Machine learning based method for classifying bird migration progress actions

Abstract

This invention discloses a machine learning-based method for classifying bird migration movements. To address the problem of ambiguous movement classification caused by multiple behaviors corresponding to the same posture during monitoring, this invention acquires various observational data, extracts temporal features of individual bird movements using a rhythm-aware state-space model, constructs a heterogeneous group graph by combining trajectory information, introduces migration stage priors, and learns group context features in a group graph Transformer. Then, the individual features, group context features, and migration stage priors are input into the movement classification network to obtain the probability distribution and prediction uncertainty of each bird's movement. By utilizing the difference between the group behavior distribution and the expected behavior distribution during migration, and combining the prediction uncertainty to adjust the prior weights, the movement probability distribution is jointly corrected, making the group behavior distribution approach the stage prior. This achieves the technical effect of alleviating posture ambiguity and improving the accuracy of bird migration movement classification in migration scenarios.

Description

Technical Field

[0001] This invention relates to the field of bird migration process action classification, and more particularly to a method for bird migration process action classification based on machine learning. Background Technology

[0002] The refined monitoring and identification of bird migration behavior is an important research direction in ecological monitoring, species conservation, and environmental assessment. With the development of video acquisition equipment, inertial sensors, satellite positioning technology, and machine learning and deep learning algorithms, researchers have begun to utilize multi-source data, such as airborne or ground-based cameras, accelerometers, and GPS, to automatically identify the individual bird postures, flight paths, and behavioral states during migration. Some existing techniques achieve action classification through posture estimation and time-series models (such as Hidden Markov Models and Recurrent Neural Networks). Some works combine trajectory features to divide migration stages, and some literature attempts to use graph neural networks to model the spatial proximity relationships between individuals in a group for group behavior analysis.

[0003] However, existing technologies still have shortcomings in classifying bird movements during migration, mainly including:

[0004] First, existing methods are mostly based on the posture sequence or short-term motion features of a single individual for classification. Temporal modeling often uses a single time scale, which is insufficient to characterize the rhythmic changes of posture over time, the dominant frequency of movement, and the intensity pattern. It is difficult to effectively distinguish the situation where "the same posture corresponds to different behaviors in different rhythms or different stages", which makes it difficult to resolve the ambiguity when the same posture corresponds to multiple behaviors.

[0005] Second, although some studies have begun to utilize multi-source data such as video, inertial sensors, and satellite positioning, they mostly employ simple feature stitching or weighted fusion, lacking unified modeling of relationships between multiple birds, between birds and the environment, and between birds and potential threats. Existing graph-based group modeling is mostly based on homogeneous graphs or only considers individual-to-individual edges, rarely explicitly incorporating factors such as habitat category and potential threats into the group structure. The supporting role of group context information in determining individual behavior is not fully utilized.

[0006] Third, existing action classification methods often treat the classifier output as a deterministic result, lacking characterization and utilization of prediction uncertainty, and also lacking explicit modeling and constraints of prior knowledge such as "migration stage priors" and their corresponding expected behavior distributions. Even when some methods introduce migration stage information, it is mostly used as a static feature in classification, lacking a mechanism to jointly constrain and correct at the group level by utilizing the consistency between stage priors and group behavior distributions. Therefore, it cannot adaptively enhance the influence of group information and stage priors when individual judgments are unstable or ambiguous.

[0007] Therefore, it is necessary to provide a bird migration process action classification method that can comprehensively utilize multi-source information, group structure relationships, and migration stage priors, and can adaptively adjust the intensity of prior utilization based on prediction uncertainty, so as to alleviate the ambiguity problem of the same posture corresponding to multiple behaviors and improve the accuracy and stability of action classification in migration scenarios. Summary of the Invention

[0008] One objective of this invention is to propose a machine learning-based method for classifying bird migration movements. Addressing the problems of existing methods that rely solely on single-timescale posture sequences and lack prior modeling of group structure relationships and migration stages, leading to difficulties in eliminating ambiguity when the same posture corresponds to multiple behaviors and resulting in low accuracy and stability of movement classification, the following technical solution is proposed: Based on the time synchronization and spatial alignment of multi-source video, inertial sensor, satellite positioning, and environmental image data, a rhythm-aware state-space model is used to model posture sequences and motion intensity sequences at both slow and fast timescales, respectively. This extracts individual movement temporal characteristics including the dominant frequency of movement, rhythmic stability, and intensity variation patterns. This invention constructs a heterogeneous population graph containing bird nodes, environment nodes, threat nodes, and global population nodes based on the trajectories, habitat categories, and potential threat target information of multiple birds. A migration stage prior vector is generated using a migration stage estimation model. A graph attention mechanism conditioned on migration stage priors is introduced into the population graph Transformer to obtain the population context features and global population features of each bird. Then, by combining the prediction uncertainty index output by the action classification network and the population consistency index between the population behavior distribution and the expected behavior distribution at the migration stage, the fusion weights of individual features, population information, and stage priors are adaptively adjusted to perform two-level correction of the action probability distribution at both the individual and population levels. This invention effectively alleviates ambiguity regarding multiple behaviors corresponding to the same posture in migration scenarios, fully utilizes population structure and stage prior knowledge, and significantly improves the accuracy and robustness of action classification during bird migration.

[0009] A method for classifying bird migration movements based on machine learning according to an embodiment of the present invention is characterized by comprising:

[0010] S1. Acquire multi-source raw data to obtain the posture and movement intensity sequences, spatial location information, habitat environment and potential threat target image fragments, and satellite positioning time series for multiple birds;

[0011] S2. Input the posture and motion intensity sequence into the rhythm perception state space model for temporal modeling to obtain the individual action temporal characteristics of each bird. Based on spatial location information, habitat image fragments and potential threat target image fragments, obtain the trajectory sequence, habitat category and potential threat target information of each bird.

[0012] S3. Input the temporal features of individual actions, trajectory sequences, habitat categories, information on potential threats and spatial locations into the group graph construction module to construct a heterogeneous group graph structure containing bird nodes, environment nodes, threat nodes and global group nodes. Input the trajectory sequence into the migration stage estimation model to obtain the migration stage prior vector.

[0013] S4. Input the heterogeneous group graph structure, migration stage prior vectors and individual action temporal features into the group graph Transformer model, and use the graph attention mechanism with the migration stage prior vector as a condition for message passing and feature aggregation to obtain the group context features and group global features of each bird.

[0014] S5. Input the individual action temporal features, group context features, group global features, and migration stage prior vectors into the action classification network to obtain the initial action probability distribution and prediction uncertainty index of each bird in each preset action category. Statistically analyze the initial action probability distribution of all birds to obtain the group behavior distribution. Input the migration stage prior vectors into the behavior expectation generation module to obtain the expected behavior distribution of the migration stage. Compare the group behavior distribution with the expected behavior distribution of the migration stage to obtain the group consistency index.

[0015] S6. Input the initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, group consistency index, individual action temporal features, group context features, and migration prior vectors into the action result correction module. Based on the prediction uncertainty index, adaptively determine the fusion weights of individual action temporal features, group context features, and migration prior vectors to adjust the initial action probability distribution of each bird. Based on the group consistency index, jointly correct the adjusted action probability distribution so that the group behavior distribution approaches the expected behavior distribution during migration, obtain the final action probability distribution of each bird, and output the corresponding bird migration process action classification results.

[0016] Optionally, step S1 specifically includes:

[0017] Acquire continuous video data, inertial sensor data, satellite positioning data, and environmental image data collected within a preset time window. Based on timestamps and synchronization trigger signals, synchronize the continuous video data, inertial sensor data, satellite positioning data, and environmental image data in time, and align the multi-source raw data corresponding to the same preset time window to a unified time axis.

[0018] Camera calibration and coordinate registration are performed on the continuous video data, and the pixel coordinates in the continuous video data are spatially aligned with the geographic coordinates corresponding to the satellite positioning data to establish a unified spatial coordinate system.

[0019] Based on the completion of time synchronization and coordinate registration, target detection and multi-target tracking are performed on the continuous video data to identify and track each bird. In each frame, the key point position or contour features of the corresponding bird are extracted and arranged in chronological order to form a posture sequence corresponding to each bird. The motion amplitude is calculated based on the changes in key point position or contour between adjacent frames to form a motion intensity sequence corresponding to each bird.

[0020] Based on the bird positions in the unified spatial coordinate system, the spatial positions of each bird in each frame are connected in chronological order to obtain the spatial position information corresponding to each bird.

[0021] Centered on the bird's location, image segments containing the background area around the bird are cropped from the continuous video data and the environmental image data as habitat environment image segments corresponding to each bird. Target detection is performed on the area around the bird in the continuous video data to identify potential threat targets such as people, vehicles, and birds of prey. Image segments containing these potential threat targets are cropped as potential threat target image segments.

[0022] The satellite positioning data is time-aligned, interpolated, and transformed to align the satellite positioning data corresponding to each bird to the unified time axis and unified spatial coordinate system, forming a satellite positioning time series corresponding to each bird.

[0023] Output the attitude sequence, motion intensity sequence, spatial location information, habitat image fragments, potential threat target image fragments, and satellite positioning time series.

[0024] Terminology definition:

[0025] The preset time window is a continuous time interval pre-set for data collection and processing when classifying bird migration process actions.

[0026] The multi-source raw data is a collection of continuous video data, inertial sensor data, satellite positioning data, and environmental image data collected within a preset time window;

[0027] The continuous video data is a sequence of images containing the target bird and its surrounding environment, continuously collected by airborne or ground-based camera equipment within a preset time window;

[0028] The inertial sensor data is time-series data reflecting the motion state collected by inertial sensors such as accelerometers and gyroscopes installed on the bird or carrier within a preset time window;

[0029] The satellite positioning data is time-stamped location signal data collected within a preset time window using satellite positioning devices such as GPS and Beidou.

[0030] The environmental image data refers to static or dynamic image data collected within a preset time window that reflects the environmental background of the bird activity area.

[0031] The time synchronization is a process of aligning continuous video data, inertial sensor data, satellite positioning data, and environmental image data using timestamps and synchronization trigger signals, so that all types of data correspond to each other under the same reference time.

[0032] The unified timeline is the same reference time series shared by different data sources after time synchronization, used to align multi-source raw data to a consistent time scale;

[0033] The camera calibration is the calibration process of determining the intrinsic and extrinsic parameters of the camera to acquire continuous video data in order to establish the correspondence between pixel coordinates and spatial coordinates.

[0034] The coordinate registration is a process of spatially aligning coordinate systems of different data sources, used to unify the camera coordinate system with the geographic coordinate system corresponding to the satellite positioning data into the same coordinate frame.

[0035] The unified spatial coordinate system is a common spatial reference coordinate system established after camera calibration and coordinate registration, used to simultaneously represent the pixel position in the video and the geographical location of the satellite positioning.

[0036] The bird position refers to the coordinate information representing the spatial position of each bird in each frame or at each time point under a unified spatial coordinate system.

[0037] The habitat image segment is an image data segment obtained by cropping continuous video data and / or environmental image data with the bird's location as the center, and includes the background area around the target bird;

[0038] The potential threat targets are objects that may interfere with or threaten bird behavior within the bird activity area, including but not limited to people, vehicles and birds of prey;

[0039] The potential threat target image segment is an image data segment containing a potential threat target, obtained by cropping from continuous video data;

[0040] The pose sequence is a sequence formed by arranging the key point positions or contour features extracted in each frame of continuous video data for the same bird in chronological order.

[0041] The amplitude of the movement is a scalar or vector index used to characterize the strength of bird movements, calculated based on changes in the position or contour of key points between adjacent frames.

[0042] The motion intensity sequence is a sequence of motion amplitudes arranged in chronological order within a preset time window, used to characterize the change of bird motion intensity over time within that time window.

[0043] The spatial location information is a sequence of location information used to characterize the bird's movement trajectory, formed by connecting the spatial locations of the bird in each frame in chronological order under a unified spatial coordinate system.

[0044] The satellite positioning time series is a sequence of satellite positioning location signals corresponding to each bird, arranged in chronological order under a unified time axis and a unified spatial coordinate system after time alignment, interpolation and coordinate transformation of satellite positioning data.

[0045] The image fragments of habitat and potential threat targets are a collective term for both habitat image fragments and potential threat target image fragments.

[0046] Optionally, step S2 specifically includes:

[0047] The posture sequence and motion intensity sequence are input into the rhythm-aware state space model. A slow time-scale sub-model and a fast time-scale sub-model are set in the rhythm-aware state space model. The slow time-scale sub-model is used to model the posture sequence at a larger time step to represent the slow change of posture over time. The fast time-scale sub-model is used to model the motion intensity sequence at a smaller time step to represent the rapid change of motion intensity. The hidden states of the slow time-scale sub-model and the fast time-scale sub-model are fused to obtain the individual motion temporal features corresponding to each bird. The individual motion temporal features include at least motion rhythm features that represent the dominant frequency of motion, rhythm stability, and motion intensity change pattern.

[0048] Based on spatial location information, the spatial locations of each bird in a unified spatial coordinate system are connected in chronological order to form a trajectory sequence corresponding to that bird.

[0049] Based on habitat image fragments, an environmental classification network is used to classify the environment of each bird's location to obtain the habitat category corresponding to that bird.

[0050] Based on image fragments of potential threat targets, the category and spatial location of the potential threat targets in a unified spatial coordinate system are determined by a target detection and recognition network, and the category and spatial location of the potential threat targets are combined into potential threat target information;

[0051] Output the individual movement time sequence characteristics, trajectory sequence, habitat category, and potential threat target information for each bird.

[0052] Terminology definition:

[0053] The rhythm-aware state-space model is a probabilistic model or neural network model that introduces multi-timescale rhythm modeling capability on the basis of the state-space model structure, and is used to perform joint temporal modeling of posture sequences and motion intensity sequences simultaneously.

[0054] The slow time-scale sub-model is a sub-model in the rhythm-aware state-space model that uses a larger time step to model the posture sequence and is used to characterize the slow change trend of posture over time.

[0055] The fast time-scale sub-model is a sub-model in the rhythm perception state-space model that uses a smaller time step to model the motion intensity sequence and is used to characterize the rapid change pattern of motion intensity.

[0056] The hidden state is a state vector that is not directly observable and is maintained within each time step of the rhythm-aware state space model and its slow-time-scale sub-model and fast-time-scale sub-model. It is used to encode the dynamic influence of historical input on the current moment.

[0057] The individual action temporal features are feature vectors or feature sequences output by the rhythm-aware state space model based on posture sequences and motion intensity sequences, used to characterize the change of a single bird's actions over time within a preset time window;

[0058] The dominant frequency of the action is a feature quantity extracted from the temporal characteristics of individual actions, used to characterize the frequency of repeated actions of birds within a preset time window;

[0059] The rhythm stability is a feature quantity extracted from the temporal characteristics of an individual's actions, used to characterize the degree of stability of the action rhythm over time or the fluctuation of rhythm.

[0060] The movement intensity change pattern is a feature quantity extracted from the temporal characteristics of individual movements, used to describe the change pattern of bird movement intensity over time (including but not limited to periodic enhancement, weakening or sudden changes);

[0061] The movement rhythm features are a set of features that comprehensively describe the movement rhythm attributes of birds, including movement dominant frequency, rhythm stability, and movement intensity variation patterns.

[0062] The trajectory sequence is a sequence of location information that represents the movement path of a bird, formed by arranging the spatial positions of each bird in a unified spatial coordinate system in chronological order based on spatial location information.

[0063] The habitat category is a discrete category label used to identify the environmental type of the area where the bird is located, obtained by classifying the habitat image fragments through an environmental classification network.

[0064] The environment classification network is a machine learning or deep learning model that extracts features from habitat image fragments and outputs the corresponding habitat category.

[0065] The potential threat target information is a combination of information obtained from potential threat target image segments through a target detection and recognition network, including the potential threat target category and its spatial location in a unified spatial coordinate system;

[0066] The target detection and recognition network is a machine learning or deep learning model used to detect potential threat targets in image fragments, determine their spatial location, and output the target category.

[0067] Optionally, step S3 specifically includes:

[0068] Individual action temporal features, trajectory sequences, habitat categories, and potential threat target information and spatial location information are input into the group graph construction module. In the group graph construction module, for each bird, a corresponding bird node is generated based on the bird's spatial location information in a unified spatial coordinate system, and the individual action temporal features of the bird are used as the initial features of the bird node.

[0069] Based on the spatial location information and trajectory sequence of multiple birds in a unified spatial coordinate system, calculate the spatial distance and relative motion relationship between any two birds. When the spatial distance is less than a preset distance threshold or the relative motion relationship meets a preset similarity condition, establish a bird node connection edge between the corresponding two bird nodes, and use the spatial distance and relative motion relationship as the edge feature of the bird node connection edge.

[0070] Based on the habitat category corresponding to each bird, an environment node is generated for each type of habitat. An environment connection edge is established between the bird node active in the habitat and the corresponding environment node. The environment-related statistics, such as the bird's stay time in the habitat, are used as the edge features of the environment connection edge.

[0071] Based on the information of potential threat targets, a threat node is generated for each type of potential threat target. Based on the distance relationship between the spatial position of the potential threat target in the unified spatial coordinate system and the spatial position of the bird node, a threat connection edge is established between the bird node and the corresponding threat node when the distance is less than the preset threat impact threshold. The spatial distance and relative approach speed between the bird node and the threat node are used as the edge features of the threat connection edge.

[0072] A global group node is set up to represent the overall group state, and a global connection edge is established between the global group node and all bird nodes. The individual action time sequence statistics and trajectory sequence statistics of bird nodes within a preset time window are used as the edge features of the global connection edge, thereby constructing a heterogeneous group graph structure that includes bird nodes, environment nodes, threat nodes and global group nodes.

[0073] Simultaneously, the trajectory sequences of multiple birds are input into the migration stage estimation model. In the migration stage estimation model, the migration distance, rest duration, and speed changes of each bird within the preset time window are calculated based on the trajectory sequences. The trajectory features are then aggregated at the group level to obtain a migration feature vector representing the current overall migration state. The migration feature vector is then mapped to a migration stage prior vector corresponding to the preset time window.

[0074] Output the heterogeneous population graph structure and the prior vectors for the migration stages.

[0075] Terminology definition:

[0076] The group graph construction module is a functional module or program unit used to automatically generate heterogeneous group graph structures based on the temporal characteristics of individual actions of multiple birds, trajectory sequences, habitat categories, potential threat target information and spatial location information.

[0077] The bird node is a node used to represent a single bird individual in a heterogeneous population graph structure, and its node features include at least the temporal features of the individual bird's actions and related statistics.

[0078] The environmental nodes are nodes set for each type of habitat in the heterogeneous population graph structure according to the habitat category, and are used to represent the corresponding habitat type;

[0079] The threat node is a node set for each type of potential threat target in the heterogeneous population graph structure based on the potential threat target information, used to represent the corresponding threat target type;

[0080] The global node of the group is a node used to represent the current overall state of the bird flock in the heterogeneous group graph structure. Its node features include at least the individual action time sequence statistics and trajectory sequence statistics of multiple birds within a preset time window.

[0081] The heterogeneous community graph structure is a graph structure that includes at least bird nodes, environment nodes, threat nodes, and global community nodes, as well as various types of connecting edges between different types of nodes, and attaches corresponding feature information to different nodes and edges.

[0082] The spatial distance is a measure of the spatial distance between any two birds or between a bird and a potential threat target, calculated based on a unified spatial coordinate system.

[0083] The relative motion relationship is the relative motion characteristics of any two birds in terms of speed, direction of motion, etc., calculated based on the trajectory sequences of multiple birds, and is used to characterize the similarity of their motion patterns.

[0084] The preset distance threshold is a pre-set distance threshold parameter used to determine whether two birds are close enough in space to establish a bird node connection edge;

[0085] The preset similarity condition is a pre-set similarity judgment rule or threshold used to determine whether the relative motion relationship between two birds meets the conditions for establishing bird node connection edges.

[0086] The bird node connection edge is an edge in the heterogeneous population graph structure that connects any two bird nodes that satisfy the conditions of spatial distance and / or relative motion relationship, and is used to represent the spatial proximity or motion association relationship between two birds;

[0087] The environmental connection edge is an edge in the heterogeneous population graph structure that connects bird nodes active in a certain type of habitat with the corresponding environmental node, and is used to represent the relationship between birds and the habitat.

[0088] The threat connection edge is an edge in the heterogeneous population graph structure that connects bird nodes within the potential threat's influence range to the corresponding threat node, and is used to represent the threat relationship between the bird and the potential threat target;

[0089] The global connection edge is the edge connecting the global node of the group and each bird node in the heterogeneous group graph structure, which is used to represent the contribution relationship of a single bird to the overall group state.

[0090] The edge features are the feature quantities attached to the bird node connection edges, environment connection edges, threat connection edges and global connection edges, which are used to characterize the quantitative relationship attributes between corresponding node pairs;

[0091] The dwell time is the cumulative dwell time of a bird in a specific habitat category within a preset time window;

[0092] The environmental statistics are a set of statistical features used to quantify the relationship between birds and their habitat, calculated based on data such as the length of time birds spend in a particular habitat category.

[0093] The preset threat impact threshold is a pre-set distance or comprehensive index threshold parameter used to determine whether bird nodes and threat nodes are spatially within the threat impact range, and thus whether to establish a threat connection edge.

[0094] The relative approach velocity is the relative velocity between the bird node and the threat node in the direction of mutual approach, calculated based on the change of their positions in a unified spatial coordinate system over time.

[0095] The individual action temporal feature statistics are the feature quantities obtained by performing time-dimensional or sample-dimensional statistics (including but not limited to mean, variance, extreme values ​​or histograms, etc.) on the individual action temporal features of a single bird within a preset time window;

[0096] The trajectory sequence statistics are feature quantities obtained by statistically analyzing the trajectory sequence of a single bird within a preset time window (including but not limited to average speed, maximum height, number of turns, etc.).

[0097] The migration stage estimation model is a machine learning or deep learning model that takes the trajectory sequences of multiple birds as input, extracts trajectory features, and performs group-level aggregation to estimate the current overall migration stage.

[0098] The migration distance is the cumulative distance traveled by a single bird along the migration path within a preset time window, calculated based on the trajectory sequence.

[0099] The pause duration is the duration during which the bird is in a low-speed or stationary state, determined based on the trajectory sequence, within a preset time window;

[0100] The speed change is a quantitative indicator of the change in the magnitude or direction of the speed of a single bird over time within a preset time window, calculated based on the trajectory sequence.

[0101] The trajectory features are a set of features used to characterize bird migration patterns, consisting of migration distance, rest duration, speed changes, and other statistics extracted from trajectory sequences.

[0102] The group-level aggregation is a process of summarizing, statistically analyzing, or learning and fusing the trajectory features of multiple birds at the group level to obtain the overall group trajectory features.

[0103] The migration feature vector is a vectorized feature representation obtained through group-level aggregation, used to characterize the current overall group migration state;

[0104] The migration stage prior vector is a vector representation obtained by mapping the migration feature vector, used to represent the prior information of the migration stage corresponding to the current preset time window.

[0105] Optionally, step S4 specifically includes:

[0106] The initial node features and edge features of bird nodes, environment nodes, threat nodes, and global nodes in the heterogeneous population graph structure are input together with the migration stage prior vector and individual action temporal features into the population graph Transformer model. Specifically, the individual action temporal features corresponding to each bird are assigned to the corresponding bird node as the initial feature of that bird node, the environment category features obtained by encoding the habitat category are assigned to the environment node as the initial feature of the environment node, the threat features obtained by encoding the potential threat target information are assigned to the threat node as the initial feature of the threat node, and the population statistical features obtained by aggregating the individual action temporal features and trajectory sequence statistics of all bird nodes in the heterogeneous population graph structure are assigned to the global nodes as the initial feature of the global nodes.

[0107] The migration stage prior vector is input into the migration stage embedding network and mapped to a migration stage embedding vector. In each layer of the population graph Transformer model, the migration stage embedding vector is combined with node features and edge features to calculate the attention weights of bird node connection edges, environment connection edges, threat connection edges and global connection edges. Different attention parameters are used for different types of connection edges, thereby forming a message passing method based on the migration stage conditional graph attention mechanism.

[0108] In each layer, the features of adjacent nodes are weighted and summed according to the attention weight and fused with the features of the current node to update the feature representations of bird nodes, environment nodes, threat nodes and group global nodes. After multi-layer message passing and feature aggregation, the updated feature representation of each bird node is used as the group context feature corresponding to that bird, and the updated feature representation of the group global node is used as the group global feature.

[0109] Output the group context features and the group global features.

[0110] Terminology definition:

[0111] The aforementioned Transformer model for the heterogeneous group graph structure employs the Transformer architecture for message passing and feature aggregation, enabling simultaneous processing of different types of node and edge features.

[0112] The initial node features are the initial feature vectors assigned to various types of nodes during the input phase of the population graph Transformer model, which are used as the initial representation for subsequent message passing and feature updates.

[0113] The node features are feature vectors associated with each node in the Transformer model of the population graph, including the initial node features and the node representations updated after message passing and feature aggregation at each layer.

[0114] The environmental category features are vectorized features obtained by encoding or embedding habitat categories, and are used as the initial node features of environmental nodes;

[0115] The threat features are vectorized features obtained by encoding or embedding potential threat target information, and are used as the initial node features of the threat node;

[0116] The group statistical features are feature vectors obtained by group-level aggregation of individual action temporal features and trajectory sequence statistics of all bird nodes in the heterogeneous group graph structure, which are used as the initial node features of the global nodes of the group.

[0117] The migration stage embedding network is a neural network or other nonlinear mapping model that takes the migration stage prior vector as input and maps it to an embedding vector suitable for attention calculation in the population graph Transformer model.

[0118] The migration stage embedding vector is a vectorized representation generated by the migration stage embedding network based on the migration stage prior vector, and is used as conditional information for calculating attention weights in each layer of the population graph Transformer model;

[0119] The graph attention mechanism is an attention calculation method that calculates the importance coefficient of each connecting edge based on the features of adjacent nodes and edge features in the graph structure, and then performs weighted aggregation of the features of adjacent nodes accordingly.

[0120] The migration stage-based conditional graph attention mechanism is an attention calculation method that uses the migration stage embedding vector as an additional conditional input in the graph attention mechanism, so that the attention weight is simultaneously modulated by node features, edge features and migration stage information.

[0121] The attention weights are non-negative weight coefficients calculated by the graph attention mechanism for each connected edge, used to represent the relative importance of the corresponding edge in the current message passing step;

[0122] The message passing is a computational process in the Transformer model of the group graph that passes the features of adjacent nodes along the connecting edges according to attention weights and converges them to the target node;

[0123] The feature aggregation is a computational process in which the features of adjacent nodes are weighted and summed according to attention weights during message passing and then fused with the features of the current node to generate updated node features.

[0124] The group context features are the final feature representations of each bird node obtained after multi-layer message passing and feature aggregation, used to characterize the comprehensive state of the bird in multi-dimensional contexts such as group structure, habitat environment and potential threats;

[0125] The group global feature is the final feature representation of the group global node obtained after multi-layer message passing and feature aggregation, which is used to characterize the global behavior and environmental state of the entire bird flock within the current preset time window.

[0126] Optionally, step S5 specifically includes:

[0127] Individual action temporal features, group context features, group global features, and migration stage prior vectors are used as input features and fed into the action classification network. In the action classification network, for each bird, the individual action temporal features corresponding to the bird are concatenated and nonlinearly transformed with the group context features corresponding to the bird, the group global features, and the migration stage prior vectors to obtain the bird's action discrimination feature vector. Based on the action discrimination feature vector, the initial action probability distribution of the bird in multiple preset action categories is calculated through a classification structure with a normalized output layer.

[0128] The prediction uncertainty index of the bird is calculated based on the probability value distribution of the initial action probability distribution. The prediction uncertainty index is used to characterize the degree of dispersion of the initial action probability distribution.

[0129] The initial action probability distributions of all birds are weighted and summed or averaged in the group dimension to obtain the group behavior distribution in multiple preset action categories within the current preset time window;

[0130] The migration stage prior vector is input into the behavior expectation generation module, and the migration stage prior vector is nonlinearly mapped in the behavior expectation generation module to obtain the migration stage expected behavior distribution corresponding to the migration stage prior vector on multiple preset action categories.

[0131] The distribution of group behavior is compared with the expected distribution of behavior during the migration stage, and a group consistency index reflecting the degree of difference between the two is calculated based on the difference, ratio or distance between the two in multiple preset action categories.

[0132] The initial action probability distribution and prediction uncertainty index are output, and the group behavior distribution, expected behavior distribution during the migration stage, and group consistency index are obtained at the group level.

[0133] Terminology definition:

[0134] The action classification network is a machine learning or deep learning model that takes individual action temporal features, group context features, group global features, and migration stage prior vectors as inputs, determines the action category of a single bird, and outputs the probability values ​​of each preset action category.

[0135] The action discrimination feature vector is an intermediate feature representation used for action category determination, obtained by concatenating and nonlinearly transforming the individual action temporal features, group context features, group global features, and migration stage prior vectors corresponding to a single bird in the action classification network.

[0136] The classification structure with a normalized output layer is a substructure in the action classification network that takes the action discrimination feature vector as input, first obtains the original score of each action category through several fully connected layers or other forms of feature transformation structure, and then performs normalization processing through the output layer that can normalize the scores of each category into probability values.

[0137] The preset action category is a set of discrete action types that are predetermined during the system design phase and used to distinguish different action states during the bird migration process.

[0138] The initial action probability distribution is the probability distribution of multiple preset action categories output by the action classification network based on the action discrimination feature vector. It is used to characterize the confidence level of each action category as the current bird action before subsequent correction.

[0139] The prediction uncertainty index is a scalar index calculated based on the probability value distribution of the initial action probability distribution. It is used to quantify the degree of dispersion or concentration of the initial action probability distribution and reflects the magnitude of uncertainty of the current action classification result.

[0140] The group behavior distribution is the overall probability distribution of the initial action probability distribution of all birds in the current preset time window, obtained by weighted summation or averaging in the group dimension, and is used to characterize the overall behavior proportion of the entire flock in each action category within the time window.

[0141] The behavior expectation generation module is a functional module or program unit that takes the migration stage prior vector as input and generates the expected behavior distribution on multiple preset action categories corresponding to the migration stage prior vector through nonlinear mapping.

[0142] The expected behavior distribution for the migration stage is a probability distribution of multiple preset action categories output by the behavior expectation generation module based on the prior vector of the migration stage. It is used to characterize the proportion of each action category that is theoretically or prior to appear in the current migration stage.

[0143] The group consistency index is an indicator calculated based on the difference, ratio, or distance between the group behavior distribution and the expected behavior distribution during the migration stage across multiple preset action categories. It is used to quantify the degree of consistency or deviation between the current group behavior distribution and the expected behavior distribution.

[0144] Optionally, step S6 specifically includes:

[0145] The initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, and group consistency index, along with individual action temporal features, group context features, and migration stage prior vectors, are input into the action result correction module. In the action result correction module, for each bird, the prediction uncertainty index of that bird, along with the corresponding individual action temporal features, group context features, and migration stage prior vectors, are input into the fusion weight calculation unit. Through nonlinear mapping, the individual action temporal feature weights, group context feature weights, and migration stage prior vector weights used for the bird's action determination are obtained respectively. The individual action temporal feature weights, group context feature weights, and migration stage prior vector weights are normalized to obtain the fusion weights corresponding to that bird in action determination.

[0146] The individual action temporal features, group context features, and migration stage prior vectors of the bird are weighted and summed using the fusion weights to obtain the corrected action discrimination feature vector of the bird. Based on the corrected action discrimination feature vector, the corrected action probability distribution of the bird in each preset action category is calculated through a corrected classification sub-network. The initial action probability distribution of the bird is weighted and adjusted using the corrected action probability distribution to obtain the individual corrected action probability distribution of the bird.

[0147] At the group level, based on the group consistency index and the expected behavior distribution during the migration stage, the individual-corrected action probability distribution of all birds is jointly adjusted in the group dimension. When the group consistency index indicates that the difference between the current group behavior distribution and the expected behavior distribution during the migration stage is greater than a preset threshold, the individual-corrected action probability distribution of each bird is corrected iteratively so that the group behavior distribution obtained by statistical analysis in the group dimension after correction gradually approaches the expected behavior distribution during the migration stage. After the iteration is completed, the group-adjusted action probability distribution of each bird is taken as the final action probability distribution of that bird within the preset time window, and the corresponding bird migration process action classification result is output according to the final action probability distribution.

[0148] Terminology definition:

[0149] The action result correction module is a functional module or program unit that takes the initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, group consistency index, individual action temporal characteristics, group contextual characteristics and migration stage prior vector as input, performs individual-level and group-level joint correction on the action probability distribution of each bird to output the final action probability distribution.

[0150] The fusion weight calculation unit is a sub-unit set inside the action result correction module. It takes the prediction uncertainty index of a single bird and the corresponding individual action temporal features, group context features and migration stage prior vector as input, and calculates the weights that the bird should be assigned to the individual action temporal features, group context features and migration stage prior vector in action determination through nonlinear mapping and normalizes them.

[0151] The individual action temporal feature weight is a weight coefficient output by the fusion weight calculation unit, used to characterize the importance of the individual action temporal feature relative to other feature information in the bird's action determination;

[0152] The group context feature weight is a weight coefficient output by the fusion weight calculation unit, used to characterize the importance of the group context feature relative to other feature information in the bird action determination;

[0153] The migration phase prior vector weights are weight coefficients output by the fusion weight calculation unit, used to characterize the importance of the migration phase prior vectors relative to other feature information in the bird's action determination.

[0154] The fusion weight is a set of weights obtained by normalizing the individual action temporal feature weight, the group context feature weight, and the migration stage prior vector weight, and is used to perform weighted fusion of the above three types of features in action determination.

[0155] The corrected action discrimination feature vector is an intermediate feature representation used to generate the corrected action probability distribution, obtained by weighting and summing the individual action temporal features, group context features, and migration stage prior vectors of the bird using fusion weights in the action result correction module.

[0156] The correction classification sub-network is a classification sub-model set inside the action result correction module, which takes the correction action discrimination feature vector as input and outputs the corrected action probability distribution of the bird in each preset action category.

[0157] The corrected action probability distribution is a probability distribution on each preset action category calculated by the correction classification sub-network based on the correction action discrimination feature vector, and is used to weight and adjust the initial action probability distribution of the bird.

[0158] The individual-corrected action probability distribution is the action probability distribution obtained at the level of a single bird by weighting the initial action probability distribution with the corrected action probability distribution, and has taken into account individual uncertainty and feature fusion weights.

[0159] The preset threshold is a pre-set numerical threshold parameter used to determine whether the group consistency index indicates that the difference between the current group behavior distribution and the expected behavior distribution during the migration stage is so large that group-level joint adjustment needs to be performed.

[0160] The group-adjusted action probability distribution is obtained by iteratively adjusting the individual-corrected action probability distribution of all birds at the group level, so that the group behavior distribution gradually approaches the expected behavior distribution during the migration stage, and then obtaining the final action probability distribution for each bird.

[0161] The final action probability distribution is the probability distribution of each preset action category used to output the corresponding bird migration process action classification results within a preset time window after individual correction and group-level joint adjustment.

[0162] The beneficial effects of this invention are:

[0163] 1. By using a rhythm-aware state-space model, we can model posture sequences on a slow time scale and motion intensity sequences on a fast time scale. We can also explicitly extract action rhythm features such as action frequency, rhythm stability, and intensity change patterns. This eliminates reliance on single-frame postures or single time-scale sequences and effectively characterizes the rhythm differences of the same posture under different behaviors. This significantly alleviates the ambiguity of "the same posture corresponding to multiple behaviors" and improves the accuracy of action recognition.

[0164] 2. By constructing a heterogeneous population graph containing bird nodes, environment nodes, threat nodes, and global population nodes, and introducing a graph attention mechanism conditioned on the migration stage prior vector in the population graph Transformer, it is possible to simultaneously model the spatial / movement relationships between multiple birds as well as the interaction between birds and their habitat and potential threats. Furthermore, by assigning differentiated weights to different relationships based on the current migration stage, a more realistic population context feature can be obtained, thereby improving the scene adaptability and overall accuracy of action classification during the migration process.

[0165] 3. By introducing prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, and group consistency index in the action classification stage, the fusion weights of individual action features, group context features, and migration stage priors are adaptively adjusted using uncertainty at the individual level. At the group level, the action probability distribution of each bird is iteratively corrected based on the group consistency index, so that the group behavior distribution approaches the expected behavior distribution during migration. Thus, when the observation noise is large and the posture judgment is unstable, the results can be automatically corrected by relying on group information and stage priors, which significantly improves the robustness and stability of action classification results. Attached Figure Description

[0166] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0167] Figure 1 This is a flowchart of a machine learning-based method for classifying bird migration processes, as proposed in this invention. Detailed Implementation

[0168] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0169] refer to Figure 1 A method for classifying bird migration movements based on machine learning, characterized by comprising:

[0170] S1. Acquire multi-source raw data to obtain the posture and movement intensity sequences, spatial location information, habitat environment and potential threat target image fragments, and satellite positioning time series for multiple birds;

[0171] S2. Input the posture and motion intensity sequence into the rhythm perception state space model for temporal modeling to obtain the individual action temporal characteristics of each bird. Based on spatial location information, habitat image fragments and potential threat target image fragments, obtain the trajectory sequence, habitat category and potential threat target information of each bird.

[0172] S3. Input the temporal features of individual actions, trajectory sequences, habitat categories, information on potential threats and spatial locations into the group graph construction module to construct a heterogeneous group graph structure containing bird nodes, environment nodes, threat nodes and global group nodes. Input the trajectory sequence into the migration stage estimation model to obtain the migration stage prior vector.

[0173] S4. Input the heterogeneous group graph structure, migration stage prior vectors and individual action temporal features into the group graph Transformer model, and use the graph attention mechanism with the migration stage prior vector as a condition for message passing and feature aggregation to obtain the group context features and group global features of each bird.

[0174] S5. Input the individual action temporal features, group context features, group global features, and migration stage prior vectors into the action classification network to obtain the initial action probability distribution and prediction uncertainty index of each bird in each preset action category. Statistically analyze the initial action probability distribution of all birds to obtain the group behavior distribution. Input the migration stage prior vectors into the behavior expectation generation module to obtain the expected behavior distribution of the migration stage. Compare the group behavior distribution with the expected behavior distribution of the migration stage to obtain the group consistency index.

[0175] S6. Input the initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, group consistency index, individual action temporal features, group context features, and migration prior vectors into the action result correction module. Based on the prediction uncertainty index, adaptively determine the fusion weights of individual action temporal features, group context features, and migration prior vectors to adjust the initial action probability distribution of each bird. Based on the group consistency index, jointly correct the adjusted action probability distribution so that the group behavior distribution approaches the expected behavior distribution during migration, obtain the final action probability distribution of each bird, and output the corresponding bird migration process action classification results.

[0176] In this specific embodiment, S1 specifically refers to:

[0177] First, within the preset time window It synchronously acquires continuous video data, inertial sensor data, satellite positioning data, and environmental image data, among which... Indicates the start time of the preset time window. This indicates the end time of the preset time window. The system constructs a unified timeline based on the timestamps accompanying the data collection from each data source and the synchronization trigger signal. ,in Indicates a unified timeline, Represents the first on the unified timeline Each discrete-time sampling moment This represents the total number of uniform time sampling points selected within a preset time window. By interpolating or resampling continuous video data, inertial sensor data, satellite positioning data, and environmental image data along the time dimension, the observations from each data source at the original timestamp are aligned to a uniform time axis. Up, so that at every moment All of them can obtain multi-source observations that are synchronized with each other in time;

[0178] Secondly, camera calibration and coordinate registration are performed on the continuous video data that has been synchronized in time to establish a unified spatial coordinate system. And using the camera intrinsic parameter matrix obtained from calibration Rotation matrix Translation vector Describing a unified spatial coordinate system The geometric relationship between the camera and the imaging plane makes it possible for any one of them to be in a unified spatial coordinate system. The three-dimensional position vector below Its homogeneous pixel position vector in the image plane Between satisfy ,in This represents the scaling factor used to represent pixel coordinates as homogeneous coordinates. This represents the homogeneous pixel coordinate vector in the camera image plane. This represents the camera intrinsic parameter matrix, which consists of parameters such as focal length and principal point. Indicates from a unified spatial coordinate system Rotation matrix to the camera coordinate system. Indicates from a unified spatial coordinate system The translation vector to the camera coordinate system. In a unified spatial coordinate system The three-dimensional spatial position vector below, through the above geometric relationship, can map satellite positioning data and video pixel coordinates to the same unified spatial coordinate system after coordinate transformation. This achieves time synchronization and spatial alignment.

[0179] After completing time synchronization and coordinate registration, for a unified timeline Every moment on The corresponding video frames are used for object detection and multi-object tracking to identify all individual birds appearing in the field of view within a preset time window, and an index is assigned to each bird. To uniquely identify the individual bird;

[0180] For each bird The system at every moment Extract the set of key point locations from the corresponding video frame, denoted as . ,in The bird is a high-dimensional vector, concatenated in a predefined order within a unified spatial coordinate system. The three-dimensional coordinates of multiple anatomical key points are used to characterize birds. At any moment The overall attitude state, by Arranging them chronologically will form a pattern related to birds. The corresponding attitude sequence;

[0181] To quantify the range of motion of birds between adjacent time points, the system calculates bird movement based on posture sequences. At every moment scalar of exercise intensity It can be defined as:

[0182] ;

[0183] in Represents birds At any moment Compared to the previous moment Overall intensity of movement The L2 norm of a vector is used to measure the Euclidean distance between keypoint position vectors at adjacent time points. According to a unified timeline By arranging the times in sequence, we can obtain the relationship with birds. The corresponding motion intensity sequence;

[0184] While extracting the pose, the system also relies on a unified spatial coordinate system. The key points below, from Calculating birds At every moment Representative spatial location vector This representative spatial location can be selected as the center of the torso, the center of the chest, or the weighted centroid of all key points, connected in chronological order. It can form with birds The corresponding spatial location information sequence is used to characterize the movement trajectory of birds within a preset time window;

[0185] Based on the spatial location information sequence, for each bird At every moment spatial location To perform spatial positioning, in order to Centered on the projection location in the image, image patches containing birds and their surrounding background areas are cropped from continuous video data and environmental image data to form a pattern resembling the bird. The corresponding collection of habitat image fragments, while focusing on birds Within a predefined spatial neighborhood centered on the target, potential threat targets are detected, identifying potential threat targets such as people, vehicles, and birds of prey. The potential threat target image fragments are then cropped based on the bounding boxes of the detected targets in the image.

[0186] For satellite positioning data, the system also uses a unified timeline. Perform time alignment for any bird and any uniform time sampling moment ,like Located at the original timestamps of two adjacent satellite positioning operations and Between these points, linear interpolation is used to estimate the bird's position at time [time value missing]. satellite positioning vector , can be represented as:

[0187] ;

[0188] in Represents birds At any moment The satellite positioning position vector, and They represent birds Original timestamp of satellite positioning and The satellite positioning vector at that location. and This represents two adjacent raw timestamps in a satellite positioning data sequence. This represents the interpolation weighting coefficients, whose values ​​are determined by time step (time). In the time interval The relative positions within are determined;

[0189] Subsequently, the geographic coordinate system, pre-established based on camera calibration and coordinate registration parameters, is converted to a unified spatial coordinate system. conversion function , satellite positioning location vector Mapped to a unified spatial coordinate system The three-dimensional position vector below ,Right now ,in Birds estimated from satellite positioning data In a unified spatial coordinate system The spatial location below, thus on a unified timeline and unified spatial coordinate system The formation of birds The corresponding satellite positioning time series; through the above steps of time synchronization, spatial alignment, target detection and multi-target tracking, key point location extraction, motion intensity calculation, spatial location information generation, habitat and potential threat target image cropping, and satellite positioning interpolation and coordinate transformation, within a preset time window You can obtain each bird at the same time. The corresponding attitude sequences, motion intensity sequences, spatial location information sequences, habitat image fragment sets, potential threat target image fragment sets, and satellite positioning time series provide unified and high-quality multi-source input data for subsequent rhythm-sensing state space modeling and population graph construction.

[0190] In this specific embodiment, S2 specifically refers to:

[0191] For each individual bird On a unified timeline Above, the posture sequence corresponding to this bird. and motion intensity sequence Input rhythm sensing state space model ,in This represents a unified timeline within a preset time window. Indicates the first Each time sampling moment, Indicates the number of time sampling points. Represents birds At any moment The attitude vector is a concatenated representation of the three-dimensional coordinates of multiple key points in a unified spatial coordinate system. Represents birds At any moment The scalar of exercise intensity;

[0192] Rhythm-aware state-space model This includes a slow-time-scale sub-model and a fast-time-scale sub-model. In the slow-time-scale sub-model, the preset time window is divided according to the larger time step. The sequence is divided into several slow time segments, and the pose sequence is downsampled or pooled within each time segment. Normalized pose features are extracted from each slow time segment. ,in Represents birds In the A pose statistical feature vector over a slow time segment This represents the index of a time segment at a slow time scale. Indicates the number of time segments on a slow time scale;

[0193] Slow-timescale submodel maintenance and birds The corresponding slow-timescale hidden state vector And based on the hidden state of the previous slow time segment and the pose statistics of the current segment, it is recursively updated, which can be represented as ,in Represents birds In the Slow-time-scale hidden state vectors over slow time segments This represents the nonlinear state transition function of the slow-time-scale sub-model, used to model the trend of slow attitude change over time at larger time steps;

[0194] In the fast timescale sub-model, a unified time axis is used. Each sampling moment on the time sequence is a time step, and the motion intensity sequence is... Perform fine-grained modeling and maintain relationships with birds The corresponding fast-timescale hidden state vector And based on the hidden state of the previous time step and the motion intensity of the current time step, it is recursively updated, which can be represented as ,in Represents birds At any moment The fast-timescale hidden state vector, The nonlinear state transition function represents the fast-time-scale sub-model, used to characterize the rapid change pattern of action intensity over a smaller time step;

[0195] After the preset time window is completed, the system fuses the hidden state sequences from the two time scales to obtain the sequence for each bird. Corresponding individual action temporal feature vector The slow-timescale hidden state sequence and the fast-timescale hidden state sequence can be jointly mapped using a feature fusion function as follows:

[0196] ;

[0197] in Represents birds Individual action temporal feature vectors within a preset time window This represents a fusion function used to weight and aggregate hidden states at two time scales and compress them into fixed-dimensional features. By performing frequency domain analysis and statistical feature extraction, we can obtain rhythmic features that characterize the dominant frequency of the movement, temporal consistency features that characterize the stability of the rhythm, and shape features that characterize the intensity variation pattern of the movement. Thus, the temporal features of an individual movement not only include posture and shape information, but also encode temporal structure information such as the frequency of movement occurrence, rhythm stability, and intensity fluctuation pattern.

[0198] In obtaining individual action temporal characteristics At the same time, the system also relies on spatial location information sequences Building with birds The corresponding trajectory sequence, where Represents birds At time in a unified spatial coordinate system The representative three-dimensional position vector, by... By connecting them in chronological order, a continuous trajectory curve can be obtained, which can be used as trajectory input for migration stage estimation and population map construction in subsequent steps;

[0199] In addition, in order to obtain semantic information about the habitat environment related to individual actions, the system will compare it with that of birds. The corresponding collection of habitat image fragments Input Environment Classification Network Obtain the habitat category label for this bird. , can be represented as ,in Represents birds A collection of habitat image fragments acquired within a preset time window. Represents an environmental classification network. Represents birds Discrete category labels of the environment;

[0200] Simultaneously, in order to obtain information on the impact of potential threats on the bird's behavior, the system will interact with the bird. A set of relevant potential threat target image fragments is input into a pre-trained target detection and recognition network. For each potential threat target instance, the network outputs its category label and spatial location vector in a unified spatial coordinate system, and also identifies the same bird species. The combination of all potential threat target categories and locations observed within a preset time window constitutes the potential threat target information set for the bird, which is used to construct the threat nodes and their relationship edges with bird nodes in the heterogeneous population graph in subsequent steps.

[0201] Through the above-mentioned processes such as rhythmic perception state space modeling, trajectory sequence generation, environmental classification, and extraction of potential threat target information, the final output is for each bird. Individual action timing characteristics Trajectory sequence, habitat category It also provides information on potential threat targets, offering fundamental input data with both temporal rhythm attributes and scene semantic attributes for subsequent swarm graph construction and swarm graph Transformer feature learning.

[0202] In this specific embodiment, S3 specifically refers to:

[0203] For each individual bird On a unified timeline and unified spatial coordinate system Below, utilizing its spatial location information sequence Individual action sequence characteristics Habitat category In addition to information on potential threat targets, corresponding bird nodes and their relationship edges with other nodes are constructed. This represents a unified timeline within a preset time window. Represents the first on the unified timeline Each time sampling moment, This indicates the number of time sampling points on a unified time axis. Represents birds At any moment In a unified spatial coordinate system The three-dimensional spatial position vector below, Represents birds Individual action temporal feature vectors extracted by a rhythm-aware state-space model within a preset time window. Represents birds output by the environmental classification network. Habitat category label;

[0204] In the group graph construction module, the first step is to create a graph for each bird. Create a bird node and position it in a unified spatial coordinate system. Spatial location and temporal characteristics of individual actions The initial node features of the bird node are used to characterize the individual spatial state and rhythmic attributes of the bird within the current preset time window;

[0205] Then, for any two different bird individuals and Based on its spatial location information sequence and Calculate at each time step spatial distance , can be defined ,in Represents birds and birds At any moment The spatial Euclidean distance between them This represents the L2 norm, used to measure the straight-line distance between two three-dimensional position vectors. It can also be used to calculate the relative motion relationship feature vectors of two birds in terms of speed, direction of movement, etc., based on a sequence of spatial position information, to characterize the similarity of their movement patterns. A preset distance threshold can then be set based on this. And the preset similarity conditions, when The statistic within the preset time window is less than the threshold. When the relative motion relationship between two birds meets the preset similarity condition, a bird node connection edge is established between the corresponding two bird nodes, and the feature quantity representing the average spatial distance and relative motion similarity is used as the edge feature of the bird node connection edge.

[0206] Secondly, based on the habitat category labels of all birds. By statistically analyzing different environment types, an environment node is created for each type of habitat in the heterogeneous population graph. The category encoding vector of that habitat type is used as the initial node feature of the environment node. When birds... When activity within the specified environment category exists within a preset time window, an environment connection edge is established between the bird node and the corresponding environment node, and based on the bird... The cumulative dwell time and the intensity of actions during the dwell time under this environmental category are used to construct the edge features of the environmental connection edges, which are used to quantify the degree of association between birds and this habitat.

[0207] Secondly, based on the potential threat target information, all potential threat targets are categorized. A threat node is set for each category of potential threat targets in the heterogeneous population graph. The encoding vector representing the threat target category is used as the initial node feature of the threat node. For each specific instance of a potential threat target, its coordinates in a unified spatial coordinate system are used... The spatial position vector of birds The spatial distance and relative approach speed between birds are measured when a bird When the bird node is less than a preset threat impact threshold in space and the relative approach speed is greater than a preset speed threshold, a threat connection edge is established between the bird node and the corresponding threat node. The statistical characteristics such as the minimum spatial distance, average spatial distance and relative approach speed between the bird and the threat target are used as edge features of the threat connection edge to characterize the threat intensity of the bird.

[0208] Furthermore, in order to explicitly represent the overall flock status within the current preset time window in the flock graph, the flock graph construction module sets up a global flock node to display the temporal characteristics of all individual bird actions. and spatial location information sequence Perform statistical aggregation at the group level to construct a group statistical feature vector. The characteristics of movement rhythm and trajectory can be summarized using statistical methods such as mean, variance, and extreme values. As the initial node features of the global nodes of the group, global connection edges are established between the global nodes of the group and each bird node. The individual action time sequence feature statistics and trajectory statistics of each bird within a preset time window are used as the edge features of the corresponding global connection edges to characterize the contribution relationship of a single bird to the overall group state.

[0209] By creating bird nodes, environment nodes, threat nodes, and global community nodes, and establishing various types of connecting edges, a heterogeneous community graph structure can be formed. ,in This represents a heterogeneous population map constructed within the current preset time window based on information about multiple birds, their environment, and threats.

[0210] While constructing the heterogeneous population graph structure, the migration stages of the trajectory sequences of multiple birds were also estimated. Specifically, the migration stages of each bird could be estimated. The trajectory sequence is represented as ,in Represents birds The trajectory sequence within a preset time window is then used to estimate the migration stage. Extract migration-related trajectory features and migration distance. , rest duration and velocity change characteristics etc., where migration distance It can be obtained by accumulating the displacements between adjacent time steps. ,in Represents birds The cumulative spatial movement distance within a preset time window and They represent birds Spatial location vectors at adjacent sampling times, The L2 norm is used to measure the Euclidean distance between adjacent position vectors, and the dwell time. Statistical analysis of velocity change characteristics can be performed based on the length of time segments where the velocity is below a preset threshold. Encoding can be based on the time-varying characteristics of velocity magnitude and direction;

[0211] Set of trajectory features of all individual birds Aggregation at the group level yields a migration feature vector representing the current overall migration status. ,in This indicates the number of individual birds detected and included in the model within the current preset time window. This represents the population migration feature vector obtained after weighted averaging or other aggregation operations on all bird trajectory features;

[0212] Finally, the migration feature vector Input the mapping subnetwork of the migration stage estimation model, and pass it through a nonlinear mapping function. Obtain the prior vector of the migration stage ,in This represents the prior vector of the migration stage corresponding to the current preset time window. This represents a nonlinear mapping function that maps group migration features to the prior space of migration stages. This function is used as conditional prior information to modulate the attention distribution and behavior expectation distribution during subsequent group graph Transformer modeling and action classification, thereby completing step S3 on the heterogeneous group graph structure. Construction and migration stage prior vectors The generation of .

[0213] In this specific embodiment, S4 specifically refers to:

[0214] heterogeneous population graph structure and migration stage prior vector and the temporal characteristics of individual bird movements. Input the population graph Transformer model together, where This represents a heterogeneous community graph constructed within the current preset time window, containing a set of nodes and a set of edges. This represents the prior vector of the migration stage corresponding to the current preset time window. Represents birds extracted from a rhythm-aware state-space model. Individual action temporal feature vector;

[0215] In heterogeneous population graph structure In this context, a set of nodes can be represented as ,in Represents the set of indices for all bird nodes. This represents the set of indexes for environment nodes. A set of indexes representing threat nodes. Represents a unique global node index for the group;

[0216] For each bird node Individual action sequence characteristics Assign this node its initial node characteristics, and denote it as... ,in This represents the bird nodes in the 0th layer (i.e., the input layer) of the group graph Transformer. eigenvectors;

[0217] For each environmental node The habitat category labels are encoded using a pre-trained embedding network to obtain the environment category feature vector, denoted as . and will Assigning environmental nodes As the initial node feature, let it be denoted as ,in Represents environment nodes Vectorized representation of the corresponding habitat category Represents environment nodes The feature vector at layer 0;

[0218] For each threat node The categories and statistical attributes of potential threat targets are encoded into threat feature vectors, denoted as... and will Assign threat nodes As the initial node feature, let it be denoted as ,in Indicates a threat node Vectorized representation of threat target type and intensity information. Indicates a threat node Feature vector at layer 0; for global nodes in the population The population statistical feature vector is obtained by aggregating the temporal features of individual movements and trajectory statistics of all birds. Assign this node as the initial node characteristic, denoted as ,in This indicates the statistical characteristics of the overall bird flock within the current preset time window. This represents the feature vector of the global node in the group at layer 0;

[0219] Meanwhile, in order to explicitly introduce prior information about the migration stage into the attention computation, a migration stage embedding network is set up inside the population graph Transformer model to embed the prior vector of the migration stage. Perform a nonlinear mapping to obtain the migration stage embedding vector. , can be represented as ,in This represents the nonlinear mapping function implemented by the embedded network during the migration stage. This represents the vectorized representation of prior information about the migration stage in the embedding space;

[0220] The group graph Transformer uses a hierarchical message-passing structure, assuming a total of Layers, each layer For any node Maintaining feature vectors ,in Representation layer index, Represents a node In the Layer feature representation;

[0221] In the During the layer message passing phase, the group graph Transformer performs attention calculations on bird node connections, environment connections, threat connections, and global connections according to edge type. Mark the edge type. Where bird represents the type of connection edge between bird nodes, env represents the type of connection edge between bird nodes and environment nodes, thr represents the type of connection edge between bird nodes and threat nodes, and global represents the type of connection edge between bird nodes and global nodes of the group;

[0222] For nodes and through which the type is The set of neighboring nodes directly connected by the edge ,in Indicates the edge type Next node The neighbor set of the group graph Transformer calculates attention scores and attention weights based on the current layer node features, edge features, and migration stage embedding vectors. For edge type Time connection node and nodes The edge feature vectors, The attention score, derived from the equilateral features of spatial distance, dwell time, and threat intensity constructed in step S3, can be expressed as: ,in Indicates the first Layer and edge types are Time from node To the node Unnormalized attention score Indicates the type of edge The learnable scoring function is designed using several fully connected layers and nonlinear activations. This represents a vector concatenation operation. and Representing nodes respectively and nodes In the The feature vector of the layer, Representing an edge In type The edge features below, This represents the migration stage embedding vector, which is used to incorporate migration stage priors as conditional information into attention computation.

[0223] Subsequently, for the same node In the same side type The scores of all neighboring nodes are normalized to obtain the attention weight. , can be represented as ,in Indicates the first Layer from node To the node Normalized attention weights Indicates at node Neighbor set The operation of normalizing by the exponential function;

[0224] After obtaining the attention weights on each type of edge, the community graph Transformer performs weighted aggregation of information from different types of neighbors and updates the node features. In the The feature update of a layer can be represented as ,in Represents a node In the The feature vector of the layer, This represents a non-linear activation function (ReLU or LeakyReLU). Represents edge type The corresponding linear transformation matrix is ​​used to map the features of neighboring nodes to a unified aggregation space. For the corresponding attention weights, by different types The neighbor node features are weighted and summed to achieve message passing and feature fusion;

[0225] Due to attention scoring function The middle explicitly includes the migration stage embedding vector Therefore, under different periods and migration stages, the population graph Transformer can automatically adjust the degree of attention to different types of connecting edges and different neighbor nodes, thereby realizing a graph attention mechanism based on migration stage conditions, and differentially modeling various relationships such as bird-bird, bird-environment, bird-threat, and bird-population global relationships; after After layer message passing and feature aggregation, the final feature representation of each node can be obtained, for example, the bird node. , and put it in the first eigenvectors of the layer As a group context feature of this bird within the current preset time window, let it be denoted as ,in This indicates a comprehensive consideration of individual movement rhythm information, surrounding bird behavior, habitat environment and potential threats, as well as the prior group context characteristics during the migration phase;

[0226] For group global nodes , and put it in the first eigenvectors of the layer As the global feature of the entire flock within the current preset time window, it is denoted as ,in This represents a global representation of group behavior patterns and environmental states after fusing multi-layer message passing results. Finally, step S4 outputs the results for each bird. Corresponding group context features and group global characteristics This provides high-level semantic features that combine group structure information and prior constraints of migration stages for subsequent action classification networks when determining individual behavior.

[0227] In this specific embodiment, S5 specifically includes:

[0228] For each individual bird detected and included in the model within the current preset time window ,in , This indicates the number of individual birds within the current preset time window. The system will use the individual action time sequence feature vector obtained in step S2. The group context feature vector obtained in step S4 The global feature vector of the population obtained in step S4 and the migration stage prior vector obtained in step S3 By performing feature concatenation, a comprehensive input feature vector for the bird's action classification is constructed. , can be represented as ,in Represents birds The comprehensive feature vector in the input layer of the action classification network This represents a vector concatenation operation. Represents birds extracted from a rhythm-aware state-space model. The temporal characteristics of individual actions, This indicates that the Transformer graph incorporates both group structure relationships and prior information about birds during migration stages. Group context features, This represents the global features of the population output by the global nodes in the population graph Transformer. This represents the migration stage prior vector output by the migration stage estimation model;

[0229] Then, the integrated input feature vectors will be... Input into motion classification network Action classification network Internally, it includes several fully connected layers and nonlinear activation structures, and at the output end, it sets a classification substructure with a normalized output layer. First, it... Mapped to unnormalized score vectors of action categories , can be represented as ,in Represents birds The original score vector for each preset action category; the set of preset action categories is denoted as... ,in This represents the set of indices for all preset action categories. This indicates the total number of preset action categories. Indicates the first One preset action category;

[0230] In the normalized output layer Applying the softmax function to the original score vector Convert to initial action probability distribution vector , can be represented as ,in Represents birds The initial action probability distribution vector across all preset action categories. Represents birds Belongs to the The initial probability of a preset action category, Represents birds In the The original scores on each preset action category This represents an exponential function, with the denominator being the sum of the original score indices over all action categories, used to normalize the scores into a probability distribution;

[0231] After obtaining the initial action probability distribution Then, the system calculates the probability distribution for each bird based on its degree of dispersion. Prediction uncertainty index Prediction uncertainty index The degree of concentration or fuzziness in the probability distribution of the initial action can be defined using a form based on information entropy. ,like ,in Represents birds The prediction uncertainty of the action classification results within the current preset time window. The information entropy represents the logarithmic function, usually the natural logarithm. The value of this information entropy is larger when the probability distribution is more uniform and smaller when the probability distribution is highly concentrated, thus reflecting whether there is significant ambiguity in the action judgment.

[0232] Obtain the initial movement probability distribution of all birds at the individual level. Subsequently, the system statistically aggregates these probability distributions at the group level to construct a group behavior distribution vector within the current preset time window. It can be obtained by using a simple average method. ,in This represents the overall probability distribution of the entire flock of birds across all preset action categories within the current preset time window. Indicates to The average weighting coefficient for each bird species. This represents the element-wise summation operation over the probability distribution vectors of the initial actions of all birds;

[0233] To introduce migration-stage priors as constraints on behavioral proportions at the group level, the system uses migration-stage prior vectors... Input behavior expectation generation module The expected behavior distribution vector of the migration stage corresponding to the current migration stage is obtained through nonlinear mapping. , can be represented as , in This represents the nonlinear mapping function implemented by the behavior expectation generation module. This represents the expected distribution of the proportion of behaviors in each preset action category at the current migration stage. This distribution can be derived from historical statistical data or ecological prior knowledge and learned through network parameters.

[0234] Subsequently, the system distributed the group behavior. Distribution of expected behaviors during migration stages By comparing the two groups, a population consistency index reflecting the degree of difference between them is calculated. It can be defined using a metric based on vector distance. ,like ,in An indicator representing the consistency between the current distribution of group behavior and the expected distribution of behavior during the migration phase. The first norm of a vector is used to measure the sum of the absolute differences between two behavior distribution vectors across all preset action categories. The larger the value, the greater the deviation between the actual behavior of the group and the expected behavior at each stage; the smaller the value, the more consistent the behavior of the group with the prior of the migration stage.

[0235] Through the above processing, step S5 outputs each bird at the individual level. Initial action probability distribution and the corresponding prediction uncertainty index Output the distribution of group behavior within the current preset time window at the group level. Distribution of expected behaviors during migration stages based on prior knowledge of the migration stages. and the group consistency index that characterizes the degree of difference between the two. This provides the necessary input for the action result correction module in the subsequent step S6.

[0236] In this specific embodiment, S6 specifically refers to:

[0237] For each individual bird within the current preset time window ,in This indicates the total number of bird individuals modeled within the current preset time window. The motion result correction module receives the initial motion probability distribution of the bird from step S5. Prediction uncertainty index Receive the group behavior distribution from step S5 Distribution of expected behaviors during migration stages and group consistency index Simultaneously, the individual action timing features corresponding to the bird are received from steps S2 and S4 respectively. Group context features and the migration stage prior vector received from step S3 ,in Represents birds The initial action probability distribution vector for each preset action category. Indicates based on The calculated prediction uncertainty scalar is used to characterize the degree of dispersion of the initial distribution. This represents the group behavior distribution vector obtained by averaging the initial action probability distributions of all birds along the group dimension. It is used to characterize the proportion of the overall bird flock's behavior across each action category within the current preset time window. This represents the expected behavior distribution vector for the migration stage, generated by the behavior expectation generation module based on the prior vector of the migration stage. It is used to characterize the proportion of each action category expected to occur from the perspective of prior knowledge at the current migration stage. Indicates the use of quantization and A scalar measure of group consistency, indicating the degree of difference between actual and expected behavior. A larger value indicates a more significant deviation between the group's actual and expected behavior. This represents birds obtained by joint modeling on both slow and fast time scales using a rhythm-aware state-space model. The individual action temporal feature vector contains information such as action dominant frequency, rhythm stability, and action intensity variation pattern. This indicates that the heterogeneous population graph and the population graph Transformer combine population structure relationships and a priori and subsequent information about birds during migration. The group context feature vector, This represents the migration stage prior vector output by the migration stage estimation model, used to represent the vectorized prior information of the current overall migration state;

[0238] In the individual-level correction stage, the fusion weight calculation unit within the action result correction module first calculates the prediction uncertainty. With features and Input together into the fusion weight calculation function In this process, the unnormalized feature weight vector is obtained. , can be represented as:

[0239] ;

[0240] in This represents a fusion weight calculation function composed of several fully connected layers and nonlinear activation units, used to adaptively estimate the importance of each feature based on prediction uncertainty and the content of the three types of features. This represents a vector of length three, whose three components are as follows: and The system indicates that it is for birds. Unnormalized weights assigned to individual action temporal features. Represented as unnormalized weights assigned to group context features. Represented as the unnormalized weights assigned to the prior vectors during the migration phase;

[0241] To facilitate the weighted summation of the three types of information in subsequent feature fusion, the system... Applying a softmax normalization operation yields a normalized fusion weight vector. , can be represented as ,in This represents a normalized weight vector of length three, whose three components sum to one. Each component is denoted as... and This represents the fusion weight of the temporal features of an individual action in action determination. The fusion weights represent the group context features. This represents the fusion weights of the prior vectors in the migration phase, and the sum of the three is equal to one;

[0242] After obtaining the normalized fusion weight vector The system then uses this vector to perform a weighted summation of the three types of features, forming a correction action discrimination feature vector for individual-level action correction. , can be represented as:

[0243] ;

[0244] in This represents the corrected action discrimination feature vector after fusing individual action information, group context information, and prior information about the migration stage, driven by prediction uncertainty. When the value is large, the fusion weight calculation function The tendency is to assign higher weights to group context features and migration stage priors, thereby... and Increase the size of the group and prior information to enhance the influence of group information and prior information on action decision-making. When smaller, This relies more on the temporal characteristics of individual actions, increasing ;

[0245] Subsequently, the correction classification subnetwork within the action result correction module by As input, output the modified action probability distribution vector of the bird for each preset action category. , can be represented as:

[0246] ;

[0247] in This represents a correction classification subnetwork with a similar structure to the action classification network in step S5, but with independent parameters. Internally, it also includes several fully connected layers and a normalized output layer, used to re-estimate birds in the feature space after adaptive adjustment of the fused weights. Action category probability, This represents the corrected action probability distribution vector obtained based on the corrected action discrimination features;

[0248] In obtaining the corrected action probability distribution Then, the system compares it with the initial action probability distribution. The birds were obtained by weighting and fusing according to the weighting coefficients of the uncertainty. Individually corrected action probability distribution vector Weighting coefficients can be set. To predict uncertainty Monotonically decreasing function:

[0249] And order ;

[0250] in This represents the weighting coefficient of the initial action probability distribution in the fusion process. This represents the weighting coefficients of the modified action probability distribution in the fusion process. This represents a hyperparameter greater than zero, used to control prediction uncertainty. The strength of the influence on the weighting coefficients, when When it is large, Reduce As the number of features increases, the system becomes more reliant on the corrected action probability distribution obtained through feature fusion. ,when When smaller, Increased, the system retains more of the initial action probability distribution. ;

[0251] After completing individual-level calibration for all birds, the system is based on The distribution of group behavior within the current preset time window is statistically analyzed again at the group level, and denoted as... This can be obtained using a simple averaging method:

[0252] ;

[0253] in This represents the behavioral distribution of the entire flock across various preset action categories after individual-level corrections, used to compare with the expected behavioral distribution during migration. Compare;

[0254] During the group-level joint adjustment phase, the system first performs adjustments based on the calculation method of the group consistency index defined in step S5. and The differences are assessed, and if the updated group consistency index is still greater than the preset threshold, the outcome is determined. If the actual behavior distribution of the current group deviates significantly from the expected behavior distribution during the migration stage, it is necessary to further adjust the individual corrected action probability distribution of each bird at the group level. In this case, the system adopts an iterative approach to gradually guide the group behavior distribution toward the expected behavior distribution during the migration stage.

[0255] Specifically, the iteration index can be set as , , This represents the maximum number of iterations for group-level joint adjustment, where the probability distribution of each bird's action at the initial iteration is set to... ,in This represents the initial distribution in the group-level iterative correction process, in each iteration. In the process, the probability distribution of the current actions of all birds is first averaged to obtain the iteration step. Corresponding group behavior distribution , can be represented as ,in Indicates the first Birds in the next iteration The action probability distribution vector, Indicates the first The behavioral distribution of the entire flock across all action categories during the next iteration;

[0256] In estimate Then, based on the requirement of group consistency, the system transforms the probability distribution of each bird's actions into the expected behavior distribution for the migration phase. The direction can be updated using linear interpolation as follows:

[0257] ;

[0258] in Indicates the first After the second iteration, birds The action probability distribution vector, This represents the preset group adjustment step size coefficient, which takes a value between zero and one. It is used to control the degree to which the individual distribution converges to the desired behavioral distribution during the migration stage in each iteration. If we want to achieve a higher group consistency index... Adjusting more quickly when the value is larger allows for adjustments based on... Adaptive settings The value of ;

[0259] Through the above iterative updates, the average behavior distribution of the group increases with the number of iterations. Expected Behavior Distribution as We Approach the Migration Phase When the number of iterations reaches the maximum value Or the group consistency index drops below the threshold. At that time, the group-wide joint adjustment was stopped, and each bird was placed under control. The action probability distribution obtained in the last iteration As the probability distribution of the bird's final action within the current preset time window, It can be taken as the iteration step corresponding to the stopping iteration. Below Finally, the system determines the outcome based on the probability distribution of the final action. The values ​​for all preset action categories are used to output the corresponding bird migration process action classification results for each bird.

[0260] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0261] This invention constructs a closed-loop link at the algorithm level, from "individual rhythm modeling - group structure modeling - stage prior constraints - uncertainty correction", by organically combining rhythm perception state space model, heterogeneous population graph and population graph Transformer, migration stage prior and uncertainty-driven correction mechanism. This enables each module to play a role around the core technical problem of "ambiguity resolution of multiple behaviors corresponding to the same posture". The rhythmic perception state-space model models posture and motion intensity at both slow and fast time scales, extracting temporal features of individual actions that can distinguish different behavioral rhythmic patterns. This enhances the ability to characterize behavioral differences hidden behind similar postures from the source. The heterogeneous group graph and its group graph Transformer introduce structural relationships between multiple birds and between them and the environment and threats. Graph attention propagation is performed under the prior conditions of the migration stage, enabling individual determination to fully utilize the group context and stage semantics. Finally, through the joint constraints of prediction uncertainty, group behavior distribution, and expected behavior distribution during the migration stage, two-level corrections are made to the initial classification results at the individual and group levels, thereby effectively improving the overall accuracy and robustness of action classification during the migration process.

[0262] In terms of algorithm structure, this invention has made targeted structural improvements and coupling designs for each sub-module to address the pose-behavior ambiguity problem:

[0263] On the one hand, a rhythmic perception state space model coupled with slow / fast time scales is introduced at the individual representation level, so that the output individual action temporal features naturally contain key discriminative information such as action dominance frequency, rhythm stability and intensity change pattern;

[0264] On the other hand, at the group modeling level, a heterogeneous group graph is constructed by introducing environmental nodes, threat nodes, and global group nodes. In the graph Transformer, migration stage conditional graph attention is adopted to assign adaptive weights to different types of edges at different times, which structurally enhances the control of the direction and intensity of group information propagation by the stage prior.

[0265] Furthermore, at the decision-making level, prediction uncertainty, group consistency index, and behavior expectation distribution based on stage prior are introduced to form a two-level feedback structure of "first making a decision, and then adaptively correcting according to uncertainty and group consistency". This enables the algorithm to automatically increase its dependence on the group and stage prior when there is insufficient attitude information or a lot of noise.

[0266] The aforementioned structural improvements work together to enable this case to better achieve improved classification accuracy, stability, and ecological adaptability in resolving pose-behavior ambiguity issues in migration scenarios, compared to traditional methods that rely solely on single-unit sequences or simple feature fusion.

Claims

1. A method for classifying bird migration process actions based on machine learning, characterized by, include: S1. Acquire multi-source raw data to obtain the posture and movement intensity sequences, spatial location information, habitat environment and potential threat target image fragments, and satellite positioning time series for multiple birds; S2. Input the posture and motion intensity sequence into the rhythm perception state space model for temporal modeling to obtain the individual action temporal characteristics of each bird. Based on spatial location information, habitat image fragments and potential threat target image fragments, obtain the trajectory sequence, habitat category and potential threat target information of each bird. S3. Input the temporal features of individual actions, trajectory sequences, habitat categories, information on potential threats and spatial locations into the group graph construction module to construct a heterogeneous group graph structure containing bird nodes, environment nodes, threat nodes and global group nodes. Input the trajectory sequence into the migration stage estimation model to obtain the migration stage prior vector. S4. Input the heterogeneous group graph structure, migration stage prior vectors and individual action temporal features into the group graph Transformer model, and use the graph attention mechanism with the migration stage prior vector as a condition for message passing and feature aggregation to obtain the group context features and group global features of each bird. S5. Input the individual action temporal features, group context features, group global features, and migration stage prior vectors into the action classification network to obtain the initial action probability distribution and prediction uncertainty index of each bird in each preset action category. Statistically analyze the initial action probability distribution of all birds to obtain the group behavior distribution. Input the migration stage prior vectors into the behavior expectation generation module to obtain the expected behavior distribution of the migration stage. Compare the group behavior distribution with the expected behavior distribution of the migration stage to obtain the group consistency index. S6. Input the initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, group consistency index, individual action temporal features, group context features, and migration prior vectors into the action result correction module. Based on the prediction uncertainty index, adaptively determine the fusion weights of individual action temporal features, group context features, and migration prior vectors to adjust the initial action probability distribution of each bird. Based on the group consistency index, jointly correct the adjusted action probability distribution so that the group behavior distribution approaches the expected behavior distribution during migration, obtain the final action probability distribution of each bird, and output the corresponding bird migration process action classification results.

2. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S1 specifically refers to: Acquire continuous video data, inertial sensor data, satellite positioning data, and environmental image data collected within a preset time window. Based on timestamps and synchronization trigger signals, synchronize the continuous video data, inertial sensor data, satellite positioning data, and environmental image data in time, and align the multi-source raw data corresponding to the same preset time window to a unified time axis. Camera calibration and coordinate registration are performed on the continuous video data, and the pixel coordinates in the continuous video data are spatially aligned with the geographic coordinates corresponding to the satellite positioning data to establish a unified spatial coordinate system. Based on the completion of time synchronization and coordinate registration, target detection and multi-target tracking are performed on the continuous video data to identify and track each bird. In each frame, the key point position or contour features of the corresponding bird are extracted and arranged in chronological order to form a posture sequence corresponding to each bird. The motion amplitude is calculated based on the changes in key point position or contour between adjacent frames to form a motion intensity sequence corresponding to each bird. Based on the bird positions in the unified spatial coordinate system, the spatial positions of each bird in each frame are connected in chronological order to obtain the spatial position information corresponding to each bird. Centered on the bird's location, image segments containing the background area around the bird are cropped from the continuous video data and the environmental image data as habitat environment image segments corresponding to each bird. Target detection is performed on the area around the bird in the continuous video data to identify potential threat targets such as people, vehicles, and birds of prey. Image segments containing these potential threat targets are cropped as potential threat target image segments. The satellite positioning data is time-aligned, interpolated, and transformed to align the satellite positioning data corresponding to each bird to the unified time axis and unified spatial coordinate system, forming a satellite positioning time series corresponding to each bird. Output the attitude sequence, motion intensity sequence, spatial location information, habitat image fragments, potential threat target image fragments, and satellite positioning time series.

3. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S2 specifically refers to: The posture sequence and motion intensity sequence are input into the rhythm-aware state space model. A slow time-scale sub-model and a fast time-scale sub-model are set in the rhythm-aware state space model. The slow time-scale sub-model is used to model the posture sequence at a larger time step to represent the slow change of posture over time. The fast time-scale sub-model is used to model the motion intensity sequence at a smaller time step to represent the rapid change of motion intensity. The hidden states of the slow time-scale sub-model and the fast time-scale sub-model are fused to obtain the individual motion temporal features corresponding to each bird. The individual motion temporal features include at least motion rhythm features that represent the dominant frequency of motion, rhythm stability, and motion intensity change pattern. Based on spatial location information, the spatial locations of each bird in a unified spatial coordinate system are connected in chronological order to form a trajectory sequence corresponding to that bird. Based on habitat image fragments, an environmental classification network is used to classify the environment of each bird's location to obtain the habitat category corresponding to that bird. Based on image fragments of potential threat targets, the category and spatial location of the potential threat targets in a unified spatial coordinate system are determined by a target detection and recognition network, and the category and spatial location of the potential threat targets are combined into potential threat target information; Output the individual movement time sequence characteristics, trajectory sequence, habitat category, and potential threat target information for each bird.

4. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S3 specifically refers to: Individual action temporal features, trajectory sequences, habitat categories, and potential threat target information and spatial location information are input into the group graph construction module. In the group graph construction module, for each bird, a corresponding bird node is generated based on the bird's spatial location information in a unified spatial coordinate system, and the individual action temporal features of the bird are used as the initial features of the bird node. Based on the spatial location information and trajectory sequence of multiple birds in a unified spatial coordinate system, calculate the spatial distance and relative motion relationship between any two birds. When the spatial distance is less than a preset distance threshold or the relative motion relationship meets a preset similarity condition, establish a bird node connection edge between the corresponding two bird nodes, and use the spatial distance and relative motion relationship as the edge feature of the bird node connection edge. Based on the habitat category corresponding to each bird, an environment node is generated for each type of habitat. An environment connection edge is established between the bird node active in the habitat and the corresponding environment node. The environment-related statistics, such as the bird's stay time in the habitat, are used as the edge features of the environment connection edge. Based on the information of potential threat targets, a threat node is generated for each type of potential threat target. Based on the distance relationship between the spatial position of the potential threat target in the unified spatial coordinate system and the spatial position of the bird node, a threat connection edge is established between the bird node and the corresponding threat node when the distance is less than the preset threat impact threshold. The spatial distance and relative approach speed between the bird node and the threat node are used as the edge features of the threat connection edge. A global group node is set up to represent the overall group state, and a global connection edge is established between the global group node and all bird nodes. The individual action time sequence statistics and trajectory sequence statistics of bird nodes within a preset time window are used as the edge features of the global connection edge, thereby constructing a heterogeneous group graph structure that includes bird nodes, environment nodes, threat nodes and global group nodes. Simultaneously, the trajectory sequences of multiple birds are input into the migration stage estimation model. In the migration stage estimation model, the migration distance, rest duration, and speed changes of each bird within the preset time window are calculated based on the trajectory sequences. The trajectory features are then aggregated at the group level to obtain a migration feature vector representing the current overall migration state. The migration feature vector is then mapped to a migration stage prior vector corresponding to the preset time window. Output the heterogeneous population graph structure and the prior vectors for the migration stages.

5. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S4 specifically refers to: The initial node features and edge features of bird nodes, environment nodes, threat nodes, and global nodes in the heterogeneous population graph structure are input together with the migration stage prior vector and individual action temporal features into the population graph Transformer model. Specifically, the individual action temporal features corresponding to each bird are assigned to the corresponding bird node as the initial feature of that bird node, the environment category features obtained by encoding the habitat category are assigned to the environment node as the initial feature of the environment node, the threat features obtained by encoding the potential threat target information are assigned to the threat node as the initial feature of the threat node, and the population statistical features obtained by aggregating the individual action temporal features and trajectory sequence statistics of all bird nodes in the heterogeneous population graph structure are assigned to the global nodes as the initial feature of the global nodes. The migration stage prior vector is input into the migration stage embedding network and mapped to a migration stage embedding vector. In each layer of the population graph Transformer model, the migration stage embedding vector is combined with node features and edge features to calculate the attention weights of bird node connection edges, environment connection edges, threat connection edges and global connection edges. Different attention parameters are used for different types of connection edges, thereby forming a message passing method based on the migration stage conditional graph attention mechanism. In each layer, the features of adjacent nodes are weighted and summed according to the attention weight and fused with the features of the current node to update the feature representations of bird nodes, environment nodes, threat nodes and group global nodes. After multi-layer message passing and feature aggregation, the updated feature representation of each bird node is used as the group context feature corresponding to that bird, and the updated feature representation of the group global node is used as the group global feature. Output the group context features and the group global features.

6. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S5 specifically refers to: Individual action temporal features, group context features, group global features, and migration stage prior vectors are used as input features and fed into the action classification network. In the action classification network, for each bird, the individual action temporal features corresponding to the bird are concatenated and nonlinearly transformed with the group context features corresponding to the bird, the group global features, and the migration stage prior vectors to obtain the bird's action discrimination feature vector. Based on the action discrimination feature vector, the initial action probability distribution of the bird in multiple preset action categories is calculated through a classification structure with a normalized output layer. The prediction uncertainty index of the bird is calculated based on the probability value distribution of the initial action probability distribution. The prediction uncertainty index is used to characterize the degree of dispersion of the initial action probability distribution. The initial action probability distributions of all birds are weighted and summed or averaged in the group dimension to obtain the group behavior distribution in multiple preset action categories within the current preset time window; The migration stage prior vector is input into the behavior expectation generation module, and the migration stage prior vector is nonlinearly mapped in the behavior expectation generation module to obtain the migration stage expected behavior distribution corresponding to the migration stage prior vector on multiple preset action categories. The distribution of group behavior is compared with the expected distribution of behavior during the migration stage, and a group consistency index reflecting the degree of difference between the two is calculated based on the difference, ratio or distance between the two in multiple preset action categories. The initial action probability distribution and prediction uncertainty index are output, and the group behavior distribution, expected behavior distribution during the migration stage, and group consistency index are obtained at the group level.

7. The method for classifying bird migration movements based on machine learning according to claim 1, characterized in that, S6 specifically refers to: The initial action probability distribution, prediction uncertainty index, group behavior distribution, expected behavior distribution during migration, and group consistency index, along with individual action temporal features, group context features, and migration stage prior vectors, are input into the action result correction module. In the action result correction module, for each bird, the prediction uncertainty index of that bird, along with the corresponding individual action temporal features, group context features, and migration stage prior vectors, are input into the fusion weight calculation unit. Through nonlinear mapping, the individual action temporal feature weights, group context feature weights, and migration stage prior vector weights used for the bird's action determination are obtained respectively. The individual action temporal feature weights, group context feature weights, and migration stage prior vector weights are normalized to obtain the fusion weights corresponding to that bird in action determination. The individual action temporal features, group context features, and migration stage prior vectors of the bird are weighted and summed using the fusion weights to obtain the corrected action discrimination feature vector of the bird. Based on the corrected action discrimination feature vector, the corrected action probability distribution of the bird in each preset action category is calculated through a corrected classification sub-network. The initial action probability distribution of the bird is weighted and adjusted using the corrected action probability distribution to obtain the individual corrected action probability distribution of the bird. At the group level, based on the group consistency index and the expected behavior distribution during the migration stage, the individual-corrected action probability distribution of all birds is jointly adjusted in the group dimension. When the group consistency index indicates that the difference between the current group behavior distribution and the expected behavior distribution during the migration stage is greater than a preset threshold, the individual-corrected action probability distribution of each bird is corrected iteratively so that the group behavior distribution obtained by statistical analysis in the group dimension after correction gradually approaches the expected behavior distribution during the migration stage. After the iteration is completed, the group-adjusted action probability distribution of each bird is taken as the final action probability distribution of that bird within a preset time window, and the corresponding bird migration process action classification result is output according to the final action probability distribution.

Images