node, or the prior prediction value of that node remains unchanged, and an "uncorrected node" mark is added to the results table so that the interval for such nodes can be expanded during subsequent uncertainty calculations.

[0078] After completing the calibration of all nodes, the local prediction terminal generates a calibrated prediction sequence. The calibrated prediction sequence includes at least: monitoring point number, node number, date identifier, time period identifier, adult population size calibration value, trapping response status calibration value, phototactic response status calibration value, trapping channel calibration count value, phototactic channel calibration count value, trapping channel observation residual, phototactic channel observation residual, update weight flag, calibration status flag, and result generation time. For uncalibrated nodes, single-source calibrated nodes, and dual-source calibrated nodes, the local prediction terminal preferably writes different node type identifiers to identify the quality source of the corresponding node when outputting risk levels and intervention windows.

[0079] While generating the corrected prediction sequence, the local prediction terminal generates assimilated correction level uncertainty information. Specifically, the assimilated correction level uncertainty information is generated based on at least the following factors: The magnitude of the observation residual of the trapping channel at the current node; the magnitude of the observation residual of the phototactic channel at the current node; The update weights of the two sources; the width of the observation confidence interval corresponding to the node; whether the node is a valid node from two sources, a valid node from a single source, or an uncorrected node; and whether the changes in correction results between adjacent nodes are stable.

[0080] In one implementation, the local prediction terminal first generates a node correction error for each node. The node correction error can be obtained by weighting the observation residuals of the trapping channel and the phototactic channel according to update weights, and then performing amplification or reduction processing based on the observation interval width for that node. When a node is a single-source correction node, an additional source missing penalty is added; when a node is an uncorrected node, the node correction error is directly set to a high uncertainty level. Subsequently, all node correction errors are arranged in chronological order to form an assimilation correction error sequence, and an assimilation correction interval width is further generated for each node.

[0081] The assimilation correction interval width can be generated as follows: Centered on the adult population size correction value of the current node, the assimilation correction interval for the current node is expanded upwards and downwards according to the node correction error. If the change in the correction result between the current node and the previous node exceeds a preset change threshold, a stationarity penalty is added to the assimilation correction interval of the current node to prevent local mutations from being misjudged as high-confidence changes. If the current node is corrected by effective observations from both sources, and the residuals of both channels are small, the assimilation correction interval width is reduced. If the current node is corrected by only a single source, or if it is a dual-source effective node but the residuals of the two channels differ significantly, the assimilation correction interval width is increased. This method can reflect the correction effect of the current node observations on the model, and also reflect factors such as the stability of the current node correction result and whether it is dominated by a single source in the uncertainty information.

[0082] Assimilation correction level uncertainty information can also be generated through replay correction. Specifically, the local prediction terminal fixes the same set of prior prediction sequences and performs multiple perturbation replays on the update weights, observation residuals, and source validity labels of each observation node within the current weekly sampling period. Each replay resamples the trap observations and insect lamp observations within the observation confidence interval and performs node correction again accordingly. After multiple repetitions, multiple adult population size correction value samples are obtained for the same node, and the assimilation correction interval for that node is generated based on the upper and lower quantiles of the sample distribution. This method allows the assimilation correction level uncertainty information to more directly reflect the impact of observation quality fluctuations and correction weight fluctuations on the final prediction results.

[0083] To facilitate direct use by the subsequent uncertainty forward propagation module, the local prediction terminal outputs an assimilation correction result table and an assimilation correction level uncertainty result table after completing the assimilation correction. The assimilation correction result table includes at least: monitoring point number, node number, prior predicted adult population size, corrected adult population size, prior count value of the trapping channel, observed value of the trapping channel, prior count value of the phototactic channel, observed value of the phototactic channel, first update weight, second update weight, observed residual of the trapping channel, observed residual of the phototactic channel, and node type identifier. The assimilation correction level uncertainty result table includes at least: monitoring point number, node number, node correction error, lower bound of the assimilation correction interval, upper bound of the assimilation correction interval, width of the assimilation correction interval, source missing penalty flag, stationarity penalty flag, and result generation time. The subsequent uncertainty propagation module can directly use the node correction error and assimilation correction interval width from the result table to generate quantitative uncertainty intervals corresponding to the occurrence index, outbreak probability, risk level, and warning window.

[0084] The uncertainty information at the input data level, the uncertainty information at the model parameter level, and the uncertainty information at the assimilation correction level are forward propagated using the error propagation formula or Monte Carlo sampling method to obtain the quantitative uncertainty intervals of the occurrence index, outbreak probability, risk level, and warning window corresponding to the correction prediction sequence. Specifically, this can be done as follows.

[0085] First, the local prediction terminal reads the aforementioned input data level uncertainty result table, model parameter level uncertainty result table, assimilation correction result table, and assimilation correction level uncertainty result table, and completes the association according to the monitoring point number, week number, and node number to form the forward propagation input dataset for the current monitoring point in the current prediction period. The forward propagation input dataset includes at least: the observation confidence coefficient, lower bound of the observation confidence interval, upper bound of the observation confidence interval, estimated value of the state transition parameter, lower bound of the state transition parameter, upper bound of the state transition parameter, estimated value of the trapping response parameter, lower bound of the trapping response parameter, upper bound of the trapping response parameter, estimated value of the phototactic response parameter, lower bound of the phototactic response parameter, upper bound of the phototactic response parameter, adult population size correction value, trapping response state correction value, phototactic response state correction value, node correction error, lower bound of the assimilation correction interval, upper bound of the assimilation correction interval, source missing penalty label, and stationarity penalty label for each node. If a node has a missing single source, insufficient validity of dual sources, or a parameter credibility marker indicating low reliability, then the corresponding node will be marked with a high uncertainty marker for use in subsequent propagation stages.

[0086] After obtaining the forward propagation input dataset, the local prediction terminal first performs input processing and scale unification. Specifically, since the input data-level uncertainty information, model parameter-level uncertainty information, and assimilation correction-level uncertainty information have different sources, dimensions, and interval widths, a unified scale transformation is first performed on all types of uncertainty measures. For input data-level uncertainty information, the local prediction terminal extracts the observation confidence interval width and observation confidence coefficient, and scales the observation confidence interval width according to the average observation level of the current monitoring point within the current prediction period. For model parameter-level uncertainty information, the local prediction terminal extracts the parameter interval widths corresponding to the state transition parameters, trapping response parameters, and phototactic response parameters, and normalizes the parameter interval widths according to the corresponding parameter estimates. For assimilation correction-level uncertainty information, the local prediction terminal extracts the node correction error and assimilation correction interval width, and scales them according to the correction values ​​for the corresponding node adult population size. After scaling, the three types of uncertainty results are uniformly mapped to the same numerical scale for subsequent weighted propagation or random sampling.

[0087] In one embodiment, the model parameter level uncertainty information includes at least the parameter confidence intervals corresponding to the state transition parameters, trapping response parameters, and phototactic response parameters. The parameter confidence intervals are obtained by estimating the parameters of the adult population dynamics model based on the effective adult observation data within the current weekly sampling period, and are used to constrain the fluctuation range of the prior prediction sequence. The specific implementation method is as follows: When generating the prior prediction sequence, the local prediction terminal not only calls the estimated values ​​of each parameter but also reads the corresponding lower and upper bounds of each parameter. When progressively extrapolating the adult population size, trapping response, and phototactic response states for future nodes, the parameter values ​​used in the extrapolation are limited to their respective confidence intervals. If the parameter correction trend corresponding to the theoretical extrapolation result for a certain node would cause the parameter to exceed its lower or upper bound, the parameter value exceeding the interval is not directly adopted. Instead, the parameter is truncated to the corresponding interval boundary or extrapolated using a decay method close to the interval boundary. This method ensures that the node fluctuation range of the prior prediction sequence is always constrained by the parameter interval supported by effective observations within the current weekly sampling period, thereby avoiding excessive fluctuations that deviate from the current observation support range due to the amplification of local anomalies during future extrapolation.

[0088] After completing input processing and parameter constraint preparation, the local prediction terminal generates four types of output results corresponding to the corrected prediction sequence. These four types of output results include at least an occurrence index, outbreak probability, risk level, and warning window. Specifically, the local prediction terminal first extracts the adult population size correction value, trapping channel correction count value, phototactic channel correction count value, and node type identifier corresponding to all future nodes from the corrected prediction sequence. Then, it generates the four types of output results according to preset business rules. The occurrence index can be generated based on the cumulative, average, peak, or weighted combination of the adult population size correction values ​​within the future target prediction period. The outbreak probability can be generated based on the proportion of nodes whose adult population size correction values ​​or channel correction count values ​​exceed the preset outbreak threshold, the cumulative exceedance, or the frequency of interval crossings within the future target prediction period. The risk level can be generated by mapping the occurrence index and outbreak probability together to preset level classification rules. The warning window can be determined based on the continuously rising range of the adult population size correction value in future nodes, the outbreak probability crossing the threshold range, and the continuous range where the risk level reaches the preset warning level. Therefore, the local prediction terminal first obtains the central estimates of the four types of output results, and then generates the corresponding quantitative uncertainty intervals around the central estimates.

[0089] In one embodiment, the process of calculating the occurrence index, outbreak probability, risk level, and the quantitative uncertainty interval corresponding to the warning window using the error propagation formula may include the following steps.

[0090] First, propagation relationships of four types of output results relative to three types of uncertainty inputs are established based on the corrected prediction sequence. Specifically, the occurrence index is expressed as a function of the corrected adult population size values ​​of multiple nodes within the future target prediction period; the outbreak probability is expressed as a cross-threshold function of the corrected adult population size values ​​of multiple nodes relative to the outbreak threshold; the risk level is expressed as a level mapping function under the combined effect of the occurrence index and the outbreak probability; and the warning window is expressed as a function of the time period in which the risk status of multiple nodes continuously meets preset conditions.

[0091] Then, the central values ​​and interval widths of the input data-level uncertainty information, model parameter-level uncertainty information, and assimilation correction-level uncertainty information are extracted respectively, and the influence strength of each input term on each output term is calculated according to the propagation relationship. Preferably, the local prediction terminal obtains the sensitivity of changes in each input term to changes in the occurrence index, outbreak probability, risk level, and warning window by performing local linearization processing on the correlation function near the current central estimation point.

[0092] Subsequently, based on the sensitivity and the magnitude of uncertainty corresponding to each input item, the occurrence exponential variance, outbreak probability variance, risk level fluctuation, and warning window offset are calculated respectively, and these items are combined to generate the propagation interval width corresponding to each of the four types of output results.

[0093] Finally, using the central estimate of each of the four types of output results as the center, the corresponding propagation interval widths are expanded upwards and downwards to form the occurrence index interval, outbreak probability interval, risk level interval, and early warning window interval, respectively.

[0094] When calculating the occurrence index interval using the error propagation formula, the local prediction terminal first accumulates the fluctuations in the correction values ​​of adult population sizes at all nodes within the future target prediction period. Then, it maps the observational fluctuations caused by input data level uncertainty, the recursive fluctuations caused by model parameter level uncertainty, and the correction fluctuations caused by assimilation correction level uncertainty onto this accumulated result, obtaining the total fluctuation corresponding to the occurrence index. Subsequently, the lower and upper bounds of the occurrence index are formed based on the total fluctuation. This method allows the uncertainty interval of the occurrence index to simultaneously reflect the impact of observation quality, model stability, and correction stability.

[0095] When calculating the outbreak probability interval using the error propagation formula, the local prediction terminal first calculates the central outbreak probability based on the estimated value of the occurrence index center or the center estimate of the future node adult population size correction value. Then, based on the fluctuation of nodes adjacent to the outbreak threshold, it calculates the possible range in which the outbreak conditions are met when the interval fluctuates above and below. Specifically, if the center estimate is much higher than the outbreak threshold, both the upper and lower bounds of the outbreak probability interval are higher and the interval is narrower; if the center estimate is close to the outbreak threshold, the input fluctuation has a more significant impact on the cross-threshold result, corresponding to a wider outbreak probability interval. In this way, the outbreak probability interval can more realistically reflect the impact of "nodes close to the threshold" on the stability of the final outbreak judgment.

[0096] In one implementation, when calculating the risk level interval using the error propagation formula, since the risk level itself is a discrete output, the local prediction terminal first maps the quantitative intervals of the occurrence index and outbreak probability to the level threshold system, and statistically analyzes the possible level ranges covered within the corresponding intervals. If the interval falls within a single level range, a single-level risk interval is output; if the interval spans two or more level thresholds, a multi-level coverage range is output, and the coverage ratio of each level within the interval is used as supplementary information for level uncertainty. This method preserves the uncertainty boundary of the risk level results without changing the discrete level output format.

[0097] When calculating the early warning window interval using the error propagation formula, the local forecasting terminal first determines the start and end nodes of the central early warning window. Then, based on the propagation interval width corresponding to each node and the stability of the node's risk status, it calculates the range of nodes that the early warning window start point may move forward or backward, as well as the range of nodes that the early warning window end point may move forward or backward, ultimately forming the early warning window start interval and the early warning window end interval. If multiple discrete high-risk segments exist within the future target prediction period, a corresponding early warning window interval is generated for each high-risk segment. In this way, the early warning window is no longer just a single fixed time period, but rather gives its possible start and end ranges in the form of intervals, which is more suitable for subsequent risk stratification in decision-making.

[0098] Monte Carlo sampling was performed on the observation confidence coefficient, parameter confidence interval, and assimilation correction error to obtain the interval distribution results corresponding to the occurrence index, outbreak probability, risk level, and warning window. The specific process is as follows.

[0099] First, the local prediction terminal establishes sampling rules for input data-level uncertainty, model parameter-level uncertainty, and assimilation correction-level uncertainty. For input data-level uncertainty, random perturbation samples are generated for each node's observations based on the observation confidence coefficient and observation confidence interval. For model parameter-level uncertainty, random sampled values ​​are generated for each parameter, using the lower and upper bounds of the state transition parameters, trap response parameters, and phototactic response parameters as ranges. For assimilation correction-level uncertainty, random perturbation samples are generated for each node's correction values ​​based on the node correction error and the assimilation correction interval width. Then, in each sampling iteration, the local prediction terminal randomly selects a set of observation perturbation values, a set of parameter sampled values, and a set of correction perturbation values, and regenerates a corresponding future output path based on this set of sampling results. The future output path includes at least one complete occurrence index calculation result, one complete outbreak probability calculation result, one complete risk level mapping result, and one complete warning window determination result. Subsequently, the sampling iteration is repeated a preset number of times to obtain multiple future output path samples. Finally, the occurrence index results in all sample paths are sorted and their upper and lower quantile values ​​are extracted to generate occurrence index intervals; the outbreak probability results in all sample paths are sorted and their upper and lower quantile values ​​are extracted to generate outbreak probability intervals; the frequency of occurrence of each risk level in all sample paths is statistically analyzed to generate risk level coverage intervals or level distribution results; the start and end points of the warning window in all sample paths are sorted and their upper and lower quantile values ​​are extracted to generate warning window start and end intervals. In this way, output interval distribution results matching the combined effect of the three types of uncertainty can be directly obtained without explicitly deriving complex propagation relationships.

[0100] To improve the stability of Monte Carlo sampling results, the local prediction terminal preferably adopts a stratified sampling method. Specifically, the sample space is first divided into multiple sampling layers based on node type, source validity, parameter reliability label, and interval width level. Then, random sampling is performed within each sampling layer to avoid the sampling results being overly dominated by a small number of extreme nodes. If a monitoring point has a large number of nodes with missing sources or low-reliability parameters within the current weekly sampling period, the sampling proportion of the corresponding sampling layer is increased to make the output interval more fully reflect the impact of such unstable factors.

[0101] After completing the error propagation formula calculation or Monte Carlo sampling, the local prediction terminal performs result shaping on the four types of output interval results. Specifically, for occurrence index intervals and outbreak probability intervals, if the lower bound of the interval is lower than a preset lower limit, it is truncated to the corresponding lower limit; if the upper bound of the interval is higher than a preset upper limit, it is truncated to the corresponding upper limit. For risk level intervals, if the interval spans multiple levels, it is output in ascending order of level; for warning window intervals, if the starting interval is later than the ending interval, or if there is overlap or confusion between two intervals, the start and end boundaries are recalibrated according to the central warning window order, and interval reshaping markers are added to the results. This method ensures that the output results maintain business interpretability and format consistency.

[0102] To facilitate direct access by the subsequent decision-making risk stratification module, the local forecasting terminal outputs an uncertainty propagation result table after completing forward propagation. The uncertainty propagation result table includes at least the following: monitoring point number, week number, occurrence index center value, occurrence index lower bound, occurrence index upper bound, outbreak probability center value, outbreak probability lower bound, outbreak probability upper bound, risk level center value, risk level coverage area, warning window center start point, warning window center end point, warning window start interval, warning window end interval, propagation method identifier, interval renormalization marker, and result generation time. If Monte Carlo sampling is used, additional fields for sampling number, sampling layer label, and sample distribution summary can be added; if error propagation formula is used, an additional field for propagation sensitivity summary can be added.

[0103] Based on the width of the quantitative uncertainty interval and its changing trend, decision risk stratification is performed on the prediction results, and backtracking prediction results are generated when preset abnormal conditions are met. Specifically, it can be performed as follows.

[0104] First, the local forecasting terminal reads the aforementioned uncertainty propagation result table, input data level uncertainty result table, unified weekly observation record table, assimilation correction result table, and assimilation correction level uncertainty result table, and associates them according to the monitoring point number, week sequence number, and future forecast node number to form a risk-stratified input dataset for the current monitoring point within the current forecast period. The risk-stratified input dataset includes at least: occurrence index center value, occurrence index lower bound, occurrence index upper bound, outbreak probability center value, outbreak probability lower bound, outbreak probability upper bound, risk level center value, risk level coverage, warning window center start point, warning window center end point, warning window start interval, warning window end interval, observation confidence coefficient, key field completeness, cross-source deviation value, source missing marker, node correction error, and stationarity penalty marker. For a monitoring point containing multiple future forecast nodes or multiple warning segments, the local forecasting terminal first sorts them chronologically, then generates corresponding node-level risk records and segment-level risk records for use in node judgment and overall warning judgment, respectively.

[0105] After obtaining the risk stratification input dataset, the local prediction terminal first extracts the interval width information used for stratification determination. Specifically, since the occurrence index, outbreak probability, risk level, and warning window each have their own corresponding interval results, the local prediction terminal first calculates the interval width corresponding to the four types of outputs. The occurrence index interval width is obtained by subtracting the lower bound from its upper bound; the outbreak probability interval width is obtained by subtracting the lower bound from its upper bound; the risk level interval width is obtained by mapping the level coverage range to a continuous level span; and the warning window interval width is determined by the starting interval length, ending interval length, and overall window start and end fluctuation range of the warning window. After completing the calculation of the four types of widths, the local prediction terminal performs uniform scaling on the four types of widths. Preferably, the occurrence index interval width is scaled according to the current monitoring point's occurrence index center value, the outbreak probability interval width is scaled according to the probability upper limit, the risk level interval width is scaled according to the preset total number of levels, and the warning window interval width is scaled according to the total duration of the current prediction period. Then, the four types of widths after processing are weighted and synthesized according to preset weights to obtain the comprehensive interval width result corresponding to the current monitoring point and the current prediction period.

[0106] In one embodiment, in addition to calculating the overall interval width, the local prediction terminal also calculates the trend of change in the interval width. Specifically, the trend of change can be generated according to at least one of the following methods: Perform differential calculation on the comprehensive interval width of adjacent future nodes within the current prediction period to obtain the interval change between nodes; A moving average is applied to the comprehensive interval width of multiple consecutive nodes within the current forecast period to obtain the local width change trend; The comprehensive interval width of the current prediction period is compared with the comprehensive interval width corresponding to the previous prediction period or the previous valid week sampling period to obtain the cross-week change result.

[0107] Preferably, the local forecasting terminal simultaneously calculates intra-node trends and cross-week trends, and outputs three types of trends: "continuous widening," "basically stable," and "continuous narrowing." Specifically, when the comprehensive interval width continuously increases across multiple future nodes, or when the comprehensive interval width of the current forecast period expands by more than a preset proportion relative to the previous effective week sampling period, it is determined as continuously widening. When the comprehensive interval width fluctuation is within a preset fluctuation range, it is determined as basically stable. When the comprehensive interval width continuously decreases across multiple future nodes, or when the comprehensive interval width of the current forecast period shrinks by more than a preset proportion relative to the previous effective week sampling period, it is determined as continuously narrowing. In this way, risk stratification no longer relies solely on the width at a single point in time, but simultaneously considers whether the current uncertainty is in a state of diffusion or convergence.

[0108] After calculating the comprehensive interval width and trend, the local prediction terminal performs risk stratification based on a first preset threshold and a second preset threshold. Specifically, the first and second preset thresholds can be pre-stored in the threshold configuration table of the local prediction terminal and configured according to monitoring point type, deployment scenario, or business risk preference. Preferably, the first and second preset thresholds are determined by statistical results of the interval width during historical low-error prediction periods, and the first preset threshold is less than the second preset threshold. Then, the local prediction terminal compares the current comprehensive interval width result with the first and second preset thresholds: When the width of the comprehensive interval is less than the first preset threshold, the corresponding prediction result is marked as a high-confidence prediction; When the width of the comprehensive interval is greater than or equal to the first preset threshold and less than or equal to the second preset threshold, the corresponding prediction result is marked as a medium confidence prediction. When the width of the comprehensive interval is greater than the second preset threshold, the corresponding prediction result will be marked as a low confidence prediction.

[0109] Based on this, if the overall interval width is within the medium confidence range but its trend shows a continuous widening, the local prediction terminal can upgrade the result to a "low medium confidence" state and add stronger supplementary monitoring suggestions; if the overall interval width is within the low confidence range and its trend continues to widen, the rollback mode or manual intervention request will be triggered first.

[0110] For high-confidence forecasts, the local forecasting terminal directly outputs the corresponding early warning window result after marking. Specifically, it reads the center start point, center end point, start interval, and end interval of the early warning window corresponding to the current forecast result, preferentially outputs the central early warning window, and includes the start and end floating ranges if necessary. If multiple discrete early warning windows exist for the current monitoring point within the future forecast period, they are output sequentially in descending order of warning level and from earliest to latest window start time. The output record for high-confidence forecasts does not include supplementary monitoring suggestions or trigger manual review requests; it only writes a "high-confidence forecast" mark in the output record so that the terminal interface, early warning log, or upper-level control module can directly read it.

[0111] For medium-confidence predictions, the local prediction terminal generates supplementary monitoring suggestions while outputting the early warning window. Specifically, it first reads the comprehensive interval width, trend, source missing marker, cross-source deviation results, and effective observation coverage within the current weekly sampling period corresponding to the current prediction result, and then generates auxiliary prompts based on preset suggestion rules. Preferably, when the medium-confidence prediction is mainly caused by a large cross-source deviation, it outputs the auxiliary prompt "It is recommended to increase the number of comparison samplings of traps and insect lamps on the same day and in the same period"; when the medium-confidence prediction is mainly caused by insufficient effective observation coverage, it outputs the auxiliary prompt "It is recommended to increase the number of supplementary observations within the current weekly sampling period"; when the medium-confidence prediction is mainly caused by a continuous widening of the interval width, it outputs the auxiliary prompt "It is recommended to review the observation results of the next sampling node in advance". For the local deployment method in the tea garden scenario, the local prediction terminal preferably specifies the supplementary monitoring suggestions as increasing the number of preset supplementary sampling nodes within the current week, increasing the coverage density of insect lamp observation periods, or increasing the frequency of trap data verification, so that the auxiliary prompts can directly affect the next round of observation execution.

[0112] For low-confidence predictions, the local prediction terminal first determines the cause of the low confidence before deciding whether to trigger a rollback mode or request manual intervention. Specifically, it first checks whether the current low-confidence result is accompanied by at least one of the following abnormal conditions: the cross-source deviation between the trap counting sequence and the insect lamp counting sequence exceeds a preset deviation threshold; the completeness of key fields is lower than a preset completeness threshold; there are valid observations missing in consecutive weekly sampling periods; or the proportion of uncorrected nodes in the current prediction period exceeds a preset proportion. If at least one of the above abnormal conditions is met, the rollback mode is triggered first. If the current low-confidence result is mainly caused by excessively wide model parameter ranges, excessive node fluctuations, or the cumulative effect of continuous multi-node stationarity penalties, rather than by abnormal data acquisition, a manual verification prompt is output first, allowing on-site personnel to manually verify the status of monitoring equipment, parameter configuration status, or recent operation records. In this way, low-confidence predictions are not uniformly treated as the same type of anomaly, but different subsequent actions are taken based on the different sources of low confidence.

[0113] In one embodiment, when at least one of the following situations is detected: the cross-source deviation between the trap counting sequence and the insect lamp counting sequence exceeds a preset deviation threshold, the completeness of key fields is lower than a preset completeness threshold, or there are missing valid observations in consecutive weekly sampling periods, a regression prediction result is generated using the correction prediction sequence corresponding to the previous valid weekly sampling period and the valid adult insect observation data corresponding to the current weekly sampling period. The specific process is as follows.

[0114] First, the local prediction terminal reads the corrected prediction sequence corresponding to the previous valid weekly sampling period from the historical results storage area. The previous valid weekly sampling period refers to the weekly sampling period closest to the current week that simultaneously meets the following conditions: key field completeness is not lower than a preset completeness threshold, cross-source deviation is not higher than a preset deviation threshold, and the proportion of valid observations from two sources is not lower than a preset proportion. The read previous valid weekly corrected prediction sequence includes at least: adult population size correction values, trapping response status correction values, phototactic response status correction values, trapping channel correction count values, phototactic channel correction count values, and the early warning window results corresponding to the previous valid week for each node.

[0115] Subsequently, the local prediction terminal reads the still valid adult insect observation data within the current weekly sampling period. This valid adult insect observation data includes at least the adult insect counts from traps and insect-infesting lamps that have passed field and time verification within the current week, along with source validity markers and corresponding observation confidence coefficients. Then, the valid adult insect observation data for the current week is aligned with the same time node structure as the previous valid week's corrected prediction sequence, forming a backtracking modeling input of "historical stable prediction baseline + current valid observation correction amount".

[0116] After aligning the time nodes, the local prediction terminal generates backtracking correction coefficients. Specifically, for nodes with valid observations in the current week, the valid observation values ​​of the current week are compared with the correction count values ​​of the corresponding nodes in the previous valid week to obtain the node offset or proportional correction amount relative to the previous valid week. Then, based on the observation confidence coefficient and source validity label corresponding to the node, the offset or proportional correction amount is weighted to obtain the node-level correction coefficient. The node-level correction coefficients of multiple valid nodes in the current week are then processed by averaging, time proximity weighting, or source confidence weighting to obtain the overall backtracking correction coefficient applicable to the current monitoring point. If there are only a few valid nodes in the current week, the inheritance ratio of the previous valid week's correction prediction sequence in the backtracking result is increased; if there are many valid nodes in the current week and they are evenly distributed, the influence ratio of the current week's correction coefficient on the backtracking result is increased. In this way, the backtracking result will neither completely copy the stable path of the previous valid week nor overly rely on local observation points when the observations in the current week are incomplete.

[0117] In one implementation, the local prediction terminal corrects the calibration prediction sequence corresponding to the previous valid week's sampling period based on the overall backtracking correction coefficient, generating backtracking prediction results. Specifically, translation correction, scaling correction, or segmented weighted correction is applied to the adult population size correction value, trapping channel correction count value, and phototactic channel correction count value corresponding to each future node in the previous valid week's calibration prediction sequence to obtain the backtracking node results for the current week. For nodes with valid observations in the current week, the backtracking node observations of the corresponding channels can be directly replaced with the valid observation values ​​of the current week, and the corrected historical stable path is used only for the missing parts. For nodes that are completely missing in the current week, the corrected historical stable path is used entirely. After processing all nodes, a complete backtracking prediction sequence corresponding to the current week is formed, and the backtracking occurrence index, backtracking outbreak probability, backtracking risk level, and backtracking warning window are recalculated based on this sequence.

[0118] In another implementation, to ensure that the output results in rollback mode fully reflect the decrease in reliability, the local prediction terminal automatically amplifies the uncertainty range while generating the rollback prediction results. Specifically, rollback amplification is added to the rollback occurrence index range, rollback outbreak probability range, rollback risk level coverage range, and rollback warning window start and end range, respectively. The rollback amplification is determined based on at least the following factors: the proportion of effective observation nodes in the current week, the degree of cross-source bias exceeding the threshold, the degree of decrease in the integrity of key fields, the number of consecutive missing weeks, and the length of the interval between the previous effective week and the current week. The lower the proportion of effective observation nodes in the current week, the greater the cross-source bias, the lower the integrity of key fields, the longer the number of consecutive missing weeks, or the farther the previous effective week is from the current week, the greater the corresponding rollback amplification. In this way, while the rollback results remain usable, their uncertainty range can truly reflect the stability decrease caused by "insufficient data + historical inheritance".

[0119] After completing the decision risk stratification and rollback processing, the local forecasting terminal generates a risk stratification result table and a rollback output result table. The risk stratification result table includes at least: monitoring point number, week number, comprehensive interval width, interval width change trend, first preset threshold, second preset threshold, confidence level marker, warning window center value, warning window interval value, supplementary monitoring suggestion marker, manual review prompt marker, and result generation time. The rollback output result table includes at least: monitoring point number, week number, previous valid week identifier, current week valid observation node ratio, overall rollback correction coefficient, rollback occurrence index center value, rollback occurrence index interval, rollback outbreak probability center value, rollback outbreak probability interval, rollback risk level center value, rollback risk level coverage, rollback warning window center value, rollback warning window interval, low confidence marker, rollback trigger reason marker, and result generation time. Finally, the local forecasting terminal sends the risk stratification results and rollback output results to the warning display interface, local alarm module, or upper-level business system to complete this round of warning prompt output.

[0120] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the growth of the tea geometrid moth based on a population dynamics model, characterized in that, include: Acquire adult insect observation data collected at each monitoring point in the target tea garden experimental site according to a preset weekly sampling cycle. The adult insect observation data includes at least the trap counting sequence and the insect lamp counting sequence, and construct a unified weekly observation record sequence based on the trap counting sequence and the insect lamp counting sequence. Based on the completeness of key fields in the unified weekly observation record sequence, the consistency of collection time, and the cross-source deviation between the trap counting sequence and the insect lamp counting sequence, input data-level uncertainty information is generated. The input data-level uncertainty information includes at least the observation confidence coefficient and the corresponding observation confidence interval. Based on the trap counting sequence, insect lamp counting sequence and observation confidence coefficient, an adult insect population dynamic model is constructed. The adult insect population dynamic model includes at least the adult insect population size state quantity, the trapping response state quantity and the phototactic response state quantity. The model parameter level uncertainty information corresponding to the state transition parameters, trapping response parameters and phototactic response parameters in the adult insect population dynamic model is determined. The adult population dynamics model is used to perform prior evolution on the adult occurrence trend during the future target prediction period to obtain the prior prediction sequence; Based on the adult insect observation data and observation confidence coefficient within the current weekly sampling period, the prior prediction sequence is assimilated and corrected to obtain the corrected prediction sequence, and assimilation and correction level uncertainty information is generated. The uncertainty information at the input data level, the uncertainty information at the model parameter level, and the uncertainty information at the assimilation correction level are forward propagated through the error propagation formula or the Monte Carlo sampling method to obtain the quantitative uncertainty interval of the occurrence index, outbreak probability, risk level, and warning window corresponding to the correction prediction sequence; Based on the width and trend of the quantitative uncertainty interval, the prediction results are stratified into decision risks. Specifically, when the width of the quantitative uncertainty interval is less than a first preset threshold, it is determined to be a high-confidence prediction and an early warning window is directly output. When the width of the quantitative uncertainty interval is between the first and second preset thresholds, it is determined to be a medium-confidence prediction, and supplementary monitoring suggestions are output along with the early warning window. When the width of the quantitative uncertainty interval is greater than the second preset threshold, it is determined to be a low-confidence prediction, and a rollback mode or manual intervention request is triggered. When at least one of the following situations is detected: the cross-source deviation between the trap counting sequence and the insect lamp counting sequence exceeds a preset deviation threshold, the completeness of key fields is lower than a preset completeness threshold, or there are missing valid observations in consecutive weekly sampling periods, a regression prediction result is generated using the corrected prediction sequence corresponding to the previous valid weekly sampling period and the valid adult insect observation data corresponding to the current weekly sampling period. After attaching a low confidence marker to the regression prediction result, an early warning prompt is output.

2. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: Construct a unified weekly observation record sequence, including: Using a one-week rolling sampling period, the trap counting sequence and the insect lamp counting sequence are merged according to a unified time benchmark, and a unified weekly observation record table is generated according to the monitoring point number, collection date, collection time period, adult insect count value and source identifier, so that adult insect observation data from different sources can form a unified observation input that can be compared within the same weekly sampling period.

3. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: The input data level uncertainty information is determined by at least three of the following: completeness of key fields, consistency of collection time, and cross-source bias. Among them, cross-source bias is used to characterize the degree of difference between the trap counting sequence and the insect lamp counting sequence in terms of adult occurrence trend at the same monitoring point and within the same weekly sampling period. The observation confidence coefficient is generated by weighting the completeness of key fields, consistency of collection time, and cross-source bias.

4. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: The adult population dynamics model is constructed using a source-separated observation structure. The adult population size state quantity is used to characterize the potential population size of the adult moths corresponding to the target monitoring point. The trapping response state quantity is used to characterize the trapping response intensity of the trappers to the adults. The phototactic response state quantity is used to characterize the phototactic response intensity of the insect lamps to the adults. The trapping count sequence and the insect lamp count sequence are used as different observation channels to constrain the adult population size state quantity.

5. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: The model parameter level uncertainty information includes at least the parameter confidence intervals corresponding to the state transition parameters, trapping response parameters, and phototaxis response parameters. The parameter confidence intervals are obtained by estimating the parameters of the adult population dynamic model based on the effective adult observation data within the current weekly sampling period, and are used to constrain the fluctuation range of the prior prediction sequence.

6. The method for predicting the grey tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: Assimilation correction is performed on the prior predicted sequence, including: The trap counting sequence and the insect lamp counting sequence are assigned observation update weights according to the observation confidence coefficient, and the adult population size and state of the prior prediction sequence are corrected according to the observation update weight to obtain the corrected prediction sequence; among them, the higher the observation confidence coefficient, the greater the influence of the corresponding observation channel on the update of the adult population size and state.

7. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: The input data level uncertainty information, model parameter level uncertainty information, and assimilation correction level uncertainty information are forward propagated, including: The quantitative uncertainty intervals corresponding to the occurrence index, outbreak probability, risk level, and warning window can be calculated using the error propagation formula, or the interval distribution results corresponding to the occurrence index, outbreak probability, risk level, and warning window can be obtained by performing Monte Carlo sampling on the observation confidence coefficient, parameter confidence interval, and assimilation correction error.

8. The method for predicting the gray tea geometrid moth based on a population dynamics model according to claim 1, characterized in that: Decision risk stratification includes: Prediction results with an interval width less than the first preset threshold are marked as high-confidence predictions; Prediction results with an interval width greater than or equal to the first preset threshold and less than or equal to the second preset threshold are marked as medium confidence predictions, and an auxiliary prompt is output suggesting that the number of supplementary observations be increased in the current weekly sampling period. Prediction results with an interval width greater than the second preset threshold are marked as low-confidence predictions, and a prompt for manual review or a control command to switch to rollback mode is output.

9. A prediction system for the brown tea geometrid moth based on a population dynamics model, characterized in that: include: The data access module is used to acquire the trap counting sequence and insect lamp counting sequence collected at each monitoring point in the target tea garden experimental site according to a preset weekly sampling cycle; A unified observation construction module is used to construct a unified weekly observation record sequence based on the trap counting sequence and the insect lamp counting sequence. The input uncertainty assessment module is used to generate input data-level uncertainty information based on the completeness of key fields, consistency of acquisition time, and cross-source deviation. The dynamic model building module is used to construct a dynamic model of adult insect populations, including the population size state, the trapping response state, and the phototactic response state, and to determine the uncertainty information of the model parameters. The prediction and assimilation module is used to generate prior prediction sequences, perform assimilation corrections, and generate assimilation correction level uncertainty information. The uncertainty propagation module is used to forward propagate input data-level uncertainty information, model parameter-level uncertainty information, and assimilation correction-level uncertainty information to obtain the quantitative uncertainty range of occurrence index, outbreak probability, risk level, and warning window. The risk stratification output module is used to perform decision risk stratification based on quantitative uncertainty intervals and output early warning prompts. The rollback processing module is used to generate rollback prediction results with low confidence flags when preset abnormal conditions are met.

Images