Pest outbreak trend prediction method fusing multi-source data and pest situation forecasting instrument

Abstract

The present application relates to a pest outbreak trend prediction method and a pest situation forecasting instrument for fusing multi-source data. The method unifies and denoises multi-dimensional monitoring data such as ground trapping count, meteorology, satellite remote sensing and unmanned aerial vehicle infrared, and performs spatio-temporal standardization processing to generate a normalized feature sequence. Based on historical pest situation reports and environmental data, a spatio-temporal anchor point knowledge base of pest situation evolution stages is constructed, and the current pest situation state attribution probability is quantified through unsupervised clustering and high-dimensional feature matching. The dynamic calibration of the characteristic weight of the ecological driving factor is combined with the rule-driven and residual feedback mechanism to improve the response capability of the dominant feature. Finally, the multi-source features with dynamic weights are input into the prediction model to realize the calculation of the pest situation outbreak probability, and the spatial consistency feedback is used to automatically optimize the model parameters. The scheme significantly improves the accuracy, generalization ability and real-time adjustment effect of pest situation trend prediction.

Description

Technical Field

[0001] This invention relates to the field of pest monitoring and prediction systems, and in particular to a method for predicting pest outbreak trends by integrating multi-source data and a pest monitoring instrument. Background Technology

[0002] Existing pest monitoring and outbreak prediction technologies typically employ multi-source data fusion analysis, combining various data sources such as ground-based trapping and counting, meteorological records, satellite remote sensing vegetation indices, and UAV remote sensing images to model and predict pest occurrence and spread. In traditional approaches, mainstream technological trends focus on using machine learning models, neural network structures, or expert rule systems to extract features and weighted learn from multidimensional ecological factors. Especially driven by artificial intelligence and big data environments, existing systems often employ static feature selection, attention mechanisms, or weighting strategies based on expert experience, aiming to improve prediction accuracy under high-dimensional, uncertain data.

[0003] In actual insect infestation scenarios, the dynamic sequence and intensity of dominant factors such as insect population density, spatial diffusion, and meteorological disturbances exhibit significant uncertainty and high time-varying characteristics. For example, at different stages of an insect outbreak (such as initial migration, accelerated proliferation, and peak diffusion), the data response patterns of the dominant disaster process shift significantly. Traditional models, using static feature weighting, struggle to capture this dynamic change, often leading to oversensitivity or suppression of certain features in model predictions, thus affecting the numerical stability and generalization ability of the overall outbreak probability. Furthermore, most current systems suffer from response delays in the cumulative optimization of historical feature contributions or residual feedback adjustment measures. Faced with complex scenarios such as crop rotation areas and regional microclimate variations, the models lack the adaptive adjustment capability to dominant driving factors.

[0004] Therefore, there is an urgent need for an innovative method that can start from the semantic stage of pest evolution, combine the response characteristics of multi-source data, and dynamically and adaptively calibrate the contribution of feature weights. This method can reduce the dependence on prior knowledge, enhance the model's cross-scenario generalization and self-learning ability, and effectively support the upgrade of the pest outbreak trend prediction system in complex ecological environments. Summary of the Invention

[0005] This application provides a method for predicting pest outbreak trends by integrating multi-source data and a pest monitoring instrument, aiming to solve one of the problems or problems of the existing technology mentioned in the background.

[0006] The pest outbreak trend prediction method provided by this invention, which integrates multi-source data, specifically includes: S1: Acquire multi-source insect infestation monitoring data, and perform noise reduction and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences; S2: Based on historical insect infestation reports and retrospective environmental data, an unsupervised clustering algorithm is used to mine data on key stages of insect population growth, spread, and diapause, and a spatiotemporal anchor knowledge base is constructed. The spatiotemporal anchor knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. S3: Map the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculate the similarity distribution between the current data and the multi-source data response fingerprints corresponding to each typical time stage, and generate a stage membership vector. S4: Based on the typical time stage pointed to by the peak in the stage membership vector, retrieve the preset feature sensitivity rule set and generate a coarse-grained feature channel activation instruction for the current ecological scenario. S5: Obtain the adaptive feature weight matrix by performing operations based on the coarse-grained feature channel activation instruction; S6: Input the standardized multi-source feature sequence into the prediction model loaded with the adaptive feature weight matrix, perform weighted fusion operation, and output the calculation result of the insect outbreak probability in the current area; S7: Compare the spatial consistency between the calculated insect outbreak probability and the actual location of the outbreak. If a significant shift in centroid distance exceeding 200 meters is detected three times consecutively, generate a characteristic response gain correction coefficient. S8: Update the spatiotemporal anchor knowledge base using the feature response gain correction coefficient to complete the unsupervised feature contribution dynamic calibration optimization.

[0007] An insect pest monitoring instrument is used to execute the aforementioned method for predicting pest outbreak trends by fusing multi-source data, comprising: Data preprocessing module: used to acquire multi-source insect infestation monitoring data, and to perform noise reduction and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences; Knowledge base construction module: Based on historical insect infestation reports and retrospective environmental data, it uses unsupervised clustering algorithms to mine data on key stages of insect population growth, spread, and diapause, and constructs a spatiotemporal anchor knowledge base. The spatiotemporal anchor knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. Stage membership calculation module: used to map the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculate the similarity distribution between the current time data and the multi-source data response fingerprints corresponding to each typical time stage, and generate a stage membership vector; Feature channel activation module: Based on the typical time stage pointed to by the peak in the stage membership vector, it retrieves a preset feature sensitivity rule set and generates a coarse-grained feature channel activation instruction for the current ecological scenario. The weight adaptive generation module is used to generate an adaptive feature weight matrix based on the coarse-grained feature channel activation instruction. Insect Infestation Prediction Output Module: Connected to the data preprocessing module and the weight adaptive generation module respectively, it is used to input the standardized multi-source feature sequence into the prediction model loaded with the adaptive feature weight matrix, perform weighted fusion operation, and output the insect outbreak probability calculation result of the current area.

[0008] The pest outbreak trend prediction method that integrates multi-source data provided in this application has the following beneficial effects: This invention constructs a spatiotemporal anchor point knowledge base based on the ecological patterns of insect outbreaks. It improves upon traditional prediction models that rely on static feature weights and data-driven attention mechanisms, making them ill-suited for complex scenarios involving cross-regional, cross-climate zones, and crop rotation. These models are prone to misjudging population behavior mutations and lagging identification of key driving factors. This scheme maps real-time data to spatially generated stage membership vectors for typical time periods, thereby driving the dynamic activation and suppression of feature-sensitive rule sets. It can reconstruct feature channel priorities in milliseconds, eliminating the need for online retraining and parameter fine-tuning. This effectively solves the problems of high response latency and weak generalization ability of traditional deep learning methods when dealing with meteorological disturbances and niche migrations, improving the robustness and deployment efficiency of models in farmland scenarios. Attached Figure Description

[0009] Figure 1 This is the main flowchart of a pest outbreak trend prediction method that integrates multi-source data.

[0010] Figure 2 This is a sub-flowchart of a pest outbreak trend prediction method that integrates multi-source data.

[0011] Figure 3 This is another sub-flowchart of the pest outbreak trend prediction method that integrates multi-source data. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0014] like Figure 1 As shown, this application provides a method for predicting pest outbreak trends by fusing multi-source data, specifically including: S1: Acquire multi-source insect infestation monitoring data, and perform noise reduction and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences; S2: Based on historical insect infestation reports and retrospective environmental data, an unsupervised clustering algorithm is used to mine data on key stages of insect population growth, spread, and diapause, and a spatiotemporal anchor knowledge base is constructed. The spatiotemporal anchor knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. S3: Map the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculate the similarity distribution between the current data and the multi-source data response fingerprints corresponding to each typical time stage, and generate a stage membership vector. S4: Based on the typical time stage pointed to by the peak in the stage membership vector, retrieve the preset feature sensitivity rule set and generate a coarse-grained feature channel activation instruction for the current ecological scenario. S5: Obtain the adaptive feature weight matrix by performing operations based on the coarse-grained feature channel activation instruction; S6: Input the standardized multi-source feature sequence into the prediction model loaded with the adaptive feature weight matrix, perform weighted fusion operation, and output the calculation result of the insect outbreak probability in the current area; S7: Compare the spatial consistency between the calculated insect outbreak probability and the actual location of the outbreak. If a significant shift in centroid distance exceeding 200 meters is detected three times consecutively, generate a characteristic response gain correction coefficient. S8: Update the spatiotemporal anchor knowledge base using the feature response gain correction coefficient to complete the unsupervised feature contribution dynamic calibration optimization.

[0015] Step S1: Acquire multi-source insect infestation monitoring data, and perform denoising and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences. Specifically, this includes: S1.1: Obtain the original multi-source insect infestation monitoring data stream, which includes ground trap counting time series data, meteorological record grid data, satellite remote sensing vegetation index image data, and UAV thermal infrared image sequence data. Use multi-protocol data interface parsing technology to perform format unification processing on the original multi-source insect infestation monitoring data stream to generate a multi-source dataset to be cleaned with a unified data encapsulation structure.

[0016] S1.1 Obtain the original multi-source insect infestation monitoring data stream, which includes ground trap counting time-series data, meteorological record grid data, satellite remote sensing vegetation index image data, and UAV thermal infrared image sequence data. Use multi-protocol data interface parsing technology to perform format unification processing on the original multi-source insect infestation monitoring data stream to generate a multi-source dataset to be cleaned with a unified data encapsulation structure.

[0017] Establish a multi-source heterogeneous data access gateway, and configure an MQTT protocol interface to receive real-time counting pulse signals from ground-based intelligent traps, an HTTP / RESTful interface to pull gridded temperature, humidity and wind speed data from meteorological stations, a WCS service interface to subscribe to satellite remote sensing NDVI image slices, and an RTSP streaming media protocol to capture thermal infrared video streams from UAV inspections.

[0018] The protocol header is parsed and payload is extracted from the raw byte streams input from each channel. For ground-trapped data, repeated trigger pulses caused by insect struggles are removed. For meteorological data, the validity of latitude and longitude coordinates is verified. For remote sensing images, GeoTIFF header files are parsed to obtain geographic reference information. For thermal infrared video, key frames are extracted and invalid frames with severe motion blur are removed.

[0019] A unified data encapsulation object model is constructed, defining a standard JSON structure that includes a timestamp field, a spatial coordinate field, a sensor type identifier field, and a raw numerical field. The parsed source data is mapped to this standard structure, where the timestamp is uniformly converted to the UTC standard time format and the spatial coordinates are uniformly converted to the decimal degree representation in the WGS84 coordinate system.

[0020] The encapsulated data is subjected to integrity verification, checking for null values ​​or abnormal extreme values ​​in each field. Missing time slices are marked as needing to be interpolated, and values ​​that exceed the range of physical meaning, such as negative trap counts or surface temperatures exceeding 60 degrees Celsius, are marked as anomalies. A multi-source dataset to be cleaned is generated, containing valid data and anomaly marker bits.

[0021] Through the above multi-protocol parsing and standardized encapsulation processing, heterogeneous raw monitoring data are transformed into a multi-source dataset to be cleaned with consistent structure and preliminary unified spatiotemporal benchmarks, eliminating format barriers between data sources and providing a standardized input basis for subsequent denoising and spatiotemporal alignment.

[0022] S1.2: Based on the multi-source dataset to be cleaned, the high-frequency random noise and sensor drift error in each data channel are filtered out using an adaptive wavelet threshold denoising method to generate a clean multi-source feature data set with a significantly improved signal-to-noise ratio. The clean multi-source feature data set retains the abrupt change characteristics of ground trap counts and the texture details of thermal infrared images.

[0023] Read the original signals of each channel in the multi-source dataset to be cleaned, identify the one-dimensional time series of ground trap counting, the two-dimensional grid matrix of meteorological records, the three-dimensional image cube of satellite remote sensing, and the four-dimensional image sequence of UAV thermal infrared, and establish the noise distribution characteristics of different data modes.

[0024] For the ground-trapped count time series data, discrete wavelet transform decomposition is performed, and Symlets wavelet basis functions are selected to decompose the signal into low-frequency approximation coefficients and high-frequency detail coefficients, separating the low-frequency component representing the mutation trend of insect population and the high-frequency component containing random counting errors.

[0025] An adaptive threshold is calculated based on the Stein unbiased risk estimation criterion. The high-frequency detail coefficients are subjected to soft threshold shrinkage to suppress outlier impulse noise caused by sensor jitter or environmental interference, while retaining the step characteristics when the insect population increases sharply.

[0026] For meteorological records and remote sensing image data, a two-dimensional wavelet pyramid structure is constructed. The noise variance is statistically analyzed using the high-frequency subband coefficients in the horizontal, vertical and diagonal directions to generate a spatially adaptive denoising mask, which filters out pixel loss caused by cloud cover and background clutter introduced by atmospheric scattering.

[0027] For the UAV thermal infrared image sequence, a spatiotemporal joint wavelet filtering strategy is adopted to correlate the wavelet coefficients of adjacent frames in the time dimension. The texture blur caused by the UAV attitude tremor is removed by three-dimensional threshold judgment, thereby enhancing the clarity of the thermal radiation edge of the insect aggregation area.

[0028] By integrating the denoised data segments from each channel, a clean multi-source feature data set with a significantly improved signal-to-noise ratio is obtained, ensuring that the abrupt change points of ground counting and the texture details of thermal infrared images are fully preserved.

[0029] By using an adaptive wavelet threshold denoising method, the data to be cleaned in the previous step is transformed into a high-fidelity pure multi-source feature data set, which realizes the efficient separation of effective ecological signals and random noise in multi-source heterogeneous data, and provides a high-quality data foundation for subsequent spatiotemporal alignment.

[0030] For example, a ground-trapping counting sequence in a cotton field was decomposed into 5 levels using the db4 wavelet basis. The median absolute deviation of the high-frequency coefficients from levels 3 to 5 was calculated, estimating the noise standard deviation to be 2.1. A global threshold of 4.8 was set. After soft thresholding, the peak insect population was clearly shown to have increased from 120 insects / day, and the background fluctuation was reduced to ±3 insects / day. For the same period's Sentinel-2 remote sensing NDVI imagery, a 2D wavelet decomposition tree was constructed to identify high-frequency energy anomalies in cloud-covered areas. Spatial adaptive masking was applied to remove noisy pixels, improving vegetation index smoothness and achieving an edge gradient preservation rate of over 92%. For the DJI Mavic thermal infrared video stream, a 3D wavelet transform was performed with a time window of 5 frames. Motion artifacts were removed through coefficient correlation analysis, enhancing the temperature contrast of insect swarm hotspots by 1.5 times and improving the texture detail signal-to-noise ratio to 28dB. The final output was a standardized, clean dataset.

[0031] S1.3: To address the spatiotemporal inconsistency in the pure multi-source feature data set, a bicubic interpolation resampling method and a linear time synchronization mechanism are used to map satellite remote sensing vegetation index image data with different spatial resolutions and meteorological record grid data with different sampling frequencies to a unified standard geographic grid and standard time slice, so as to generate an aligned multi-source feature data cube with strict spatiotemporal alignment attributes.

[0032] To address the heterogeneity in spatial resolution and temporal sampling frequency between satellite remote sensing imagery and meteorological grid data in a clean multi-source feature dataset, spatiotemporal benchmark unification processing is performed. Geographic coordinate reference system parameters are extracted from the satellite remote sensing vegetation index imagery data to determine the projection type and pixel size of the target standard geographic grid, establishing a spatial mapping matrix from the original sensor coordinate system to the standard geographic grid. A bicubic interpolation resampling method is applied, and the weighted contribution values ​​of the 16 neighboring pixels around each standard grid node are calculated based on the spatial mapping matrix to reconstruct a high-resolution vegetation index distribution map, eliminating edge jaggedness caused by resampling while maintaining image grayscale continuity.

[0033] The timestamp sequence of meteorological record grid data is acquired synchronously, and non-uniform sampling intervals between different sensor nodes are identified to construct a time synchronization axis based on linear interpolation. For meteorological element values ​​at missing or misaligned times, the two nearest valid observation points are selected, and the time weighting coefficient is calculated using a linear time synchronization mechanism to generate a continuous meteorological time-series data stream after filling in the missing data, ensuring that meteorological variables strictly correspond to the insect monitoring times.

[0034] The resampled satellite remote sensing vegetation index imagery is spatially overlaid with time-synchronized meteorological record grid data. Using a standard geographic grid as the index, data from each dimension are populated into a unified three-dimensional data structure. Through tensor stacking operations, the latitude and longitude coordinates of the spatial dimension are combined with the slice index of the temporal dimension to generate an aligned multi-source feature data cube with strict spatiotemporal alignment properties.

[0035] By using bicubic interpolation resampling and linear time synchronization mechanisms, heterogeneous multi-source data are mapped to a unified spatiotemporal benchmark, generating an aligned multi-source feature data cube. This achieves accurate registration of multi-source data in the spatiotemporal dimension, providing a consistent data foundation for subsequent feature fusion.

[0036] S1.4: Based on the aligned multi-source feature data cube, the physical dimensions of each dimension are made dimensionless using the range normalization transformation method to eliminate the numerical magnitude difference between the absolute value of the ground trap count and the relative value of the vegetation index, so as to generate a normalized multi-source feature tensor with consistent numerical distribution interval.

[0037] S1.5: Perform a multidimensional array serialization encapsulation operation on the normalized multi-source feature tensor, and reorganize the data arrangement structure according to the preset time step and spatial block index order to generate a standardized multi-source feature sequence that can be directly input into the semantic anchor matching module.

[0038] Step S2: Based on historical insect infestation reports and retrospective environmental data, an unsupervised clustering algorithm is used to mine data on key stages of insect population growth, spread, and diapause, constructing a spatiotemporal anchor knowledge base. This knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. Specifically, this includes: S2.1: Obtain long-term historical insect infestation report data and synchronous retrospective environment data, and perform spatiotemporal gridded resampling processing on the long-term historical insect infestation report data and the synchronous retrospective environment data to eliminate the spatiotemporal resolution differences of multi-source heterogeneous data and generate a multi-dimensional time-series fusion dataset with a unified spatiotemporal benchmark.

[0039] Sub-step S2.1 aims to construct a multi-dimensional time-series fusion dataset with a unified spatiotemporal benchmark, providing standardized input for subsequent unsupervised clustering. It acquires long-term historical pest report data and synchronous retrospective environmental data. The historical pest report data includes pest species, occurrence dates, and damage levels. The synchronous retrospective environmental data covers meteorological element grid data, remote sensing vegetation index time-series imagery, and soil moisture inversion data for the corresponding time period. Spatiotemporal coordinate system unification processing is performed on the multi-source heterogeneous data, converting the geographic coordinate system to a projected coordinate system to eliminate area calculation errors caused by latitude and longitude distortion, establishing a unified Cartesian spatial benchmark. Addressing the spatial resolution differences between meteorological records and remote sensing imagery, a bilinear interpolation resampling method is used to aggregate high-resolution remote sensing pixels or refine low-resolution meteorological grids to preset standard grid units, ensuring that all environmental variables have consistent pixel alignment attributes in the spatial dimension. A time indexing mechanism based on UTC timestamps is established to map discontinuous historical pest report discrete points to a continuous time axis. Linear interpolation is used to fill in short-term missing values ​​in the meteorological data, achieving strict synchronization of multi-source data in the time dimension. A four-dimensional data cube structure is constructed, with three dimensions representing spatial latitude, longitude, and elevation, and the fourth dimension representing the time series. The cleaned insect pest tags and environmental variables are encapsulated in a three-dimensional manner according to grid cells and time series slices. Through the aforementioned spatiotemporal gridded resampling process, the original heterogeneous data from the previous step is transformed into a multi-dimensional time-series fusion dataset with a unified spatiotemporal resolution. This achieves complete alignment of multi-source data on a spatiotemporal benchmark, eliminates feature matching bias caused by resolution inconsistencies, and lays a data foundation for subsequent mining of response patterns in insect pest evolution stages.

[0040] For example, long-term data from 2018 to 2023 of an agricultural demonstration area were selected. Historical insect infestation reports included trap counts twice a week, meteorological data were hourly updated ERA5 reanalysis data, and remote sensing data were Sentinel-2 NDVI images synthesized every 5 days. The standard grid resolution was set to 1km × 1km, and the time step was 1 day. Bilinear interpolation was used to resample the 0.25° resolution temperature data to a 1km grid, and the 10m resolution NDVI images were aggregated to a 1km grid. Missing air pressure data were filled in using linear interpolation between consecutive time points. Finally, a four-dimensional tensor with dimensions of [number of longitude grids, number of latitude grids, number of days, number of feature channels] was generated, where the feature channels included temperature, precipitation, wind speed, NDVI, and insect population density. This dataset eliminated spatiotemporal misalignment, ensuring a strict correspondence between the environment and insect infestation data within the same grid cell, significantly improving the accuracy of subsequent clustering algorithms in identifying insect infestation evolution stages.

[0041] S2.2: Based on the multidimensional time-series fusion dataset, the sliding window segmentation method is used to extract time segments covering the insect population growth, diffusion and diapause processes, and the time segments are subjected to standardized difference operations to remove seasonal trend terms, generating a set of detrended local evolutionary sequences that characterize the rate of change of population dynamics.

[0042] Read the multidimensional time-series fusion dataset generated by S2.1 and extract time-series slices containing insect population density, meteorological elements, and vegetation indices. Construct a sliding time window mechanism and set the window length to [value missing]. Step size is By extracting continuous data along the time axis, a set of local evolutionary segments covering the entire process of insect population growth, spread, and diapause is generated.

[0043] For each local evolutionary segment, meteorological and vegetation components with significant seasonal cycles are identified. A first-order difference operator is used to perform standardized differencing on the original time-series data to eliminate interference from long-term trend terms caused by seasonal changes. Specifically, for the eigenvalues ​​at time t... Calculate the difference between its eigenvalue and the eigenvalue at the previous time t-1, and generate the detrended incremental sequence using the following formula: in, This represents the detrended rate of change at time t. The current observation value. This is the value observed at the previous moment.

[0044] For discrete data such as ground-based trapping counts, logarithmic differencing is applied to stabilize variance and suppress the impact of sudden impulse noise on trend judgment. For continuous data such as remotely sensed vegetation indices, centering is performed followed by differencing to retain high-frequency fluctuations reflecting abrupt changes in crop growth status.

[0045] The differentiald data from each channel were recombined to construct a detrended local evolutionary sequence set representing the rate of change in population dynamics. This set eliminated seasonal and normal fluctuations in the environmental background, highlighting anomalous evolutionary patterns during insect outbreaks.

[0046] By using sliding window segmentation and standardized difference processing, the original multidimensional time series data is transformed into a set of detrended local evolutionary sequences focusing on the dynamic rate of change. This enhances the extraction of features at key stages of insect infestation evolution, eliminates the interference of seasonal trends on unsupervised clustering analysis, and provides a high signal-to-noise ratio input basis for subsequent semantic anchor mining.

[0047] S2.3: Apply an unsupervised clustering algorithm based on density peaks to the set of detrended local evolutionary sequences, calculate the high-dimensional Euclidean distance distribution between sequences and identify cluster centers, and map the converged clusters to five typical semantic labels: initial immigration period, population acceleration and proliferation period, spatial diffusion peak period, environmental inhibition and stabilization period, and natural decline period, to generate insect evolution stage clusters with semantic annotations.

[0048] The S2.3 sub-step aims to perform unsupervised clustering based on density peaks on the detrended local evolution sequence set to identify typical stages of insect infestation evolution and generate semantically labeled clusters. The detrended local evolution sequence set generated in the preceding steps is read; this set contains multidimensional time segment data after seasonality has been removed through differencing operations, establishing a sample distribution benchmark in the high-dimensional feature space.

[0049] The high-dimensional Euclidean distance between any two sample points in the sequence set is calculated, a symmetric distance matrix is ​​constructed, and the geometric dispersion of each evolutionary segment in the multidimensional ecological feature space is quantified, providing a metric basis for density estimation.

[0050] Based on the distance matrix, the local density value of each data point is calculated using a Gaussian kernel function. The local density of the i-th sample point is determined by the following formula. : in, Let be the Euclidean distance between samples i and j. The cutoff distance parameter is used to control the neighborhood range and ensure that the density calculation reflects local clustering characteristics.

[0051] Calculate the minimum distance from each data point to higher density points as a decision variable, identify points with extremely high local density that are far from other high density points as potential cluster centers, and eliminate false centers formed by noise interference.

[0052] A decision map is drawn, with local density as the horizontal axis and minimum distance as the vertical axis. Significant deviation points located in the upper right corner are selected manually or automatically as the initial cluster centers, corresponding to key turning points in the evolution of insect infestations.

[0053] Based on the principle of proximity, the remaining non-central samples are assigned to the clusters of the nearest cluster centers to form a preliminary division of insect evolution stages, ensuring that samples within the same cluster have similar dynamic change patterns.

[0054] A cluster merging mechanism is introduced, which calculates the distance threshold between adjacent cluster centers. If the distance between two centers is less than the preset merging threshold and the density difference is not significant, they are merged into a single cluster, thereby optimizing the compactness and separation of the cluster structure.

[0055] The five main clusters after convergence are mapped to five typical semantic labels: initial immigration period, population acceleration and proliferation period, peak period of spatial diffusion, environmental inhibition and stabilization period, and natural decline period, respectively, thus completing the logical connection from data clusters to ecological semantics.

[0056] By using an unsupervised clustering algorithm based on density peaks, detrended local evolution sequences are transformed into clusters of insect infestation stages with semantic annotations, enabling automated identification and structured characterization of key stages throughout the insect outbreak process, and providing a clear stage division basis for subsequent extraction of quantitative response fingerprints.

[0057] S2.4: For each insect evolution stage cluster with semantic annotation, statistically analyze the Pearson correlation coefficient matrix and time-frequency energy spectrum characteristics among multi-source monitoring variables within the cluster, extract a combination of quantitative indicators that can distinguish the significant differences between this stage and other stages, and generate multi-source data response fingerprints corresponding to each typical time stage.

[0058] Substep S2.4 extracts multi-source data response fingerprints to distinguish each stage of insect evolutionary stages from semantically labeled insect population clusters. Standardized multi-source feature sequences within five stage clusters (initial migration stage, accelerated population growth stage, etc.) generated by clustering are read to establish each cluster as an independent analysis unit.

[0059] Pearson correlation coefficients are calculated for the multi-source monitoring variables within each cluster to construct a symmetric correlation matrix representing the linear dependence between variables. The Pearson correlation coefficient between any two feature channels is then calculated.

[0060] Feature pairs with an absolute correlation coefficient greater than 0.7 were selected to identify strongly coupled environment-insect-driven combinations in this semantic stage, such as the positive correlation between air temperature and trap count or the negative correlation between humidity and diffusion rate.

[0061] Perform a Fast Fourier Transform on the time-series data of each channel to convert the time-domain signal into a frequency-domain energy spectrum, and extract the energy ratio of the dominant frequency component and the spectral entropy index.

[0062] The complexity of calculating spectral entropy to quantify the frequency distribution of a signal is given by the following formula: in, For spectral entropy, For frequency Normalized energy power at the location, This is the index for the highest frequency.

[0063] By comparing the spectral characteristics of different stages, we can identify the low-frequency dominant characteristics of the accelerated proliferation period or the high-frequency fluctuation characteristics of the diffusion period, and form a time-frequency domain identification index.

[0064] By integrating highly correlated variable combinations with time-frequency domain identification indicators, a multi-dimensional vector containing statistical moments, correlation structures, and spectral features is constructed to generate multi-source data response fingerprints corresponding to the typical time stage.

[0065] Through the above statistical and time-frequency analysis, the abstract clusters are transformed into concrete quantitative response fingerprints, enabling a precise digital description of the essential characteristics of each insect infestation stage, and providing a highly discriminative comparison benchmark for subsequent semantic anchor matching.

[0066] For example, for a population in the accelerated propagation phase, the Pearson correlation coefficient between ground-based trap counts and daily average temperature was calculated to be 0.85, indicating a strong positive correlation. An FFT transformation was performed on the trap count sequence, revealing that the energy proportion in the 0.01-0.05 Hz frequency band was 72%, with a spectral entropy of 1.2 bits, significantly lower than the 2.8 bits in the natural decline phase. Integrating the correlation coefficient of 0.85, the dominant frequency energy proportion of 72%, and the spectral entropy of 1.2 bits, a quantitative fingerprint vector for this stage was generated [0.85, 0.72, 1.2, ...]. This fingerprint accurately distinguishes the propagation phase from other phases, significantly improving matching accuracy.

[0067] S2.5: The tag information of the five typical time stages and their corresponding multi-source data response fingerprints are key-value associated and encapsulated to construct a spatiotemporal anchor knowledge base with fast retrieval function, so as to provide a standardized comparison benchmark for calculating the stage membership degree of any input insect pest data.

[0068] like Figure 2 As shown, step S3 involves mapping the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculating the similarity distribution between the current data and the multi-source data response fingerprints corresponding to each typical time stage, and generating a stage membership vector. Specifically, this includes: S3.1: Obtain standardized multi-source feature sequences and multi-source data response fingerprints from the spatiotemporal anchor knowledge base. Based on the high-dimensional space projection method, map the standardized multi-source feature sequences and multi-source data response fingerprints to a unified insect evolution semantic feature space, and generate an isomorphic feature coordinate set containing the current ecological representation vector and the benchmark fingerprint vectors of each stage.

[0069] The S3.1 sub-step aims to map the standardized multi-source feature sequences and the quantized fingerprints in the spatiotemporal anchor knowledge base to a unified high-dimensional semantic space, constructing an isomorphic feature coordinate set to provide a geometric benchmark for subsequent similarity calculations. It reads the standardized multi-source feature tensor generated by the S1 preprocessing and extracts multi-dimensional ecological factor values ​​within a specific geographic grid cell at the current time, including the ground trapping count growth rate, diurnal temperature range, NDVI change rate, and thermal infrared texture entropy value, forming the original input vector. It then calls the spatiotemporal anchor knowledge base interface to retrieve the multi-source data response fingerprint vectors corresponding to five pre-stored typical insect infestation evolution stages (initial migration stage, accelerated proliferation stage, peak diffusion stage, suppression and stabilization stage, and natural decline stage). Each fingerprint vector contains feature statistical mean and variance parameters consistent with the dimension of the input vector.

[0070] A linear projection transformation is performed on the original input vector and the fingerprint vectors at each stage to eliminate residual differences in physical dimensions between different sensor data, ensuring that all features are comparable in the semantic space. A principal component analysis dimensionality reduction strategy is employed to calculate the covariance matrix of the entire historical data set, extract the eigenvectors corresponding to the top K largest eigenvalues ​​to form a projection matrix, and map the high-dimensional original data to a low-dimensional latent semantic subspace, retaining more than 95% of the information energy.

[0071] In the latent semantic subspace, an isomorphic feature coordinate set is constructed, where the current ecological representation vector serves as the dynamic query point, and the baseline fingerprint vectors of the five stages serve as the static reference anchor points, and the feature vector coordinates are calculated.

[0072] The projected coordinate vectors are normalized using the L2 norm, mapping all vectors to a unit hypersphere. This eliminates the interference of vector magnitude on directional similarity calculations, generating a standardized isomorphic feature coordinate set. This coordinate set clarifies the precise location of the current ecological state in the semantic space, as well as its relative geometric relationship with the anchor points of each typical evolutionary stage.

[0073] By using high-dimensional spatial projection and normalization processing, the heterogeneous multi-source data from the previous step is transformed into a set of isomorphic feature coordinates with unified geometric metrics, achieving seamless integration between real-time monitoring data and historical evolution patterns, and laying a rigorous mathematical foundation for distance-based similarity matching.

[0074] S3.2: Based on the isomorphic feature coordinate set, the spatial dispersion of the current ecological characterization vector and the corresponding benchmark fingerprint vectors for the initial migration period, the population acceleration and proliferation period, the peak period of spatial diffusion, the environmental inhibition and stabilization period, and the natural decline period are calculated using a weighted Euclidean distance metric model, thereby generating an original distance scalar set representing the degree of difference in data distribution.

[0075] S3.3: Perform kernel density estimation transformation on the original distance scalar set to convert the spatial dispersion values ​​into non-negative similarity scores that reflect the matching probability, and generate a preliminary similarity distribution matrix that characterizes the degree of fit between the current insect situation data and each typical time stage.

[0076] Step S3.3 aims to transform the original set of distance scalars into non-negative similarity scores reflecting the probability of matching, generating a preliminary similarity distribution matrix. The original set of distance scalars generated in the previous step is obtained, containing weighted Euclidean distance values ​​between the current ecological representation vector and the baseline fingerprint vectors of five typical time stages. Adaptive bandwidth parameter selection is performed on the original set of distance scalars. The optimal bandwidth for kernel density estimation is calculated using the Silverman rule to balance the estimation bias and variance, ensuring the smoothness of the probability density function and the ability to preserve local features. Based on the selected optimal bandwidth, a Gaussian kernel function is constructed as the basic operator for density estimation. Each distance scalar is treated as an independent sample point, and kernel superposition operations are performed in a multi-dimensional space. For each distance scalar, its probability density contribution value in the zero neighborhood is calculated using the Gaussian kernel function. An integral transformation maps the discrete distance values ​​to continuous non-negative similarity scores. Non-negativity constraint verification is performed on the calculated initial similarity scores, eliminating abnormal negative values ​​caused by numerical overflow, and restricting all scores to a non-negative real number interval. The processed similarity scores are arranged in semantic stage label order and encapsulated into a row vector structure to form a preliminary similarity distribution matrix. Through kernel density estimation transformation, the spatial dispersion values ​​from the previous step are converted into non-negative similarity scores reflecting the matching probability, generating a preliminary similarity distribution matrix characterizing the degree of fit between the current insect infestation data and each typical time stage. This achieves a semantic mapping from geometric distance to probability density, providing a physically meaningful similarity metric foundation for subsequent normalization processing.

[0077] S3.4: Based on the preliminary similarity distribution matrix, apply the soft maximum normalization function to perform probabilistic reconstruction of the similarity scores of each row, eliminate the difference in dimensions and constrain the numerical range, and generate a normalized state matching probability sequence with a sum of one and each value range between zero and one.

[0078] S3.5: Based on the semantic stage labels corresponding to each element in the normalized state matching probability sequence, the probability values ​​are arranged in order to construct a multidimensional state attribution descriptor, and the probability vector representing the current insect situation evolution state attribution, i.e., the stage membership vector, is generated as the meta-control signal for triggering the feature sensitivity rule set.

[0079] The system receives the normalized state matching probability sequence generated by S3.4, which contains similarity scores between the current data and five typical time stages. A five-dimensional state attribution descriptor vector structure is constructed, mapping the labels for the initial migration period, population acceleration period, spatial diffusion peak period, environmental inhibition stabilization period, and natural decline period to vector index positions according to a preset ecological evolution logic. An ordered probability value arrangement operation is performed to ensure that the vector element values ​​are strictly constrained to the interval between zero and one and that the sum is always equal to one, forming a standardized stage membership distribution. This probability vector is defined as a meta-control signal, with each component directly representing the confidence level of the current insect situation at a specific evolutionary stage. Through vector encapsulation processing, the similarity distribution from the previous step is transformed into a stage membership vector with clear semantic direction, achieving a quantitative mapping from data statistical features to ecological behavioral states, providing a deterministic control basis for the precise triggering of subsequent feature-sensitive rule sets.

[0080] like Figure 3 As shown, step S4 involves: based on the typical time stage pointed to by the peak value in the stage membership vector, retrieving a preset set of feature sensitivity rules, and generating a coarse-grained feature channel activation instruction for the current ecological scenario. Specifically, this includes: S4.1: Perform peak detection processing on the stage membership vector to extract the typical time stage label that represents the highest probability of belonging to the current insect situation evolution state, and generate the dominant ecological stage identifier.

[0081] Receive the stage membership vector output from step S3, which is a normalized sequence containing the probability values ​​of five typical time stages.

[0082] Perform a traversal scan operation on the stage membership vector to sequentially read the probability values ​​corresponding to the initial immigration period, the population acceleration and proliferation period, the peak period of spatial diffusion, the environmental inhibition and stabilization period, and the natural decline period.

[0083] Construct a maximum value retrieval logic, compare the currently read probability value with the preset temporary maximum threshold, and if the current value is greater than the temporary maximum threshold, update the temporary maximum threshold and record the current index position.

[0084] After completing the full vector traversal, the index position with the highest probability value is locked, and this index position uniquely corresponds to a specific insect evolution semantic stage label.

[0085] Extract the semantic stage name corresponding to the index position, encapsulate it as the dominant ecosystem stage identifier, and use it as the unique key value for subsequent rule retrieval.

[0086] By using peak detection processing, the probability distribution results from the previous step are transformed into clear dominant ecological stage identifier data, realizing the transformation from fuzzy probability state to definite control signal, and providing accurate scenario basis for feature channel activation.

[0087] For example, the stage membership vector is [0.05, 0.12, 0.78, 0.03, 0.02]. The system iterates through this vector, compares the values ​​of each element, and finds that the value at index 2, 0.78, is the maximum. Index 2 corresponds to "spatial diffusion peak period" in the spatiotemporal anchor knowledge base. The system extracts this tag and generates it as the dominant ecological stage identifier. This identifier directly points to the feature rules related to wind speed and terrain connectivity, avoiding weight dilution caused by multi-stage mixing and ensuring that the model's sensitivity to meteorological factors is significantly improved during the diffusion period.

[0088] S4.2: Based on the dominant ecological stage identifier, retrieve the feature sensitivity rule set to match the pre-stored multi-source data response fingerprint association rules corresponding to the typical time stage, and generate an initial list of dominant driving factors.

[0089] The system receives a dominant ecological stage identifier, which represents the typical time stage with the highest probability of belonging to the current insect infestation evolution state. Based on the dominant ecological stage identifier, a key-value retrieval operation is performed in the feature sensitivity rule set to locate the pre-stored set of multi-source data response fingerprint association rules bound to this typical time stage. The set of multi-source data response fingerprint association rules is parsed to extract a list of defined high-sensitivity ecological driving factor names, which includes key factors such as wind speed vector fields, topographic connectivity maps, or nighttime infrared radiation differences. The extracted list of high-sensitivity ecological driving factor names is deduplicated and validated, and invalid factor references that do not exist in the current monitoring data dimension are removed. The validated high-sensitivity ecological driving factor names are classified and aggregated according to their physical meaning in the standardized multi-source feature sequence to generate a structured initial list of dominant driving factors. Through the above retrieval and parsing processing, the abstract stage identifier is transformed into a specific set of ecological factors to be enhanced, realizing a logical mapping from semantic state recognition to feature channel selection, and providing a clear target object for subsequent channel index mapping.

[0090] S4.3: Perform channel mapping parsing processing on each driving factor in the initial list of dominant driving factors to convert the abstract factor names such as wind speed vector field, terrain connectivity map or nighttime infrared radiation difference into the corresponding specific feature channel index numbers in the standardized multi-source feature sequence, and generate a set of feature channel indexes to be activated.

[0091] The system receives an initial list of dominant driving factors, containing abstract ecological factor names such as wind speed vector fields, topographic connectivity maps, or nighttime infrared radiation differences. A semantic-channel mapping dictionary is constructed, pre-storing the correspondence between abstract factor names and specific data dimensions in the standardized multi-source feature sequences. A hash lookup operation is performed on each driving factor name in the list to retrieve the corresponding unique identifier from the mapping dictionary. The feature channel index number bound to this unique identifier is extracted, pointing to a specific dimension position in the standardized multi-source feature tensor. The extracted index number is verified to be within the valid data dimension range; if it exceeds the range, it is marked as invalid and an exception handling log is triggered. The valid index numbers that pass verification are stored sequentially in the set of feature channel indexes to be activated. Through the above channel mapping parsing process, abstract ecological driving factors are transformed into specific data dimension indices, achieving precise positioning from the semantic space to the feature space, providing a precise addressing basis for subsequent gain coefficient configuration.

[0092] S4.4: Based on the set of indexes of the feature channels to be activated, query the preset gain configuration table to obtain the baseline response gain coefficient of each feature channel to be activated under the current typical time stage, and generate a coarse-grained channel gain parameter set.

[0093] It receives the set of feature channel indexes to be activated and the identifier of the current dominant ecological stage as input data.

[0094] Based on the dominant ecological stage identifier, the corresponding semantic stage partition record is retrieved from the preset gain configuration table. This record stores the baseline response intensity setting value of each feature channel under different insect evolution stages.

[0095] Iterate through each channel index number in the set of feature channel indexes to be activated, and use it as a key to perform an exact match search in the retrieved semantic stage partition records.

[0096] Extract the baseline response gain coefficient value corresponding to the successfully matched channel. This value represents the baseline level of contribution of the feature to the prediction of insect outbreaks under a specific ecological semantic background.

[0097] The extracted baseline response gain coefficients are assembled into a structured array according to the channel index order, forming a list of key-value pairs containing channel IDs and gain values.

[0098] Perform an integrity check on the key-value pair list to ensure that all channels to be activated have obtained valid gain parameters. If any are missing, fill in the default neutral gain value of 1.0 to maintain weight balance.

[0099] The verified list of key-value pairs is encapsulated into a standardized coarse-grained channel gain parameter group data structure, which includes stage labels, channel index sequences, and corresponding gain coefficient sequences.

[0100] By querying the preset gain configuration table and assembling the benchmark response gain coefficients, the channel index of the previous step is transformed into a specific quantization gain parameter set, thereby achieving a preliminary calibration of coarse-grained feature sensitivity for the current ecological scenario.

[0101] S4.5: Based on the coarse-grained channel gain parameter group, the feature channel activation instruction is encapsulated to explicitly specify the target channel whose response gain needs to be increased and its gain magnitude, and generate coarse-grained feature channel activation instructions for the current ecological scenario.

[0102] Step S5: Obtain the adaptive feature weight matrix by performing operations based on the coarse-grained feature channel activation instruction. Specifically, this includes: S5.1: Based on the dominant driving factor channel specified in the coarse-grained feature channel activation instruction, the pre-stored static reference feature weight vector is masked to generate an initial reconstructed feature weight vector containing high-gain markers. This initial reconstructed feature weight vector clarifies the priority enhancement status of key channels such as wind speed vector field, terrain connectivity map and nighttime infrared radiation difference.

[0103] The system receives coarse-grained feature channel activation instructions and parses the set of feature channel indices to be activated and their corresponding baseline response gain coefficients. It constructs a static baseline feature weight vector with dimensions consistent with the standardized multi-source feature sequence, storing the basic contribution values ​​of each ecological factor in its default state. An all-zero mask matrix is ​​initialized, its dimension strictly matching the number of feature channels, to identify the target channel positions requiring weight intervention at the current moment. Based on the set of feature channel indices to be activated, unit markers are placed at the corresponding index positions in the mask matrix to form a binarized channel selection mask. Element-wise multiplication is performed between the binarized channel selection mask and the static baseline feature weight vector, retaining the original baseline weights of non-dominant channels while temporarily clearing the weights of dominant driving factor channels to zero in preparation for injecting new gains. The baseline response gain coefficients are extracted from the coarse-grained channel gain parameter set and mapped to the corresponding weight positions of the dominant driving factor channels to generate high-gain weight components. The high-gain weight components are vector-superimposed with the retained weights of non-dominant channels to synthesize an initial reconstructed feature weight vector containing high-gain markers. By using mask mapping and gain injection, the abstract activation instructions from the previous step are transformed into an initial reconstructed feature weight vector with a specific numerical distribution, thereby explicitly improving the priority of key ecological driving factors.

[0104] For example, in a scenario with peak spatial diffusion, the set of feature channel indices to be activated includes wind speed vector field channel 3, terrain connectivity map channel 7, and nighttime infrared radiation difference channel 12. The static baseline feature weight vector is a 15-dimensional uniformly distributed vector, with each element having an initial value of 0.067. A 15-dimensional binarized channel selection mask is constructed, with only indices 3, 7, and 12 set to 1, and the rest to 0. After performing element-wise multiplication, non-dominant channels retain a base weight of 0.067, while the dominant channel weights are set to zero. The preset gain configuration table is queried to obtain the baseline response gain coefficients for wind speed, terrain, and infrared channels, which are 0.25, 0.20, and 0.18, respectively. These gain coefficients are filled into the corresponding positions and superimposed with the retained weights of other channels to generate the initial reconstructed feature weight vector. In this vector, the weight of the wind speed channel is increased from 0.067 to 0.25, significantly enhancing the model's sensitivity to airflow diffusion effects and ensuring that subsequent predictions focus on the dominant driving factors.

[0105] S5.2: Using the initial reconstructed feature weight vector as the input benchmark, the sliding window statistics module is called to perform multivariate covariance matrix calculation on the standardized multi-source feature sequences within the past fourteen days, so as to extract the covariance stability constraint parameter that characterizes the fluctuation amplitude of the linear correlation between each feature channel. This covariance stability constraint parameter quantifies the degree of coupling between different ecological factors in the time dimension.

[0106] The system receives the initial reconstructed feature weight vector generated by S5.1 and the standardized multi-source feature sequence of the past fourteen days output by S1.5 as the input data source for covariance stability calculation.

[0107] A sliding time window with a length of 336 hours was constructed, and backtracking from the current moment to extract multi-dimensional time-series data segments including ground trap counts, meteorological records, satellite remote sensing vegetation indices, and UAV thermal infrared images.

[0108] Zero-mean processing is performed on the feature channel data within the sliding window to eliminate the benchmark bias caused by different physical dimensions and ensure the statistical unbiasedness of subsequent correlation calculations.

[0109] Based on the processed time-series data fragments, the Pearson correlation coefficient between any two feature channels is calculated, and a symmetric multivariate covariance matrix is ​​constructed to quantify the linear coupling strength between ecological factors.

[0110] The standard deviation of the diagonal elements of the covariance matrix is ​​extracted and used as an indicator of the self-fluctuation amplitude of each feature channel, reflecting the stability of a single ecological factor in the short term.

[0111] The average absolute value of the off-diagonal elements is calculated as an index of the inter-channel mutual disturbance intensity, characterizing the degree of cross-influence of environmental factor changes on insect monitoring data.

[0112] The self-fluctuation amplitude and mutual disturbance intensity are weighted and fused to generate a covariance stability constraint parameter vector. The lower the value of this vector, the more stable the characteristic relationship is, and the higher the value, the greater the risk of violent oscillation.

[0113] Through the above multivariate statistical analysis, the original time-series data is transformed into covariance stability constraint parameters that quantify the tightness of coupling between ecological factors. This enables an accurate measurement of the characteristic noise level under uncertain data distribution, providing a robust constraint boundary for subsequent residual feedback fine-tuning.

[0114] For example, in a cotton field area, 336 sample points were selected from hourly data collected over the past 14 days. The covariance of the wind speed vector field and the difference in nighttime infrared radiation was calculated. If their means are 3.5 m / s and 2.1 K respectively, and their standard deviations are 0.8 and 0.5 respectively, the covariance value calculated using the formula is 0.32. Further calculation of the covariance matrix for all 12 feature channels revealed a mean standard deviation of 0.45 for the diagonal and a mean absolute standard deviation of 0.12 for the off-diagonal. After weighted fusion, the stability constraint parameter is 0.57. This parameter indicates that the coupling between current meteorological and insect data is moderate, with no significant abnormal oscillations. It can be directly used for constructing the fine-tuning gain function in S5.3, significantly improving the adaptability of the weight reconstruction process to sudden meteorological disturbances.

[0115] S5.3: Based on the covariance stability constraint parameters and the historical prediction residual sequence, a residual feedback-driven fine-tuning gain function is constructed. Nonlinear decay or enhancement operations are performed on each channel element in the initial reconstructed feature weight vector to generate a fine-grained modified feature weight vector with anti-interference capability. This fine-grained modified feature weight vector eliminates the risk of weight oscillation caused by sudden changes in data distribution.

[0116] The initial reconstructed feature weight vector, covariance stability constraint parameters, and historical prediction residual sequence are obtained as the input data source for fine-tuning the gain function.

[0117] The mean absolute error at each time step is calculated based on the historical prediction residual sequence, and a residual energy index is constructed to characterize the degree of recent prediction bias of the model.

[0118] A diagonalized regularization matrix is ​​constructed using covariance stability constraint parameters, and a weighted mapping is performed on the residual energy index to generate a channel-specific sensitivity factor.

[0119] The channel-specific sensitivity factor is mapped to the range of zero to one using the Sigmoid activation function, generating a basic fine-tuning coefficient vector to ensure the smoothness of the weight adjustment.

[0120] A nonlinear decay operator is introduced to perform an exponential decay operation on the basic fine-tuning coefficient vector based on the duration of the current semantic stage, thereby suppressing over-adjustment in the long-term stable state.

[0121] The attenuated fine-tuning coefficients are combined with the initial reconstructed feature weight vector using the Hadamard product to generate a fine-grained modified feature weight vector with anti-interference capabilities.

[0122] By using a nonlinear fine-tuning mechanism driven by residual feedback, coarse-grained weights are transformed into fine-grained weights that adapt to fluctuations in real-time data distribution, thus eliminating the risk of weight oscillations caused by sudden environmental changes.

[0123] For example, setting the sliding window length to 14 days, the calculated covariance stability parameter of the wind speed channel is 0.85, and the historical residual mean is 0.12. After constructing the regularization matrix, the sensitivity factor of the wind speed channel is mapped to 1.35. After Sigmoid function mapping and duration decay processing, a fine-tuning coefficient of 0.92 is generated. This coefficient is then multiplied by the initial weight of 0.6 using the Hadamard product to obtain a corrected weight of 0.552. This process significantly reduces the impact of single meteorological abrupt changes on the overall prediction model and improves the robustness of weight allocation.

[0124] S5.4: Input the fine-grained modified feature weight vector into the weight normalization engine and perform proportional redistribution processing under the sum constraint to generate a dynamic calibration intermediate feature weight matrix that satisfies the probability distribution characteristics. This dynamic calibration intermediate feature weight matrix ensures that the sum of the contributions of all feature channels is constant and maintains the relative sensitivity difference.

[0125] The fine-grained modified feature weight vector output from S5.3 is received. This vector contains the original gain values ​​of each channel after residual feedback fine-tuning. A non-negativity constraint check is performed on the fine-grained modified feature weight vector, forcing the negative elements less than zero to zero, eliminating the physically invalid weights caused by excessive reverse adjustment, and generating a non-negative intermediate weight vector.

[0126] Based on the non-negative intermediate weight vector, the arithmetic sum of all channel weight values ​​is calculated, and a normalized denominator scalar is constructed to ensure that subsequent proportional redistribution has a unified benchmark reference.

[0127] The element-wise division operation is used to divide each channel weight value in the non-negative intermediate weight vector by the normalized denominator scalar, and a global scaling operation is performed to map each channel weight value to an open interval between zero and one.

[0128] Perform a sum consistency check on the normalized weight vector to verify whether the sum of all elements is strictly equal to one. If there is a floating-point operation error, fine-tune the maximum weight value using the least squares method to compensate for the precision loss and generate a dynamic calibration intermediate feature weight matrix that satisfies the probability distribution characteristics.

[0129] Through the above normalization process, the fine-grained correction results of the previous step are transformed into a standardized weight distribution that conforms to the probability axiom system, so as to maintain the relative sensitivity of feature contribution under the constraint of constant total amount, and provide unbiased and stable input parameters for the prediction model.

[0130] For example, in a pest monitoring scenario including 12 feature channels such as wind speed, air temperature, and vegetation index, the fine-grained corrected weight vector output by S5.3 is [0.15, -0.02, 0.30, ..., 0.05]. After applying the non-negativity constraint, -0.02 is set to 0, resulting in a non-negative vector. The sum of the weights of all 12 channels is calculated to be 1.48. A division operation is performed on each element; for example, the weight of the wind speed channel becomes 0.15 / 1.48≈0.1014. After verification and fine-tuning, the sum of all elements in the generated dynamic calibration intermediate feature weight matrix is ​​strictly 1.0000. This processing ensures that while increasing the weights of the dominant factors, it suppresses the interference of secondary factors, significantly improving the model's sensitivity to wind speed changes during the peak diffusion period and effectively eliminating prediction bias caused by dimensional differences.

[0131] S5.5: Based on the matching result of the dynamic calibration intermediate feature weight matrix and the semantic anchor point of the current insect evolution stage, perform the final matrix encapsulation and version iteration operation to output a dynamically calibrated adaptive feature weight matrix that can be directly loaded into the prediction model. This dynamically calibrated adaptive feature weight matrix realizes the real-time dynamic calibration of feature contribution under unsupervised driving.

[0132] Step S6: Input the standardized multi-source feature sequence into a prediction model loaded with an adaptive feature weight matrix, perform a weighted fusion operation, and output the calculated probability of insect outbreak in the current region. Specifically, this includes: S6.1: Obtain the dynamically calibrated adaptive feature weight matrix and the standardized multi-source feature sequence. Based on the tensor broadcasting mechanism, map the channel gain coefficients in the adaptive feature weight matrix to the corresponding feature dimensions of the standardized multi-source feature sequence to generate a weighted feature tensor with channel sensitivity identifier, so as to establish the basic contribution strength of each ecological driving factor in the current semantic stage.

[0133] It receives the dynamically calibrated adaptive feature weight matrix and the standardized multi-source feature sequence, parses the channel dimension index of the weight matrix and the data structure of the feature sequence, and establishes the mapping relationship between the two.

[0134] Based on the tensor broadcasting mechanism, the adaptive feature weight matrix is ​​extended to the same spatiotemporal dimension as the standardized multi-source feature sequence, generating a full-dimensional weight mask tensor, ensuring that each spatial grid point and time slice corresponds to an independent channel gain coefficient.

[0135] Perform element-wise Hadamard product operation, multiplying and fusing the expanded weight mask tensor with the normalized multi-source feature sequence. The calculation formula is as follows: in For the weighted feature tensor, For adaptive weight matrix, For standardized feature sequences, This represents element-wise product.

[0136] The channel sensitivity identifier embedding operation is performed on the calculation result, and the semantic stage affiliation and gain level of each channel are marked in the metadata layer of the weighted feature tensor, forming a multidimensional data structure with interpretability.

[0137] By using the tensor broadcasting and element-wise fusion processing methods described above, the dynamic weight parameters generated in the previous step are transformed into weighted feature tensors with channel sensitivity identifiers, thereby establishing the basic contribution strength of each ecological driving factor in the current semantic stage and eliminating the interference of dimensional differences on feature importance.

[0138] For example, based on insect infestation monitoring data from a certain farmland area, the standardized multi-source feature sequence includes four channels: ground trap count, diurnal temperature range, vegetation index change rate, and thermal infrared texture entropy, with a spatial resolution of 100x100 grids and a time step of 1 hour. The adaptive feature weight matrix outputs that the current stage is "accelerated population growth," with corresponding channel gain coefficients of [0.4, 0.3, 0.1, 0.2]. Using a tensor broadcasting mechanism, the weight vector of length 4 is expanded into a 100x100x1x4 four-dimensional weight tensor. After performing the Hadamard product operation, the value of the ground trap count channel is amplified by 0.4 times, the diurnal temperature range channel by 0.3 times, while the vegetation index channel retains only a 0.1-fold contribution. In the generated weighted feature tensor, the signal strength of the trap count and temperature features is significantly higher than that of other channels, accurately reflecting the evolutionary pattern dominated by biological factors at this stage, and providing a high signal-to-noise ratio input for subsequent spatiotemporal aggregation.

[0139] S6.2: Perform spatiotemporal context aggregation operation on the weighted feature tensor with channel sensitivity identifier, and use three-dimensional convolution kernel to extract the spatial correlation patterns of dominant driving factors such as wind speed vector field, terrain connectivity map and nighttime infrared radiation difference in local neighborhood, and generate a high-dimensional potential state vector containing spatiotemporal evolution law to capture the nonlinear coupling characteristics in the insect spread process.

[0140] The system receives a weighted feature tensor with channel sensitivity identifiers. This tensor contains multi-source insect monitoring data calibrated with dynamic weights, in which dominant driving factors such as wind speed vector field, topographic connectivity map and nighttime infrared radiation difference have been given high response gain.

[0141] A three-dimensional convolutional neural network architecture is constructed, with the convolutional kernel size set to 3×3×3, corresponding to the height, width and temporal dimensions of continuous frames respectively. The stride is set to 1, and the padding mode is zero padding to preserve edge spatial information.

[0142] The weighted feature tensor is input into a three-dimensional convolutional layer, and a sliding window convolution operation is performed to extract multi-dimensional feature combinations in the local spatiotemporal neighborhood, capturing the coupling relationship between pest migration paths and meteorological factors at the microscale.

[0143] The convolutional output is standardized by using a batch normalization layer, which eliminates the dimensional differences between different feature channels, accelerates model convergence, and improves robustness to noisy data.

[0144] By applying a modified linear unit activation function to perform nonlinear mapping on the normalized feature map, negative responses are suppressed, thereby enhancing the feature expression intensity of key positive driving factors for pest outbreaks.

[0145] By stacking multiple 3D convolutional modules, the receptive field is expanded layer by layer, and from local texture details to global spatial distribution, high-dimensional potential state vectors containing the evolutionary laws of insects are abstracted layer by layer.

[0146] By using three-dimensional convolution and spatiotemporal aggregation processing, the weighted feature tensor from the previous step is transformed into a high-dimensional latent state vector, enabling the accurate capture of nonlinear coupling features during insect infestation.

[0147] S6.3: Construct an uncertainty suppression gating unit based on the high-dimensional latent state vector, calculate the covariance stability index of each feature channel in the past fourteen days and convert it into a gating attenuation factor, perform adaptive masking processing on the unstable components in the high-dimensional latent state vector that are greatly affected by noise, and generate a denoising-enhanced spatiotemporal feature representation to eliminate feature drift interference caused by uncertain data distribution.

[0148] The high-dimensional latent state vector output by S6.2 is received. This vector contains the spatiotemporal coupling characteristics of the insect situation evolution at the current moment and serves as the input object for uncertainty suppression processing.

[0149] Historical data fragments from each channel in the standardized multi-source feature sequence over the past fourteen days were extracted, and a sliding time window was constructed to capture the dynamic fluctuation characteristics of ecological factors.

[0150] Calculate the covariance matrix of each feature channel within the sliding window to quantify the linear correlation between different ecological factors and the stability of their changes over time.

[0151] Based on the eigenvalue decomposition results of the covariance matrix, the ratio of the largest eigenvalue to the smallest eigenvalue is extracted as the condition number, which is used to characterize the numerical stability of the channel data.

[0152] The Sigmoid activation function is used to map the condition number to the interval between zero and one, generating a gated attenuation factor that reflects the reliability of each channel. The larger the condition number, the closer the attenuation factor is to zero.

[0153] The generated gating attenuation factor is broadcast to the corresponding channel dimension of the high-dimensional latent state vector, and element-wise multiplication is performed to adaptively mask the unstable components.

[0154] Through the above processing method, the high-dimensional latent state vector of the previous step is transformed into a denoising-enhanced spatiotemporal feature representation, thereby achieving the expected technical effect of eliminating feature drift interference caused by uncertain data distribution.

[0155] S6.4: Input the denoised and enhanced spatiotemporal feature representation into the pre-trained insect outbreak probability estimation network, perform nonlinear transformation and normalized activation operation of the fully connected layer, map the high-dimensional abstract features to a continuous probability interval from zero to one, and generate an initial insect outbreak probability distribution map to quantify the probability level of pest population outbreaks in the current ecological scenario.

[0156] The denoised and enhanced spatiotemporal feature representation is received and flattened into a one-dimensional high-dimensional feature vector, which serves as the input layer data for the insect outbreak probability estimation network. A deep neural network architecture with three fully connected layers is constructed. The first fully connected layer receives the input vector, performs a linear weighted transformation, and introduces the non-linear activation function ReLU to capture high-order interactions between features. The second fully connected layer further compresses the feature dimension, extracts abstract semantic features strongly correlated with insect population density, and randomly discards some neuron outputs through the Dropout mechanism to suppress the risk of overfitting. The third fully connected layer maps the high-dimensional abstract features to the single-node output space, generating the original logits value, which represents the unnormalized confidence of the insect outbreak in the current ecological scenario. The Sigmoid activation function is used to perform a non-linear mapping on the logits value, constraining the value within a continuous interval from zero to one, generating pixel-level initial insect outbreak probability values. The outbreak probability of a single grid point is calculated using the following formula: in, This represents the probability of an insect infestation. The logits value is the output of the fully connected layer, and e is the natural constant. The above calculation process is repeated for all spatial grid points to generate a two-dimensional probability matrix covering the monitored area. This probability matrix is ​​then upsampled using bilinear interpolation to ensure its spatial resolution matches the original satellite remote sensing image, resulting in a smooth initial insect outbreak probability distribution map. Through nonlinear transformation and normalized activation operations of the fully connected layer, high-dimensional abstract features are mapped to a continuous probability interval from zero to one, generating the initial insect outbreak probability distribution map and achieving a quantitative expression of the probability level of pest population outbreaks in the current ecological scenario.

[0157] For example, the input denoising-enhanced spatiotemporal features are represented as a one-dimensional vector of length 512. The first fully connected layer contains 256 neurons, with a weight matrix size of 512×256 and a bias term of 256 dimensions, outputting a 256-dimensional feature vector after activation. The second fully connected layer contains 64 neurons, with a weight matrix size of 256×64 and a Dropout rate of 0.5, outputting a 64-dimensional abstract feature. The third fully connected layer contains 1 neuron, with a weight matrix size of 64×1, outputting a logits value z of 2.3. Substituting into the Sigmoid formula, the probability of a single-point outbreak is approximately 0.909. The same calculation is performed on 100×100 grid points within the monitoring area to generate a 100×100 probability matrix. After bilinear interpolation and upsampling to a resolution of 1000×1000, an initial insect outbreak probability distribution map is formed. The distribution map clearly identifies high-probability outbreak areas, and the probability value distribution conforms to the pest spread pattern, providing an accurate data foundation for subsequent geospatial constraint smoothing.

[0158] S6.5: Apply geospatial constraint smoothing method to the initial insect outbreak probability distribution map, and perform edge correction and regional normalization on the probability values ​​by combining crop planting boundary and administrative division vector data, and output the final determined insect outbreak probability calculation results for the current area, so as to form high-precision insect trend prediction data that can be directly used for decision support.

[0159] Step S7: Compare the spatial consistency between the calculated insect outbreak probability and the actual location of the outbreak. If a significant shift in centroid distance exceeding 200 meters is detected three consecutive times, a characteristic response gain correction coefficient is generated. Specifically, this includes: S7.1: Perform spatial geometric analysis on the predicted high-probability regions generated from the insect outbreak probability calculation results to extract the centroid coordinate data of the predicted high-probability regions. The centroid coordinate data of the predicted high-probability regions serves as the benchmark reference object for subsequent spatial consistency comparison.

[0160] S7.2: Perform Euclidean distance metric calculation based on historical data of the actual outbreak location and centroid coordinate data of the predicted high-probability area to generate a single prediction spatial offset value. The single prediction spatial offset value characterizes the instantaneous geographical deviation between the model prediction center and the actual disaster center.

[0161] Obtain the centroid coordinates of the predicted high-probability area output from step S7.1. This data includes two-dimensional geographic information of longitude and latitude, serving as the spatial reference for the model's predicted center. Read historical data of the actual outbreak location, sourced from field pest reports from ground-based plant protection monitoring stations or disaster center points confirmed by drone inspections, and extract their corresponding precise geographic coordinates as the true reference. Map the predicted centroid coordinates and the actual outbreak location coordinates to the same plane projection coordinate system to eliminate nonlinear deformation errors in longitude and latitude caused by the Earth's curvature, ensuring geometric consistency of distance measurements. Based on the projected two-dimensional coordinate vectors, construct a Euclidean distance measurement model to calculate the linear spatial offset between the predicted center and the true center. Validate the calculated offset values, removing outliers caused by GPS signal drift and retaining valid deviation data that conforms to the physical diffusion laws. Through the above Euclidean distance metric calculation, the coordinate data from the previous step is transformed into a single prediction spatial offset value that characterizes the instantaneous geographical deviation between the model's predicted center and the actual disaster center. This enables a quantitative assessment of the prediction accuracy and provides a basic input indicator for subsequent time-series sliding window cumulative analysis.

[0162] S7.3: Perform time-series sliding window cumulative analysis on the single predicted spatial offset values ​​generated within three consecutive time steps to construct a spatial offset sequence status flag. The spatial offset sequence status flag is used to identify whether a significant offset anomaly pattern with a centroid distance exceeding 200 meters is detected three times consecutively.

[0163] S7.4: Based on the trigger signal of the spatial offset sequence state flag, retrieve the list of unstable feature channels in the current semantic stage to locate the set of dominant driving factor channels that cause spatial prediction deviation. The set of dominant driving factor channels explicitly specifies the specific feature dimensions such as wind speed vector field or terrain connectivity map that need to be reversed.

[0164] The system receives trigger signals from the spatial offset sequence status flags to confirm that the current prediction model is in a state of continuous and significant deviation. It retrieves the feature sensitivity rule set corresponding to the current dominant ecological stage from the spatiotemporal anchor knowledge base, extracting the indexes of all feature channels marked with high response gain under this stage. It calls the historical covariance stability constraint parameter library to read the Pearson correlation coefficient matrix and variance volatility data between each feature channel over the past fourteen days. Based on a preset instability threshold, it filters out feature dimensions with variance volatility exceeding 0.15 and correlation with other dominant channels below 0.3, constructing an initial unstable candidate set. It calculates the azimuth angle of the offset vector between the centroid of the predicted high-probability region and the actual outbreak location in the geographic coordinate system. It performs cosine similarity matching between the azimuth angle of the offset vector and the spatial gradient direction of each unstable candidate feature channel, identifying the feature channel with the highest consistency with the offset direction. It merges the feature channel with the highest consistency and its associated auxiliary channels to generate a set of dominant driving factor channels causing spatial prediction deviation. Through the above chain derivation, the abstract spatial offset anomaly is transformed into specific feature dimension indexes such as wind speed vector field or terrain connectivity map that need to be adjusted in reverse, thus realizing the accurate location of the source of prediction deviation.

[0165] For example, when the spatial offset sequence status flag shows that the centroid distance exceeds 200 meters for three consecutive times, the system retrieves the semantic stage rule set currently in the "spatial diffusion peak period". Three high-gain channel indices are extracted: wind speed vector field, terrain connectivity map, and nighttime infrared radiation difference. Reading the covariance data for the past 14 days, it is found that the variance volatility of the wind speed vector field is 0.18, higher than the threshold of 0.15, and its correlation coefficient with terrain connectivity is 0.25, lower than the threshold of 0.3; therefore, it is included in the initial unstable candidate set. The calculated predicted centroid is biased towards the northeast, with an offset vector azimuth of 45 degrees. The spatial gradient direction of the wind speed vector field in this area is 40 degrees, and their cosine similarity is 0.996, significantly higher than other channels. Therefore, the wind speed vector field is determined to be the dominant driving factor causing the bias, and a dominant driving factor channel set containing the wind speed vector field channel index is generated for subsequent gain correction.

[0166] S7.5: Use the spatial offset direction vector of the dominant driving factor channel set and the preset covariance stability constraint matrix to perform gradient back-calculation to generate feature response gain correction coefficients. The feature response gain correction coefficients are used as closed-loop control signals to update the parameter configuration of the corresponding channel in the spatiotemporal anchor knowledge base.

[0167] The spatial offset direction vector of the dominant driving factor channel set and the preset covariance stability constraint matrix are obtained to construct a gradient backpropagation calculation model based on the error backpropagation principle. According to the deviation direction determined by the state flag of the spatial offset sequence, the difference between the two-dimensional geographic coordinates of the predicted centroid and the actual outbreak location is mapped to the error gradient vector in the feature space, quantifying the contribution of each dominant driving factor to the spatial prediction deviation.

[0168] The pre-defined covariance stability constraint matrix is ​​invoked. This matrix stores the Pearson correlation coefficients and inverse variances of each feature channel over the past fourteen days, used to characterize the robustness of the feature responses. Matrix inverse operations are used to decouple the linear coupling effects between features, ensuring that the correction coefficients are adjusted only for unstable channels that cause deviations, avoiding erroneous perturbations to stable channels.

[0169] The calculated correction coefficient vector is subjected to amplitude limiting to prevent drastic weight oscillations due to excessive single deviations. The amplitude-limited correction coefficients are then algebraically superimposed with the baseline gain coefficients under the current semantic stage to generate updated feature response gain parameters.

[0170] By using gradient back-calculation, the spatial offset anomaly from the previous step is transformed into quantitative feature response gain correction coefficient data, thereby achieving unsupervised dynamic calibration closed-loop control of feature contribution and significantly improving the model's generalization ability and prediction accuracy in complex ecological scenarios.

[0171] Step S8: Update the spatiotemporal anchor knowledge base using the feature response gain correction coefficient to complete the unsupervised-driven dynamic calibration and optimization of feature contribution. Specifically, this includes: S8.1: Obtain the feature response gain correction coefficient generated by the previous spatial consistency comparison step and the target typical time stage identifier that triggers the reverse adjustment mechanism. Based on the target typical time stage identifier, retrieve the corresponding multi-source data response fingerprint set in the spatiotemporal anchor knowledge base to locate the storage address of the unstable feature channel parameter to be updated.

[0172] S8.2: Read the historical feature response gain coefficient from the storage address of the unstable feature channel parameter to be updated, and perform a weighted fusion operation on the feature response gain correction coefficient and the historical feature response gain coefficient to generate a real-time feature response gain coefficient optimized for the current ecological scenario.

[0173] The storage address of the unstable feature channel parameter to be updated is read from the spatiotemporal anchor knowledge base, and the historical feature response gain coefficient of the channel under the current typical time stage is extracted. This historical coefficient represents the model's sensitivity benchmark to the ecological driving factor in past prediction periods, and serves as an inertial reference for weight updates.

[0174] Obtain the feature response gain correction coefficient generated by the preceding spatial consistency comparison step. This correction coefficient is a dynamic adjustment calculated by gradient backpropagation based on the spatial offset vector between the predicted centroid and the actual outbreak location, reflecting the instantaneous deviation direction and magnitude of the feature contribution in the current ecological scenario.

[0175] A weighted fusion computation model is constructed, which nonlinearly combines the historical feature response gain coefficients and the feature response gain correction coefficients. An exponential moving average method is used to smooth the update process, thereby suppressing drastic weight oscillations caused by single prediction errors and ensuring the continuity and stability of feature weight evolution.

[0176] The weighted fusion calculation is performed using the following formula: in, This refers to the real-time feature response gain coefficient optimized for the current ecological scenario. The historical characteristic response gain coefficient; This is the characteristic response gain correction coefficient; The forgetting factor, with a value ranging from 0 to 1, is used to balance the weight of historical memory and recent error feedback.

[0177] Dynamically adjust the forgetting factor based on the covariance stability constraint parameter The value. When the covariance of the feature channel fluctuates significantly over the past fourteen days, reduce... Values ​​are adjusted to enhance the response speed to recent error feedback; conversely, values ​​are increased. The values ​​are set to maintain the robustness of the weights and prevent overfitting caused by noise interference.

[0178] The calculated real-time feature response gain coefficients are truncated to limit them to a pre-defined, physically reasonable range. This operation eliminates the risk of gain coefficient divergence caused by extreme weather disturbances or sensor anomalies, ensuring the numerical stability of the input features for subsequent prediction models.

[0179] By using weighted fusion operations and dynamic boundary constraint processing, the historical parameters and correction signals from the previous step are transformed into real-time feature response gain coefficients with spatiotemporal adaptability. This enables refined dynamic calibration of feature contribution under uncertain data distribution, significantly improving the model's generalization ability and prediction accuracy in complex ecological scenarios.

[0180] S8.3: Based on the real-time feature response gain coefficient, the numerical distribution interval of the corresponding unstable feature channel in the multi-source data response fingerprint set is linearly mapped and reconstructed to generate an updated version of the multi-source data response fingerprint with dynamic adaptability.

[0181] Obtain the real-time feature response gain coefficient and the historical numerical distribution interval parameters of the corresponding unstable feature channels in the spatiotemporal anchor knowledge base. The parameters include the mean, standard deviation and extreme value boundaries, which serve as the benchmark data for linear mapping reconstruction.

[0182] Based on the historical numerical distribution interval parameters, a linear transformation matrix is ​​constructed from the original feature space to the dynamic calibration space. The initial calculation basis for the mapping slope and intercept is determined to ensure that the relative order of the data remains unchanged during the mapping process.

[0183] The mapping slope is nonlinearly scaled using the real-time feature response gain coefficient. When the gain coefficient is greater than 1, the distribution interval is expanded to enhance feature sensitivity. When the gain coefficient is less than 1, the distribution interval is compressed to suppress noise interference, thus generating an adaptive mapping slope parameter.

[0184] Substitute the adaptive mapping slope parameter into the linear transformation formula and combine it with the historical mean to calculate the new distribution center point.

[0185] Based on the updated distribution center point and adaptive mapping slope, the upper and lower bound thresholds of the distribution interval are recalculated, and the new effective data range is determined using the three-standard-deviation principle, eliminating the influence of outliers that exceed the new boundaries.

[0186] The original data of the corresponding unstable feature channels in the multi-source data response fingerprint set are standardized and remapped to transform the original values ​​into the standard normal distribution space according to the new distribution interval ratio, thereby generating an intermediate feature vector with dynamic adaptability.

[0187] The intermediate feature vector is denormalized and mapped back to the physical dimension space to preserve the actual physical meaning of the insect monitoring data. At the same time, the distribution characteristics after gain correction are embedded to generate an updated version of the multi-source data response fingerprint.

[0188] By using a linear mapping reconstruction process, the real-time feature response gain coefficients generated in the previous step are transformed into the distribution pattern adjustment of multi-source data response fingerprints, thereby enhancing the dynamic adaptability of feature channels to the current ecological scenario and improving the robustness of subsequent prediction models to uncertain data distributions.

[0189] S8.4: Replace the original multi-source data response fingerprint in the spatiotemporal anchor knowledge base with the updated multi-source data response fingerprint, and synchronously update the feature sensitivity rule set association index under the typical time stage to complete the unsupervised feature contribution dynamic calibration closed-loop operation.

[0190] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0191] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0192] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting pest outbreak trends by integrating multi-source data, characterized in that, Specifically, it includes: S1: Acquire multi-source insect infestation monitoring data, and perform noise reduction and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences; S2: Based on historical insect infestation reports and retrospective environmental data, an unsupervised clustering algorithm is used to mine data on key stages of insect population growth, spread, and diapause, and a spatiotemporal anchor knowledge base is constructed. The spatiotemporal anchor knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. S3: Map the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculate the similarity distribution between the current data and the multi-source data response fingerprints corresponding to each typical time stage, and generate a stage membership vector. S4: Based on the typical time stage pointed to by the peak in the stage membership vector, retrieve the preset feature sensitivity rule set and generate a coarse-grained feature channel activation instruction for the current ecological scenario. S5: Obtain the adaptive feature weight matrix by performing operations based on the coarse-grained feature channel activation instruction; S6: Input the standardized multi-source feature sequence into the prediction model loaded with the adaptive feature weight matrix, perform weighted fusion operation, and output the probability calculation result of the insect outbreak in the current area.

2. The method for predicting pest outbreak trends by fusing multi-source data according to claim 1, characterized in that, S6 is followed by: S7: Compare the spatial consistency between the calculated insect outbreak probability and the actual location of the outbreak. If a significant shift in centroid distance exceeding 200 meters is detected three times consecutively, generate a characteristic response gain correction coefficient. S8: Update the spatiotemporal anchor knowledge base using the feature response gain correction coefficient to complete the unsupervised feature contribution dynamic calibration optimization.

3. The method for predicting pest outbreak trends by fusing multi-source data according to claim 1, characterized in that, The multi-source insect monitoring data includes ground trapping counts, meteorological records, satellite remote sensing vegetation indices, and UAV thermal infrared images.

4. The method for predicting pest outbreak trends by fusing multi-source data according to claim 1, characterized in that, The process of obtaining the adaptive feature weight matrix based on the coarse-grained feature channel activation instruction includes: The original feature weights are reconstructed based on the coarse-grained feature channel activation instructions, and combined with the covariance stability constraints between feature channels over the past fourteen days, a fine-grained fine-tuning operation driven by residual feedback is performed to generate a dynamically calibrated adaptive feature weight matrix.

5. The method for predicting pest outbreak trends by fusing multi-source data according to claim 1, characterized in that, The feature sensitivity rule set stores association rules for multi-source data response fingerprints.

6. The method for predicting pest outbreak trends by fusing multi-source data according to claim 1, characterized in that, The S2 specifically includes: Historical insect infestation reports and retrospective environmental data are obtained, and the historical insect infestation reports and retrospective environmental data are subjected to spatiotemporal gridded resampling processing to generate a multidimensional time-series fusion dataset with a unified spatiotemporal benchmark. A set of detrended local evolution sequences was obtained by processing a multidimensional time-series fusion dataset. Unsupervised clustering algorithm is applied to the detrended local evolution sequence set to obtain insect evolution stage clusters with semantic annotations; Statistical analysis and extraction of multi-source data response fingerprints were performed on insect populations at different stages of insect infestation evolution. By encapsulating key-value associations of multi-source data response fingerprints, a spatiotemporal anchor knowledge base with fast retrieval capabilities is constructed.

7. The method for predicting pest outbreak trends by fusing multi-source data according to claim 6, characterized in that, The detrended local evolution sequence set obtained by processing the multidimensional time-series fusion dataset includes: Based on the multidimensional time-series fusion dataset, a sliding window segmentation method is used to extract time segments covering insect population growth, diffusion, and diapause processes. Standardized difference operations are then performed on the time segments to remove seasonal trend terms, generating a set of detrended local evolutionary sequences that characterize the rate of change in population dynamics.

8. The method for predicting pest outbreak trends by fusing multi-source data according to claim 6, characterized in that, The unsupervised clustering algorithm applied to the detrended local evolutionary sequence set to obtain semantically labeled insect evolutionary stage clusters includes: An unsupervised clustering algorithm based on density peaks is applied to the detrended local evolution sequence set to calculate the high-dimensional Euclidean distance distribution between sequences and identify cluster centers. The converged clusters are then mapped to five typical semantic labels: initial migration period, population acceleration and proliferation period, spatial diffusion peak period, environmental inhibition and stabilization period, and natural decline period, generating insect evolution stage clusters with semantic annotations.

9. The method for predicting pest outbreak trends by fusing multi-source data according to claim 8, characterized in that, The statistical analysis and extraction of multi-source data response fingerprints for insect populations at different stages of insect infestation evolution includes: For each insect infestation stage cluster with semantic annotation, the Pearson correlation coefficient matrix and time-frequency energy spectrum characteristics among multi-source monitoring variables within the cluster are statistically analyzed. A combination of quantitative indicators that can distinguish the significant differences between this stage and other stages is extracted to generate multi-source data response fingerprints corresponding to each typical time stage.

10. An insect pest monitoring instrument, characterized in that, A method for predicting pest outbreak trends by fusing multi-source data as described in any one of claims 1-9, comprising: Data preprocessing module: used to acquire multi-source insect infestation monitoring data, and to perform noise reduction and spatiotemporal alignment preprocessing on the multi-source insect infestation monitoring data to generate standardized multi-source feature sequences; Knowledge base construction module: Based on historical insect infestation reports and retrospective environmental data, it uses unsupervised clustering algorithms to mine data on key stages of insect population growth, spread, and diapause, and constructs a spatiotemporal anchor knowledge base. The spatiotemporal anchor knowledge base stores multi-source data response fingerprints corresponding to each typical time stage. Stage membership calculation module: used to map the standardized multi-source feature sequence to the spatial dimension of the spatiotemporal anchor knowledge base, calculate the similarity distribution between the current time data and the multi-source data response fingerprints corresponding to each typical time stage, and generate a stage membership vector; Feature channel activation module: Based on the typical time stage pointed to by the peak in the stage membership vector, it retrieves a preset feature sensitivity rule set and generates a coarse-grained feature channel activation instruction for the current ecological scenario. The weight adaptive generation module is used to generate an adaptive feature weight matrix based on the coarse-grained feature channel activation instruction. Insect Infestation Prediction Output Module: Connected to the data preprocessing module and the weight adaptive generation module respectively, it is used to input the standardized multi-source feature sequence into the prediction model loaded with the adaptive feature weight matrix, perform weighted fusion operation, and output the insect outbreak probability calculation result of the current area.

Images