A method for predicting insect aggregation pheromone behavior

Abstract

This application discloses a method for predicting insect stratification behavior, relating to the field of insect radar technology. The method includes: acquiring phased array insect radar echo data in the target airspace, as well as vertical profile meteorological data collected by laser wind radar and microwave radiometer; generating a spatiotemporal distribution heatmap based on the echo data; mapping the vertical profile meteorological data to the spatiotemporal distribution heatmap using a spatiotemporal interpolation algorithm to obtain a spatiotemporally aligned multidimensional meteorological feature heatmap; labeling the multidimensional meteorological feature heatmap to obtain a labeled sample dataset, and training an XGBoost prediction model using the sample dataset; and using the trained XGBoost prediction model to perform forward inference on real-time meteorological data to obtain prediction results, including the probability of stratification, predicted height, and estimated intensity. This application achieves a leap from empirical guesswork to intelligent and interpretable prediction of insect stratification behavior through the above method.

Description

Technical Field

[0001] This application relates to the field of insect radar technology, and in particular to a method for predicting insect aggregation and stratification behavior. Background Technology

[0002] Insect migration, especially nocturnal aggregation and stratification, is a crucial biometeorological process affecting agricultural ecosystems. Effective monitoring and prediction of this behavior are key prerequisites for precise control of migratory pests. Phased array insect radar can effectively detect the distribution and dynamics of aerial insects, providing important data support for research.

[0003] However, existing technical solutions often have the following limitations: insect radar data and high-resolution meteorological observation data are often difficult to effectively integrate on a spatiotemporal scale; existing prediction methods mostly rely on empirical models or simple statistical regression, which are difficult to characterize the nonlinear, multi-factor coupling relationship between meteorological factors and complex insect behavior, resulting in insufficient prediction accuracy and reliability; most models are black boxes, lacking explanations of the physical mechanisms of prediction results and quantitative analysis of key driving factors, making it difficult to use prediction results to guide actual prevention and control decisions.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] Existing technologies for predicting insect aggregation behavior suffer from difficulties in data fusion, low model prediction accuracy, and poor interpretability, further hindering accurate prediction and effective early warning of insect aggregation behavior. Summary of the Invention

[0006] This application provides a method for predicting insect aggregation behavior, which can solve the problems in the prior art of predicting insect aggregation behavior, such as difficulty in data fusion, low model prediction accuracy and poor interpretability, which further makes it difficult to accurately predict and effectively warn of insect aggregation behavior.

[0007] In a first aspect, embodiments of this application provide a method for predicting insect aggregation and stratification behavior. The method includes: acquiring phased array insect radar echo data in the target airspace, as well as vertical profile meteorological data collected by laser wind radar and microwave radiometer; generating a spatiotemporal distribution heatmap based on the echo data; mapping the vertical profile meteorological data to the spatiotemporal distribution heatmap using a spatiotemporal interpolation algorithm to obtain a spatiotemporally aligned multidimensional meteorological feature heatmap; labeling the multidimensional meteorological feature heatmap to obtain a labeled sample dataset, and training an XGBoost prediction model using the sample dataset; and using the trained XGBoost prediction model to perform forward inference on real-time meteorological data to obtain prediction results, including the probability of stratification, predicted height, and estimated intensity.

[0008] In one implementation of this application, a spatiotemporal interpolation algorithm is used to map vertical profile meteorological data to a spatiotemporal distribution heatmap to obtain a spatiotemporally aligned multidimensional meteorological feature heatmap. Specifically, this includes: performing bilinear interpolation on the vertical profile meteorological data based on the timestamp and height coordinates of each pixel unit in the spatiotemporal distribution heatmap to obtain the wind speed vector, temperature value, and humidity value at the same height coordinate; and combining the wind speed vector, temperature value, and humidity value to form the basic meteorological feature vector of the current pixel unit.

[0009] In one implementation of this application, the method further includes: calculating the vertical temperature lapse rate and the vertical relative humidity gradient based on the timestamp and height coordinates of each pixel unit; calculating the vertical wind shear intensity and direction based on the wind speed vector of adjacent height pixel units, and calculating the temperature-humidity combination index in combination with the ambient temperature and saturated water vapor pressure; and concatenating the vertical temperature lapse rate, humidity gradient, wind shear intensity, and temperature-humidity combination index with the basic meteorological feature vector to obtain an enhanced feature vector, which includes low-altitude enhanced features and full vertical profile enhanced features.

[0010] In one implementation of this application, the XGBoost model includes a global filtering model based on low-altitude enhancement features and a local refinement model activated by the global filtering model. A multi-dimensional meteorological feature heatmap is labeled to obtain a labeled sample dataset, and the XGBoost prediction model is trained using this sample dataset. Specifically, this includes: dividing the sample dataset into training and validation sets according to spatiotemporal blocks; training the global filtering model using the training set, which takes low-altitude enhancement features as input and outputs the probability of the first stratification; selecting training sets with the probability of the first stratification occurring higher than an activation threshold for training the local refinement model, which takes full vertical profile enhancement features as input and outputs the probability of the second stratification, as well as altitude and intensity regression values; wherein, in the training of the local refinement model, spatiotemporal smoothness constraints and flight altitude physical boundary constraints are introduced into the objective function to regularize the model prediction.

[0011] In one implementation of this application, the method further includes: defining hyperparameter search spaces for the global screening model and the local refinement model respectively, wherein the hyperparameters include the maximum depth of the tree, the learning rate, and the subsampling ratio; and using a Bayesian optimization method, with the average performance index of spatiotemporal block cross-validation as the optimization objective, to automatically optimize within the search space.

[0012] In one implementation of this application, after using the trained XGBoost prediction model to perform forward inference on real-time meteorological data and obtain the prediction result, the specific steps include: calculating the variance of the prediction results of all base learners in the XGBoost prediction model as an uncertainty estimate of the current sample prediction result; and weighting and fusing the prediction results of multiple spatially adjacent pixel units according to the uncertainty estimate to generate an overall regional prediction.

[0013] In one implementation of this application, the method further includes: presetting multi-layer thresholds for stratification occurrence probability, prediction intensity, and uncertainty; classifying warning levels according to the matching relationship between the prediction results and the multi-layer thresholds, and generating warning information, which includes warning area, expected altitude, dominant meteorological factor, and confidence level description.

[0014] In one implementation of this application, a swarm segmentation and behavior quantification labeling are performed on a spatiotemporal distribution heatmap to obtain a labeled sample dataset. Specifically, this includes: performing temporal slicing and spatial clustering analysis on the spatiotemporal distribution heatmap to identify candidate swarm regions; calculating the vertical profile morphology index and multi-frame temporal stability index of the candidate swarm regions, and classifying the candidate swarm regions into discrete noise, diffuse swarms, or stable stratified swarms according to preset morphology and stability discrimination rules; labeling the pixel units within the stable stratified swarm regions with category labels indicating stratification, and recording the centroid height and average density attribute labels.

[0015] In one implementation of this application, the method further includes: introducing atmospheric electric field instrument data to obtain time-series data on atmospheric electric field intensity and polarity changes in the target airspace; spatiotemporally aligning and fusing the time-series data on atmospheric electric field intensity and polarity changes with a multidimensional meteorological feature heatmap; extracting electric field risk feature vectors through a lightweight neural network encoder; and fusing the electric field risk feature vectors and enhancement feature vectors in the early or mid-term stages of forward inference in the XGBoost prediction model, using them together as model inputs.

[0016] In one implementation of this application, the method further includes: constructing an insect and path knowledge graph, the graph including insect flight paths, source and sink information, and historical stratification heights in different seasons; after the XGBoost prediction model outputs prediction results, matching the prediction results with the knowledge graph to obtain the prior distribution of species tendencies and behaviors under the given spatiotemporal conditions; and using a Bayesian correction framework, fusing the prediction results output by the model as the likelihood and the prior distribution of species tendencies and behaviors provided by the knowledge graph as the prior probability to obtain the posterior prediction result.

[0017] This application provides a method for predicting insect aggregation and stratification behavior. It maps meteorological data to a radar observation grid using a rigorous spatiotemporal interpolation algorithm and constructs derived features with clear biophysical significance, including temperature gradients, wind shear, and temperature-humidity indices. This provides the model with inputs more directly related to insect behavioral mechanisms, laying a data foundation for high-precision prediction. A cascaded multi-task XGBoost model framework is employed, with initial screening followed by refinement, improving computational efficiency and specificity. The introduction of spatiotemporal smoothness and flight altitude physical boundary constraints into the model ensures that the prediction results are not only statistically accurate but also physically accurate. It is also more biologically reasonable and credible; by quantifying the uncertainty of the model prediction and combining it with multi-level thresholds to generate dynamic hierarchical early warning information, the prediction output is no longer a single value, but a decision basis with confidence, which significantly improves the scientificity and practicality of the early warning information; by introducing multimodal data such as atmospheric electric field and insect migration knowledge graph, and using neural network encoding and Bayesian correction, the sudden weather risk and historical ecological laws are integrated into the prediction framework, which enables the model to have reasoning and correction capabilities that go beyond conventional meteorological correlations and are closer to real ecological scenarios, significantly improving the prediction robustness and ecological interpretability in complex situations. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 A flowchart illustrating a method for predicting insect aggregation and stratification behavior provided in an embodiment of this application;

[0020] Figure 2 A heat map of the spatiotemporal distribution of insect numbers, provided for a method of predicting insect aggregation and stratification behavior in an embodiment of this application;

[0021] Figure 3 A multidimensional meteorological characteristic heat map of insect aggregation stratification behavior provided in an embodiment of this application for predicting insect aggregation stratification behavior;

[0022] Figure 4 A meteorological factor-driven ranking diagram illustrating a method for predicting insect aggregation and stratification behavior provided in this application embodiment;

[0023] Figure 5 This is a schematic diagram illustrating the meteorological factor-driven ranking of a method for predicting insect aggregation and stratification behavior provided in this application embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] This application provides a method for predicting insect aggregation behavior, which solves the problems in the prior art of predicting insect aggregation behavior, such as difficulty in data fusion, low model prediction accuracy and poor interpretability, which further makes it difficult to accurately predict and effectively warn of insect aggregation behavior.

[0026] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0027] Figure 1 A flowchart illustrating a method for predicting insect aggregation and stratification behavior provided in an embodiment of this application. Figure 1 As shown in the embodiments of this application, a method for predicting insect aggregation and stratification behavior specifically includes the following steps:

[0028] Step 10: Acquire phased array insect radar echo data of the target airspace, as well as vertical profile meteorological data collected by laser wind radar and microwave radiometer.

[0029] In this step, high-resolution phased array insect radar is used to acquire radar observation data including the phenomenon of insects agglomerating in layers at night, covering the entire process of insect activity in the night sky. At the same time, laser wind-measuring radar and microwave radiometers deployed at the same site as the radar system are used to synchronously acquire data on key meteorological factors such as temperature, humidity, and wind speed, ensuring a high degree of consistency between radar data and meteorological data in the time dimension.

[0030] Step 20: Generate a spatiotemporal distribution heatmap based on the echo data.

[0031] In this step, a spatiotemporal distribution heatmap of insect populations is first constructed based on high-resolution phased array radar data, reflecting the population distribution at different times and altitudes. For example... Figure 2 As shown, this heatmap illustrates the aggregation trend of insect swarms throughout the night.

[0032] As an optional embodiment, insect swarm segmentation and behavior quantification annotation are performed on the spatiotemporal distribution heatmap to obtain a labeled sample dataset. Specifically, it may include: Step 201: Perform temporal slicing and spatial clustering analysis on the spatiotemporal distribution heatmap to identify candidate insect swarm regions.

[0033] In this step, the continuous spatiotemporal heatmap is divided into a series of discrete time slices along the time axis, according to the original time resolution or multiple time units merged according to the analysis requirements. Each time slice is a two-dimensional image, with the horizontal axis representing distance and the vertical axis representing height. The pixel value represents the number of insects within that spatiotemporal unit. On each time slice, a density-based spatial clustering algorithm is applied, treating the number of insects per pixel as density. This identifies local clusters where the number of insects is significantly higher than the surrounding background, while filtering out scattered isolated points that may be caused by noise. In this way, one or more candidate insect swarm regions are obtained on each time slice, and each candidate region consists of a series of spatially adjacent and density-connected pixel units.

[0034] Step 202: Calculate the vertical profile morphology index and multi-frame temporal stability index of the candidate swarm region, and classify the candidate swarm region into discrete noise, diffuse swarm or stable stratified swarm according to the preset morphology and stability discrimination rules.

[0035] In this step, a vertical profile of insect population distribution in the candidate region is extracted, and the peak intensity, full width at half maximum (FWHM), and distribution skewness of the profile are calculated. Peak intensity reflects the core density of the aggregation; FWHM characterizes the vertical thickness of the insect layer, with smaller values ​​indicating a thinner and more concentrated layer; distribution skewness describes the symmetry of the profile shape and can be used to determine whether the insect layer is symmetrically distributed or spreading upwards or downwards. The evolution of the candidate region in multiple consecutive time slices is tracked, and the centroid height, i.e., the time series of the region-weighted average height, is calculated. Based on this time series, the variance of the centroid height and the autocorrelation coefficient between the centroid heights of adjacent time slices are further calculated. A small variance indicates small fluctuations in the insect layer height; a high autocorrelation coefficient indicates smooth and consistent height changes, rather than random jumps.

[0036] Furthermore, the calculated core density, vertical profile morphology indices (peak value, full width at half maximum, skewness), and temporal stability indices (height variance, autocorrelation) are input into a pre-defined set of discrimination rules based on domain knowledge. If the core density of a region is extremely low and spatially dispersed, it is judged as discrete noise. If the full width at half maximum of the vertical profile of a region is large, or the centroid height variance is large and the autocorrelation is low, it indicates that the region is in a diffusion, migration, or unstable state and is judged as a diffusion swarm. Only when the region simultaneously satisfies the following conditions—core density above a threshold, full width at half maximum of the vertical profile below a threshold, centroid height variance below a threshold, and autocorrelation coefficient above a threshold—is it judged as a stable stratified swarm. This indicates that the insects have formed a typical stratified structure with high density, thin thickness, and stable position at a specific height.

[0037] Step 203: Label the pixel units in the stable stratified swarm region with the category labels of the aggregation and stratification, and record the centroid height and average density attribute labels.

[0038] In this step, each pixel unit within the region is assigned a uniform category label, for example, using a binary value of 1 to represent stratification. Simultaneously, two key continuous attribute labels are recorded for each pixel unit: Centroid Height: Records the overall centroid height value of the stable stratified insect colony region to which the pixel unit belongs, serving as a supervisory signal for the model learning the "stratification height" regression task; Average Density: Records the normalized average insect population density of the region to which the pixel unit belongs, serving as a supervisory signal for the model learning the stratification strength regression task.

[0039] In this way, the original spatiotemporal distribution heatmap is transformed into a structured sample dataset. In this dataset, each valid spatiotemporal pixel unit corresponds to a data record, which includes spatiotemporal coordinates, corresponding multidimensional meteorological features, and regression labels such as whether it forms a layer, layer height, and layer intensity.

[0040] Step 30: Using a spatiotemporal interpolation algorithm, the vertical profile meteorological data is mapped to a spatiotemporal distribution heatmap to obtain a spatiotemporally aligned multidimensional meteorological feature heatmap.

[0041] In this step, meteorological data collected by lidar and microwave radiometer can be mapped to each pixel unit using data interpolation and nearest neighbor algorithms, resulting in a meteorological variable heatmap corresponding to the spatiotemporal distribution heatmap of insect populations. Simultaneously, using insect swarm segmentation methods, behavioral labels are assigned to each pixel unit based on the spatiotemporal distribution heatmap of insect populations, indicating whether it belongs to a migratory area, whether it has formed stratification, etc. Figure 3 As shown in the figure. Finally, the spatiotemporal information of each pixel, insect quantity information, behavioral label information, and multidimensional meteorological information are integrated to form a dataset for modeling.

[0042] As an optional embodiment, the vertical profile meteorological data is mapped to the spatiotemporal distribution heat map through a spatiotemporal interpolation algorithm to obtain a spatiotemporally aligned multidimensional meteorological feature heat map. Specifically, it may include: Step 301: Based on the timestamp and height coordinate of each pixel unit in the spatiotemporal distribution heat map, bilinear interpolation is performed on the vertical profile meteorological data to obtain the wind speed vector, temperature value and humidity value at the same height coordinate.

[0043] In this step, each pixel unit of the spatiotemporal distribution heatmap has a unique timestamp and height coordinates. The wind speed and direction data provided by the laser wind radar, as well as the temperature and humidity vertical profile data retrieved by the microwave radiometer, all have their own timestamps and height layering information, and all data sources are strictly synchronized in time.

[0044] For any target pixel unit P(t, h) in the heatmap (t being time and h being altitude), the system searches for available meteorological observation points in the vicinity of the spatiotemporal range. In the time dimension, it finds the two nearest wind field scanning times before and after t; in the vertical dimension, it finds the two nearest wind field observation altitude layers above and below h. Based on the wind speed vectors at these four spatiotemporal points—previous time-lower layer, previous time-upper layer, subsequent time-lower layer, and subsequent time-upper layer—which can be decomposed into u and v components, linear interpolation is first performed in the vertical direction to obtain the wind vectors at altitude h at the two nearest times. Then, linear interpolation is performed in the time direction to finally obtain the horizontal wind speed and direction of pixel unit P at time t and altitude h. The microwave radiometer provides temporally continuous and vertically discrete temperature and relative humidity profiles. For the altitude h of pixel unit P, the two nearest data layers above and below altitude h are directly found in the temperature and humidity profile corresponding to time t. Through one-dimensional linear interpolation, the temperature and relative humidity values ​​at altitude h are accurately calculated.

[0045] Step 302: Combine the wind speed vector, temperature value, and humidity value to form the basic meteorological feature vector of the current pixel unit.

[0046] In this step, the horizontal wind speed obtained by the above interpolation calculation can be represented as the east-west component u and the north-south component v, wind direction, temperature, and relative humidity, which are combined into a basic feature vector, denoted as [u, v, T, RH].

[0047] As an optional embodiment, the method may further include: step 303: calculating the vertical temperature lapse rate and the vertical relative humidity gradient based on the timestamp and height coordinates of each pixel unit.

[0048] In this step, for pixel unit P(t,h), the local vertical temperature gradient dT / dh near that height is calculated using the temperature interpolation results of neighboring pixels in the vertical direction, h-Δh and h+Δh. This parameter is directly related to atmospheric stability. Similarly, the rate of change of relative humidity with height dRH / dh is calculated. The humidity gradient affects the loss of moisture from the insect's body surface and may be related to the insect's hygroscopic behavior. The wind vector difference is calculated using the horizontal wind vectors of the adjacent layers above and below pixel unit P. The vertical wind shear intensity is defined as |ΔV| / Δh, and the direction is determined by the direction of ΔV. Strong wind shear may hinder the formation of the insect layer or affect the structure.

[0049] Step 304: Based on the wind speed vector of adjacent height pixel units, calculate the vertical wind shear intensity and direction, and combine the ambient temperature and saturated water vapor pressure to calculate the temperature and humidity combination index.

[0050] In this step, an index that can comprehensively reflect the feeling of stuffiness or dryness is constructed by combining ambient temperature T and relative humidity. First, the saturated water vapor pressure is calculated based on temperature T, then the actual water vapor pressure is calculated based on relative humidity, and finally the saturation difference or temperature-humidity ratio T / RH is calculated.

[0051] Step 305: Concatenate the temperature vertical lapse rate, humidity gradient, wind shear intensity, and temperature-humidity combination index with the basic meteorological feature vector to obtain the enhanced feature vector, which includes low-altitude enhanced features and full vertical profile enhanced features.

[0052] In this step, the derived features such as the temperature vertical lapse rate, humidity vertical gradient, vertical wind shear intensity, and temperature-humidity combination index obtained above are concatenated with the basic feature vector [u,v,T,RH] to form an enhanced feature vector with higher dimensions and richer physical information.

[0053] Furthermore, the enhancement vector is divided into low-altitude enhancement features, which refers to the set of enhancement features of all pixel units within a certain critical height range from the ground, mainly used to quickly screen low-altitude areas that may form layers; and full vertical profile enhancement features, which refers to the complete set of enhancement features of all pixel units within a range from the ground to the maximum detection height of the radar, used for fine-tuning the selected key areas.

[0054] In this way, the originally independent multi-source data are integrated into a unified, pixel-level feature dataset rich in physical information, which corresponds one-to-one with the behavior label dataset through spatiotemporal coordinates, and together constitute the training samples for the subsequent XGBoost model to learn complex mapping relationships.

[0055] Step 40: Label the multidimensional meteorological feature heat map to obtain a labeled sample dataset, and train the XGBoost prediction model using the sample dataset.

[0056] In this step, based on the constructed dataset, a layered behavior prediction model is built using the XGBoost algorithm. Multidimensional meteorological environmental factors for each spatiotemporal unit are used as input features, and a series of insect swarm labels for the corresponding spatiotemporal unit, such as whether they cluster in layers, are used as response variables. The model learns the correspondence between meteorological features and insect swarm labels to predict insect aggregation behavior. After training, XGBoost can obtain the importance ranking of each meteorological factor through feature contribution calculation, reflecting its role in improving model performance.

[0057] As an optional embodiment, the XGBoost model includes a global filtering model based on low-altitude enhanced features and a local refinement model activated by the global filtering model; the multi-dimensional meteorological feature heat map is labeled to obtain a labeled sample dataset, and the XGBoost prediction model is trained through the sample dataset, which may specifically include: Step 401: Divide the sample dataset into a training set and a validation set according to spatiotemporal blocks.

[0058] In this step, considering the temporal and spatial continuity (autocorrelation) of insect behavior and meteorological conditions, a completely random partitioning method is not adopted to prevent information leakage. The entire observation period can be divided into several non-overlapping blocks based on consecutive nighttime periods and different altitude ranges. Data from most of these blocks is randomly allocated as the training set for learning model parameters; the remaining blocks serve as the validation set for hyperparameter tuning, model selection, and early stopping to monitor the model's generalization ability.

[0059] Step 402: Train the global screening model using the training set. The global screening model takes the low-level enhancement features as input and outputs the probability of the first layer formation.

[0060] In this step, the low-altitude enhanced feature vectors of all samples in the training set are used as input. These features are mainly concentrated in the lower troposphere, such as 0-1000 meters. The model is trained as a binary classification task. The learning objective is to determine whether stratification is likely to occur in a given low-altitude region, given its meteorological characteristics. The model output is the probability of the first stratification, a continuous value between 0 and 1, with higher values ​​indicating a greater likelihood of stratification. By minimizing the loss function between the predicted probability and the true binary label, the model learns to identify environmental signals favorable to stratification from low-altitude meteorological patterns.

[0061] Step 403: Select the training set where the probability of the first layering is higher than the activation threshold to train the local refinement model. The local refinement model takes the full vertical profile enhancement feature as input and outputs the probability of the second layering, as well as the altitude regression value and intensity regression value. In the training of the local refinement model, spatiotemporal smoothness constraints and flight altitude physical boundary constraints are introduced into the objective function to regularize the model prediction.

[0062] In this step, the trained global screening model is used to perform forward inference on all samples in the training set, calculating the probability of first stratification for each sample. An activation threshold is set, and only samples with probabilities higher than this threshold are retained, forming a new, high-quality, refined training subset. This subset automatically focuses on potential stratified regions with more complex meteorological conditions that require more detailed modeling. For each sample in the refined training subset, a full vertical profile enhanced feature vector is used as input. This differs from the input of the first-level model; it includes all meteorological features and their gradient information across the entire vertical profile from the ground to the maximum radar detection height, providing a more comprehensive description of the environmental conditions.

[0063] Understandably, the model is designed as a multi-task learning model, performing three tasks simultaneously: Task 1: Refined classification: outputting the probability of the second stratification occurring; Task 2: Height regression: outputting the predicted height value, i.e., the stratification center height estimated by the model; Task 3: Intensity regression: outputting the predicted intensity value, i.e., the relative stratification intensity estimated by the model, such as normalized density.

[0064] Furthermore, to ensure that model predictions are not only data-driven but also conform to physical common sense and spatiotemporal laws, a custom regularization term is introduced into the objective function of the locally refined model: a spatiotemporal smoothness constraint, which penalizes the model for giving excessively different prediction results for spatiotemporally adjacent samples, such as pixels that are sequential in time and at similar heights. This forces the model to learn the temporal and spatial continuity of insect stratification behavior, resulting in a smooth and reasonable transition in the predicted stratification height and intensity on the spatiotemporal map, rather than producing isolated, discontinuous noise points; and a flight altitude physical boundary constraint, which defines a reasonable feasible altitude region based on the known physiological flight capabilities of the target insect species, such as the practical flight upper limit for migratory moths. During training, if the predicted altitude value significantly exceeds this feasible region, the model will be subject to additional penalties. This ensures that the model's prediction output is biologically reasonable.

[0065] As an optional embodiment, the method may further include: defining hyperparameter search spaces for the global screening model and the local refinement model respectively, wherein the hyperparameters include the maximum depth of the tree, the learning rate, and the subsampling ratio; and using a Bayesian optimization method, with the average performance index of spatiotemporal block cross-validation as the optimization objective, to automatically optimize within the search space.

[0066] In this step, search ranges are defined for the key hyperparameters of both models, mainly including the maximum tree depth, learning rate, and subsampling ratio. The overall performance metric of the model on the validation set is used as the optimization objective. To accurately assess generalization ability, spatiotemporal block cross-validation is employed for performance evaluation to prevent evaluation bias caused by the spatiotemporal correlation of data. Based on the evaluated hyperparameter combinations and performance results, a surrogate model is iteratively constructed, and the next candidate combination is selected for evaluation. This process is repeated until performance convergence or the maximum number of iterations is reached, ultimately determining the optimal hyperparameter combination.

[0067] Step 50: Using the trained XGBoost prediction model, perform forward inference on real-time meteorological data to obtain prediction results, including the probability of stratification, predicted height, and estimated intensity.

[0068] As an optional embodiment, after using the trained XGBoost prediction model to perform forward inference on real-time meteorological data and obtain the prediction results, the specific steps may include: Step 501: Calculate the variance of the prediction results of all base learners in the XGBoost prediction model as an uncertainty estimate of the prediction results of the current sample; Step 502: Perform weighted fusion on the prediction results of multiple spatially adjacent pixel units based on the uncertainty estimate to generate an overall regional prediction.

[0069] In this step, the XGBoost model is an ensemble of multiple decision trees (i.e., base learners). For the prediction of each sample, the variance of the outputs of all base learners is calculated as an estimate of the uncertainty of the prediction at that point. The larger the variance, the greater the divergence in the model's predictions for that sample, and the more uncertain the result. For the prediction results of multiple spatially adjacent pixel units, an inverse variance-weighted average is performed based on their respective uncertainty estimates. Predictions with lower uncertainty have higher weights, and vice versa, thereby generating a smoother and more reliable overall prediction of the region.

[0070] As an optional embodiment, the method may further include: step 503: preset multi-layer thresholds for stratification occurrence probability, prediction intensity, and uncertainty; step 504: classify warning levels according to the matching relationship between the prediction results and the multi-layer thresholds, and generate warning information, which includes warning area, expected altitude, dominant meteorological factor, and confidence level description.

[0071] In this step, the importance of the dominant meteorological factors can be quantified using SHAP (Shapley Additive Explanations) values ​​to analyze the nonlinear impact of each environmental factor on migration decisions:

[0072]

[0073] in For the complete set of features, For feature subset, This serves as the model's prediction function. Based on the aforementioned data preprocessing and model construction, for example, using data from 21:00 to 4:00 the next day, eight key meteorological variables are selected as independent variables, and whether clustering occurs is used as the response variable. An XGBoost model is constructed for predictive analysis, and SHAP values ​​are applied for effective feature analysis.

[0074] like Figure 4 As shown, the occurrence of nocturnal stratification behavior is mainly affected by temperature, with a SAP value of 0.20, which is higher than the absolute humidity SAP value of 0.16 and the normalized value of daily temperature variation SAP value of 0.10. This indicates that temperature, as an absolute controlling factor, plays a dominant role in triggering migration behavior.

[0075] Furthermore, such as Figure 5 As shown, once aggregation occurs, the cruising altitude is primarily determined by the relative horizontal wind speed (SHAP value = 0.40), a much greater influence than temperature (SHAP value = 0.14). This result indicates that the mechanisms of action of meteorological factors differ significantly at different migration stages.

[0076] As an optional embodiment, the method may further include: introducing atmospheric electric field meter data to obtain time-series data on atmospheric electric field intensity and polarity changes in the target airspace; spatiotemporally aligning and fusing the time-series data on atmospheric electric field intensity and polarity changes with a multidimensional meteorological feature heatmap; extracting electric field risk feature vectors through a lightweight neural network encoder; and fusing the electric field risk feature vectors and enhancement feature vectors in the early or mid-term stages of forward inference in the XGBoost prediction model, using them together as model inputs.

[0077] In this step, an atmospheric electric field meter is deployed at the same location as the observation station to continuously monitor the temporal changes in the intensity and polarity of the atmospheric electric field in the target airspace. The raw electric field data undergoes preprocessing such as denoising and filtering, and the timestamps are standardized to the same reference as radar and meteorological data. The preprocessed temporal data of the atmospheric electric field is synchronized with a spatiotemporal distribution heatmap or a multidimensional meteorological feature heatmap. Since the electric field meter typically provides ground or single-point vertical electric field data, it needs to be treated as an environmental risk signal that is weakly correlated with altitude but changes significantly over time, and correlated with each time slice of the heatmap. A lightweight neural network encoder, such as a one-dimensional convolutional neural network or a gated recurrent unit network, is designed specifically for processing the temporal sequence of atmospheric electric field intensity. This encoder can automatically learn and extract abstract features representing risk patterns such as abrupt changes and persistent anomalies in the sequence, outputting a fixed-dimensional electric field risk feature vector.

[0078] During forward inference in the XGBoost prediction model, the extracted electric field risk feature vector is fused with the original enhanced feature vector. This fusion can be achieved through early fusion, where the electric field risk feature vector is directly concatenated with the enhanced feature vector before being input into the model, forming a longer joint feature vector. Alternatively, mid-term fusion can be used, employing a dual-branch structure where the meteorological enhanced features and electric field risk features first undergo preliminary interaction and dimensionality reduction through different shallow networks or fully connected layers. The processed feature vectors are then concatenated and finally input into the main body of the XGBoost model.

[0079] In this way, through this fusion, the model can learn the relationship between meteorological behavior and perceive and utilize abnormal signals of the atmospheric electric field. It may learn that under specific temperature, humidity and wind conditions, if there is a sharp increase in the negative polarity of the electric field, it may indicate the development of convection, and the probability of insect stratification will be significantly reduced or dissipated; conversely, under a stable electric field background, stratification behavior is more likely to be maintained.

[0080] As an optional embodiment, the method may further include: constructing an insect and path knowledge graph, the graph including insect flight paths, source and sink information, and historical stratification heights in different seasons; after the XGBoost prediction model outputs prediction results, matching the prediction results with the knowledge graph to obtain the prior distribution of species predisposition and behavior under the given spatiotemporal conditions; and using a Bayesian correction framework, fusing the prediction results output by the model as the likelihood and the prior distribution of species predisposition and behavior provided by the knowledge graph as the prior probability to obtain the posterior prediction result.

[0081] In this step, to integrate long-accumulated insect ecology knowledge into the data-driven model and improve prediction, a structured insect migration knowledge graph is constructed based on historical documents, observation reports, and long-term monitoring data. This graph uses spatiotemporal and species dimensions to show the typical migration periods and peak periods of different insect species in different geographical regions and seasons; the known or inferred major migration corridors, the spatial relationships between origin and destination; and the statistical characteristics such as the typical stratification height range and intensity level observed in different regions and under different meteorological backgrounds throughout history.

[0082] When the XGBoost prediction model outputs a preliminary prediction result for a specific spatiotemporal region, i.e., a likelihood distribution, the system automatically uses the predicted spatiotemporal parameters—season, month, geographic coordinates, and information on possible target insect species—as query conditions. This data is then matched against a constructed insect migration knowledge graph to obtain prior knowledge that matches the spatiotemporal and species conditions. Using a Bayesian inference framework, the XGBoost model's prediction output is considered a likelihood distribution based on current meteorological data, while statistical information from the knowledge graph, reflecting historical patterns and ecological constraints, is used as the prior distribution. The Bayesian formula is then used to fuse the likelihood and prior information to obtain the posterior distribution. For stratification probability, the original probability output by the model can be calibrated based on the background probability provided by the knowledge graph, making it more accurate in the long term.

[0083] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0084] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0085] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0090] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0091] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0092] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0093] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method of predicting the layering behavior of insects, characterized by, The method comprises: acquiring phased array insect radar echo data of a target airspace, and vertical profile meteorological data collected by a laser wind radar and a microwave radiometer; generating a spatiotemporal distribution heat map based on the echo data; mapping the vertical profile meteorological data to the spatiotemporal distribution heat map by a spatiotemporal interpolation algorithm to obtain a multi-dimensional meteorological feature heat map aligned in space and time; the XGBoost model comprises a global screening model based on low-altitude enhanced features and a local fine-tuning model activated by the global screening model, which are used to label the multi-dimensional meteorological feature heat map to obtain a sample data set with labels, and the XGBoost prediction model is trained through the sample data set, specifically comprising: dividing the sample data set into a training set and a validation set according to spatiotemporal blocks; training the global screening model using the training set, which takes the low-altitude layer enhanced features as input and outputs a first stratification occurrence probability; screening the training set with a first stratification occurrence probability higher than an activation threshold for training the local fine-tuning model, which takes full vertical profile enhanced features as input and outputs a second stratification occurrence probability, as well as a height regression value and an intensity regression value; wherein, in the training of the local fine-tuning model, spatiotemporal smoothness constraints and flight height physical boundary constraints are introduced into the objective function to regularize model prediction; using the trained XGBoost prediction model to perform forward inference on real-time meteorological data to obtain a prediction result, which includes a stratification occurrence probability, a predicted height, and an estimated intensity.

2. The method of claim 1, wherein the method is for predicting the formation of a layer of insects. The mapping of the vertical profile meteorological data to the spatiotemporal distribution heat map by the spatiotemporal interpolation algorithm to obtain a multi-dimensional meteorological feature heat map aligned in space and time specifically comprises: performing bilinear interpolation on the vertical profile meteorological data according to the timestamp and height coordinates of each pixel unit in the spatiotemporal distribution heat map to obtain wind speed vectors, temperature values, and humidity values at the same height coordinates; combining the wind speed vectors, temperature values, and humidity values to form a basic meteorological feature vector of the current pixel unit.

3. The method of claim 2, wherein the method is for predicting the formation of a layer of insects. The method further comprises: calculating the temperature vertical decrement rate and the relative humidity vertical gradient according to the timestamp and height coordinates of each pixel unit; calculating the vertical wind shear intensity and direction based on the wind speed vectors of adjacent height pixel units, and calculating the temperature-humidity combination index in combination with the ambient temperature and the saturated water vapor pressure; concatenating the temperature vertical decrement rate, humidity gradient, wind shear intensity, and temperature-humidity combination index with the basic meteorological feature vector to obtain an enhanced feature vector, which includes low-altitude enhanced features and full vertical profile enhanced features.

4. The method of claim 1, wherein the method is a method of predicting the formation of a layer in a group of insects, characterized in that, The method further comprises: defining a hyperparameter search space for the global screening model and the local fine-tuning model, respectively, the hyperparameters including the maximum depth of the tree, the learning rate, and the subsampling ratio; using the Bayesian optimization method to automatically optimize in the search space with the average performance indicators of spatiotemporal block cross-validation as the optimization target.

5. The method of claim 1, wherein the method is for predicting the formation of a layer in a group of insects. The method further comprises: presetting multi-layer threshold values of occurrence probability, predicted intensity, and uncertainty; dividing early warning levels and generating early warning information according to the matching relationship between the prediction result and the multi-layer threshold values, wherein the early warning information comprises a warning area, an expected height, a dominant meteorological factor, and a confidence level.

6. The method of claim 5, wherein the method is for predicting the formation of a layer of insects. The method further comprises: performing temporal slicing and spatial clustering analysis on the spatiotemporal distribution heat map to identify candidate insect swarm areas; calculating the vertical profile shape index and multi-frame temporal stability index of the candidate insect swarm areas, and classifying the candidate insect swarm areas into discrete noise, diffused insect swarm, or stable stratified insect swarm according to preset shape and stability discrimination rules; 7. The method of claim 1, wherein the method is a method of predicting the formation of a layer in a group of insects, characterized by, labeling the pixel units in the stable stratified insect swarm area with a class label indicating the occurrence of stratification, and recording the centroid height and average density attribute labels. The method further comprises: introducing atmospheric electric field instrument data to obtain atmospheric electric field intensity and polarity change time series data of the target space; spatiotemporally aligning and fusing the atmospheric electric field intensity and polarity change time series data with the multi-dimensional meteorological feature heat map, and extracting an electric field risk feature vector through a lightweight neural network encoder; 8. The method of claim 3, wherein the method is characterized by, early or intermediate fusing the electric field risk feature vector and the enhanced feature vector as model inputs when the XGBoost prediction model performs forward inference. The method further comprises: constructing the insect and path knowledge graph, wherein the graph comprises the flight path of insects in different seasons, source and sink area information, and historical stratification height; after the XGBoost prediction model outputs the prediction result, performing matching query on the prediction result and the knowledge graph to obtain the species tendency and behavior prior distribution under the spatiotemporal condition; 9. The method of claim 1, wherein the method is a method of predicting the formation of a layer in a group of insects, characterized in that, through a Bayesian correction framework, taking the prediction result output by the model as a likelihood and taking the species tendency and behavior prior distribution provided by the knowledge graph as a prior probability to perform fusion calculation to obtain a posterior prediction result. ​ ​ ​

Images