A method for predicting the spatiotemporal propagation of multi-subtype avian influenza driven by migratory birds

Abstract

This application discloses a method for predicting the spatiotemporal transmission of multi-subtype avian influenza driven by migratory birds. The method includes: acquiring migratory bird movement trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data; identifying movement events of each type of migratory bird population on a population-specific subgraph based on the migratory bird movement trajectory data and key stopover location data; calculating the dynamic edge weights of directed edges in the population-specific subgraph of each type of migratory bird population based on edge weight calculation rules and movement events, obtaining a dynamically weighted population-specific subgraph; generating a node feature matrix for each historical time window based on environmental factor data and avian influenza virus detection data; inputting the node feature matrices of multiple consecutive historical time windows and the dynamically weighted population-specific subgraph into a pre-trained migratory bird multi-subtype avian influenza prediction model, and outputting the transmission probability of each avian influenza virus subtype at each key stopover during the target prediction period.

Description

Technical Field

[0001] This application relates to the fields of bioinformatics technology and animal disease prevention and control technology, and relates to, but is not limited to, a method for predicting the spatiotemporal spread of multi-subtype avian influenza driven by migratory birds. Background Technology

[0002] Migratory birds are the core vectors for the cross-regional transmission of avian influenza viruses. Globally, avian influenza outbreaks caused by migratory birds occur frequently each year, seriously threatening the healthy development of the livestock and poultry farming industry and public health security. Accurately predicting the spatiotemporal transmission risk of multi-subtype avian influenza viruses driven by migratory birds is a core prerequisite for conducting proactive epidemic prevention and control and reducing control costs. Currently, avian influenza transmission risk warnings are mostly based on discrete monitoring data from fixed-point sampling, which is difficult to adapt to the dynamic cross-regional transmission characteristics brought about by long-distance migration of birds and cannot meet the needs of refined prevention and control.

[0003] Current technologies primarily rely on static migration models or single-dimensional statistical analysis methods for risk assessment. When processing multi-source heterogeneous data, these methods exhibit significant differences in their ability to identify transmission chains among different migratory bird populations, and the epidemiological signals of some populations are difficult to capture effectively. Furthermore, the risk profiles output by existing models show significant lag or volatility over time, misaligning with the actual seasonal migration rhythms of birds. In addition, in complex scenarios involving the co-existence and transmission of multiple avian influenza virus subtypes, the risk quantification results output by existing methods show low agreement with actual detection levels, resulting in insufficient detection accuracy.

[0004] Therefore, there is an urgent need to propose a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, in order to improve the accuracy of detecting the risk of transmission of multiple subtypes of avian influenza. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, which at least solves the problem of low detection accuracy of the risk of transmission of multiple subtypes of avian influenza.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, including: Acquire migratory bird movement trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data across multiple consecutive historical time windows; Based on the migratory bird movement trajectory data and the key stopover location data, the movement events of each migratory bird population within each historical time window on a pre-constructed population-specific subgraph are identified; the nodes of the pre-constructed population-specific subgraph are key stopovers, and the directed edges are the connections between two key stopovers that have direct movement paths; the movement event is the behavior of each migratory bird population flying directly from one key stopover to another without passing through other key stopovers in between; Based on the predetermined edge weight calculation rules, according to the movement event, the dynamic edge weight of each directed edge in the population-specific subgraph of each migratory bird population in each historical time window is calculated to obtain the dynamically weighted population-specific subgraph of each historical time window; the dynamic edge weight represents the movement probability of each migratory bird population between two key stopover sites. Based on the environmental factor data and the avian influenza virus detection data, a node feature matrix is ​​generated for each of the historical time windows. The node feature matrices of multiple consecutive historical time windows and the dynamically weighted population-specific subgraphs of each migratory bird population in multiple consecutive historical time windows are input into a pre-trained migratory bird multi-subtype avian influenza prediction model, which outputs the transmission probability of each avian influenza virus subtype at each key stopover point during the target prediction period.

[0007] The beneficial effects of the technical solutions provided in this application include at least the following: This application acquires migratory bird migration trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data across multiple consecutive historical time windows. Based on the migratory bird migration trajectory data and key stopover location data, it identifies the movement events of each migratory bird population within each historical time window on a pre-constructed population-specific subgraph. The nodes of the pre-constructed population-specific subgraph represent key stopover locations, and directed edges indicate direct movement paths between two key stopover locations. Movement events are defined as the behavior of each migratory bird population flying directly from one key stopover location to another without passing through other key stopover locations. Because the nodes of the population-specific subgraph are key stopover locations rather than administrative regions, the ecological continuity of migratory bird migration routes is preserved, avoiding the disruption of transmission paths by administrative boundaries. Furthermore, by constructing separate population-specific subgraphs and identifying corresponding movement events for each migratory bird population, the differences in migration routes, movement rhythms, and multi-subtype avian influenza virus carrying patterns among different migratory bird populations can be captured separately, avoiding information confusion caused by mixed modeling of multiple populations. Based on predetermined edge weight calculation rules, the dynamic edge weights of each directed edge in the population-specific subgraph of each migratory bird population within each historical time window are calculated according to migration events, resulting in a dynamically weighted population-specific subgraph for each historical time window. The dynamic edge weights represent the probability of each migratory bird population moving between two key stopover sites. By dynamically adjusting the edge weights of the population-specific subgraph according to the seasonal changes in migratory bird movement behavior—increasing edge weights in months with active migratory bird migration and decreasing them in winter when migratory birds do not move—the bias of static edge weights predicting high transmission risk in winter is avoided. Since dynamic edge weights are calculated separately for each migratory bird population, the differences in migration rhythms, movement rates, and stopover durations among different migratory bird populations are reflected separately, improving the accuracy of migratory bird movement probability calculations. The dynamically weighted population-specific subgraph provides a time-varying graph structure input for subsequent spatiotemporal graph neural networks, enabling the model to capture the seasonal fluctuations in avian influenza virus transmission risk. Based on environmental factor data and avian influenza virus detection data, a node feature matrix for each historical time window is generated. This method transforms environmental factor data and avian influenza virus detection data into numerical inputs that can be processed by a spatiotemporal graph neural network, preserving environmental conditions and historical epidemic information at key stopover sites. Each row in the node feature matrix corresponds to a feature vector of a key stopover site, providing a data foundation for subsequent models to simultaneously utilize environmental factors and historical virus transmission patterns for prediction. The node feature matrices of multiple consecutive historical time windows, along with dynamically weighted population-specific subgraphs of each migratory bird population across multiple consecutive historical time windows, are input into a pre-trained multi-subtype avian influenza prediction model for migratory birds. The model outputs the transmission probability of each avian influenza virus subtype at each key stopover site within the target prediction period. By utilizing a pre-trained spatiotemporal graph neural network to output the transmission probability of multiple subtype avian influenza viruses, multi-subtype collaborative prediction is achieved.Because the model input includes a dynamically weighted population-specific subgraph, the prediction results can reflect the impact of seasonal variations in migratory bird behavior on the spread of different subtypes. The output transmission probability can be directly used to classify risk levels, providing a quantitative basis for avian influenza virus monitoring and early warning. Therefore, it solves the problem of low detection accuracy for the risk of multi-subtype avian influenza transmission. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, provided in an embodiment of this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0011] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0012] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0013] This application provides a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds. Figure 1 A flowchart illustrating a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes at least the following steps: Step S110: Obtain migratory bird movement trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data from multiple consecutive historical time windows.

[0014] Migratory bird movement trajectory data consists of the location sequences of individual migratory birds collected by satellite trackers. This data includes longitude, latitude, altitude, instantaneous speed, heading angle, and location timestamp. The data determines the spatial and temporal movement paths of migratory birds. The migratory bird population includes at least two species, such as the bar-headed goose and the common gull. At least 10 individuals of each species are tracked, and the data collection period continuously covers at least two full years of migratory bird migration cycles.

[0015] Environmental factor data, acquired through remote sensing platforms, includes monthly average temperature, monthly cumulative precipitation, monthly normalized difference vegetation index (NDI), and monthly average water coverage index for the location of key stopover sites. This environmental factor data describes the ecological and environmental conditions of the key stopover sites.

[0016] Avian influenza virus detection data were collected monthly at key stopover sites along migratory bird flyways. Each stopover included fresh fecal swabs and / or water samples, with a minimum of 30 samples collected per stopover. Avian influenza virus nucleic acid was detected using quantitative real-time polymerase chain reaction (qPCR) in a biosafety laboratory to identify subtypes such as H5, H9, H4, H13, and H3. Results were recorded in the format of key stopover site, sampling month, virus subtype, and number of positive samples, along with the total number of samples collected each time. The avian influenza virus detection data reflects the historical prevalence of each subtype of avian influenza virus at each key stopover site.

[0017] The location data of key stopover sites include the name, longitude coordinates, latitude coordinates, and ecological function type of each key stopover site.

[0018] Consecutive historical time windows are a series of consecutive months divided according to calendar months, with each historical time window corresponding to one calendar month, such as January, February, or March. These consecutive historical time windows are sequential in time, for example, the first, second, third, and so on up to the 24th month of the past 24 months. Each historical time window contains migratory bird movement data, environmental factor data, and avian influenza virus detection data for that month. This data organization of consecutive historical time windows provides time-series input for subsequent models, enabling them to learn the changing patterns of multi-subtype avian influenza transmission risk over time.

[0019] Step S120: Based on the migratory bird movement trajectory data and the key stopover location data, identify the movement events of each migratory bird population on a pre-constructed population-specific subgraph within each historical time window; the nodes of the pre-constructed population-specific subgraph are key stopovers, and the directed edges are the connections between two key stopovers with direct movement paths; the movement event is the behavior of each migratory bird population flying directly from one key stopover to another without passing through other key stopovers in between.

[0020] The pre-constructed population-specific subgraph node set consists of all key stopover sites actually used by each migratory bird population during at least two full-year migration cycles, and remains fixed across all historical time windows. The node sets of the population-specific subgraphs for different migratory bird populations are constructed and maintained independently, and partial overlap is allowed. If a key stopover site is not actually visited by a migratory bird population within a certain historical time window, the node feature vector corresponding to that key stopover site is filled with zero values.

[0021] Key stopover sites are breeding grounds, intermediate stopover sites, or wintering grounds with clearly defined geographical coordinates used by migratory birds during their migration, such as Bird Island in Qinghai Lake and Wuliangsuhai Lake in Inner Mongolia. Each type of migratory bird population refers to a group of migratory birds classified by species.

[0022] The pre-constructed population-specific subgraph is a graph network built separately for each type of migratory bird population. The nodes of the population-specific subgraph are the key stopover sites actually used by that type of migratory bird population.

[0023] A directed edge is a connection between two key stopover sites identified based on the migratory bird movement trajectory data, indicating a direct movement path between them. Specifically, if, in the migratory bird movement trajectory data of a certain bird population, there exists at least one individual whose trajectory point moves directly from key stopover site A to key stopover site B without passing through other key stopover sites along the way, then a directed edge from node A to node B is established in the corresponding population-specific subgraph. Once established, this directed edge maintains its topological existence across all historical time windows and does not change with the time window.

[0024] By identifying migration events based on migratory bird trajectory data and key stopover location data, continuous migratory bird trajectories are transformed into migration events on a population-specific subgraph. These migration events record the behavior of each migratory bird population flying directly from one key stopover to another within a specific historical time window, providing a statistically valid event count basis for subsequent calculation of the dynamic edge weights of directed edges.

[0025] Because the nodes of the population-specific subgraph are key stopover sites rather than administrative regions, the ecological continuity of migratory bird routes is preserved, avoiding the disruption of transmission paths by administrative boundaries. Furthermore, by constructing separate population-specific subgraphs and identifying corresponding movement events for each migratory bird population, the differences in migration routes, movement rhythms, and multi-subtype avian influenza virus carrying patterns among different migratory bird populations can be captured separately, avoiding information confusion caused by modeling multiple populations together.

[0026] Step S130: Based on the predetermined edge weight calculation rules, according to the movement event, calculate the dynamic edge weight of each directed edge in the population-specific subgraph of each migratory bird population within each historical time window, and obtain the dynamically weighted population-specific subgraph of each historical time window; the dynamic edge weight represents the movement probability of each migratory bird population between two key stopover sites.

[0027] The predetermined rule for calculating edge weights is to count the total number of movement events originating from the starting key rest stop within each time window, and the number of directional movement events reaching the destination key rest stop. The weight of the directed edge is obtained by dividing the latter by the former. If the total number of events is zero, then all outgoing edges from the starting key rest stop are set to the same weight.

[0028] Dynamic edge weights are moving probability values ​​that change over a time window. For example, during the spring migration months of migratory birds, the edge weight from Bird Island in Qinghai Lake to Wuliangsuhai Lake is relatively high, while in winter, the edge weight from Bird Island in Qinghai Lake to Wuliangsuhai Lake is close to zero.

[0029] The dynamically weighted population-specific subgraph is a subgraph for each migratory bird population within each time window, with a set of edge weights corresponding to that time window.

[0030] By dynamically adjusting the edge weights of the population-specific subgraph according to the seasonal changes in migratory bird behavior—increasing edge weights during months of active migration and decreasing them during the winter when migratory birds are not moving—the bias of static edge weights predicting high transmission risk in winter is avoided. Since dynamic edge weights are calculated separately for each migratory bird population, differences in migration rhythms, movement rates, and dwell times among different populations are reflected separately, improving the accuracy of migratory bird movement probability calculations. The dynamically weighted population-specific subgraph provides a time-varying graph structure input for subsequent spatiotemporal graph neural networks, enabling the model to capture the seasonal fluctuations in avian influenza virus transmission risk.

[0031] Step S140: Based on the environmental factor data and the avian influenza virus detection data, generate a node feature matrix for each of the historical time windows.

[0032] The node feature matrix is ​​a matrix formed by concatenating the node feature vectors of all key stopover sites within that time window. Each node feature vector is a fixed-dimensional numerical vector obtained by encoding environmental factor data and avian influenza virus detection data of the key stopover sites. For example, for a key stopover site in a certain month, the temperature, precipitation, vegetation index, water coverage index of that month, and the positivity rate of each subtype in the past few months are concatenated and mapped to a 128-dimensional vector through a fully connected layer.

[0033] This approach transforms environmental factor data and avian influenza virus detection data into numerical inputs that can be processed by a spatiotemporal graph neural network, while preserving environmental conditions and historical epidemic information of key stopover sites. Each row in the node feature matrix corresponds to a feature vector of a key stopover site, providing a data foundation for subsequent models to make predictions using both environmental factors and historical virus transmission patterns.

[0034] Step S150: Input the node feature matrix of multiple consecutive historical time windows and the dynamically weighted population-specific subgraph of each type of migratory bird population in multiple consecutive historical time windows into the pre-trained migratory bird multi-subtype avian influenza prediction model, and output the transmission probability of each avian influenza virus subtype at each key stopover point during the target prediction period.

[0035] The pre-trained migratory bird multi-subtype avian influenza prediction model is a spatiotemporal graph neural network. The model takes node feature matrices from multiple consecutive historical time windows and dynamically weighted population-specific subgraphs as input. After spatial and temporal feature extraction, it outputs the transmission probability of each avian influenza virus subtype at each key stopover point in the next time window, with the probability ranging from 0 to 1. The target prediction period is the time window following the input historical time window; for example, if the input is data from the previous 12 months, the output is data for the next month. The transmission probability of each avian influenza virus subtype refers to the probability that subtypes such as H5, H9, and H4 may be prevalent at that key stopover point.

[0036] By utilizing a pre-trained spatiotemporal graph neural network, the transmission probabilities of multiple subtypes of avian influenza virus are output, achieving collaborative prediction of multiple subtypes. Since the model input includes dynamically weighted population-specific subgraphs, the prediction results can reflect the impact of seasonal variations in migratory bird behavior on the transmission of different subtypes. The output transmission probabilities can be directly used to classify risk levels, providing a quantitative basis for avian influenza virus monitoring and early warning.

[0037] This application acquires migratory bird migration trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data across multiple consecutive historical time windows. Based on the migratory bird migration trajectory data and key stopover location data, it identifies the movement events of each migratory bird population within each historical time window on a pre-constructed population-specific subgraph. The nodes of the pre-constructed population-specific subgraph represent key stopover sites, and directed edges indicate direct movement paths between two key stopover sites. Movement events are defined as the behavior of each migratory bird population flying directly from one key stopover site to another without passing through other key stopover sites. Because the nodes of the population-specific subgraph are key stopover sites rather than administrative regions, the ecological continuity of migratory bird migration routes is preserved, avoiding the disruption of transmission paths by administrative boundaries. Furthermore, by constructing separate population-specific subgraphs for each migratory bird population and identifying corresponding movement events, the differences in migration routes, movement rhythms, and multi-subtype avian influenza virus carrying patterns among different migratory bird populations can be captured separately, avoiding information confusion caused by mixed modeling of multiple populations. Based on predetermined edge weight calculation rules, the dynamic edge weights of each directed edge in the population-specific subgraph of each migratory bird population within each historical time window are calculated according to migration events, resulting in a dynamically weighted population-specific subgraph for each historical time window. The dynamic edge weights represent the probability of each migratory bird population moving between two key stopover sites. By dynamically adjusting the edge weights of the population-specific subgraph according to the seasonal changes in migratory bird movement behavior—increasing edge weights in months with active migratory bird migration and decreasing them in winter when migratory birds do not move—the bias of static edge weights predicting high transmission risk in winter is avoided. Since dynamic edge weights are calculated separately for each migratory bird population, the differences in migration rhythms, movement rates, and stopover durations among different migratory bird populations are reflected separately, improving the accuracy of migratory bird movement probability calculations. The dynamically weighted population-specific subgraph provides a time-varying graph structure input for subsequent spatiotemporal graph neural networks, enabling the model to capture the seasonal fluctuations in avian influenza virus transmission risk. Based on environmental factor data and avian influenza virus detection data, a node feature matrix for each historical time window is generated. Environmental factor data and avian influenza virus detection data were transformed into numerical inputs that could be processed by a spatiotemporal graph neural network, preserving environmental conditions and historical epidemic information of key stopover sites. Each row in the node feature matrix corresponds to a feature vector of a key stopover site, providing a data foundation for subsequent models to simultaneously utilize environmental factors and historical virus transmission patterns for prediction. The node feature matrices of multiple consecutive historical time windows, along with dynamically weighted population-specific subgraphs of each migratory bird population across multiple consecutive historical time windows, were input into a pre-trained multi-subtype avian influenza prediction model for migratory birds. This model outputs the transmission probability of each avian influenza virus subtype at each key stopover site within the target prediction period. By utilizing a pre-trained spatiotemporal graph neural network to output the transmission probability of multiple subtype avian influenza viruses, multi-subtype collaborative prediction was achieved.Because the model input includes a dynamically weighted population-specific subgraph, the prediction results reflect the impact of seasonal variations in migratory bird behavior on the spread of different subtypes. The output transmission probability can be directly used to classify risk levels, providing a quantitative basis for avian influenza virus monitoring and early warning. Therefore, it solves the problem of low detection accuracy for the risk of multi-subtype avian influenza transmission.

[0038] Optionally, before identifying the movement events of each migratory bird population on a pre-constructed population-specific submap within each historical time window based on the migratory bird movement trajectory data and the key stopover location data, the method further includes: matching each trajectory point in the migratory bird movement trajectory data to the nearest key stopover location based on the key stopover location data; if the spatial distance between the trajectory point and any key stopover location exceeds a preset threshold, then the trajectory point is excluded; otherwise, the trajectory point is assigned to the nearest key stopover location; summarizing the avian influenza virus detection data according to the historical time window and key stopover locations based on the key stopover location data, and counting the number of positive detections of each avian influenza virus subtype at each key stopover location within each historical time window; and extracting the raster values ​​of the location of each key stopover location within each historical time window from the environmental factor data based on the key stopover location data.

[0039] Optionally, the directed edge connection between any two nodes in the population-specific subgraph is predetermined based on historical migratory bird movement trajectory data, and the topology remains fixed within all historical time windows, with only the dynamic edge weight of each directed edge being dynamically adjusted.

[0040] Because migratory bird routes exhibit interannual stability, the same migratory bird population utilizes the same migration routes and key stopover sites during the same season each year. Therefore, by analyzing migratory bird movement trajectory data over at least one complete annual migration cycle, the direct movement relationships between key stopover sites for that migratory bird population can be determined. These direct movement relationships do not fundamentally change in subsequent time windows. If the actual movement trajectory of a migratory bird population does not cover certain directed edges within a certain time window, it only indicates that the population did not utilize that migration route during that time window, but the topological connectivity of the route still exists and can still be used when the migratory birds migrate again.

[0041] Therefore, by fixing the topology of directed edges and only dynamically adjusting their weights, the number of model parameters is reduced, the stability of the graph network structure is ensured, and the problem of frequent additions and deletions of directed edges due to data sparsity is avoided.

[0042] Optionally, the step of calculating the dynamic edge weight of each directed edge in the population-specific subgraph of each migratory bird population within each historical time window, based on a predetermined edge weight calculation rule and according to the movement events, includes: for each directed edge in the population-specific subgraph, the starting point of each directed edge is the departure key resting place, and the ending point of each directed edge is the destination key resting place. The total number of all movement events starting from the departure key resting place within the historical time window is counted; the number of orientations of movement events starting from the departure key resting place and ending at the destination key resting place within the historical time window is counted; the ratio of the orientation count to the total count is used as the dynamic edge weight of the directed edge; when the total count is zero, the dynamic edge weights of all outgoing edges from the departure key resting place are set to equal values, and the sum of the dynamic edge weights of all outgoing edges is 1.

[0043] In a population-specific subgraph, directed edges are unidirectional connections from one node to another. For example, a directed edge from node A to node B and a directed edge from node A to node C are both outgoing edges originating from node A.

[0044] For each directed edge in the population-specific subgraph that points from the starting key resting point to the destination key resting point, the dynamic edge weight is calculated independently for each edge. The dynamic edge weight of each directed edge is calculated only based on the movement event statistics corresponding to that directed edge itself, and does not affect the calculation of other directed edges.

[0045] This counts the total number of movement events originating from this key stop within the given historical time window. This total count refers to the number of all movement events originating from this key stop within the given historical time window, regardless of the destination key stop. For example, for a directed edge originating from node A, all movement events originating from node A are counted, regardless of whether the destination is node B, node C, or any other node.

[0046] From the total number of counts mentioned above, the number of directed movement events whose destination is the critical rest stop is selected. This number is the number of movement events in the total count where the critical rest stop is exactly the critical rest stop pointed to by the directed edge. For example, for a directed edge from node A to node B, the number of directed movement events whose destination is node B is selected from the total count.

[0047] The dynamic edge weight of a directed edge is obtained by dividing the number of migration events whose destination is the key stopover point by the total number of migration events originating from that key stopover point. This dynamic edge weight is a probability value ranging from 0 to 1, representing the proportion of migration events that choose the migration path corresponding to that directed edge among all migration events originating from that key stopover point within the current historical time window. A higher dynamic edge weight indicates a higher frequency of use of that migration path within the current time window, and more active migratory bird movement.

[0048] When the total number of movement events originating from a critical stop is zero, it means that no movement events were observed from that critical stop within the historical time window. In this case, the dynamic edge weights of all outgoing edges from that critical stop are set to equal values, and the sum of the dynamic edge weights of all outgoing edges is 1. For example, if the total number of movement events originating from node A is zero, and node A has k outgoing edges, then the dynamic edge weight of each outgoing edge is set to 1 divided by k, and the sum of the weights of all outgoing edges equals 1. This approach avoids graph structure breakage due to lack of observation data, preserves the topological integrity of the migration path, satisfies the probability normalization requirement, and enables the model to maintain a reasonable propagation risk assessment even during periods without observation, avoiding the misjudgment of zero propagation risk due to lack of observation data.

[0049] Optionally, generating a node feature matrix for each historical time window based on the environmental factor data and the avian influenza virus detection data includes: for each historical time window, generating a node feature vector for each key stopover site within the historical time window based on the environmental factor data and the avian influenza virus detection data and a predetermined node feature coding rule; and forming a node feature matrix for the historical time window based on each node feature vector; wherein the node feature vector is used to represent the environmental conditions and historical avian influenza virus detection status of the key stopover site.

[0050] Optionally, for each historical time window, generating a node feature vector for each key stopover site within the historical time window based on the environmental factor data and the avian influenza virus detection data, according to a predetermined node feature encoding rule, includes: extracting the environmental factor vector of the key stopover site within the historical time window based on the environmental factor data; extracting the positivity rate of each avian influenza virus subtype in the key stopover site within a predetermined number of consecutive historical time windows prior to the historical time window based on the avian influenza virus detection data, wherein the positivity rate is the ratio of the number of positive detections of that subtype in the corresponding window to the total number of samples, to obtain a historical virus detection vector; if there are missing samples, filling them with the positivity rate of the previous effective window of the key stopover site or the average positivity rate of the historical windows of the same period of the key stopover site; concatenating the environmental factor vector and the historical virus detection vector into an original feature vector; and transforming the original feature vector into a fixed-dimensional node embedding vector through a nonlinear mapping, which serves as the node feature vector of the key stopover site.

[0051] Preferably, the environmental factor vector is composed of four quantitative indicators: temperature, precipitation, normalized vegetation index, and water coverage index. Each quantitative indicator is obtained from the monthly value of the corresponding grid cell of the key stopover site through a remote sensing data platform, which can fully characterize the climate, vegetation and hydrological conditions of the key stopover site and improve the accuracy of model prediction.

[0052] The historical virus detection vector consists of the positivity rate of each avian influenza virus subtype. The positivity rate is obtained by statistically analyzing the ratio of the number of positive detections of each subtype to the total number of samples within a predetermined number of consecutive historical time windows at the key stopover site. This reflects the recent level of virus prevalence at the key stopover site. Due to potential sampling gaps, to ensure data continuity, the positivity rate of the previous valid historical time window or the average positivity rate of the same historical window is used to fill in the gaps, avoiding incomplete features due to missing data.

[0053] The environmental factor vector and the historical virus detection vector are concatenated to form the original feature vector, which contains both environmental and epidemic information. This original feature vector is then non-linearly mapped through a fully connected layer and an activation function, transforming it into a node feature vector of a pre-defined dimension. The purpose of this mapping is to unify the original data of different dimensions into a fixed-dimensional feature space, facilitating subsequent model processing. The weight matrix and bias vector of the fully connected layer automatically learn the optimal mapping method during model training.

[0054] For each historical time window, the above operations are performed on all key rest stops to obtain the node feature vector for each key rest stop. The node feature vectors of all key rest stops are then concatenated row-wise to form the node feature matrix for that historical time window. The number of rows in the node feature matrix is ​​the total number of key rest stops, and the number of columns is the preset embedding dimension. This matrix serves as one of the inputs to the subsequent graph convolutional spatial feature extraction layer to describe the state of each key rest stop in the current time window.

[0055] Optionally, the pre-trained migratory bird multi-subtype avian influenza prediction model includes a graph convolutional spatial feature extraction layer, a gated recurrent unit temporal feature extraction layer, and a multi-subtype transmission probability prediction layer. Within each historical time window, the graph convolutional spatial feature extraction layer performs graph convolution operations based on the adjacency matrix of the dynamically weighted population-specific subgraph and the node feature matrix to aggregate the neighbor node information of each key stopover site, generating a spatial feature vector for each key stopover site. The spatial feature vectors corresponding to the same key stopover site within multiple consecutive historical time windows are then sorted by time. The input sequence is composed of sequential components. The temporal features of the input sequence are extracted through the gated recurrent unit temporal feature extraction layer to generate a spatiotemporal fusion feature vector for each key stopover site. Through the multi-subtype propagation probability prediction layer, the parameters of the graph convolutional spatial feature extraction layer and the gated recurrent unit temporal feature extraction layer are reused using a hard parameter sharing mechanism. The spatiotemporal fusion feature vector of each key stopover site is mapped to the probability distribution of each avian influenza virus subtype. After normalization, the propagation probability of each avian influenza virus subtype at each key stopover site within the target prediction period is output.

[0056] Based on predetermined node feature encoding rules, node feature vectors for each key resting point in each historical time window are generated for subsequent network layers to process.

[0057] For a key resting point v within a historical time window t, extract the environmental factor vector for that historical time window. Simultaneously, extract the historical virus detection vector. The historical virus detection vector consists of the positivity rates of M avian influenza virus subtypes.

[0058] The positivity rate is calculated as follows: The number of positive detections of each avian influenza subtype detected at the key stopover location v within K consecutive historical time windows prior to historical time window t is counted, where K is the preset backtesting window length. The total number of samples within each historical time window is also recorded. The positivity rate for each subtype is the number of positive detections divided by the total number of samples. The dimension of the historical virus detection vector is M, where M is the total number of virus subtypes detected. If no sampling is performed within a historical time window, the positivity rate for the corresponding subtype is taken from the previous valid historical time window. If consecutive missing samples exceed a preset threshold (the preset threshold is 3 historical time windows), the average positivity rate of the same historical time window at that key stopover location is used.

[0059] The environmental factor vector and the historical virus detection vector are concatenated to form a 4+M dimensional original feature vector. This original feature vector is then mapped to a d-dimensional node feature vector through a fully connected layer and a ReLU activation function, where d is a preset embedding dimension. The weight matrix and bias vector of the fully connected layer are learnable parameters.

[0060] The graph convolutional spatial feature extraction layer uses graph convolution operations to aggregate neighbor node information of key stopover sites within each historical time window, extracting spatial features of virus spread along the migratory bird network.

[0061] Within each historical time window t, the node feature vectors of all generated key resting places are concatenated row-wise to form the node feature matrix for that historical time window, denoted as X raised to the power of t. The node feature matrix and the adjacency matrix of the dynamically weighted population-specific subgraph for that historical time window, denoted as A raised to the power of t, are then input into the graph convolutional network.

[0062] The graph convolution operation is calculated as follows: The adjacency matrix A raised to the power of t of the current historical time window is added to the identity matrix I, allowing the key resting points' own features to participate in the aggregation, resulting in a matrix containing self-loops, denoted as A raised to the power of t plus I. The matrix A raised to the power of t plus I is then normalized using the inverse of the out-degree diagonal matrix D of A raised to the power of t plus I. The normalized matrix is ​​multiplied by the node feature matrix X raised to the power of t and the learnable weight matrix W, and finally output through an activation function to obtain the spatial feature matrix Z raised to the power of t. Each row of the spatial feature matrix Z raised to the power of t corresponds to a spatial feature vector of a key resting point after fusing neighbor information. The weight matrix W of the graph convolutional layer consists of learnable parameters.

[0063] The role of the gated recurrent unit temporal feature extraction layer is to receive a matrix sequence composed of spatial feature vectors from multiple consecutive historical time windows, and to learn the temporal dependency of the propagation risk over time through the gated recurrent unit network.

[0064] The spatial feature matrices of L consecutive historical time windows are organized into a time series based on key rest locations. For each key rest location v, the spatial feature vectors of the L consecutive historical time windows are arranged chronologically into a sequence and input into a gated recurrent unit (GRU) network. The GRU adaptively retains or forgets information from different historical time steps through reset and update gate mechanisms, outputting the spatiotemporal fusion feature vector of that key rest location in the last historical time window. All weight matrices and bias vectors of the GRU are learnable parameters.

[0065] The multi-subtype transmission probability prediction layer maps the spatiotemporal fusion feature vectors of each key stopover site to the transmission probability of each avian influenza virus subtype. Each virus subtype shares the aforementioned spatiotemporal feature extraction layer. Through a hard parameter sharing mechanism of multi-task learning, the prediction tasks of different subtypes are mutually regularized during the learning process, implicitly capturing the epidemiological association patterns of each subtype in the same ecological network, such as co-circulation or substitution relationships.

[0066] The spatiotemporal fusion feature vectors of each key stopover point, output from the time feature extraction layer of the gated recurrent unit, are input into the fully connected layer, outputting an M-dimensional vector. After normalization by the sigmoid function, each element of this M-dimensional vector is a value between 0 and 1, corresponding to the predicted transmission probability of M avian influenza subtypes in the next historical time window at that key stopover point. The weight matrix and bias vector of this fully connected layer are learnable parameters.

[0067] Optionally, the training process of the pre-trained migratory bird multi-subtype avian influenza prediction model includes: acquiring historical migratory bird movement trajectory data, historical environmental factor data, and historical avian influenza virus detection data for historical training periods as training datasets; the historical training periods are earlier than the historical time window; constructing an initial framework for the migratory bird multi-subtype avian influenza prediction model; the initial framework is a multi-task spatiotemporal graph neural network oriented towards dynamically weighted population-specific subgraphs, including a graph convolutional spatial feature extraction layer, a gated recurrent unit temporal feature extraction layer, and a multi-subtype transmission probability prediction layer, with each avian influenza virus subtype sharing the parameters of the spatial feature extraction layer and the temporal feature extraction layer; preprocessing the training dataset to obtain a historical node feature matrix within the historical training period, and the corresponding dynamically weighted historical population-specific subgraph for each type of migratory bird population; constructing training samples using a sliding window method, and... The historical node feature matrix and dynamically weighted historical population-specific subgraph of the historical training period are used as input features. The true value of the transmission probability of each virus subtype at each key stopover point in the next time window of the historical training period is used as the corresponding label. The training samples are divided into a training subset and a validation subset. The initial framework is trained using the training subset: the binary cross-entropy loss between the predicted probability of each subtype output by the initial framework and the corresponding label is calculated, and the binary cross-entropy loss of each virus subtype is summed to obtain the total loss. The gradient of the total loss with respect to each model parameter of the initial framework is calculated using the backpropagation algorithm, and the parameters of the initial framework are updated using the gradient. The performance of the initial framework is evaluated using the validation subset. The above training steps are repeated until the decreasing trend of the total loss on the validation subset meets the preset convergence condition, thus obtaining the migratory bird multi-subtype avian influenza prediction model.

[0068] The labels for the training samples are determined based on historical avian influenza virus detection data. For each avian influenza virus subtype at each key stopover point during the target prediction period, if at least one positive sample is detected at that key stopover point during that period, the corresponding label is set to 1; otherwise, it is set to 0. The labels are used to calculate the binary cross-entropy loss between the predicted probability and the actual probability.

[0069] During model training, the Adam optimizer is used to update the model parameters. Each iteration through all training samples is called a training epoch. The training process iterates through multiple training epochs until the loss function stops decreasing on the validation set, at which point the model has reached convergence.

[0070] After training, the parameter values ​​of each layer and the model structure configuration are saved as a model file. The parameter values ​​of each layer include the weight matrix and bias vector of the fully connected layer during node feature vector generation, the weight matrix of the graph convolutional spatial feature extraction layer, the weight matrix and bias vector of the gated recurrent unit temporal feature extraction layer, and the weight matrix and bias vector of the fully connected layer in the multi-subtype propagation probability prediction layer. The model structure configuration includes the embedding dimension, the number of graph convolutional layers, the hidden layer dimension of the gated recurrent unit, the input sequence length, the backtracking window length, and the number of virus subtypes.

[0071] The trained model can be deployed to real-world application environments. Each time a prediction is executed, the latest input data is prepared according to steps S110 to S140, and the model is called to perform a forward calculation once to obtain the transmission probability of each avian influenza virus subtype at each key resting point during the target prediction period, without the need for retraining.

[0072] The following describes a method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, based on a specific embodiment. This specific embodiment is merely for illustrating the invention and does not constitute an undue limitation thereof.

[0073] To verify the validity of this application, four comparative experiments were conducted based on GPS tracking data of 15 bar-headed geese and 12 gulls in the Qinghai Lake area from 2020 to 2022, at least 30 environmental samples per month from each key stopover site, and remote sensing data. Time-series cross-validation was used, and the evaluation metrics included precision, recall, F1 score, AUC, and consistency of propagation path.

[0074] The first set of comparative experiments compared ecological nodes and administrative region nodes. The ecological resting place nodes in this application were replaced with county-level administrative region nodes, with edge weights dynamically updated and the rest of the structure remaining unchanged. As shown in Table 1, this application outperforms administrative region nodes in both path consistency and cross-regional recall.

[0075] Table 1. Comparison of propagation paths and recall rates between key stopover nodes and administrative region nodes.

[0076] The second set of comparative experiments compares dynamic edge weights with static edge weights. The dynamic moving probability in this application is replaced with the interannual average static probability, while the rest of the structure remains unchanged. As shown in Table 2, compared to the static edge weight scheme, the dynamic moving probability introduced in this application significantly improves the peak period recall rate and effectively reduces the false positive rate in winter.

[0077] Table 2 Performance Comparison of Dynamic Edge Weight and Static Edge Weight Models

[0078] This application constructs separate population-specific subgraphs for the migration routes of bar-headed geese and black-necked gulls; while the control group experiment merges the GPS tracking data of the two types of migratory birds to construct a single hybrid graph network, with the remaining network structure, feature extraction, and prediction process remaining consistent. As shown in Table 3, this application can effectively deconstruct the migration differences between species and achieve higher F1 scores in the prediction of migration routes for both types of migratory birds.

[0079] Table 3. Comparison of prediction performance between population-specific subgraphs and mixed single graphs.

[0080] Table 3 shows the difference in predictive performance between the two network construction strategies across different species populations. The experimental results show that the prediction performance using population-specific sub-maps is significantly better than that using mixed single-maps. Specifically, for the bar-headed goose population, the average F1 score of this application reached 0.895, an improvement of 19.3% compared to the mixed single-map; for the gull population, the F1 score of this application was 0.85, an improvement of 25.9% compared to the mixed single-map.

[0081] The above results indicate significant differences among different migratory bird species in their migration route selection, stopover site preferences, and activity rhythms. This application, by deconstructing the heterogeneity among populations and constructing targeted migration networks, can more accurately characterize the risk transmission patterns of specific species, thereby avoiding the information masking effect caused by data merging and verifying the ecological rationality of this application at the spatial modeling level.

[0082] It should be noted that, in the embodiments of this application, if the above-mentioned method for predicting the spatiotemporal transmission of multi-subtype avian influenza driven by migratory birds is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0083] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the steps in any of the migratory bird migration-driven spatiotemporal transmission prediction methods for avian influenza in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in any of the migratory bird migration-driven spatiotemporal transmission prediction methods for avian influenza in the above embodiments.

[0084] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0085] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0087] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0088] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0089] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0090] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the spatiotemporal transmission of multiple subtypes of avian influenza driven by migratory birds, characterized in that, The method includes: Acquire migratory bird movement trajectory data, environmental factor data, key stopover location data, and avian influenza virus detection data across multiple consecutive historical time windows; Based on the migratory bird movement trajectory data and the key stopover location data, the movement events of each migratory bird population within each historical time window on a pre-constructed population-specific subgraph are identified; the nodes of the pre-constructed population-specific subgraph are key stopovers, and the directed edges are the connections between two key stopovers that have direct movement paths; the movement event is the behavior of each migratory bird population flying directly from one key stopover to another without passing through other key stopovers in between; Based on the predetermined edge weight calculation rules, according to the movement event, the dynamic edge weight of each directed edge in the population-specific subgraph of each migratory bird population in each historical time window is calculated to obtain the dynamically weighted population-specific subgraph of each historical time window; the dynamic edge weight represents the movement probability of each migratory bird population between two key stopover sites. Based on the environmental factor data and the avian influenza virus detection data, a node feature matrix is ​​generated for each of the historical time windows. The node feature matrices of multiple consecutive historical time windows and the dynamically weighted population-specific subgraphs of each migratory bird population in multiple consecutive historical time windows are input into a pre-trained migratory bird multi-subtype avian influenza prediction model, which outputs the transmission probability of each avian influenza virus subtype at each key stopover point during the target prediction period.

2. The method according to claim 1, characterized in that, Before identifying the movement events of each migratory bird population on a pre-constructed population-specific submap within each historical time window based on the migratory bird movement trajectory data and the key stopover location data, the method further includes: Based on the key rest stop location data, each trajectory point in the migratory bird movement trajectory data is matched to the nearest key rest stop. If the spatial distance between the trajectory point and any key rest stop exceeds a preset threshold, the trajectory point is excluded; otherwise, the trajectory point is assigned to the nearest key rest stop. Based on the key stopover location data, the avian influenza virus detection data are summarized according to the historical time window and key stopover location, and the number of positive detections of each avian influenza virus subtype in each key stopover location within each historical time window is counted. Based on the key rest stop location data, the raster values ​​of the location of each key rest stop within each historical time window are extracted from the environmental factor data.

3. The method according to claim 1, characterized in that, The directed edge connection between any two nodes in the population-specific subgraph is predetermined based on historical migratory bird movement trajectory data, and the topology remains fixed throughout all historical time windows, with only the dynamic edge weight of each directed edge being dynamically adjusted.

4. The method according to claim 1, characterized in that, The method, based on a pre-determined edge weight calculation rule, calculates the dynamic edge weight of each directed edge in the population-specific subgraph of each migratory bird population within each historical time window, according to the movement event, including: For each directed edge in the population-specific subgraph, the starting point of each directed edge is the starting critical resting point, and the ending point of each directed edge is the destination critical resting point: The total number of all movement events originating from the aforementioned key stop point within the historical time window is counted. The number of orientation events for movement events that start from the departure key stop and end at the destination key stop within the historical time window is counted. The ratio of the number of directed edges to the total number of directed edges is used as the dynamic edge weight of the directed edge; When the total number of times is zero, the dynamic edge weights of all outgoing edges of the starting key resting point are set to equal values, and the sum of the dynamic edge weights of all outgoing edges is 1.

5. The method according to claim 1, characterized in that, The generation of a node feature matrix for each historical time window based on the environmental factor data and the avian influenza virus detection data includes: For each historical time window, based on the environmental factor data and the avian influenza virus detection data, and according to a predetermined node feature encoding rule, a node feature vector is generated for each of the key rest stops within the historical time window. The node feature matrix of the historical time window is composed of the feature vectors of each node; wherein the node feature vectors are used to represent the environmental conditions and historical avian influenza virus detection status of the key stopover sites.

6. The method according to claim 5, characterized in that, For each historical time window, based on the environmental factor data and the avian influenza virus detection data, and according to a pre-determined node feature encoding rule, a node feature vector is generated for each of the key rest stops within the historical time window, including: Based on the environmental factor data, extract the environmental factor vector of the key stopover sites within the historical time window; Based on the avian influenza virus detection data, the positivity rate of each avian influenza virus subtype in the key stopover sites within a predetermined number of consecutive historical time windows prior to the historical time window is extracted. The positivity rate is the ratio of the number of positive detections of that subtype in the corresponding window to the total number of samples, thus obtaining a historical virus detection vector. If there are missing samples, the positivity rate of the previous effective window of the key stopover site or the average positivity rate of the historical windows of the same period in the key stopover site is used to fill the gaps. The environmental factor vector is concatenated with the historical virus detection vector to form the original feature vector; The original feature vector is transformed into a fixed-dimensional node embedding vector through a nonlinear mapping, which serves as the node feature vector of the key resting place.

7. The method according to claim 1, characterized in that, The pre-trained migratory bird multi-subtype avian influenza prediction model includes a graph convolutional spatial feature extraction layer, a gated recurrent unit temporal feature extraction layer, and a multi-subtype transmission probability prediction layer. Within each historical time window, the graph convolution spatial feature extraction layer performs graph convolution operations based on the adjacency matrix of the dynamically weighted population-specific subgraph and the node feature matrix to aggregate the neighbor node information of each key resting place and generate the spatial feature vector of each key resting place. The spatial feature vectors corresponding to the same key stopover point within multiple consecutive historical time windows are arranged in chronological order to form an input sequence. The temporal feature extraction layer of the gated loop unit is used to extract temporal features from the input sequence to generate a spatiotemporal fusion feature vector for each key stopover point. Through the multi-subtype transmission probability prediction layer, the parameters of the graph convolutional spatial feature extraction layer and the gated recurrent unit temporal feature extraction layer are reused using a hard parameter sharing mechanism. The spatiotemporal fusion feature vector of each key stopover site is mapped to the probability distribution of each avian influenza virus subtype. After normalization, the transmission probability of each avian influenza virus subtype in each key stopover site during the target prediction period is output.

8. The method according to claim 1, characterized in that, The training process of the pre-trained migratory bird multi-subtype avian influenza prediction model includes: Historical migratory bird movement trajectory data, historical environmental factor data, and historical avian influenza virus detection data for historical training periods are acquired as training datasets; the historical training periods are earlier than the historical time windows. An initial framework for constructing a multi-subtype avian influenza prediction model for migratory birds is provided. The initial framework is a multi-task spatiotemporal graph neural network oriented to a dynamically weighted population-specific subgraph, which includes a graph convolutional spatial feature extraction layer, a gated recurrent unit temporal feature extraction layer, and a multi-subtype transmission probability prediction layer. Each avian influenza virus subtype shares the parameters of the spatial feature extraction layer and the temporal feature extraction layer. The training dataset is preprocessed to obtain the historical node feature matrix within the historical training period, and the corresponding dynamically weighted historical population-specific subgraph for each type of migratory bird population. Training samples are constructed using a sliding window approach, with the historical node feature matrix and dynamically weighted historical population-specific subgraph of the historical training period as input features, and the true value of the transmission probability of each virus subtype at each key stopover point within the next time window of the historical training period as the corresponding label; the training samples are divided into a training subset and a validation subset. The initial framework is trained using the training subset: the binary cross-entropy loss between the predicted probability of each subtype output by the initial framework and the corresponding label is calculated, and the total loss is obtained by summing the binary cross-entropy losses of each virus subtype; the gradient of the total loss with respect to each model parameter of the initial framework is calculated using the backpropagation algorithm, and the parameters of the initial framework are updated using the gradient. The performance of the initial framework is evaluated using the validation subset. The above training steps are repeated until the decreasing trend of the total loss on the validation subset meets the preset convergence condition, thus obtaining the migratory bird multi-subtype avian influenza prediction model.

Images