Insect inhabitation density prediction method

Abstract

The invention relates to an insect inhabitation density prediction method. Comprising the steps that a research area is divided into a plurality of micro-inhabitation units MHU, and the division granularity is based on the activity radius of a target insect; introducing a density-driven causal model of a behavior and climate coupling map; aiming at specific insects such as aphids and pine moths, constructing a day and night behavior rhythm map and a climate response preference curve; extracting a high-weight path in the coupling map, and deducing a causal path of a density lifting chain: a certain climate trend-certain behavior enhancement-inhabitation probability increase-density increase; performing spatial diffusion prediction by adopting a density transfer mechanism; defining an ecological communication weight between the MHUs, and outputting insect inhabitation density distribution at a future time point; time sequence uncertainty evaluation of a density fluctuation confidence interval is introduced; performing multi-step prediction on the inhabitation density of each MHU by using a time sequence model, and introducing uncertainty by using a Dropout regularization depth model; and outputting a confidence interval of a predicted value, and correcting a confidence boundary in combination with real-time information of trapping data and remote sensing data.

Description

Technical Field

[0001] The invention relates to a method for predicting insect habitat density. Background Art

[0002] There are currently many methods for predicting insect habitat density, including common empirical statistical methods, grid overlay methods, meteorologically driven distribution models, species distribution models (SDMs), and traditional neural network prediction models. Although these methods have achieved density trend estimation in local areas or specific time periods to a certain extent, they still have many shortcomings and obvious drawbacks in their application in actual complex ecosystems, which restricts their further development in terms of refined prediction, dynamic evolution simulation, and ecological mechanism restoration.

[0003] First, traditional empirical statistical methods are highly dependent on regression relationship modeling of historical observation data, and often use linear, polynomial, ridge regression and other means to fit meteorological factors and habitat density. Although this type of model is simple to construct and easy to adjust parameters, the biggest problem is that it lacks the characterization of the ecological behavior laws of insects themselves, and cannot reflect the causal logic between "behavior-climate-density". The prediction results tend to be static and limited to the distribution of existing data. It lacks the ability to respond to emergencies or nonlinear ecological changes. In the face of situations such as drastic changes in day and night temperature differences and sudden changes in micro-scale shading conditions, the prediction results fluctuate greatly and have poor stability. Secondly, common grid overlay or spatial distribution methods are usually based on remote sensing data or DEM information, dividing the area into equidistant grids, and constructing ecological suitability scores based on factors such as temperature, humidity, and vegetation, and then superimposing weighted prediction density hotspots. The essential problems of this type of method lie in "simplification of ecological heterogeneity" and "absence of biological behavior", that is, the actual activity radius and behavioral scale of insects cannot be simulated during grid division, resulting in multiple areas with completely different ecological structures being forcibly merged into the same analysis unit, seriously affecting the accuracy. At the same time, insect migration, habitat, stay and other behaviors are not included in the modeling system, and "high density" is only inferred based on "environmental suitability", which lacks ecological mechanism support.

[0004] Furthermore, species distribution models driven by meteorology (such as MaxEnt or CLIMEX, etc.) are mainly used to predict the potential distribution range of a certain insect in a large area. The model logic pays more attention to "where the species may appear" rather than "how it will spread" or "when the habitat density will increase". Therefore, its spatial accuracy and temporal continuity cannot meet the needs of short-term density evolution. Moreover, this type of model highly relies on long-term average climate background data and has extremely weak performance in the case of short-term climate variations and sudden weather conditions, making it unable to conduct hourly-level dynamic simulations. In addition, although traditional neural network prediction models such as BP neural network and support vector regression (SVR) have certain data fitting capabilities, they often exist as "black box models" with poor model interpretability and are difficult to reflect the ecological path of "why the density increases and how the behavior changes", lacking operational guiding significance in actual ecological governance or field management. Moreover, most of these models cannot effectively handle spatial structure relationships, especially lacking the ability to describe ecological connectivity, migration paths, and boundary resistance in the simulation of spatial diffusion processes. As a result, the results are often limited to point predictions of a certain unit and are difficult to support regional-level or path-level control strategies. In addition, most existing models do not consider "the impact of behavioral rhythm changes on density drive". The activities of insects such as foraging, mating, and migration have typical diurnal periodicity and climate condition dependence, but these high-frequency change factors are simplified or even ignored in most traditional models, resulting in the inability to utilize the "precursor signal" where behavior changes precede density changes, causing prediction delays and response lags. At the same time, existing methods generally lack a modeling mechanism for ecological uncertainty factors, do not introduce perturbation correction or confidence interval expression, only output a single value, lack a risk analysis dimension, and cannot distinguish which areas have reliable prediction results and which areas have observation blind spots in actual deployment. Summary of the Invention

[0005] The object of the present invention is to provide a method for predicting the habitat density of insects, so as to solve some of the drawbacks and deficiencies pointed out in the background technology.

[0006] The following technical solutions are adopted by the present invention to solve its above-mentioned technical problems: A method for predicting the habitat density of insects, including:

[0007] S1. Fine-grained structured modeling of the habitat environment using micro-habitat unit division: Divide the research area into multiple micro-habitat units (MHUs), and the division granularity is based on the activity radius of the target insect; collect environmental factors for each MHU and retain the internal gradient information of the factors including high humidity in the southeast corner and low humidity in the north side to construct a spatial gradient tensor;

[0008] S2. Introduce a density-driven causal model of the behavior and climate coupling map:

[0009] S2.1. Construct the circadian behavior rhythm map and climate response preference curve for specific insects including aphids and pine caterpillars; use the graph neural network GNN to map the climate and behavior driving relationships between MHUs, with nodes as units and edges as the association strengths of climate and behavior factors;

[0010] S2.2. Extract the high-weight paths in the coupled map and deduce the density increase chain: the causal path of a certain climate trend → a certain behavior enhancement → an increase in habitat probability → density growth;

[0011] S3. Adopt spatial diffusion prediction with a density transfer mechanism:

[0012] S3.1. Define the ecological connectivity weights between each MHU, indicating whether the insects are predicted to migrate based on vegetation continuity and temperature corridors;

[0013] S3.2. Construct a spatial diffusion convolutional layer to model the density values at each moment in two directions of active diffusion and passive diffusion: Active diffusion: Simulate the density flowing out from highlands based on the migration intention of insects; Passive diffusion: Simulate the attraction of neighboring areas to the habitat density based on environmental adaptability;

[0014] S3.3. Output the insect habitat density distribution at future time points;

[0015] S4. Introduce the temporal uncertainty assessment of the density fluctuation confidence interval:

[0016] S4.1. Make multi-step predictions on the habitat density of each MHU using a time series model, and introduce uncertainty using a Dropout regularized deep model;

[0017] S4.2. Output the confidence interval of the predicted values, distinguishing: normal fluctuation interval, abnormal high density warning interval, data confidence decline area; Combine trap data and remote sensing data real-time information to correct the confidence boundary.

[0018] Furthermore, the method for constructing the density-driven causal model of the behavior and climate coupling map:

[0019] Based on the ecological characteristics of the target insect species, construct a circadian behavior rhythm map. The rhythm map observes the activity frequency of insects within a 24-hour cycle and combines the behavior response characteristics under different climate conditions to establish a dynamic behavior response surface, which is used to reflect the change in behavior activity driven by time + climate two-dimensional; The rhythm map is described by the following formula:

[0020]

[0021] Where:

[0022] B s(t, ω) represents the behavioral intensity value of insect species s at time t and climate variable ω; α s is the basic activity coefficient of the behavior of a species of insect in the suitable ecological state; sin(πt / 12) is used to simulate the typical diurnal activity cycle; tanh(ω - ω0) represents the non-linear enhancement or inhibition response of insect behavior under climate offset, where ω0 is the climate threshold at which the behavior is most sensitive, and η is the climate impact amplification coefficient, reflecting the amplification or buffering effect of climate change on the degree of behavioral response.

[0023] Furthermore, the method for constructing the density-driven causal model of the behavior-climate coupling map:

[0024] The study area is divided into multiple micro-habitat units (MHUs), and each unit represents a locally complete habitat space in terms of ecology; in the modeling, each MHU is regarded as a single node in the graph neural network, and the information contained in the node includes: the micro-climate state of the area, the historical habitat density record, the ecological structure attributes, and the behavioral response state calculated through the rhythm map; the connection between each node is modeled by edge weights to reflect the coupling strength between adjacent units in terms of the linkage between the ecological environment and behavior. The edge weight is defined as follows:

[0025] W ij = δ·exp(-λ·|Ψ i - Ψ j |)

[0026] Where:

[0027] W ij represents the behavioral coupling weight between the i-th and j-th micro-habitat units, indicating whether the behavioral influence can be effectively transmitted between the two units; δ is the weight scaling coefficient, adjusting the overall edge connection strength; λ is the response difference sensitivity factor, regulating the attenuation degree of the edge weight due to the difference in behavioral states; |Ψ i - Ψ j | represents the distance difference between the two units on the current behavioral response tensor, which is jointly composed of the behavioral intensity, direction, and rhythm pattern of the node under the drive of climate change, and serves as the characteristic distance in the graph structure.

[0028] Furthermore, the method for constructing the density-driven causal model of the behavior-climate coupling map:

[0029] Perform density propagation and prediction on the graph structure, and identify and extract the causal chain path with a significant increase in habitat density through the evolution trend of the state of each node in the path over time; the density change trend function is defined as:

[0030]

[0031] Where:

[0032] D(P) represents the cumulative change contribution value of density along the path P = {v1, v2, …, v n}, reflecting the density increase potential on the path; Ω k is the ecological suitability factor of the k-th node, indicating whether insects tend to inhabit this unit under the current climate and environmental conditions; At time t k the behavior change rate in the rhythm map, representing the trend change of behavior activity; φ(W k,k+1 ) is the mapping function of the edge weight W, used to adjust the influence weight of the coupling strength in the density propagation path; By identifying the path with the D(P) value in the graph, the logical chain of climate trend → behavior enhancement → increased habitat probability → density growth is extracted, enabling density prediction with causal interpretation ability.

[0033] Furthermore, the spatial diffusion prediction method of the density transfer mechanism includes:

[0034] Define multiple micro-habitat units MHU in the study area, and construct the ecological connectivity weight value between every two units according to the ecological function elements, which is used to characterize the feasibility basis of insect population migration. The connectivity takes the ecological path as the actual coherence, including: whether there is a continuous vegetation structure for passage between regions, whether there is a locally suitable temperature channel, and the airflow compliance and shelter coverage conditions. The following connectivity weight function is comprehensively constructed:

[0035] L ij = ξ·[χ(v i , v j ) + τ·κ(θ i , θ j )] ρ

[0036] where:

[0037] L ij represents the ecological connectivity weight between the micro-habitat unit i and the unit j; ξ is the ecological accessibility adjustment coefficient, controlling the overall connectivity intensity; χ(v i , v j ) is the vegetation continuity function, measuring the vegetation cover similarity and connection smoothness between the units i and j; τ is the climate factor weight, used to adjust the contribution ratio of the temperature corridor to the connectivity; κ(θ i , θ j ) is the temperature adaptation function, measuring the microclimate compatibility between the two units, and θ i , θ j are the representative temperature indicators of the two units respectively; ρ is the ecological coupling index, used to control the non-linear enhancement effect.

[0038] Further, the spatial diffusion prediction method of the density transfer mechanism includes: constructing a spatial diffusion mechanism to simulate the dynamic changes in the mobility of insect populations in the ecological network. The diffusion mechanism consists of active diffusion and passive diffusion: the active diffusion is the spontaneous overflow phenomenon that occurs when the population density reaches the upper limit of the ecological capacity, driven by the local density difference, and the diffusion intensity can be expressed as:

[0039]

[0040] Where:

[0041] represents the active diffusion output amount occurring at the i-th unit at time point t; μ is the diffusion efficiency factor, controlling the upper limit ratio of the density overflow speed; D i (t) is the insect habitat density of unit i at the current moment; λ i is the ecological carrying capacity upper limit of unit i, indicating that when the density exceeds this value, overcompetition may occur, leading to individual migration; max[D i (t) - λ i , 0] indicates that diffusion will only be triggered when the density exceeds the ecological capacity, otherwise it is zero; ∑ j∈N(i) L ij represents the total connection weight sum from unit i to all its connected neighbor units j;

[0042] The passive diffusion is based on the density import effect caused by the sudden suitability of the ecological environment in the neighboring area, attracting insects to inhabit. The modeling method is expressed according to the habitat attraction gradient in the neighboring area as follows:

[0043]

[0044] Where:

[0045] represents the passive diffusion input amount of unit j at time point t; σ is the passive diffusion attraction coefficient, controlling the drainage intensity of the neighboring area to the target area; ∑ i∈N(j) represents the set of all units adjacent to unit j; L ij is the ecological connection weight; 1 - represents the difference degree between the current density and the ecological maximum suitability ζ j ; α ij (t) is the habitat attraction function from neighboring unit i to target unit j, depending on the difference between the thrust of the current environment of i and the pull of the environment of j.

[0046] Further, the spatial diffusion prediction method of the density transfer mechanism includes:

[0047] Overlay the results of active and passive diffusion, and combine the ecological driving variables of climate, light, and wind direction in the historical time series to construct a sequence of response functions for density evolution to predict the trend of habitat density distribution at multiple future time steps. The density update function is expressed as:

[0048]

[0049] Where:

[0050] D j (t + 1) represents the predicted density of cell j at the next time step t + 1; D j (t) is the actual density at the current moment; Δt is the time advancement step size used to control the prediction granularity; are the density input and output amounts caused by passive attraction and active diffusion respectively; ∈ j (t) is the perturbation correction term used to simulate the perturbation effect of random or unobserved ecological factors on density changes.

[0051] The prediction method of insect habitat density proposed by the present invention has the technical advantages of multi-dimensional integration, high interpretability, and dynamic adaptability, and can achieve refined, spatio-temporal, and visual prediction of the habitat density of the target insect population in natural or agroforestry systems with complex ecology and strong spatial heterogeneity. The specific beneficial effects include the following aspects:

[0052] Incorporate the diurnal behavioral rhythm and climate response preference of insects into the density prediction system. By constructing a "behavior-climate coupling map", the density prediction is transformed from a traditional statistical fitting model to the identification of a "behavior-environment-density" causal path with biological interpretability, improving the ecological rationality and decision traceability of the model;

[0053] By dividing the study area based on the activity radius of the target insect and retaining the environmental gradients and spatial heterogeneity characteristics within each microcell, the over-simplification of ecological information by traditional grid methods is effectively avoided, greatly enhancing the simulation ability of the actual perception scale of insects, and significantly improving the prediction accuracy; distinguish and model the migration mechanisms of active diffusion (based on density pressure) and passive diffusion (based on ecological attraction), and dynamically adjust the diffusion flux in combination with the ecological connectivity weight, which can truly reproduce the population diffusion process of insects under complex terrain and climate conditions, identify density aggregation points and propagation paths, and provide spatial support for precise intervention; by constructing a density update response function and adding external driving variables such as climate, wind direction, and micro-ecology and perturbation terms, the present invention can achieve density prediction at the hourly level or even higher frequencies, effectively coping with the drastic density fluctuations caused by sudden ecological perturbations or environmental mutations, and improving the timeliness and stability of the system.

[0054] By embedding an uncertainty mechanism into the deep learning model, the density confidence intervals for multiple future time periods of each unit are generated, realizing the leap from "predicted value" to "prediction bandwidth". This enables managers to not only know "where the density will increase", but also know "the credibility of the change", significantly improving the practicality of the system in plant protection early warning and pest outbreak prediction. It is applicable to multiple ecosystems (such as forests, farmlands, wetlands, urban green belts) and multiple types of insects (such as piercing-sucking insects, lepidopterans, hymenopterans, etc.), can be seamlessly docked with multi-source information such as remote sensing images, sensor networks, and trapping data, and has good potential for cross-regional promotion and platform deployment capabilities. Brief Description of the Drawings

[0055] Figure 1 This is a flowchart of the prediction method for the insect habitat density of the present invention.

[0056] Figure 2 This is a flowchart of the method for constructing the density-driven causal model of the behavior-climate coupling map of the present invention.

[0057] Figure 3 This is a flowchart of the spatial diffusion prediction method for the density transfer mechanism of the present invention. Detailed Embodiments

[0058] The following provides a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0059] Refer to the appendix Figure 1, in the method for predicting the insect habitat density of the present invention, step S1 takes the fine-grained structured modeling of the habitat environment divided by micro-habitat units as the core content. Its main purpose is to reconstruct the spatial structure with behavioral ecological significance for the research area starting from the true perception scale of the ecological space. Specifically, first, according to the activity radius of the target insect - that is, the maximum spatial range it can complete foraging, reproduction, or migration within a unit of time - the entire research area is divided into multiple micro-habitat units (MicroHabitatUnit, abbreviated as MHU). For example, the activity range of aphids is in units of 2 meters, while that of pine caterpillars can extend to 10 meters or more. Therefore, the MHU division granularity must be set differently according to the insect species, ecological habits, and life cycle characteristics. After the division, environmental factors are collected for each MHU. This process is not limited to obtaining single values of conventional parameters such as temperature, humidity, light, soil moisture, etc., but further delves into the spatial difference dimension within each unit. Specifically, different from the previous way of simplifying the modeling by taking the average or median of the whole MHU, this method retains the spatial gradient information within the MHU, that is, the difference in environmental factors at different positions within a unit. For example, for the humidity factor, if the humidity is higher in the southeast corner and lower in the north of a certain MHU, this information will be recorded as it is and encoded into a three-dimensional spatial gradient tensor, enabling the model to "see" the micro-scale ecological heterogeneity within the unit. This tensor is usually represented by a multi-channel matrix composed of spatial coordinates (such as x, y) and environmental factor values, which is the basic data structure for subsequent construction of the ecological suitability distribution map, identification of behavioral hotspots, and establishment of the behavior-environment response model. At the same time, it also lays a modeling foundation with extremely high spatial resolution for the accurate prediction of insect habitat density, avoiding meaningless compression or pattern loss of information in the data input stage, thereby enhancing the interpretability and spatial migration ability of the prediction.

[0060] Step S2 involves introducing a density-driven causal model of the behavior and climate coupling map. Its core is to establish an explanatory mechanism that reflects how climate change drives changes in insect behavior and thus affects habitat density from the perspective of combining ecology and data structure. Specifically, in the S2.1 stage, first, for different types of target insect populations, such as aphids and pine caterpillars, based on field observations, literature data or intelligent sensing equipment, collect their behavioral activity data during the diurnal cycle, including the temporal distribution patterns of behaviors such as foraging, perching, mating, and flying within a day, and then construct their own diurnal behavioral rhythm map to characterize the changing trends of insect activity under natural rhythms; at the same time, statistically analyze the behavioral response intensity of insects under different climatic conditions (such as increased temperature, decreased humidity, extended daylight, etc.), and extract their climate response preference curve from it, that is, reveal at what environmental threshold the behavior of insects will be significantly activated or inhibited. Next, at the modeling level, the aforementioned microhabitat units (MHUs) are used as nodes in a graph neural network (GNN) to construct a graph structure. Each node in the graph represents a spatial unit and embeds characteristics such as its current climate state and the intensity of rhythmic behavior. Edges between nodes connect adjacent or ecologically coupled units. Edge weights represent the strength of the correlation between climatic changes and behavioral linkages. For example, when rising temperatures lead to a simultaneous increase in foraging behavior in adjacent units, the system assigns a higher weight to that edge, reflecting the actual driving force behind the spread of ecological behavior. In stage S2.2, the trained GNN model is used to analyze information propagation and graph structure. The system identifies highly weighted paths in the graph, often representing the main channels for rapid diffusion or concentrated changes in insect density. Within these paths, the system can clearly extract a logically complete chain of density increases: A certain climate trend first leads to increased insect behavior (e.g., increased foraging frequency, longer roosting time), which in turn increases the ecological suitability of the unit, triggering the influx or retention of more individuals, ultimately manifesting as an increase in habitat density. This causal chain not only provides the ability to predict the density evolution process, but also provides a basis for explaining the prediction results from the perspective of ecological behavior. It enables the system to no longer rely solely on data fitting results to make black-box conclusions, but can explain the traceable process of "why the density increases and which ecological events drive it". It significantly improves the scientific nature of the density prediction model, the rationality of ecological logic and the operability of actual intervention.

[0061] Step S3 is the spatial diffusion prediction using the density transfer mechanism. Its main purpose is to simulate the density dynamic flow of insects between different microhabitat units due to behavior-driven or environmental incentives, so as to achieve a reasonable prediction of the future spatio-temporal density distribution. In S3.1, first, the ecological connectivity weight between each pair of microhabitat units (MHU) is defined. This weight is no longer simply based on geographical distance or grid proximity, but the concept of ecological function accessibility is introduced, emphasizing the practical feasibility of insects to migrate ecologically from one unit to another. This feasibility is mainly determined by two core ecological factors: one is vegetation continuity, that is, whether there is a connected forest belt, grassland, crop area or other continuous vegetation structures for insects to cross or inhabit between the two units; the other is the temperature corridor effect, that is, whether there is a suitable temperature channel that insects can perceive in time or space, enabling them to carry out small-scale migrations without excessive energy consumption or risk interference in different seasons or diurnal cycles. Therefore, this connectivity weight not only represents the existence of a spatial channel, but more importantly represents ecological traversability and survival suitability, which is a prerequisite for the active or passive transfer of insects. In S3.2, based on the above ecological connectivity structure, a spatial diffusion convolutional layer is constructed. In this layer, the dynamic propagation calculation of the habitat density at each moment is carried out, and the modeling mechanisms in two directions of active diffusion and passive diffusion are introduced. Active diffusion refers to the overflow process formed by the behavior intention of insects themselves. That is, when the insect habitat density in a certain MHU is higher than the ecological carrying capacity threshold, or due to the saturation of foraging resources and the enhancement of internal competition, some individuals will actively move out and flow to adjacent areas. This kind of flow follows the direction of the density gradient and diffuses from high to low; while passive diffusion reflects the attraction of neighboring areas to insect individuals. That is, in some units, due to factors such as suddenly suitable climate conditions, resource recovery, and enhanced shelter, an ecological attraction effect is formed, thus attracting insects from surrounding areas to move in. This mechanism is often directly proportional to the suitability degree of the target unit and inversely proportional to the current density state. The system will comprehensively integrate the quantitative results of active output and passive input at each time step to update the density state in each MHU. Finally, in S3.3, through continuous time series simulation, the system outputs the insect habitat density distribution maps at multiple future time points, presenting the trend trajectory of density evolution with time and space diffusion, so as to identify key changes such as which areas will have density concentration, which areas are in the density growth channel, and which areas form ecological barriers to block the diffusion path, providing forward-looking prediction basis and spatial decision-making support for insect population management, pest control, and ecological intervention planning.

[0062] Step S4 is the time-series uncertainty assessment that introduces the confidence interval of density fluctuations. Its core lies in solving the problem that traditional prediction models only give deterministic values and cannot reflect the prediction reliability, enabling the system to have the ability to quantify its own prediction credibility and identify abnormal situations while making habitat density predictions. In S4.1, first, based on the historical habitat density time-series data of each micro-habitat unit (MHU), a multi-step prediction model is constructed to infer the density change trend in the future for a period of time. To effectively introduce uncertainty factors and simulate the prediction errors caused by observation noise, model structure limitations, or ecological process fluctuations in reality, this method uses a deep learning network structure with a Dropout regularization mechanism. In each forward prediction process, the model randomly discards some neuron connections, thereby obtaining multiple prediction sample distributions at the output layer. This method of approximate Bayesian inference can effectively simulate the range of changes in the model output values under the same input conditions, and then characterize the confidence boundary of the density prediction value. In S4.2, based on the above multiple sampling prediction results, the system calculates the confidence interval at each time point and classifies the density fluctuations in a structured manner: one is the normal fluctuation interval, which means that the predicted density value is within the expected change range of the ecological model, without significant increase or abnormal behavior, and belongs to the acceptable ecological natural change; the second is the abnormal high-density warning interval, that is, the upper confidence bound of the prediction result breaks through the historical density extreme value or the ecological safety threshold, indicating the occurrence of population outbreaks, migration tides, or environmental load risks; the third is the data confidence decline area, which means that the confidence interval becomes abnormally wide or fluctuates violently, indicating that the prediction stability in this area is poor, caused by data loss, sudden changes in ecological conditions, or insufficient response of the model to current features. The system spatially and graphically outputs the above three types of confidence analysis results, endowing the density prediction not only with a "result" but also with a "risk state". In addition, the system also supports comparing the confidence boundary in the prediction stage with external data such as the data of insect traps, remote sensing vegetation indices, and climate sensors collected in real time. If the observed value deviates from the prediction interval, the system will dynamically correct the distribution judgment and confidence range of the current model, thereby forming a density prediction system with self-assessment and self-correction capabilities, greatly improving its adaptability and reliability under complex ecological conditions.

[0063] Example 1:

[0064] Combined with Figure 2 , in this example, in a mixed forest and agricultural ecological zone in the eastern hilly area, the research team hopes to carry out the prediction work on the evolution trend of the habitat density of Dendrolimus punctatus during the high-incidence period in spring based on the insect habitat density prediction method proposed in the present invention. The terrain of this area is complex, with a large temperature difference between morning and evening and a drastic change in humidity during the day and night. There are diverse vegetation types, which is a typical area where Dendrolimus punctatus adults are active. The researchers deployed 64 micro-habitat units (MHUs) in the area, and each unit is about 25m2 , the division granularity is set according to the average activity radius of Dendrolimus punctatus in the adult stage, and climate factors such as activity frequency, temperature, and humidity at different time periods within 24 hours (once per hour) are collected, as well as the behavioral response data recorded by the trap. First, the researchers constructed the diurnal behavioral rhythm of Dendrolimus punctatus based on the entry and exit frequencies recorded by the trap at different time points and found that its activities were mainly concentrated between 18:00 and 04:00, showing significant nocturnal characteristics. Through statistics, the team determined that the peak of the behavioral frequency during its most active period was approximately 42 entries and exits per hour, while it was almost zero during the day. In the climate response experiment, the researchers found that when the night temperature was around 22 °C, the behavioral activity was the strongest; when the temperature dropped below 15 °C or rose above 30 °C, the activity decreased significantly; combined with the analysis of the humidity factor, the range of 60%-75% was the most suitable humidity. Therefore, the researchers substituted this behavior-climate relationship into the following formula to establish a rhythm map:

[0065]

[0066] The set parameters are:

[0067] α s = 40: representing the peak of the basic behavioral intensity per hour of Dendrolimus punctatus in the ecologically optimal environment (corresponding to the trapping frequency),

[0068] η = 0.6: the climate response amplification coefficient, indicating that the climate deviation has a moderate amplification effect on behavioral changes,

[0069] ω0 = 22: the optimal air temperature is 22 degrees Celsius,

[0070] t ∈ [0, 24]: the time is measured in hours,

[0071] ω: the real-time monitored air temperature (unit: °C), provided daily by a dynamic sensor.

[0072] Taking the time of 21:00 at night (i.e., t = 21) and the air temperature ω = 25 °C as an example, substituting into the formula for calculation is as follows:

[0073]

[0074] tanh(ω - ω0) = tanh(3) ≈ 0.995;

[0075] B s (21, 25) = 40·(-0.707)·(1 + 0.6·0.995) = 40·(-0.707)·(1.597) ≈ -45.17;

[0076] Since the behavior intensity value is negative, it indicates that although it is in the nocturnal active period at this moment, the temperature is relatively high, and the behavior curve shows a decline. Further compare the behavior values at 20:00 (t = 20, air temperature 22°C), and the calculation is as follows:

[0077]

[0078] tanh(0) = 0, because the air temperature is exactly equal to the optimal threshold

[0079] So: B s (20,22) = 40·(-0.5)·(1 + 0) = -20

[0080] Although they belong to the same active time period, the degree of behavior decline at the optimal temperature is smaller, and the behavior response is more stable. This numerical change helps the system identify the "behavior peak" drifting back and forth with temperature fluctuations, providing a data base of real behavior drivers for subsequent graph neural structure mapping. Subsequently, the researchers formed a node feature matrix with the B s (t,ω) values of each MHU at the same time point, and calculated the climate-behavior coupling relationship between MHUs by combining the vegetation continuity map and the temperature and humidity flow data. A graph neural network was reconstructed through edge weight to form a spatio-temporal map. The system further extracted the behavior enhancement path propagating from a certain temperature corridor direction (such as the air temperature in the south slope area at 20:00 remains at 21.5°C) to the dense forest area in the downwind. It was found that the average behavior value in these areas increased by 27% between 21:00 and 01:00, and the corresponding trapping density increased from an average of 3 per hour to 6 per hour, and the roosting density doubled. Based on this, the system deduced a clear causal chain: the night temperature is between 21 and 23°C → significant behavior enhancement → improved habitat suitability on the south slope → increased immigration of individual groups → increased roosting density in the downwind dense forest MHU, thus not only realizing the prediction of the spatial density trend, but also providing an interpretable result of the ecological driving path. Through 3 days of comparative monitoring, the error between the system prediction and the actual trapping data was controlled within the range of ±8.3%, verifying the effectiveness, stability and decision-making support ability of the model in the actual complex ecological environment.

[0081] Based on the practice of predicting the habitat density of pine caterpillars in the aforementioned eastern hilly agroforestry mixed ecological zone, after the research team completed the modeling of the diurnal behavior rhythm map and the analysis of key behavioral climate factors, it further entered the stage of constructing the graph neural network structure. According to the density-driven causal model method of the behavior-climate coupling map of the present invention, the research team first divided the entire research area into 64 micro-habitat units (MHUs), each unit being a 5×5-meter square grid. In each unit, micro-environmental collection devices and trapping devices were arranged to collect micro-climate data such as temperature, humidity, light, wind speed, etc., as well as the trapping frequency per hour per day. Subsequently, the research team modeled each MHU as a node in the graph neural network, and the node feature vector Ψ i contains four main dimensions: one is the micro-climate state from real-time weather sensors (a 3D vector composed of air temperature, humidity, and wind speed, with units of degrees Celsius, relative humidity percentage, and m / s respectively); the second is the recorded habitat density within the past 5 days of the node, taking the average value and standard deviation (2D); the third is the current hourly behavior intensity B s (t,ω) value; the fourth is the surface structure feature vector of this area, including the vegetation index NDVI and the shelterability score (2D), totaling an 8D feature vector for the modeling input of each node. In the process of connecting and constructing the graph, the research team connected each MHU to its 8 neighboring units in the spatial 8 directions according to the ecological adjacency principle, and defined the edge weight according to the following formula:

[0082] W ij =δ·exp(-λ·|Ψ i -Ψ j |)

[0083] Among them, δ takes the value of 1, indicating that the edge weight intensity is dominated by the real behavior difference and does not require scaling adjustment; λ is the behavior response difference sensitivity factor, set to 0.35, which empirically shows that the exponential decay of the edge weight caused by each unit of behavior state difference is about 30%-40%; and |Ψ i -Ψ j | is the Euclidean distance between the 8D behavior-ecological state vectors of two nodes. Taking MHU numbers 12 and 13 as an example, they are respectively located on the north side of the hillside and the edge of the sunny slope. The former's current behavior intensity is 24.6, temperature is 21.8°C, humidity is 73%, wind speed is 1.1 m / s, the historical density average is 5.2 per hour, and NDVI is 0.68; the latter's behavior intensity is 30.4, temperature is 22.9°C, humidity is 70%, wind speed is 0.8 m / s, the historical density average is 6.8, and NDVI is 0.72. Substituting their feature vectors into the calculation, |Ψ 12 -Ψ 13 | is approximately 2.14, so there is:

[0084] W12,13 = 1·exp(-0.35·2.14) ≈ exp(-0.749) ≈ 0.473

[0085] This edge weight indicates that there is a moderately strong coupling relationship between MHU12 and 13 in terms of behavioral response and microclimate state, indicating that the behavioral influence of pine caterpillars between these two units has a certain potential for transmission. Especially during the enhanced nocturnal behavior period, the behavioral excitation of unit 13 is transmitted to unit 12 through environmental continuity, triggering a response in density change. On the other hand, when the same calculation is performed between MHU22 and 29, since one of the units is located in a sheltered valley area with significantly different temperature and humidity conditions, the difference in their behavioral vectors is 4.8, and the calculated edge weight is only 0.189, indicating that the transmission ability of the behavioral influence between them is extremely weak.

[0086] After the system completes the construction of node features and edge weight initialization for the entire graph, it can utilize the graph neural network training mechanism to perform multi-layer graph embedding propagation and feature fusion on the entire ecological coupling graph, enabling the model to not only learn the spatial behavior propagation trend in the current state but also capture the high-order association paths formed by the linkage between behavior and climate change. Thus, in subsequent density prediction, it can accurately infer which regions have their density increased due to the enhanced behavior of other regions, ultimately realizing the visualization tracking of the causal path of the density evolution chain, and generating a fine multi-temporal spatial density evolution prediction map in combination with the aforementioned rhythm response model, providing timely intervention points, path control nodes, and prediction basis for plant protection personnel. The core parameter values in the entire modeling process are all derived from regional measured data or ecological experience settings. Among them, the reasonable range of δ is [0.5, 1] for scaling the edge weight intensity, and the recommended value range of λ is between [0.2, 0.5] to ensure that the sensitivity to feature distance is neither overly amplified nor weakened. The feature difference value |Ψ i - Ψ j | usually falls within the range of [0, 6], forming a stable, sparse but diffusible graph structure in the ecological coupling graph structure. This overall modeling scheme fully demonstrates the applicability, computability, and ecological interpretability of the behavior-climate coupling causal graph in density prediction.

[0087] After the research team completes the modeling of the behavioral rhythm map and the construction of the graph neural network structure, it further applies the density-driven causal model of the behavior-climate coupling map proposed in the present invention for density propagation path analysis and evolution trend prediction. The core lies in identifying the entire process causal chain that starts from a certain ecological stimulus (such as temperature increase) and ultimately leads to a significant increase in the habitat density of a certain region. The team first conducts a time series tracking analysis on the node states in the graph structure. The node states mainly include the change trend of behavioral activity B s (t), ecological suitability Ω k and the edge weight φ(W k,k+1) Therefore, during a three - day continuous monitoring period, the status of 64 micro - habitat units in the figure was recorded at hourly time steps, with particular attention paid to the path segment P = {MHU12, MHU13, MHU18, MHU23, MHU29} that spreads from the south slope to the north woodland. The team used the following formula to calculate the cumulative density change trend function for this path:

[0088]

[0089] Among them, Ω k represents the ecological suitability factor of node k, which is comprehensively calculated by the researchers through the NDVI value, shading rate, and micro - climate deviation, and the value range is set to [0.3, 1.0]; in this path, MHU13 and MHU18 are located at the wind - mouth junction zone, with a relatively high NDVI and small humidity fluctuations. It is set that Ω 13 = 0.92, Ω 18 = 0.88, while MHU29 at the northern edge has relatively poor suitability, set as Ω 29 = 0.52. represents the change rate of the actogram behavior intensity at time t k . The team calculated its derivative based on the previously constructed B s (t, ω) curve, and the sample values are as follows: during the period from t = 20 to t = 21 (i.e., from 20:00 to 21:00), B s in MHU13 increased from 27.4 to 32.9, then while in MHU29 it only increased from 19.6 to 21.1, and the change rate is about 1.5, indicating that the behavior of the former increased more significantly. The edge weight W k,k+1 in the figure has been calculated in the previous stage. The mapping function φ(W) is set as a non - linear amplification function: φ(W)=log(1 + 5W), which can amplify the edge - weight range [0.1, 0.9] to [0.48, 1.79] to enhance the response of the edge weight to the density propagation effect. Taking the 5 nodes on path P as an example, the team substituted the values to calculate each segment: From MHU12 to 13: Ω 12 = 0.85, W = 0.64, φ(W)=log(1 + 5*0.64)=1.49, and the product is 6.05; From MHU13 to 18: Ω 13 = 0.92, W = 0.71, φ(W)=1.59, and the product is 8.04; From MHU18 to 23: Ω 18 = 0.88, W = 0.69, φ(W)=1.56, and the product is 5.63; From MHU23 to 29: Ω 23 = 0.65, W = 0.54, φ(W) = 1.40, and their product is 2.64; there is no successor node at the end of MHU29, so it no longer propagates. The final path density propagation value:

[0090] D(P) = 6.05 + 8.04 + 5.63 + 2.64 = 22.36

[0091] This result is recognized by the system as a path with high propagation potential (the set threshold value of D is 15). Based on this, the system gives a density warning prompt: the density of MHU29 will reach a density aggregation more than twice the regional average within 3 hours. It is recommended to deploy biological prevention and control points in the front section of its path (such as MHU18, 23) to weaken the propagation flux. By continuously comparing the predicted and measured trapping data, the system successfully gives an early warning of a sudden increase in the density of pine caterpillars on this path 2 hours in advance. The deviation between the actual trapping number and the predicted density by the system does not exceed ±7.9%, effectively verifying the actual response ability of the causal chain prediction model and the scientific nature of the spatial propagation mechanism modeling. The value range of the parameters in the model is as follows: ecological suitability factor Ω k ∈[0.3, 1.0], behavior derivative value edge weight value W ∈ [0.1, 0.9], the constant coefficient of the mapping function is set to 5, and the value range of φ(W) = log(1 + 5W) is approximately [0.48, 1.79]. This example further proves that the high-D path in the graph structure can not only be used as a main channel identification tool for behavior propagation, but also be used for multi-objective linkage support for density anomaly prediction, trapping deployment optimization, and ecological process interpretation.

[0092] Example 2:

[0093] Combined with the attached Figure 3 , in this example, after the research team completed the modeling of the behavior rhythm map and the density causal chain, they further deployed the spatial diffusion prediction method of the density transfer mechanism proposed in the present invention. The goal is to simulate how insect populations migrate and spread from one microhabitat unit to other regions under the conditions of climate change or enhanced group behavior, so as to identify potential high-density regions in advance. The first step of this solution is to construct an "ecological connectivity weight matrix" to evaluate the basis of the migration feasibility between any two microhabitat units (MHUs). The team continues to use the previously divided 6^{4} MHUs in the research area and obtains the vegetation index (NDVI) based on remote sensing images and drone multispectral data. At the same time, the surface temperature, wind speed, and humidity of each unit are recorded by miniature climate sensors. In the modeling, the team defines the ecological connectivity weight according to the following formula:

[0094] L ij = ξ·[χ(v i , v j ) + τ·κ(θ i , θ j )] ρ

[0095] The meanings of the variables are as follows: L ij represents the ecological migration feasibility from unit i to unit j, with a value range of [0, 1]; ξ is the ecological accessibility adjustment coefficient, set between 0.8 and 1.0, reflecting the overall regional accessibility, and the actual value is 0.9; χ(v i , v j ) is the vegetation continuity function, which calculates its coherence based on the NDVI difference between two units and is defined as 1 - |NDVI i - NDVI j |. When the NDVI similarity is high (i.e., the vegetation structures are the same), the value approaches 1; τ is the temperature factor adjustment weight, with a value range of [0.3, 0.7], used to balance the weight effects of terrain temperature continuity and vegetation structure in the migration path. In this example, it is set to 0.5; κ(θ i , θ j ) is the temperature adaptation function, defined as where T max = 32 °C, T min = 12 °C, that is, the upper and lower limits of the air temperature during the research period are set. Finally, ρ is the ecological coupling index, used to control the enhancement intensity of the overall non - linear effect. The actual set value is 1.2, so that the high - connectivity paths show stronger diffusion potential in the model.

[0096] Taking two key units MHU13 (sunny slope forest edge, NDVI = 0.74, temperature = 22.3 °C) and MHU18 (ridge top shrub belt, NDVI = 0.68, temperature = 21.2 °C) on the path as an example, χ(v 13 , v 18 ) = 1 - |0.74 - 0.68| = 0.94, κ(θ 13 , θ 18 ) = 1 - |22.3 - 21.2| / (32 - 12) = 1 - 1.1 / 20 = 0.945. Substitute into the calculation:

[0097] L 13,18 = 0.9·[0.9 + 0.5·0.945] 1.2 = 0.9·(0.9 + 0.4725)[[ID=4l]] 1.2 = 0.9·(1.4125) 1.2

[0098] In the calculation, (1.4125) 1.2 ≈1.570, then L 13,18≈0.9·1.570 = 1.413. However, since the maximum value of L is 1, this value is taken as 1, indicating that the ecological connectivity of this path is extremely high and it has complete migrability. On the contrary, between MHU13 and MHU51, it is calculated that: the NDVI difference is large (0.74 vs 0.52), the temperature deviation is 5.6 °C, χ = 0.78, κ = 0.72, and finally L 13,51 ≈0.62 is obtained, indicating that although the physical distance of this path is close, the ecological migration accessibility is moderately weak and not suitable for the rapid migration of high-density populations.

[0099] The research team further implemented the spatial diffusion prediction method of the density transfer mechanism proposed in the present invention, and achieved accurate modeling of the dynamic diffusion behavior of populations in the ecological network. The core of this method lies in simulating the spatial migration behavior of insect populations between microhabitat units (MHUs), and decomposing it into two mechanisms: active diffusion and passive diffusion.

[0100] The team first records the insect density D i (t) per hour in each MHU, and determines the ecological carrying capacity threshold λ of each unit through historical observations i . For example, in MHU18 (wet grassland in the forest), the trapping data shows that its daily maximum density can be stably maintained at 15 individuals / hour. Based on this, the research personnel set the upper limit of the ecological carrying capacity of this unit as λ 18 = 15. At 21:00 on April 11th, the trapping quantity in this unit rose to D 18 (t) = 18.3, significantly exceeding the carrying capacity threshold. The system determines that active diffusion has been triggered and uses the following model for calculation:

[0101]

[0102] Set μ = 0.6 as the active diffusion efficiency coefficient to control the maximum ratio of individuals emigrating. The recommended value range is [0.4, 0.8] to reflect the flow response driven by group behavior in the medium and high-density state; max[18.3 - 15, 0] = 3.3, which is the overloaded density; from the aforementioned graph structure, it is known that the sum of the ecological connectivity weights of MHU18 and its 5 adjacent units is ∑L ij = 3.96. Substituting it in, we get:

[0103]

[0104] This means that at the current moment, MHU18 will have individuals with a density equivalent to 7.83 units tending to emigrate, and its distribution will be jointly determined by the subsequent path connectivity structure and the ecological suitability of the target unit. At the same time, the system evaluates the passive diffusion process, that is, which areas attract surrounding insects due to a sudden improvement in ecological conditions. Taking MHU23 (open space at the forest edge) as an example, at the current moment D 23Ψ(t) = 5.1, while the ecological maximum suitability ζ of this unit 23 = 14.5, which is derived from its NDVI value (0.77), shading index (0.68) and the comprehensive score of temperature and humidity, and its surrounding adjacent units include 6 neighbors such as MHU18, 19, 28, etc. The system uses the following model to calculate passive diffusion:

[0105]

[0106] Set σ = 0.5 as the passive diffusion attraction coefficient to control the maximum attraction response of the target unit. The recommended range of this value is usually [0.3, 0.6]; the ratio of the current density to the maximum suitability is 5.1 / 14.5 = 0.352, and the remaining attraction space is 1 - 0.352 = 0.648; set the ecological connectivity weight L 18,23 = 0.81 between MHU18 and the attraction function α 18,23 (t) is rated 0.78 (the recommended range is [0.4, 1.0]) according to the current temperature difference, humidity mutation and resource status; then the density imported unidirectionally from MHU18 to this unit is:

[0107]

[0108] The total passive diffusion values brought by the remaining 5 adjacent units are 1.92, so the overall passive diffusion input is:

[0109]

[0110] Based on this, the system judges that MHU23 will show a strong density growth trend in the next time period, and the density growth rate reaches 2.124 units / hour. Superimposing its own source, it will enter the rapid density increase period around 22:00. At the same time, combined with the path analysis of active diffusion, the system identifies MHU18 as the source unit of this change and marks it as a "high diffusion point" in real time, prompting the manager to preferably deploy sex pheromone belts or biological control lines in the downwind direction of it. After three days of continuous operation, the model verification shows that this spatial diffusion mechanism accurately identifies two "secondary density aggregation sources", and the coincidence rate between the predicted diffusion path by the model and the actual trapping hot area spatial distribution reaches 84.6%. The fitting error between the density evolution curve after modeling the active and passive diffusion mechanisms and the measured curve is controlled within ±8.7%. The recommended parameter range in this model is: μ ∈ [0.4, 0.8], σ ∈ [0.3, 0.6], λ i is determined by multiplying the historical maximum trapping mean by 1.1 - 1.3 as an adjustment factor, ζ j can be set by multiplying the density maximum reference value after normalizing the environmental multi-factor index, and α ij (t) is derived from the regional comparison score, and the value range is [0.4, 1.0].

[0111] After the research team completed the ecological connectivity modeling and the construction of active and passive diffusion mechanisms, they further used the spatial diffusion prediction method based on the density transfer mechanism proposed in the present invention. For active diffusion and passive diffusion the calculation results were superimposed, and combined with the time advancement on an hourly scale, a sequence of dynamic density evolution response functions was constructed to predict the evolution trend of the habitat density in future consecutive time periods. This process is the core of the prediction output of the entire model and is achieved through the following density update function:

[0112]

[0113] In this formula, D j (t) represents the habitat density of the target micro-habitat unit j at time point t, with the unit of "individuals / hour"; Δt is the time step, which is set to 1 hour in this study to adapt to the measured trapping frequency and the sensor data upload period; and are respectively the passive diffusion input amount and the active diffusion output amount of this unit calculated previously at the current moment; ∈ j (t) is the perturbation correction term, which is used to make up for the ecological factors that are not explicitly modeled in the system but actually affect the density, such as sudden changes in the direction of the gentle breeze, predation pressure, and sudden damage to the vegetation.

[0114] Taking MHU23 as an example, at 21:00 on April 11th, the actual trapping density of this unit is D 23 (21) = 5.1 individuals / hour. The passive diffusion value calculated in the previous step is The active diffusion (diffusing outward by itself) is calculated to be The perturbation term ∈ 23 (21) is a small positive correction introduced under high humidity perturbation conditions and is set to 0.35 (unit: individuals / hour). This value is set according to the retention effect formed when the wind speed is lower than 1.0 m / s and the humidity is higher than 80%. Substituting into the formula gives:

[0115] D 23 (22) = 5.1 + 1·(2.124 - 0.86) + 0.35 = 5.1 + 1.264 + 0.35 = 6.714

[0116] Therefore, the system predicts that the density of this unit will reach 6.714 individuals / hour at 22:00, which is 32% higher than the previous moment, and the trend shows an accelerating growth. If this density evolution process is continuously iterated to t + 3 (i.e., from 23:00 to 00:00), and the diffusion inputs and outputs at each moment are updated in turn, the following sequence is formed:

[0117] t = 22:E - = 1.89, E+ = 1.01, ∈ = 0.25, then D = 6.714 + 0.88 + 0.25 = 7.844

[0118] t = 23:E - = 1.56, E + = 1.12, ∈ = 0.18, then D = 7.844 + 0.44 + 0.18 = 8.464

[0119] This sequence indicates that the population of Dendrolimus punctatus is gradually migrating and aggregating towards this area, and it is predicted that a secondary density peak in this area will form before midnight. The system judges whether to issue a local warning according to the density threshold (such as triggering a warning when exceeding 9 individuals / hour), and it is recommended to deploy biological intervention measures or physical trapping devices in advance around MHU23.

[0120] Among them, the key coefficients of the model are all set within a reasonable range to ensure ecological interpretation and numerical stability:

[0121] Δt: The time advancement granularity is 1 hour, which can be adjusted to 0.5 or 2 hours to adapt to different sampling periods;

[0122] Diffusion term E + , E - : The unit is density quantity (individuals / hour), and the measured range is mostly between [0.1, 8.0], and the extreme value can reach 10+;

[0123] Perturbation correction term ∈ j (t): The measured or expert rule setting range is recommended to be [-0.8, +0.8]. A positive value indicates enhanced conditions (such as low wind and high humidity), and a negative value indicates a sudden drop in density (such as cold front invasion, pesticide application impact);

[0124] Formula result D j (t + 1) The output is the predicted density value, which is used for comparison and verification with the measured trapping data. The system allows an error within ±10% to be regarded as a valid prediction.

[0125] Finally, this density update mechanism is not only used to predict the density change of the current unit, but can also perform parallel updates of multiple units synchronously in the graph structure, and spread the diffusion impact on the graph to form a density spatio-temporal evolution map of the entire region. In actual deployment, this mechanism successfully predicted the secondary density peaks on two migration paths, prompted the plant protection personnel to conduct manual inspections 2 hours in advance, and after verification, there was indeed a sharp increase in the trapping amount. The goodness of fit R between the system prediction result and the measured data 2 = 0.87, and the root mean square error RMSE is controlled within 1.05 individuals, which reflects the practicality, interpretability and high deployability of the density response function in actual complex ecological scenarios.

[0126] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the insect inhabitation density, characterized in that It includes the following steps: S1. Fine-grained structured modeling of the habitat environment using micro-habitat units: Divide the study area into multiple micro-habitat units (MHUs), and the division granularity is based on the activity radius of the target insects; collect environmental factors for each MHU, and retain the internal gradient information of the factors including high humidity in the southeast corner and low humidity in the north, and construct a spatial gradient tensor; S2. Introduce a density-driven causal model of the behavior-climate coupling atlas: S2.

1. For specific insects including aphids and pine caterpillars, construct a circadian behavior rhythm atlas and a climate response preference curve; use a graph neural network (GNN) to map the climate-behavior driving relationship between MHUs, with the units as nodes and the association strength between climate and behavior factors as edges; S2.

2. Extract the high-weight paths in the coupling atlas and deduce the density increase chain: the causal path of a certain climate trend → a certain behavior enhancement → an increase in the habitation probability → an increase in density; S3. Spatial diffusion prediction using a density transfer mechanism: S3.

1. Define the ecological connectivity weight between each MHU, indicating whether the insects are predicted to migrate based on vegetation continuity and temperature corridors; S3.

2. Construct a spatial diffusion convolutional layer to model the density value at each moment in two directions: active diffusion and passive diffusion: Active diffusion: Simulate the density flowing out from high ground based on the migration intention of the insects; Passive diffusion: Simulate the attraction of the neighboring area to the habitation density based on environmental adaptability; S3.

3. Output the insect habitation density distribution at future time points; S4. Introduce a temporal uncertainty assessment of the density fluctuation confidence interval: S4.

1. Use a time series model to make multi-step predictions of the habitation density of each MHU, and introduce uncertainty using a Dropout regularized deep model; S4.

2. Output the confidence interval of the predicted value, and distinguish: normal fluctuation interval, abnormal high density warning interval, data confidence decline area; combine the real-time information of trap data and remote sensing data to correct the confidence boundary.

2. The prediction method of insect habitat density according to claim 1, characterized in that The method for constructing the density-driven causal model of the behavior-climate coupling atlas: Based on the ecological characteristics of the target insect species, construct a circadian behavior rhythm atlas. The rhythm atlas is established by observing the activity frequency of the insects within a 24-hour cycle and combining the behavior response characteristics under different climate conditions, and a dynamic behavior response surface is established to reflect the change in behavior activity driven by time + climate in two dimensions.

3. The prediction method of insect habitat density according to claim 2, characterized in that The method for constructing the density-driven causal model of the behavior-climate coupling atlas: Divide the study area into multiple micro-habitat units (MHUs), and each unit represents an ecologically complete local habitation space; in the modeling, each MHU is regarded as a single node in the graph neural network, and the information contained in the node includes: the micro-climate state of the area, historical habitation density records, ecological structure attributes, and the behavior response state calculated through the rhythm atlas; the connection between each node is modeled by edge weights to reflect the coupling strength between adjacent units in terms of ecological environment and behavior linkage.

4. The prediction method of insect habitat density according to claim 3, characterized in that The method for constructing the density-driven causal model of the behavior-climate coupling atlas: Perform density propagation and prediction on the graph structure, and identify and extract the causal chain paths with significantly increasing habitat density through the evolution trend of the states of each node in the path over time. The density change trend function is defined as: Where: D(P) represents the cumulative change contribution value of density along the path P = {v1, v2, …, v n}, reflecting the density increase potential on the path; Ω k is the ecological suitability factor of the k-th node, indicating whether insects tend to inhabit this unit under the current climate and environmental conditions; At time t k is the behavior change rate in the rhythm map, representing the trend change of the behavior activity level; φ(W k,k+1 ) is the mapping function of the edge weight W, used to adjust the influence weight of the coupling strength in the density propagation path; By identifying the path with the D(P) value in the graph, the logical chain of climate trend → behavior enhancement → increased habitat probability → density growth is extracted, enabling density prediction with causal interpretation ability.

5. The prediction method of insect habitat density according to claim 1, characterized in that The spatial diffusion prediction method of the density transfer mechanism includes: Define multiple micro-habitat units (MHUs) within the study area, and construct the ecological connectivity weight values between every two units based on ecological function elements to represent the feasibility basis for the migration of insect populations. The connectivity takes the ecological path as the actual coherence, including: whether there is a continuous vegetation structure for passage between regions, whether there are local channels with suitable temperatures, and the airflow compliance and shelter coverage conditions.

6. The prediction method of insect habitat density according to claim 5, wherein The spatial diffusion prediction method of the density transfer mechanism includes: constructing a spatial diffusion mechanism to simulate the dynamic changes in the mobility of insect populations in the ecological network. The diffusion mechanism consists of active diffusion and passive diffusion: The active diffusion is the spontaneous overflow phenomenon that occurs when the population density reaches the upper limit of the ecological capacity, driven by the local density difference. The diffusion intensity can be expressed as: Where: represents the active diffusion output of the $i$-th unit at time point $t$; $\mu$ is the diffusion efficiency factor, controlling the upper limit ratio of the density spillover speed; $D$ i (t) is the insect inhabitation density of unit $i$ at the current moment; $\lambda$ i is the ecological carrying capacity upper limit of unit $i$, indicating that when the density exceeds this value, over-competition may occur, leading to individual migration; $\max[D$ i (t) - $\lambda$ i , 0]$ means that diffusion is triggered only when the density exceeds the ecological capacity, otherwise it is zero; $\sum$ j∈N(i) L ij represents the total connection weight sum of unit $i$ to all its connected neighbor units $j$; The passive diffusion is based on the density import effect caused by the sudden suitability of the ecological environment in the neighboring area, attracting insects to inhabit. The modeling method is expressed according to the habitat attraction gradient in the neighboring area as follows: Where: Denote the passive diffusion input of unit j at time point t; σ is the passive diffusion attraction coefficient, which controls the drainage intensity of adjacent areas to the target area; ∑ i∈N(j) Denote the set of all units adjacent to unit j; L ij Is the ecological connectivity weight; Denote the difference degree between the current density and the ecological maximum suitability ζ j ; α ij (t) is the habitat attraction function of adjacent unit i to target unit j, which depends on the difference between the driving force of the current environment of i and the pulling force of the environment of j.

7. The prediction method of insect habitat density according to claim 6, characterized in that The spatial diffusion prediction method of the density transfer mechanism includes: Overlay the results of active diffusion and passive diffusion, and combine the ecological driving variables such as climate, light, and wind direction in the historical time series to construct a sequence of response functions for density evolution to predict the trend of habitat density distribution at multiple future time steps. The density update function is expressed as: Where: D j (t + 1) represents the predicted density of unit j at the next time step t + 1; D j (t) is the actual density at the current moment; Δt is the time advancement step size, used to control the prediction granularity; are the density input and output amounts caused by passive attraction and active diffusion respectively; ∈ j (t) is the perturbation correction term, used to simulate the perturbation effect of random or unobserved ecological factors on density changes.

Images