Orchard multi-point network pest situation prediction method and system based on spatio-temporal gate coupling of LTC-GNN

Abstract

This invention discloses a method and system for predicting insect pest infestations in orchards using LTC-GNN spatiotemporal gating coupling. The method includes the following steps: S1: acquiring multi-source time-series data, which includes at least insect pest infestation time-series data for each monitoring node; S2: extracting features from the multi-source time-series data according to multiple preset time scales; S3: constructing static spatial associations and dynamically updating the temporal association strength between nodes, and constructing a dynamic graph topology based on the static spatial associations and the temporal association strengths; S4: synchronously aggregating the spatial neighborhood information of each monitoring node into the local state update of the current node to obtain the spatiotemporal coupling representation of each monitoring node; S5: predicting the insect pest infestation prediction results for each monitoring node in a future preset time period based on the spatiotemporal coupling representation. This invention can achieve collaborative prediction of insect pest changes in multi-point network scenarios, thereby significantly improving the prediction accuracy and adaptability of non-stationary insect pest sequences.

Description

Technical Field

[0001] This invention relates to the fields of agricultural informatization and intelligent pest monitoring and forecasting, specifically to a method and system for predicting pest infestations in orchards using a multi-point network based on LTC-GNN spatiotemporal gating coupling. Background Technology

[0002] Agriculture, as a fundamental industry of the national economy, bears the important responsibility of ensuring national food security and maintaining rural social stability. In recent years, with the deepening of agricultural modernization, agricultural production methods are accelerating their transformation towards green, intelligent, and sustainable development, and the concept of pest control is also undergoing a fundamental change. The traditional control model, which relies on chemical pesticides, has become insufficient to meet the real needs of high-quality agricultural development due to problems such as increased pesticide resistance in pests, deterioration of the farmland ecological environment, and increased pesticide residue risks in agricultural products caused by long-term use. To address these challenges, a green plant protection system guided by the principle of "prevention first, integrated pest management" has gradually become an industry consensus. This system emphasizes the synergistic application of non-chemical means such as physical control, biological control, and intelligent monitoring to build an environmentally friendly, resource-saving, precise, and efficient integrated pest management mechanism. Among these, early prediction of pest occurrence and development trends is a key link in achieving "prevention first." Only by moving the control window forward and initiating precise intervention at the early stages of pest outbreaks can the use of chemical pesticides be minimized, reducing control costs and ecological risks.

[0003] At the monitoring and sensing level of the green plant protection system, a type of green intelligent plant protection equipment, namely insect monitoring lamps (hereinafter referred to as monitoring lamps), has been widely deployed. This equipment integrates the dual functions of physical insect trapping and insect monitoring: it uses a specific wavelength light source to attract and capture phototactic agricultural pests, while also incorporating a high-definition image acquisition module, an automatic counting unit, and multi-parameter environmental sensors to identify the species and count the number of trapped insects. Some advanced models also support wireless communication and remote data upload, enabling cloud-based visual management of monitoring information. Insect monitoring lamps solve the real-time monitoring problem of "what is the current insect population?", and the historical time-series data they collect provides a data foundation for predicting "how the insect population will change in the future." For large-scale planting scenarios such as mountain orchards, dozens of insect monitoring lamps are typically deployed in a distributed manner to cover microclimate areas of different altitudes and slope aspects. Each node continuously generates daily insect population and environmental data, forming a multi-point monitoring dataset with spatiotemporal correlation characteristics.

[0004] At the level of insect infestation prediction algorithms, existing research has explored numerous avenues. Some researchers utilize convolutional networks to extract insect infestation image features to aid in species identification and counting, while others employ ensemble learning methods such as random forest models, gradient boosting tree models, and extreme random tree models to construct regression prediction models based on historical insect populations and environmental factors. Furthermore, traditional machine learning methods such as support vector regression models, K-nearest neighbor regression models, ridge regression models, Lasso regression models, and elastic network regression models have also been applied to insect infestation time-series prediction tasks.

[0005] However, existing insect infestation forecasting technologies still have the following shortcomings when addressing the long-term forecasting needs of multi-point insect infestation monitoring lamp networks in mountainous orchards: First, single-point isolated modeling lacks spatial correlation. Existing methods mostly use historical data from a single insect monitoring lamp as input for independent modeling, without incorporating spatial correlation information from surrounding monitoring points. The spread of pests in orchards has a clear spatiotemporal propagation path, and the changes in insect populations at adjacent monitoring points exhibit sequential response relationships. Single-point models cannot capture this spatial co-evolutionary pattern, resulting in prediction results being greatly affected by noise in single-point data, making it difficult to support precise control decisions in large-scale distributed scenarios. Although some studies have introduced graph neural networks to model spatial correlations, their graph topology is only statically constructed based on geographical distance and cannot dynamically evolve with insect populations. Furthermore, spatial aggregation and temporal evolution are presented in a sequential manner, making it difficult to capture the synchronous co-response patterns of insect populations in space.

[0006] Second, the models have limited capacity and feature representation capabilities, making them ill-suited for non-stationary insect infestation sequences. Existing methods, such as ensemble learning methods like random forests and gradient boosting trees, as well as traditional machine learning methods like support vector regression, have limited model capacity and feature representation capabilities, making it difficult to characterize the complex nonlinear dynamic features of insect infestation sequences. While recurrent neural network methods can model certain temporal dependencies, their fixed time constant gating mechanism is insufficient for insect infestation data with extreme non-stationary characteristics (alternating between long-term low values ​​and intermittent outbreaks), and they generally only model on a single time scale, neglecting the different granularities of insect infestation evolution, such as weekly rhythms and ten-day trends. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a multi-point network insect infestation prediction method for orchards based on LTC-GNN spatiotemporal gating coupling, which can achieve collaborative prediction of insect infestation changes in multi-point network scenarios, thereby significantly improving the prediction accuracy and adaptability of non-stationary insect infestation sequences.

[0008] To achieve one of the objectives of this invention, the following solution is adopted: A multi-point network insect infestation prediction method for orchards based on LTC-GNN spatiotemporal gating coupling includes the following steps: S1: Obtain multi-source time-series data of multiple monitoring nodes in the orchard during historical periods, wherein the multi-source time-series data includes at least the insect infestation time-series data of each monitoring node; S2: Extract features from the multi-source time-series data according to multiple preset time scales to obtain multi-scale time-series features of each monitoring node; S3: Construct a static spatial association based on the spatial positional relationship between each monitoring node, and dynamically update the temporal association strength between nodes based on the insect infestation time series data of each monitoring node, and construct a dynamic graph topology based on the static spatial association and the temporal association strength; S4: Based on the dynamic graph topology and the multi-scale temporal features, during the continuous time state evolution process, the spatial neighborhood information of each monitoring node is synchronously aggregated into the local state update of the current node to obtain the spatiotemporal coupling representation of each monitoring node. S5: Based on the spatiotemporal coupling characterization, predict the insect infestation forecast results for each monitoring node within a future preset time period.

[0009] Furthermore, in step S2, the plurality of preset time scales include at least short-period scales, medium-period scales, and long-period scales; Feature extraction is performed on the multi-source time-series data according to multiple preset time scales, specifically as follows: The insect infestation time series data are back-looked and aggregated using the window width corresponding to each preset time scale to obtain the aggregated insect population sequence at each scale. The environmental variable time series data at each scale are then aggregated using a moving average to obtain the aggregated environmental sequence at each scale. Lag statistical features reflecting the sparsity, trend direction, and fluctuation level of insect populations are extracted from each of the aggregated insect population sequences. The aggregated insect population sequences, aggregated environmental sequences, and the lag statistical features are used together as the multi-scale time series features at the corresponding scale.

[0010] Furthermore, in step S3, constructing the dynamic graph topology specifically includes: Based on the physical distance between each monitoring node and / or terrain occlusion constraints, static spatial edge weights are generated and fixed once during system deployment. Based on the statistical correlation measure between insect infestation time series data of each monitoring node in the recent historical window, dynamic time series edge weights are generated and updated daily. The static spatial edge weights and the dynamic temporal edge weights are weighted and fused, and the fused weight matrix is ​​normalized to obtain the attention weight matrix of the dynamic graph topology.

[0011] Furthermore, in step S4, spatial neighborhood information is synchronously aggregated into the local state update during the continuous-time state evolution process, specifically as follows: The spatial neighborhood aggregation operation of the graph neural network is embedded into each internal integration step of the ordinary differential equation solver of the liquid time constant network; Within each of the internal integration steps, the hidden states of all neighboring nodes of the current node are first aggregated according to the dynamic graph topology to obtain spatial context information. Then, the aggregated spatial context information is injected into the local hidden state of the current node to update the local hidden state of the current node. Subsequently, the next integration step of the ordinary differential equation is continued.

[0012] Further, in step S4, the aggregated spatial context information is injected into the local hidden state of the current node, specifically as follows: A learnable gating network is used to generate a dimension-wise gating vector based on the joint state of the current node's local hidden state and the spatial context information. The intensity of the spatial context information is adjusted using the gating vector, and the adjusted spatial context information is injected into the local hidden state of the current node in a residual superposition manner to obtain the coupled state. The coupled state is used as the state variable of the ordinary differential equation of the liquid time constant network in the current integration step for continuous evolution.

[0013] Further, in step S4, the spatiotemporal coupling characterization of each monitoring node is obtained, specifically including: The multi-scale temporal features of each monitoring node at each preset time scale are sent into multiple encoding channels with independent parameters. Each encoding channel independently performs the continuous time state evolution and spatial neighborhood information synchronous aggregation, and outputs the spatiotemporal coupled hidden state matrix corresponding to each scale. Multi-head self-attention reweighting is performed on the spatiotemporal coupled hidden state matrix corresponding to each scale to capture the long-range dependencies within each scale. The latent vectors of the last time step after attention reweighting at each scale are extracted, spliced ​​and fused to obtain a cross-scale joint spatiotemporal representation, which serves as the spatiotemporal coupling representation of each monitoring node.

[0014] Furthermore, in step S5, the insect infestation prediction result includes the predicted insect quantity for each monitoring node within a preset time period. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling further includes a trend direction determination step after obtaining the predicted insect population values: Calculate the average of the predicted insect population for each monitoring node within the preset future time period, and the average of the actual insect population over the same past time period. The difference between the predicted mean and the actual mean is compared, and the insect infestation trend is determined to be an upward trend, a flat trend, or a downward trend based on the numerical relationship between the difference and a preset tolerance threshold.

[0015] Furthermore, in step S1, the multi-source time-series data also includes environmental variable time-series data, which includes at least temperature parameters and humidity parameters; The monitoring nodes are distributed insect monitoring lamps deployed in the orchard, and the multi-source time-series data is reported via wireless communication.

[0016] Furthermore, in step S3, the statistical correlation measure includes any one of Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information; The terrain occlusion constraint is based on digital elevation model data to determine the mountain barrier relationship between each monitoring node, and is used to attenuate or set the static spatial edge weights between monitoring nodes with terrain barriers to zero.

[0017] The second objective of this invention is to provide an orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling, which can achieve collaborative prediction of insect infestation changes in multi-point network scenarios, thereby significantly improving the prediction accuracy and adaptability of non-stationary insect infestation sequences.

[0018] To achieve the second objective of this invention, the following solution is adopted: A multi-point network insect infestation prediction system for orchards based on LTC-GNN spatiotemporal gating coupling includes: The multi-source data acquisition module is used to acquire multi-source time-series data of multiple monitoring nodes in the orchard during historical periods. The multi-source time-series data includes at least the insect infestation time-series data of each monitoring node. A multi-scale temporal feature construction module is used to extract features from the multi-source time-series data according to multiple preset time scales to obtain multi-scale temporal features of each monitoring node. The dynamic graph topology construction module is used to construct static spatial associations based on the spatial positional relationships between each monitoring node, and to dynamically update the temporal association strength between nodes based on the insect infestation time-series data of each monitoring node, and to construct a dynamic graph topology based on the static spatial associations and the temporal association strength. The LTC-GNN spatiotemporal gating coupling module is used to synchronously aggregate the spatial neighborhood information of each monitoring node into the local state update of the current node during the continuous time state evolution process, based on the dynamic graph topology and the multi-scale temporal features, so as to obtain the spatiotemporal coupling representation of each monitoring node. The prediction and trend output module is used to predict the insect infestation forecast results of each monitoring node in the future preset time period based on the spatiotemporal coupling characterization.

[0019] The following is a detailed description of the orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling of the present invention.

[0020] Existing multi-point insect pest prediction systems in citrus orchards generally suffer from key problems such as neglecting spatial correlation between monitoring points, insufficient adaptability of fixed time constant models due to sparse-outbreak alternation of insect population sequences, and the inability of traditional serial spatiotemporal architectures to characterize continuous spatiotemporal information. The present invention addresses these issues by constructing a daily updated dynamic graph topology to model the distributed insect monitoring lamp array as a graph neural network and embedding the GNN spatial aggregation operation into the continuous differential state update of the liquid time constant network (LTC), thereby achieving synchronous and gradual spatial information exchange and temporal evolution.

[0021] The orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling of the present invention constructs five organically coupled core modules, as follows: The multi-source data acquisition module serves as the data foundation of the system. Multiple insect monitoring lamps are deployed in the mountainous citrus orchard. The module automatically collects data on the amount of insects trapped and killed, as well as environmental variables such as temperature and humidity, using a daily time window. The data is then reported via 4G / LoRa wireless communication. Simultaneously, a DEM digital elevation model is introduced to pre-calculate the terrain occlusion relationship between any two monitoring lamps. After data cleaning, date alignment, and missing data marking, a standardized daily multi-source spatiotemporal dataset is formed.

[0022] The multi-scale time series feature construction module uses the original daily series as a base and avoids the phase misalignment caused by traditional downsampling by using three-way back-look sliding overlap aggregation of daily, weekly, and ten-day periods (the insect population is taken as the sum within the window and the environment is taken as the mean). On the weekly and ten-day branches, three types of lag statistical features are further extracted: non-zero mean, difference between the mean of the first and second halves, and rolling mean and standard deviation. The sparsity, rapid and slow changes, and fluctuation levels of insect population data are explicitly encoded as numerical features. Finally, a fixed-length historical window is truncated as the model input.

[0023] The dynamic graph topology construction module abstracts the distributed monitoring light array as graph nodes. The edge weights are formed by the fusion of static distance edges and dynamic Pearson correlation edges. The static edges are based on the physical distance with Gaussian kernel decay and are constrained by the DEM terrain shading coefficient (the edge weights are forced to zero if there is a mountain barrier). They are calculated and fixed once during deployment, providing a stable spatial skeleton that does not change with the seasons. The dynamic edges quantify the degree of coordinated fluctuation of the insect population sequences of the two monitoring lights by the Pearson correlation coefficient of the recent insect population sequences of the two monitoring lights. They are recalculated daily based on the latest observation data to capture the dynamic fluctuation of the spatial correlation strength driven by microclimate and phenological changes. The two are weighted and fused and then normalized to form the graph attention weights.

[0024] Unlike the sequential architecture of spatial aggregation and temporal evolution in existing ST-GNN methods, the LTC-GNN spatiotemporal gating coupling module embeds the spatial aggregation operation of the graph neural network into each internal integration step of the ordinary differential equation solver of the liquid time constant network. Each step first aggregates the hidden states of neighboring nodes using graph attention weights, then adaptively adjusts the injection intensity through sigmoid gating, injecting the local hidden state as a residual before continuing ODE advancement. Spatial information exchange and local state evolution thus proceed synchronously and progressively along a continuous time axis. The weak increase in neighboring insect population signals gradually guides the rise of the local hidden state after continuous penetration through several integration steps. Simultaneously, the LTC time constant is jointly determined by the current input and the hidden state. During insect-free periods, forgetting is accelerated to filter zero-value noise, while during the pre-outbreak precursor period, memory is extended to retain key signals. This gating coupling and the adaptive time constant of LTC complement each other, enabling the model to simultaneously possess flexible adaptability to sparse insect population sequences and the ability to finely model multi-point dynamic spatial collaborative relationships. The three encoded outputs are then independently weighted by multi-head self-attention and concatenated with the cross-scale final-step latent vector to obtain a joint spatiotemporal representation.

[0025] The prediction and trend output module feeds cross-scale fusion features into a fully connected regression layer, directly outputting the daily insect population prediction value for the next 10 days from the monitoring lamp. It then compares the predicted mean with the past mean through pure post-processing to determine the three-category trend direction: rising, stable, or falling. These recommendations correspond to suggestions for prevention and control decisions such as applying pesticides in advance, maintaining existing control measures, or reducing patrols.

[0026] The end-to-end prediction system formed by the above modules upgrades insect infestation prediction from the traditional passive statistical mode of single-point independence, fixed parameters, and static space to an active prediction mode of multi-point collaboration, adaptive time constant, and dynamic graph synchronous coupling. It is particularly suitable for agricultural edge scenarios such as mountain citrus orchards that are distributed, have complex terrain, and limited resources.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves accurate modeling of multi-point spatial relationships by constructing a dynamic graph topology, significantly improving prediction accuracy. Addressing the shortcomings of existing methods, such as isolated single-point modeling, missing spatial relationships, and the inability to dynamically evolve graph topologies based solely on geographical distance, this invention constructs static spatial relationships based on the spatial location relationships between monitoring nodes. Simultaneously, it dynamically updates the temporal relationship strength between nodes based on the temporal data of insect populations at each monitoring node, fusing these two elements to construct a graph topology structure that combines a stable framework with dynamic timeliness. Static spatial relationships preserve the basic spatial dependencies between geographically adjacent monitoring nodes, providing a stable spatial framework that does not change with the seasons; dynamic temporal relationship strength captures the dynamic fluctuations in spatial relationship strength driven by microclimate and phenological changes. The weighted fusion of these two elements allows the graph topology to adaptively adjust according to insect population conditions, overcoming the limitation of traditional static graph structures in reflecting the temporal evolution of spatial relationships. This enables the model to accurately characterize the synchronous linkage and gradual transmission process of insect population fluctuations among multiple monitoring points, significantly improving the prediction accuracy of insect populations in multi-point network scenarios.

[0028] 2. This invention overcomes the limitations of discrete truncation in serial architectures by synchronously aggregating spatial neighborhood information during continuous temporal state evolution. Addressing the shortcomings of existing methods that use a sequential approach for spatial aggregation and temporal evolution, making it difficult to capture the synchronous collaborative response patterns of insect populations in space, this invention, based on dynamic graph topology and multi-scale temporal features, synchronously aggregates the spatial neighborhood information of each monitoring node into the local state update of the current node during continuous temporal state evolution. That is, spatial neighborhood information aggregation and local state updates are performed synchronously in each internal step of continuous temporal state evolution, enabling spatial information exchange and temporal evolution to proceed synchronously and progressively along a continuous time axis. Compared to the traditional serial architecture that performs spatial aggregation before or after temporal evolution, this invention fundamentally avoids the fragmentation and loss of spatiotemporal information during discretization, enabling the model to possess a more refined continuous temporal domain characterization capability for the collaborative evolution process of insect populations across multiple monitoring points.

[0029] 3. This invention enhances the adaptability to non-stationary insect infestation sequences through the synergy of multi-scale temporal feature extraction and continuous-time state evolution. Addressing the shortcomings of existing methods, such as the insufficient adaptability of fixed-time-constant gating mechanisms to insect infestation data with extremely non-stationary characteristics, and the common practice of modeling only at a single time scale while neglecting weekly rhythms and ten-day trends, this invention extracts features from multi-source temporal data across multiple preset time scales. This enables the model to simultaneously capture multi-granularity insect infestation evolution patterns, including short-term fluctuations, medium-term rhythms, and long-term trends. Furthermore, by performing state evolution on a continuous time axis, it avoids the truncation distortion of insect infestation changes caused by traditional discrete time steps. Multi-scale temporal features provide the model with multi-granular information on insect infestation evolution, while continuous-time state evolution provides the model with a fine-grained ability to characterize the insect infestation change process. The synergistic effect of these two methods significantly improves the model's adaptability and predictive robustness to the complex characteristics of non-stationary and highly sparse insect infestation data in mountainous orchards.

[0030] 4. This invention establishes an end-to-end multi-point collaborative prediction system, effectively supporting precise prevention and control decisions. The complete process of multi-source data acquisition, multi-scale feature extraction, dynamic graph topology construction, spatiotemporal coupling representation, and pest prediction output constitutes an end-to-end prediction system. This elevates pest prediction from a traditional passive statistical model of single-point independence, fixed parameters, and static space to an active intelligent prediction model of multi-point collaboration, dynamic graph topology, and spatiotemporal synchronous coupling. This system can independently output pest prediction results for a preset time period for each monitoring node in the orchard, providing a scientific basis for fruit farmers to implement precise intervention in the early stages of pest outbreaks. It helps to minimize the use of chemical pesticides, reduce control costs and ecological risks, and has significant agricultural economic benefits and ecological environmental value. It is particularly suitable for distributed, terrain-complex, and resource-constrained agricultural edge computing scenarios such as mountain orchards. Attached Figure Description

[0031] Figure 1 This is a block diagram of the orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling in an embodiment of the present invention. Figure 2 This is an architecture diagram of the LTC-GNN spatiotemporal gating coupling model in an embodiment of the present invention; Figure 3 This is a flowchart of the orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling in an embodiment of the present invention. Detailed Implementation

[0032] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0033] Example 1 like Figure 1 As shown, this embodiment of the invention provides an orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling, comprising: The data acquisition module is used to acquire multi-source time-series data of multiple monitoring nodes in the orchard during historical periods. The multi-source time-series data includes at least the insect infestation time-series data of each monitoring node.

[0034] The feature construction module is used to extract features from the multi-source time series data according to multiple preset time scales to obtain multi-scale time series features of each monitoring node.

[0035] The graph topology construction module is used to construct static spatial associations based on the spatial positional relationships between each monitoring node, and dynamically update the temporal association strength between nodes based on the insect infestation time series data of each monitoring node, and construct a dynamic graph topology based on the static spatial associations and the temporal association strength.

[0036] The spatiotemporal coupling module is used to synchronously aggregate the spatial neighborhood information of each monitoring node into the local state update of the current node during the continuous time state evolution process, based on the dynamic graph topology and the multi-scale temporal features, so as to obtain the spatiotemporal coupling representation of each monitoring node.

[0037] The prediction output module is used to predict the insect infestation prediction results of each monitoring node in the future preset time period based on the spatiotemporal coupling characterization.

[0038] The following is a more detailed description of the orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling, according to an embodiment of the present invention.

[0039] To address the issues of isolated single-point modeling and lack of spatial correlation, this invention presents a dynamic graph topology construction method for a multi-point network insect pest prediction system in orchards based on LTC-GNN spatiotemporal gating coupling. This method involves abstracting multiple distributed insect pest monitoring lamps within the orchard as graph nodes. Edge weights are derived from a dual-source fusion of static distance edges and dynamic correlation edges, updated daily. Static edges, based on physical distance and constrained by DEM terrain occlusion, are calculated and fixed once during deployment, providing a stable spatial framework unaffected by seasonal changes. Dynamic edges quantify the degree of coordinated fluctuation in insect population sequences between two lamps using the Pearson correlation coefficient of their recent insect population sequences. This is recalculated daily based on the latest observation data, capturing the dynamic fluctuations in spatial correlation strength driven by microclimate and phenological changes. The weighted fusion of these two methods, followed by normalization, forms graph attention weights, ensuring the graph topology possesses both the stability of the static framework and the timeliness of dynamic correlations.

[0040] To address the limitations of fixed-time-constant and single-scale modeling capabilities, this invention presents an orchard multi-point network pest prediction system based on LTC-GNN spatiotemporal gating coupling. This system employs an LTC-GNN spatiotemporal gating coupling mechanism: replacing the fixed-time-constant LSTM / GRU with a liquid-time-constant network as the temporal encoding unit. The LTC's forgetting rate is dynamically controlled by the current input, accelerating forgetting during long-term low-value periods to reduce historical noise interference, and slowing down forgetting during outbreak-prone periods to retain key precursor signals. Simultaneously, independent encoding channels at daily, weekly, and ten-day time scales are constructed to capture pest evolution patterns at different granularities. Furthermore, the spatial aggregation operation of the graph neural network is embedded into each internal integration step of the LTC ordinary differential equation solver. Neighbor hidden state aggregation, Sigmoid adaptive gating, and residual injection are synchronously executed at each step of the differential equation progression, enabling spatial information exchange and temporal evolution to proceed synchronously and progressively along a continuous time axis. This overcomes the discretization of the spatial synergistic effects of pests in the temporal dimension inherent in traditional serial architectures. The aforementioned gating coupling complements the adaptive time constant of LTC: LTC is responsible for adjusting the forgetting rate based on input characteristics, determining when to remember and when to forget; GNN gating is responsible for evaluating the reliability of neighbor information, determining how much spatial information to inject. The synergy of both enables the model to possess both flexible adaptability to sparse insect infestation sequences and the ability to finely model multi-point dynamic spatial relationships.

[0041] At the predictive output level, this invention uses a fully connected regression layer to map the cross-scale fusion features of each insect monitoring lamp to the predicted daily insect population for the next 10 days. That is, each monitoring lamp independently outputs a daily insect population sequence with a ten-day length. By comparing the predicted mean of each monitoring lamp with the recent actual mean through pure post-processing, the three-category trend direction of rising, flat, or falling is determined, providing fruit farmers with a simple and intuitive basis for prevention and control decisions.

[0042] The overall architecture of the orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling in this embodiment of the invention is as follows: Figure 2 As shown, the whole system consists of five organically coupled modules: The multi-source data acquisition module deploys multiple insect monitoring lamps in the mountainous citrus orchard. It automatically collects data on the amount of insects trapped and killed, as well as environmental variables such as temperature and humidity, using a daily time window. The data is then reported via 4G / LoRa wireless communication. Simultaneously, it incorporates DEM (Digital Elevation Model) data to pre-calculate the terrain occlusion relationship between any two monitoring lamps. After data cleaning, date alignment, and missing data marking, a unified and standardized daily multi-source spatiotemporal dataset is formed, providing standardized input for all subsequent modeling modules.

[0043] The multi-scale time series feature construction module uses the original daily series as a base and employs back-look sliding aggregation with physical alignment of daily, weekly, and ten-day periods. The insect population is taken as the sum within the window, and the environmental variables are taken as the mean within the window, avoiding the phase misalignment caused by traditional downsampling. On the aggregation sequences of the weekly and ten-day branches, three types of lag statistical features are further extracted: non-zero mean, difference between the mean of the first and second halves, and rolling mean and standard deviation. The sparsity, rapid and slow changes, and fluctuation levels of insect data are explicitly encoded as numerical features. Finally, fixed-length historical windows of 30 steps, 4 steps, and 3 steps are respectively extracted from the tail of the three sequences as model input.

[0044] The dynamic graph topology construction module abstracts the distributed insect monitoring lamp array as graph nodes. Edge weights are derived from two components: static distance edges, based on physical distance decayed by a Gaussian kernel and constrained by the DEM terrain occlusion coefficient (where edge weights are forced to zero if there is a mountain blocking the path between two lamps), are calculated and fixed once during deployment; dynamic correlation edges, quantifying the degree of coordinated fluctuation in insect population sequences between two lamps using the Pearson correlation coefficient of their recent insect population sequences, are recalculated daily based on the latest observation data. The weighted fusion of these two components, after normalization, forms the graph attention weights, constructing a daily updated adjacency matrix that combines the stability of the static spatial framework with the timeliness of dynamic correlations.

[0045] The LTC-GNN spatiotemporal gating coupling module embeds the spatial aggregation operation of the graph neural network into the continuous differential state update of the liquid time constant network: at each integration step, the hidden states of neighboring nodes are first aggregated with graph attention weights, the injection intensity is adaptively adjusted by sigmoid gating, and the gated neighbor information is injected into the local hidden state through residual connections before continuing the next step of the differential equation. The above coupling operation is repeated in each internal integration step of the ODE solver, so that spatial information exchange and temporal evolution proceed synchronously and progressively on a continuous time axis. This achieves the progressive fusion of spatial collaborative information between monitoring lights on a continuous time axis, accurately depicting the synchronous linkage and gradual transmission process of insect population fluctuations among multiple monitoring points. The three encoded outputs are then independently reweighted by multi-head self-attention and concatenated with the cross-scale final hidden vector to obtain a joint spatiotemporal representation.

[0046] The prediction and trend output module uses a fully connected regression layer to map the cross-scale fusion features of each monitoring lamp to the predicted daily insect population for the next 10 days. This means each monitoring lamp independently outputs a daily insect population sequence spanning one ten-day period. The trend direction is determined through post-processing: comparing the predicted mean for the next 10 days with the actual mean for the past 10 days, and combining this with a tolerance threshold, outputting three categories of trends: rising, stable, and falling.

[0047] The end-to-end prediction system formed by the above modules enables long-term insect infestation trend prediction for multi-point monitoring lamp networks, and is particularly suitable for distributed, resource-constrained agricultural edge scenarios such as mountain orchards.

[0048] In this embodiment, the specific contents of the multi-source data acquisition module are as follows: Set up in the mountain citrus orchard The insect monitoring lamps are numbered as follows: Each monitoring lamp records its latitude and longitude coordinates via GPS positioning, and then converts them into Cartesian coordinates via UTM (Universal Transverse Mercator) projection. The unit is meters, which ensures that all subsequent spatial distance calculations are consistent with the actual distances on the Earth's surface.

[0049] Meanwhile, each insect monitoring lamp automatically performs the following data collection tasks within a daily time window: Insect trapping and killing amount : Monitoring Light In the Total number of pests trapped per day (unit: head). ; Environment variable vector The corresponding daily average temperature (°C), daily maximum temperature (°C), daily minimum temperature (°C), daily temperature difference (°C), and daily average relative humidity (%).

[0050] The collected data is reported to the edge gateway or cloud server daily via 4G mobile communication or LoRa (Long Range) module.

[0051] In this embodiment, the specific contents of the multi-scale temporal feature construction module are as follows: Insect population sequences are non-stationary and sparse, with intermittent bursts interspersed within long-term zero or low values. The original sequence at a single time granularity is insufficient to simultaneously capture short-term fluctuations, medium-term rhythms, and long-term trends. This module uses the original daily sequence as a base, constructing aggregated sequences at three time granularities—daily, weekly, and ten-day periods—through sliding overlap aggregation, calculating lag statistical characteristics, and finally truncating a fixed-length window as model input.

[0052] Specifically, the three-way rearview sliding convergence: In time As the endpoint of the window, respectively with the window width Slide backwards to aggregate (step size is always 1 day) to construct branches at three scales: daily, weekly, and ten-day. Japanese branch ( (Window width is 1 day; no aggregation is performed; the original daily series is retained directly. Series length is...) The characteristic of each time step is the original insect population for that day. With environmental variables .

[0053] Weekly branch ( The window width is 7 days, and the step size is 1 day. For each window, the insect population is the sum within the window, and the environmental variables are the mean within the window. The sequence length is... .

[0054] Branches of the ten-day period ( The window width is 10 days, and the step size is 1 day. For each window, the insect population is the sum within the window, and the environmental variables are the mean. The sequence length is... .

[0055] The aggregation formula for the three branches can be uniformly expressed as: in For monitoring lights by The endpoint is [the endpoint], and the window width is [the width]. The amount of aggregate insects, This is the corresponding aggregated environment variable vector.

[0056] Specifically, lagged statistical characteristics: At each time step of the above three aggregation sequences, in addition to the original number of insects... and environment variables In addition, the following hysteresis characteristics are calculated.

[0057] A non-zero mean reflects the average intensity when insects are present. The difference between the mean values ​​of the first and second halves reflects the directional trend within the window; The rolling mean and standard deviation reflect the recent level of volatility and dispersion.

[0058] Specifically, sequence slicing and input window: For each monitoring point and each prediction time A fixed-length historical window is truncated from the end of the three aggregation sequences mentioned above and used as the three-way input. (Day branch input) Extract the daily aggregate sequence from the past 30 days, totaling 30 time steps; weekly branch input. Extract the weekly aggregate sequence from the past 4 weeks, for a total of 4 time steps; decadal branch input. Extract the aggregation sequence of the past three ten-day periods, for a total of three time steps.

[0059] The time resolution of the three inputs ranges from fine to coarse, focusing on short-term fluctuations, medium-term rhythms, and long-term trends, respectively. All three windows end at the prediction time, ensuring that each branch feature corresponds to the same target time to be predicted.

[0060] In this embodiment, the specific contents of the dynamic graph topology construction module are as follows: Within the orchard The insect monitoring lamp is abstracted as an undirected graph. ,in For a set of nodes, For the first The set of static spatial proximity edges and dynamic temporally related edges for each day. For the first The weighted adjacency matrix of the day.

[0061] Specifically, the static space adjacent edge : The geographical distance between monitoring lights reflects short-distance crawling dispersal and passive migration caused by agricultural operations. The closer the distance, the higher the risk of spread. However, in mountainous orchards, if there is a mountain barrier between two monitoring lights, pests cannot spread directly, and that side should be forcibly cut off. Therefore, DEM topographic shading pre-calculation is introduced.

[0062] For any two monitoring lights and Connect the two points to form a straight line in space, and check each DEM grid along this line. If the terrain at any point in the middle exceeds the preset terrain occlusion threshold of the smooth connection between the two points, it is determined that there is severe occlusion, and an occlusion coefficient is set. 1; otherwise .

[0063] Based on the occlusion coefficient and the UTM plane distance between the monitoring lamps, a static spatial neighbor edge weight with Gaussian kernel decay is constructed: in The distance between the two monitoring lights is the Euclidean distance in the UTM plane. This is the Gaussian kernel bandwidth parameter (the default value is the median distance between all monitoring lights). At that time, the edge weakens the false spatial connection that is blocked by the terrain. The calculation is completed and fixed once during system deployment, and will not be changed subsequently.

[0064] Specifically, dynamic temporal similarity edges : In addition to physical distance, there is also a coordinated fluctuation correlation between two monitoring lights due to local microclimate synchronization (such as being located in the same valley wind gap, or receiving similar sunshine and precipitation on the same slope aspect) or phenological periods. Even if two monitoring lights are far apart, if they are in the same microclimate unit, their insect population fluctuation rhythms will be highly synchronized. This correlation cannot be derived from a fixed geographical location and must be dynamically extracted from the statistical patterns of historical insect population sequences.

[0065] Calculation and forecasting lamp and monitoring lamp in the past Heaven (as) The Pearson correlation coefficient of the insect population sequence. The two sequences are centered (each element subtracts its own mean) and denoted as... ,but: Will Activated by softplus and mapped to non-negative edge weights The weight is greatest when there is a strong positive correlation, and the weight approaches zero when there is a weak or negative correlation. The data is recalculated daily based on the latest insect population data, reflecting the dynamic fluctuations in the spatial correlation intensity driven by meteorological conditions and phenological changes.

[0066] Specifically, dynamic adjacency matrix fusion: The two weight matrices above are summed with weights to form a dynamic adjacency matrix that is updated every 24 hours. in This is the balance coefficient (obtained through optimization on the validation set during training). It provides a stable spatial framework that does not change over time, and even if the correlation of insect population is weak during a certain period (such as the low activity period in winter), the basic connection between the nearby monitoring lights is still maintained. Dynamically capturing spatial correlation intensity fluctuations caused by microclimate and phenological changes over time, during periods of synchronous outbreaks (such as the spring warming period), even distant monitoring lights located within the same microclimate unit will be assigned high weight. Each row is normalized using softmax to obtain the graph attention weights used for subsequent LTC-GNN aggregation. .

[0067] In this embodiment, the specific contents of the LTC-GNN spatiotemporal gating coupling module are as follows: This module is the core of the method of this invention. The daily, weekly, and ten-day feature tensors are respectively fed into three encoding channels with identical structures and independent parameters. Within each channel, the spatial information aggregation of the GNN is embedded and executed synchronously in each integration step of the LTC continuous differential state update.

[0068] Specifically, GNN spatial aggregation embedding: During the LTC coding process, each monitoring light node Maintain a hidden state ( (For the hidden state dimension), this invention injects the spatial information of neighboring monitoring lights into the local hidden state through the following steps: Neighbor information aggregation: Dynamic graph attention weights derived from the dynamic graph topology building module. For nodes The hidden states of all neighboring nodes are weighted and aggregated: For nodes The received spatial context vector encodes the current overall insect infestation status in its neighborhood.

[0069] Gating signal calculation: calculating the hidden state of the node itself. Aggregating information with neighbors After concatenation, an element-wise gated vector is generated through a learnable sigmoid gated network: in This represents vector concatenation. and For trainable gating parameters, The Sigmoid activation function compresses the gate value to... Interval. Each element Independently control the injection intensity of neighbor information in the corresponding dimension.

[0070] Neighbor information transformation: After performing a linear transformation on the aggregated neighbor information, it is then passed through... Activate and map it to The interval allows it to both positively enhance and negatively suppress the local hidden state: in , These are trainable parameters.

[0071] Gated injection and state correction: The transformed neighbor information is gated and then superimposed onto the original hidden state as a residual. in This represents element-wise multiplication. The residual design ensures that: when gated... (When neighbor information is inconsistent with local status or is useless) The local dynamics of the LTC remain undisturbed; when (When neighbor information is highly relevant), neighbor information is fully injected to achieve spatial collaboration.

[0072] Specifically, LTC continuous dynamics: Coupled states are obtained through GNN spatial injection. Subsequently, its continuous-time evolution is described by the following LTC (Liquid Time Constant Network) ordinary differential equation: in: : Adaptive time constant, determined by the current input With the current hidden state Joint decision; : Nonlinear driving function (small fully connected network); : Each of their own trainable parameters.

[0073] The three LTC-GNN encoding channels do not share parameters and independently capture spatial cooperative patterns at the daily, weekly, and decadal scales, respectively outputting spatiotemporally coupled hidden state matrices: Specifically, the long position self-attention reweighting: Output sequence of each LTC The time dimension is then reweighted through a multi-head self-attention layer. The self-attention mechanism allows each time step in the sequence to directly establish a dependency with any other time step in the same sequence, without being restricted by the stepwise propagation of the LTC loop structure, thereby capturing long-range cross-step dependencies—such as the periodic influence of the insect population on the corresponding day of this week on the same day last week, or the continuation effect of the overall trend of the previous ten-day period on the current ten-day period.

[0074] Specifically, taking the daily branch as an example (the same applies to the weekly and ten-day branches), for Each row vector Calculate the scaled dot-product attention: in (Query matrix, Query) (Key matrix, Key) (Value matrix) For a trainable projection matrix, For each dimension of attention head, This refers to the number of attention heads, which involves parallel computation of four independent Query-Key-Value projections, each focusing on different dependency patterns of its sequence, and finally concatenating them for reception. Divide by Its purpose is to prevent the softmax gradient from vanishing due to excessively large dot product values.

[0075] The output of multi-head attention is The result of splicing the sizes: in The weekly and ten-day branches perform the above operations independently, and the parameters of the three attention layers are not shared.

[0076] Specifically, cross-scale feature fusion: The latent vectors of the last time step from the three attention outputs are flattened and concatenated to form a cross-scale joint representation: in for The last line (the (The attention output at each time step), and so on. This fused vector is the final output of this module and is fed into the prediction module.

[0077] In this embodiment, the specific contents of the prediction and trend output module are as follows: Cross-scale fusion vector The signal is fed into a fully connected regression layer, and the output is for each monitoring lamp. Daily insect population forecast for the next 10 days: in , These are trainable parameters. Indicator Light In the future sky( The predicted number of insects.

[0078] To further provide guidance on macroeconomic trends, the trend direction is extracted from the above forecasts using a purely post-processing method: the average forecast value for the next 10 days is calculated. Compared with the actual average of the past 10 days Compare the difference between the two and introduce a tolerance threshold. (Take 10% of the historical insect population standard deviation to avoid minor fluctuations being misjudged as trend changes), output the three-category trend direction: This invention constructs a dynamic graph topology for a multi-point insect monitoring lamp network. Multiple insect monitoring lamps distributed throughout an orchard are not isolated; pests spread and migrate between lamps, and the insect population fluctuations of lamps within the same microclimate unit are highly synchronized. Existing insect prediction methods generally model each monitoring point independently, ignoring the spatial relationships between lamps and losing crucial information that changes in insect populations at neighboring lamps can predict local trends. This invention uses a graph neural network to explicitly model the multi-point lamp array. Static edges are based on physical distances with Gaussian kernel decay, superimposed with DEM topographic occlusion filtering to adapt to the unique terrain of mountainous orchards, providing a stable spatial framework that does not change with the seasons. Dynamic edges quantify the degree of coordinated fluctuation between insect population sequences of two lamps using the Pearson correlation coefficient, recalculated daily based on the latest observation data, capturing the dynamic fluctuations in spatial correlation strength driven by microclimate and phenological changes. Through weighted fusion, the graph topology combines the stability of the static framework with the timeliness of dynamic correlations, overcoming the limitations of traditional static graphs in modeling the temporal evolution of insect spatial correlations.

[0079] This invention designs a spatiotemporally gated coupling mechanism using LTC-GNN. Insect infestation sequences exhibit sparsity and non-stationarity, with intermittent bursts interspersed within long-term low values, requiring the temporal encoder to dynamically adjust the rhythm of information retention and forgetting based on input characteristics. Existing LSTM / GRU methods employ fixed-time-constant gating mechanisms, which cannot adapt to such drastically changing temporal patterns. This invention introduces LTC (Liquid Time Constant Network) as the temporal encoding unit, whose time constant is jointly determined by the current input and the hidden state. In the spatial dimension, existing methods often employ a sequential architecture of spatial information followed by temporal information or vice versa, decoupling spatial information exchange and temporal evolution into two independent stages, resulting in discretized spatiotemporal information of insect infestations. This invention proposes a spatiotemporally gated coupling mechanism embedding GNN with LTC, embedding the spatial aggregation operation of GNN into each internal integration step of the LTC's ODE solver. At each step of the differential equation progression, neighboring hidden states are first aggregated with graph attention weights, and the injection intensity is adaptively adjusted via Sigmoid gating, injecting the local hidden state in residual form before continuing ODE progression. Spatial information exchange and local state evolution therefore proceed synchronously and gradually over a continuous time axis. The slight increase in neighboring insect populations, after several integral steps of continuous infiltration, gradually guides the local latent state to rise, macroscopically manifested as the predicted value following the neighboring trend several days later. This gating coupling complements the adaptive time constant of LTC: LTC is responsible for adjusting the forgetting rate based on input characteristics, determining when to remember and when to forget; GNN gating is responsible for evaluating the reliability of neighboring information, determining how much spatial information to inject. The synergy of both enables the model to simultaneously possess the flexible adaptability to sparse insect population sequences and the ability to continuously model the dynamic spatial coordination relationships of multi-point insect populations in the temporal domain.

[0080] The orchard multi-point network insect infestation prediction system based on LTC-GNN spatiotemporal gating coupling in this invention has the following advantages: At the spatial modeling level, existing methods process each monitoring point independently or use a static graph structure, losing the spatial correlation information brought about by the spread of insect infestation and the synchronization of microclimate between monitoring lamps. This invention uses a graph neural network to explicitly model the multi-point monitoring lamp array, fusing static distance edges and dynamic Pearson correlation edges from two sources and updating them daily, so that the graph topology has both the stability of the spatial skeleton and the timeliness of the correlation strength. At the temporal coding level, existing LSTM / GRU uses a fixed time constant, which cannot adapt to the non-stationary characteristics of sparse insect infestation sequences and alternating outbreaks. This invention replaces it with an LTC liquid time constant network, which dynamically adjusts the forgetting rate according to the input, accelerating forgetting during insect-free periods and prolonging memory during the pre-outbreak period. At the spatiotemporal coupling level, existing serial architectures decouple spatial aggregation and temporal evolution into two independent stages, severing continuous spatiotemporal information. This invention embeds GNN spatial aggregation into each internal integration step of the LTC ODE solver, so that spatial information exchange and temporal evolution proceed synchronously and progressively on a continuous time axis, fundamentally breaking through the discretization and truncation limitations of serial architectures.

[0081] Regarding dynamic graph topology construction, in other embodiments, in addition to the Pearson correlation coefficient used in this invention to quantify the statistical correlation of insect population sequences between two monitoring nodes, alternative measurement methods such as Spearman rank correlation coefficient, mutual information, or dynamic time warping (DTW) can also be used to capture the degree of synchronization between sequences.

[0082] In other embodiments, a further alternative is to abandon manually designed association metrics and instead maintain a learnable embedding vector for each monitoring node. Adaptive adjacency matrices are generated end-to-end through embedding inner products or small multilayer perceptrons (MLPs) (as in the ideas of MTGNN, GTS, etc.), so that the graph topology is learned entirely by data, eliminating the limitations of preset similarity formulas.

[0083] In other embodiments, instead of recalculating daily, multiple static map templates can be preset according to seasons or phenological stages, and the corresponding template can be automatically activated based on the current date or accumulated temperature. This method has a coarser granularity but lower computational overhead.

[0084] Regarding the spatiotemporal coupling mechanism of LTC-GNN, in other embodiments, the liquid time constant network unit can be replaced with any continuous-time state evolution model such as generalized neural ordinary differential equation (Neural ODE, NODE) ​​or GRU-ODE. The coupling framework of the spatial aggregation of the graph neural network embedded in the integral step inside the ODE solver has model independence, and replacing the underlying temporal unit does not affect the effectiveness of the coupling mechanism itself.

[0085] In other embodiments, in terms of spatial aggregation methods, in addition to graph attention-weighted aggregation, a multi-head self-attention mechanism across nodes can also be adopted, that is, all monitoring nodes perform attention calculations on each other's hidden states, which is equivalent to exchanging spatial information on a dynamic fully connected graph, and can eliminate the explicit graph topology construction steps.

[0086] In other embodiments, a learnable coupling scheduling strategy can be designed for the coupling frequency, which adaptively determines whether to perform spatial coupling in the current integration step based on the current insect situation, rather than forcibly executing it in every step, so as to achieve a dynamic balance between spatial cooperation benefits and computational overhead.

[0087] In terms of multi-scale feature construction, in other embodiments, the number of branches and window width of the three-way sliding aggregation can be adjusted according to the insect cycle characteristics of the specific orchard (e.g., changing to three branches for daily, weekly, and monthly periods, or increasing or decreasing the window size), and the lag statistical features can also be added or deleted according to domain knowledge.

[0088] Regarding the predicted output, in other embodiments, in addition to pure post-processing comparison of the mean to determine the trend, a lightweight trend classification head and a regression head can be trained to share an encoder, or quantile regression can be used to directly output the insect population prediction interval instead of point estimation.

[0089] Example 2 like Figure 3 As shown, this embodiment of the invention provides a multi-point network insect infestation prediction method for orchards based on LTC-GNN spatiotemporally gated coupling (Liquid Time-Constant-Graph Neural Network), including the following steps: S1: Obtain multi-source time-series data of multiple monitoring nodes in the orchard during historical periods, wherein the multi-source time-series data includes at least the insect infestation time-series data of each monitoring node.

[0090] S2: Extract features from the multi-source time-series data according to multiple preset time scales to obtain multi-scale time-series features of each monitoring node.

[0091] S3: Construct a static spatial association based on the spatial location relationship between each monitoring node, and dynamically update the temporal association strength between nodes based on the insect infestation time series data of each monitoring node, and construct a dynamic graph topology based on the static spatial association and the temporal association strength.

[0092] S4: Based on the dynamic graph topology and the multi-scale temporal features, during the continuous time state evolution process, the spatial neighborhood information of each monitoring node is synchronously aggregated into the local state update of the current node to obtain the spatiotemporal coupling representation of each monitoring node.

[0093] S5: Based on the spatiotemporal coupling characterization, predict the insect infestation forecast results for each monitoring node within a future preset time period.

[0094] Furthermore, in step S2, the plurality of preset time scales include at least short-period scales, medium-period scales, and long-period scales; Feature extraction is performed on the multi-source time-series data according to multiple preset time scales, specifically as follows: The insect infestation time series data are back-looked and aggregated using the window width corresponding to each preset time scale to obtain the aggregated insect population sequence at each scale. The environmental variable time series data at each scale are then aggregated using a moving average to obtain the aggregated environmental sequence at each scale. Lag statistical features reflecting the sparsity, trend direction, and fluctuation level of insect populations are extracted from each of the aggregated insect population sequences. The aggregated insect population sequences, aggregated environmental sequences, and the lag statistical features are used together as the multi-scale time series features at the corresponding scale.

[0095] Furthermore, in step S3, constructing the dynamic graph topology specifically includes: Based on the physical distance between each monitoring node and / or terrain occlusion constraints, static spatial edge weights are generated and fixed once during system deployment. Based on the statistical correlation measure between insect infestation time series data of each monitoring node in the recent historical window, dynamic time series edge weights are generated and updated daily. The static spatial edge weights and the dynamic temporal edge weights are weighted and fused, and the fused weight matrix is ​​normalized to obtain the attention weight matrix of the dynamic graph topology.

[0096] Furthermore, in step S4, spatial neighborhood information is synchronously aggregated into the local state update during the continuous-time state evolution process, specifically as follows: The spatial neighborhood aggregation operation of the graph neural network is embedded into each internal integration step of the ordinary differential equation solver of the liquid time constant network; Within each of the internal integration steps, the hidden states of all neighboring nodes of the current node are first aggregated according to the dynamic graph topology to obtain spatial context information. Then, the aggregated spatial context information is injected into the local hidden state of the current node to update the local hidden state of the current node. Subsequently, the next integration step of the ordinary differential equation is continued.

[0097] Further, in step S4, the aggregated spatial context information is injected into the local hidden state of the current node, specifically as follows: A learnable gating network is used to generate a dimension-wise gating vector based on the joint state of the current node's local hidden state and the spatial context information. The intensity of the spatial context information is adjusted using the gating vector, and the adjusted spatial context information is injected into the local hidden state of the current node in a residual superposition manner to obtain the coupled state. The coupled state is used as the state variable of the ordinary differential equation of the liquid time constant network in the current integration step for continuous evolution.

[0098] Further, in step S4, the spatiotemporal coupling characterization of each monitoring node is obtained, specifically including: The multi-scale temporal features of each monitoring node at each preset time scale are sent into multiple encoding channels with independent parameters. Each encoding channel independently performs the continuous time state evolution and spatial neighborhood information synchronous aggregation, and outputs the spatiotemporal coupled hidden state matrix corresponding to each scale. Multi-head self-attention reweighting is performed on the spatiotemporal coupled hidden state matrix corresponding to each scale to capture the long-range dependencies within each scale. The latent vectors of the last time step after attention reweighting at each scale are extracted, spliced ​​and fused to obtain a cross-scale joint spatiotemporal representation, which serves as the spatiotemporal coupling representation of each monitoring node.

[0099] Furthermore, in step S5, the insect infestation prediction result includes the predicted insect quantity for each monitoring node within a preset time period. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling further includes a trend direction determination step after obtaining the predicted insect population values: Calculate the average of the predicted insect population for each monitoring node within the preset future time period, and the average of the actual insect population over the same past time period. The difference between the predicted mean and the actual mean is compared, and the insect infestation trend is determined to be an upward trend, a flat trend, or a downward trend based on the numerical relationship between the difference and a preset tolerance threshold.

[0100] Furthermore, in step S1, the multi-source time-series data also includes environmental variable time-series data, which includes at least temperature parameters and humidity parameters; The monitoring nodes are distributed insect monitoring lamps deployed in the orchard, and the multi-source time-series data is reported via wireless communication.

[0101] Furthermore, in step S3, the statistical correlation measure includes any one of Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information; The terrain occlusion constraint is based on digital elevation model data to determine the mountain barrier relationship between each monitoring node, and is used to attenuate or set the static spatial edge weights between monitoring nodes with terrain barriers to zero.

[0102] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An orchard multi-point network pest situation prediction method based on LTC-GNN spatiotemporal gate coupling, characterized in that, Includes the following steps: S1: Obtain multi-source time-series data of multiple monitoring nodes in the orchard during historical periods, wherein the multi-source time-series data includes at least the insect infestation time-series data of each monitoring node; S2: Extract features from the multi-source time-series data according to multiple preset time scales to obtain multi-scale time-series features of each monitoring node; S3: Construct a static spatial association based on the spatial positional relationship between each monitoring node, and dynamically update the temporal association strength between nodes based on the insect infestation time series data of each monitoring node, and construct a dynamic graph topology based on the static spatial association and the temporal association strength; S4: Based on the dynamic graph topology and the multi-scale temporal features, during the continuous time state evolution process, the spatial neighborhood information of each monitoring node is synchronously aggregated into the local state update of the current node to obtain the spatiotemporal coupling representation of each monitoring node. S5: Based on the spatiotemporal coupling characterization, predict the insect infestation forecast results for each monitoring node within a future preset time period.

2. The orchard multi-point network pest situation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S2, the plurality of preset time scales include at least short-period scales, medium-period scales, and long-period scales; Feature extraction is performed on the multi-source time-series data according to multiple preset time scales, specifically as follows: The insect infestation time series data are back-looked and aggregated using the window width corresponding to each preset time scale to obtain the aggregated insect population sequence at each scale. The environmental variable time series data at each scale are then aggregated using a moving average to obtain the aggregated environmental sequence at each scale. Lag statistical features reflecting the sparsity, trend direction, and fluctuation level of insect populations are extracted from each of the aggregated insect population sequences. The aggregated insect population sequences, aggregated environmental sequences, and the lag statistical features are used together as the multi-scale time series features at the corresponding scale.

3. The orchard multi-point network pest situation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S3, constructing the dynamic graph topology specifically includes: Based on the physical distance between each monitoring node and / or terrain occlusion constraints, static spatial edge weights are generated and fixed once during system deployment. Based on the statistical correlation measure between insect infestation time series data of each monitoring node in the recent historical window, dynamic time series edge weights are generated and updated daily. The static spatial edge weights and the dynamic temporal edge weights are weighted and fused, and the fused weight matrix is ​​normalized to obtain the attention weight matrix of the dynamic graph topology.

4. The orchard multi-point network pest situation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S4, spatial neighborhood information is synchronously aggregated into the local state update during the continuous-time state evolution process, specifically as follows: The spatial neighborhood aggregation operation of the graph neural network is embedded into each internal integration step of the ordinary differential equation solver of the liquid time constant network; Within each of the internal integration steps, the hidden states of all neighboring nodes of the current node are first aggregated according to the dynamic graph topology to obtain spatial context information. Then, the aggregated spatial context information is injected into the local hidden state of the current node to update the local hidden state of the current node. Subsequently, the next integration step of the ordinary differential equation is continued.

5. The orchard multi-point network pest situation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 4, characterized in that, In step S4, the aggregated spatial context information is injected into the local hidden state of the current node, specifically as follows: A learnable gating network is used to generate a dimension-wise gating vector based on the joint state of the current node's local hidden state and the spatial context information. The intensity of the spatial context information is adjusted using the gating vector, and the adjusted spatial context information is injected into the local hidden state of the current node in a residual superposition manner to obtain the coupled state. The coupled state is used as the state variable of the ordinary differential equation of the liquid time constant network in the current integration step for continuous evolution.

6. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S4, the spatiotemporal coupling characterization of each monitoring node is obtained, specifically including: The multi-scale temporal features of each monitoring node at each preset time scale are sent into multiple encoding channels with independent parameters. Each encoding channel independently performs the continuous time state evolution and spatial neighborhood information synchronous aggregation, and outputs the spatiotemporal coupled hidden state matrix corresponding to each scale. Multi-head self-attention reweighting is performed on the spatiotemporal coupled hidden state matrix corresponding to each scale to capture the long-range dependencies within each scale. The latent vectors of the last time step after attention reweighting at each scale are extracted, spliced ​​and fused to obtain a cross-scale joint spatiotemporal representation, which serves as the spatiotemporal coupling representation of each monitoring node.

7. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S5, the insect infestation prediction results include the predicted insect quantity for each monitoring node within a preset time period. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling further includes a trend direction determination step after obtaining the predicted insect population values: Calculate the average of the predicted insect population for each monitoring node within the preset future time period, and the average of the actual insect population over the same past time period. The difference between the predicted mean and the actual mean is compared, and the insect infestation trend is determined to be an upward trend, a flat trend, or a downward trend based on the numerical relationship between the difference and a preset tolerance threshold.

8. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 1, characterized in that, In step S1, the multi-source time-series data also includes environmental variable time-series data, which includes at least temperature parameters and humidity parameters; The monitoring nodes are distributed insect monitoring lamps deployed in the orchard, and the multi-source time-series data is reported via wireless communication.

9. The orchard multi-point network insect infestation prediction method based on LTC-GNN spatiotemporal gating coupling according to claim 3, characterized in that, In step S3, the statistical correlation measure includes any one of Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information; The terrain occlusion constraint is based on digital elevation model data to determine the mountain barrier relationship between each monitoring node, and is used to attenuate or set the static spatial edge weights between monitoring nodes with terrain barriers to zero.

10. A multi-point network insect infestation prediction system for orchards based on LTC-GNN spatiotemporal gating coupling, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source time-series data of multiple monitoring nodes in the orchard during historical periods. The multi-source time-series data includes at least the insect infestation time-series data of each monitoring node. A multi-scale temporal feature construction module is used to extract features from the multi-source time-series data according to multiple preset time scales to obtain multi-scale temporal features of each monitoring node. The dynamic graph topology construction module is used to construct static spatial associations based on the spatial positional relationships between each monitoring node, and to dynamically update the temporal association strength between nodes based on the insect infestation time-series data of each monitoring node, and to construct a dynamic graph topology based on the static spatial associations and the temporal association strength. The LTC-GNN spatiotemporal gating coupling module is used to synchronously aggregate the spatial neighborhood information of each monitoring node into the local state update of the current node during the continuous time state evolution process, based on the dynamic graph topology and the multi-scale temporal features, so as to obtain the spatiotemporal coupling representation of each monitoring node. The prediction and trend output module is used to predict the insect infestation forecast results of each monitoring node in the future preset time period based on the spatiotemporal coupling characterization.

Images