A sparse scene forestry wildlife trajectory prediction method, system, device and medium

Abstract

This invention relates to a method, system, device, and medium for predicting the trajectory of forestry wildlife in sparse scenes. The method includes: dividing the study area into hexagonal grids and extracting multi-dimensional environmental features to construct an environmental state space; pairing start and end points of historical occurrence point data and mapping them to hexagonal nodes to generate a sparse movement task set; calling a random forest model to learn the relationship between environmental features and occurrence points to generate an environmental resource suitability distribution map; learning individual path preferences based on the sparse movement task set and resource suitability distribution map and iteratively training through reinforcement learning to generate an individual experience memory bank; obtaining the current start and end point positions of the target animal and mapping them to the environmental state space, and performing weighted path search; using attention weight allocation, selecting the node sequence with the highest score after weighted integration as the main prediction path, and using the individual paths generated by virtual individuals with high weight rankings as alternative prediction paths.

Description

Technical Field

[0001] This invention relates to the field of forestry wildlife conservation technology, specifically to a method, system, device, and medium for predicting the trajectory of forestry wildlife in sparse scenes. Background Technology

[0002] In the field of forestry wildlife conservation and management, predicting animal movement trajectories is a key technical means for achieving early warning of human-wildlife conflict, ecological corridor planning, and habitat protection. Traditional methods for predicting wildlife movement trajectories mainly rely on two technical approaches: one is Bayesian inference methods based on continuous GPS tracking data, which estimate animal movement paths between adjacent time points by constructing Brownian bridge movement models or hidden Markov models; the other is empirical selection models based on environmental characteristics, which infer animal movement preferences among different habitat types by analyzing the statistical relationship between animal appearance points and environmental factors. Both of these methods can achieve good predictive results in scenarios with sufficient research data and stable animal activity patterns, providing important support for wildlife management and conservation decision-making.

[0003] However, existing technologies have significant limitations when applied to trajectory prediction for forestry wildlife, particularly elusive species such as forest elephants and big cats. First, in dense canopy-covered forest environments, GPS tracking devices often suffer from signal obstruction, resulting in significant data loss and making it difficult to acquire high-temporal-resolution continuous trajectory data. This hinders the effective application of Bayesian inference methods that rely on dense time series data. Second, existing experience-based selection models based on environmental characteristics typically assume that all individuals have the same level of awareness and preference for their environment, neglecting the unique movement experiences and path memories formed by individual wildlife due to differences in age, experience, and social status. This homogenization assumption leads to highly similar predicted paths, failing to reflect the biological learning of different individuals choosing vastly different paths when faced with the same starting and ending points in actual observations. Particularly in research scenarios where only sparse data such as feces and footprints are available, existing methods cannot reconstruct continuous trajectories from sparse data, nor can they extract information on inter-individual experience differences, resulting in significant deviations between predicted results and actual movement patterns. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a method, system, device and medium for predicting the trajectory of forestry wildlife in sparse scenarios that can effectively model differences in individual movement experience under sparse data conditions, so as to improve the biological rationality and practical application value of the prediction results.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for predicting the trajectory of forestry wildlife in sparse scenes, comprising the following steps:

[0007] S1: Divide the forestry area under study into hexagonal grids, dividing the forestry area into a mosaic grid composed of multiple hexagonal nodes, and extract the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features;

[0008] S2: Pair start and end points of historical appearance data of wild animals in the forestry area. Select observation point pairs with clear start and end relationships of movement direction from the historical appearance data. Map the start and end points of each observation point pair to the nearest hexagonal node to generate a sparse movement task set composed of the start hexagonal node and the end hexagonal node.

[0009] S3: Based on the environmental state space, resource selection probability estimation is performed. The random forest model is called to learn and fit the nonlinear relationship between multidimensional environmental characteristics and the historical occurrence points of wild animals, and an environmental resource suitability distribution map covering the entire study area is generated.

[0010] S4: Based on the sparse mobile task set and the environmental resource suitability distribution map, individual path preference learning is performed. A trainable individual memory vector is constructed for each virtual individual, and the individual memory vector is iteratively trained based on the reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors.

[0011] S5: Obtain the current start and end positions of the target wild animal, map the current start and end positions to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within the preset ranking as alternative prediction paths.

[0012] Secondly, this invention provides a sparse scene forestry wildlife trajectory prediction system, which is configured with the following modules:

[0013] The hexagonal mesh construction module is used to divide the forestry area under study into hexagonal meshes, dividing the forestry area into a mosaic mesh composed of multiple hexagonal nodes, and extracting the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features.

[0014] The mobile task pairing and mapping module is used to pair start and end points of historical appearance data of wild animals in the forestry area. It selects observation point pairs with clear start and end relationships in the historical appearance data, maps the start and end points of each observation point pair to the nearest hexagonal node, and generates a sparse mobile task set composed of start hexagonal nodes and end hexagonal nodes.

[0015] The resource suitability assessment module is used to estimate the probability of resource selection based on the environmental state space. It calls the random forest model to learn and fit the nonlinear relationship between multidimensional environmental characteristics and the historical occurrence points of wild animals, and generates an environmental resource suitability distribution map covering the entire study area.

[0016] The individual path preference learning module is used to learn individual path preferences based on sparse mobile task sets and environmental resource suitability distribution maps. It constructs a trainable individual memory vector for each virtual individual and iteratively trains the individual memory vector based on a reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors.

[0017] The attention path search module is used to obtain the current start and end positions of the target wild animal, map the current start and end positions to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and the individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within the preset ranking as alternative prediction paths.

[0018] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned methods for predicting the trajectory of forestry wildlife in sparse scenes.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned methods for predicting the trajectory of forestry wildlife in sparse scenes.

[0020] In summary, the sparse scene forestry wildlife trajectory prediction method provided in this application constructs an environmental state space representing the heterogeneity of geographic space by dividing the research forestry area into hexagonal grids and extracting multidimensional environmental features of each node; it pairs the start and end points of the historical appearance data of wild animals and maps them to hexagonal nodes to generate a set of mobile tasks reflecting sparse observation information; it learns and fits the nonlinear relationship between environmental features and appearance points by calling a random forest model to generate an environmental resource suitability distribution map to quantify habitat resource distribution; it constructs trainable individual memory vectors for virtual individuals and iteratively trains them based on a reinforcement learning framework combined with the sparse mobile task set and resource suitability distribution map to generate an individual experience memory bank that can represent different path preferences within the group; it obtains the current start and end point positions of the target animal and maps them to the environmental state space; it performs weighted path search based on the resource suitability distribution map and individual experience memory bank; it assigns weights to individual memory vectors through an attention mechanism to integrate group experience; it uses the node sequence with the highest integration score as the main prediction path and uses the individual paths generated by virtual individuals with high attention weight ranking as alternative prediction paths. This invention can effectively model individual movement experience differences under the condition of only sparse occurrence point data, in order to solve the problem that existing methods rely on dense trajectory data and ignore individual preferences.

[0021] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a method for predicting the trajectory of forestry wildlife in sparse scenes, provided in an embodiment of this application;

[0023] Figure 2 This is a schematic diagram of the structure of a sparse scene forestry wildlife trajectory prediction system provided in another embodiment of this application. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] In one embodiment, such as Figure 1 As shown, a method for predicting the trajectory of forestry wildlife in sparse scenes is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and can also be applied to a system including both a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0027] S1: Divide the forestry area under study into hexagonal grids, dividing the forestry area into a mosaic grid composed of multiple hexagonal nodes, and extract the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features.

[0028] Specifically, the system divides the research forestry area into hexagonal grids, segmenting the area into a mosaic grid composed of multiple hexagonal nodes. During grid division, the system determines the granularity of the grid units based on the scale of the research area, the stride distance of the target forestry wildlife, and the data resolution. The system transforms the continuous research forestry area into a discrete set of hexagonal nodes through spatial discretization, with each hexagonal node possessing a unique spatial identifier. The system extracts multidimensional environmental features corresponding to each hexagonal node. It acquires vegetation-related features through remote sensing interpretation technology, topographic and hydrological-related features through geographic information system spatial analysis technology, and anthropogenic disturbance-related features through forestry field survey data. These multidimensional environmental features include topographic features, vegetation features, hydrological features, and anthropogenic disturbance features. Topographic features cover altitude, slope, aspect, and topographic relief; vegetation features cover canopy closure, vegetation community type, and understory vegetation coverage; hydrological features cover distance from surface water sources and water distribution density; and anthropogenic disturbance features cover distance from forest roads, frequency of human activities, and the radiation range of settlements. The system normalizes and standardizes the extracted multidimensional environmental features to eliminate dimensional and numerical differences between different environmental features, thus preventing these differences from interfering with subsequent calculations. The system uses each hexagonal node as a basic spatial unit and the processed multidimensional environmental features as attribute vectors to integrate all hexagonal nodes and their corresponding multidimensional environmental features to construct an environmental state space.

[0029] S2: Pair start and end points of historical appearance data of wild animals in the forestry area. Select observation point pairs with clear start and end relationships of movement direction from the historical appearance data. Map the start and end points of each observation point pair to the nearest hexagonal node to generate a sparse movement task set composed of the start hexagonal node and the end hexagonal node.

[0030] Specifically, the system acquires historical occurrence data of wild animals in the forestry area under study. This data includes discrete data such as feces, footprints, and infrared camera trigger points. The system performs start-end point pairing processing on the historical occurrence data. Based on the spatiotemporal sequence, spatial movement distance constraints, and habitat connectivity constraints of the historical occurrence data, the system conducts correlation analysis on all historical occurrence points, selecting observation point pairs with clear movement directions and start-end relationships. The system removes isolated and unrelated abnormal occurrence points from the historical occurrence data, ensuring that each selected observation point pair has a true movement start-end directionality, consistent with the actual movement patterns of wild animals. The system employs a spatial nearest neighbor matching algorithm, mapping the start and end points of each observation point pair to the nearest hexagonal node in the environmental state space. The system converts the continuous latitude and longitude coordinates of the historical occurrence points into unique spatial identifiers for the hexagonal nodes through coordinate transformation, completing the standardized transformation from continuous spatial locations to discrete grid nodes and eliminating the interference of continuous coordinates on the subsequent modeling process. The system integrates all the mapped start and end hexagonal nodes to generate a sparse movement task set. The sparse movement task set uses discretized start and end node pairs to represent the historical movement needs of wild animals under sparse observation conditions.

[0031] S3: Based on the environmental state space, resource selection probability estimation is performed. The random forest model is called to learn and fit the nonlinear relationship between multidimensional environmental characteristics and the historical occurrence points of wild animals, and an environmental resource suitability distribution map covering the entire study area is generated.

[0032] Specifically, the system uses a constructed environmental state space as its foundation, combined with grid node data mapped from historical occurrence points, to estimate the probability of environmental resource selection. The system invokes a random forest model, using the multidimensional environmental features of each hexagonal node in the environmental state space as model input features, and the distribution of grid nodes mapped from historical occurrence points as model fitting labels. The system learns the nonlinear relationship between multidimensional environmental features and historical occurrence points of wild animals through the random forest model, without requiring pre-set linear assumptions, thus adapting to the complex feature association patterns in forestry environments. The system optimizes the hyperparameters of the random forest model using cross-validation methods. Hyperparameters include the number of decision trees and the feature sampling ratio. Multiple rounds of cross-validation are used to adjust the hyperparameter values, ensuring that the model can accurately fit the association between multidimensional environmental features and wild animal occurrence behavior.

[0033] Furthermore, the system utilizes the trained random forest model to predict resource selection probabilities for all hexagonal nodes in the environmental state space, calculating the environmental resource suitability value corresponding to each hexagonal node. This value characterizes the suitability of the node for wildlife. The system then performs spatial visualization processing on the environmental resource suitability values ​​of all hexagonal nodes, generating an environmental resource suitability distribution map covering the entire forestry area under study. This map clearly presents the suitability distribution of each spatial unit within the study area.

[0034] S4: Based on the sparse mobile task set and the environmental resource suitability distribution map, individual path preference learning is performed. A trainable individual memory vector is constructed for each virtual individual, and the individual memory vector is iteratively trained based on the reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors.

[0035] Specifically, the system learns individual path preferences based on a sparse mobile task set and an environmental resource suitability distribution map to address the path prediction convergence problem caused by the homogeneity assumption in traditional methods. The system sets a corresponding number of virtual individuals based on the population structure differences of the target wild animals, with each virtual individual corresponding to a different type of real individual in the population. The system initializes a set of high-dimensional trainable individual memory vectors for each virtual individual. These individual memory vectors encode the virtual individual's path selection preferences, environmental avoidance habits, and historical movement experience. Initially, the system uses a normal distribution method to randomly initialize the individual memory vectors. The system uses the start-end node pairs in the sparse mobile task set as training tasks to construct a comprehensive reward function. The reward function is constructed based on environmental resource suitability, path connectivity, and path movement cost. The system uses the reward function to quantify the quality of virtual individuals in the path selection process.

[0036] Within a reinforcement learning framework, the system uses a sparse movement task set as its training basis. Training is conducted through iterative interactions between virtual individuals and the environment's state space. During the iteration process, the system continuously updates and optimizes the individual memory vectors, enabling virtual individuals to gradually develop differentiated path selection logic. The system continuously monitors the training process of the virtual individuals. Once all virtual individuals have completed iterative training and reached convergence, the system integrates and stores all the differentiated individual memory vectors to generate an individual experience memory bank. This individual experience memory bank is used to represent the differences in individual experience within a wildlife population.

[0037] S5: Obtain the current start and end positions of the target wild animal, map the current start and end positions to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within the preset ranking as alternative prediction paths.

[0038] Specifically, in the trajectory prediction stage, the system acquires the current starting and ending positions of the target wild animal, with the position data being continuous spatial coordinates. The system employs a spatial nearest neighbor matching algorithm to map the continuous spatial coordinates of the current starting and ending positions to corresponding hexagonal nodes in the environmental state space, determining the discretized start and end nodes of the prediction task, completing the conversion from continuous coordinates to discrete nodes, and ensuring that the prediction process can adapt to the discretized modeling mode of the environmental state space. Based on the suitability of each node provided by the environmental resource suitability distribution map, and combined with the differentiated individual memory vectors in the individual experience memory bank, the system introduces an attention mechanism to assign adaptive attention weights to the memory vectors of each virtual individual in the individual experience memory bank. The weight value is determined based on the matching degree between the corresponding virtual individual's experience and the target wild animal; the higher the matching degree, the larger the weight value. The system uses an improved path search algorithm to conduct a weighted path search in the environmental state space composed of a hexagonal grid. During the search process, the system integrates the environmental resource suitability score and the individual experience weighted score, comprehensively scoring each possible hexagonal node sequence, and ranking all node sequences based on their comprehensive scores. The system uses the hexagonal node sequence with the highest comprehensive score as the primary prediction path. This primary prediction path integrates environmental suitability and optimal individual experience, possessing the highest prediction reliability. The system sets a preset ranking threshold, and paths independently generated by virtual individuals whose attention weight ranking falls within this threshold are used as alternative prediction paths. These alternative paths cover potential movement trajectories under different individual experiences, enriching the diversity of prediction results. The system outputs both the primary and alternative prediction paths, enhancing the biological rationality and practical application value of forestry wildlife trajectory prediction in sparse scenes, and supporting the development of forestry wildlife protection and management.

[0039] In summary, the sparse scene forestry wildlife trajectory prediction method provided in this application constructs an environmental state space representing the heterogeneity of geographic space by dividing the research forestry area into hexagonal grids and extracting multidimensional environmental features of each node; it pairs the start and end points of the historical appearance data of wild animals and maps them to hexagonal nodes to generate a set of mobile tasks reflecting sparse observation information; it learns and fits the nonlinear relationship between environmental features and appearance points by calling a random forest model to generate an environmental resource suitability distribution map to quantify habitat resource distribution; it constructs trainable individual memory vectors for virtual individuals and iteratively trains them based on a reinforcement learning framework combined with the sparse mobile task set and resource suitability distribution map to generate an individual experience memory bank that can represent different path preferences within the group; it obtains the current start and end point positions of the target animal and maps them to the environmental state space; it performs weighted path search based on the resource suitability distribution map and individual experience memory bank; it assigns weights to individual memory vectors through an attention mechanism to integrate group experience; it uses the node sequence with the highest integration score as the main prediction path and uses the individual paths generated by virtual individuals with high attention weight ranking as alternative prediction paths. This invention can effectively model individual movement experience differences under the condition of only sparse occurrence point data, in order to solve the problem that existing methods rely on dense trajectory data and ignore individual preferences.

[0040] In one embodiment, S1 of the sparse scene forestry wildlife trajectory prediction method provided by the present invention specifically includes the following steps:

[0041] S11: Analyze the geographic spatial boundary of the forestry area under study, and perform mosaic subdivision of the forestry area under study based on the preset hexagonal unit size parameters to generate a hexagonal grid layer covering the entire study area.

[0042] Specifically, the system analyzes the geospatial boundary of the forestry area under study. It reads the corresponding vector spatial data, which includes the boundary contour coordinates and spatial topological relationships of the area. The system performs a topology check on the read vector spatial data to eliminate redundant nodes and topological errors, ensuring the integrity and continuity of the boundary contour and preventing boundary breaks or overlaps from interfering with subsequent segmentation operations. The system retrieves preset hexagonal cell size parameters from a pre-defined parameter configuration library to determine the geometric size of the hexagonal cells. These parameters are tailored to the scale of the study area and the spatial resolution requirements of subsequent trajectory prediction. The system employs a regular mosaic segmentation algorithm, performing full-area spatial segmentation of the topology-checked forestry area based on the retrieved hexagonal cell size parameters. During segmentation, the system performs the segmentation operation progressively from the boundary of the study area inwards, ensuring that the generated hexagonal cells have no overlapping areas or gaps, achieving complete coverage of the forestry area. After completing the segmentation operation, the system generates a hexagonal mesh layer containing all the generated hexagonal cells, each with an independent spatial identifier. The system establishes a spatial index for the hexagonal grid layer, and records the position information and corresponding identification of each hexagonal unit through the spatial index, which facilitates the rapid location and retrieval of the hexagonal units in the future.

[0043] S12: Perform geographic coordinate positioning processing on each hexagonal node in the hexagonal grid layer, extract the elevation value, slope value, Euclidean distance value to the nearest water source, normalized vegetation index value, Euclidean distance value to the nearest road, and canopy closure value of the center point of each hexagonal node, and generate the initial environmental feature set of each hexagonal node.

[0044] Specifically, the system performs geographic coordinate positioning processing on each hexagonal node in the hexagonal grid layer. First, the system identifies the geometric center of each hexagonal unit, determining the geometric center point as the corresponding node, ensuring that each hexagonal node accurately represents the spatial location of its corresponding hexagonal unit. The system employs a geospatial coordinate transformation algorithm to convert the geometric center coordinates of the hexagonal nodes into unified planar geographic coordinates, achieving coordinate system standardization and eliminating deviations caused by different coordinate systems. The system retrieves topographic elevation data of the study area and obtains the elevation value corresponding to the center point of each hexagonal node through spatial location matching. The elevation value directly reflects the terrain elevation of the node's location. Based on the topographic elevation data, the system calculates the slope value of the center point of each hexagonal node using a slope calculation algorithm. The slope value reflects the terrain inclination at the node's location. The system extracts water source distribution vector data of the study area and calculates the straight-line distance from the center point of each hexagonal node to the nearest water source using an Euclidean distance calculation algorithm, obtaining the corresponding Euclidean distance value.

[0045] Furthermore, the system retrieves remote sensing image data of the study area, performs radiometric and geometric corrections on the image data, and then uses the Normalized Difference Vegetation Index (NDVI) calculation logic to obtain the NDVI value corresponding to the center point of each hexagonal node. This value reflects the vegetation growth status of the area where the node is located. The system extracts road distribution vector data of the study area and uses the Euclidean distance calculation algorithm to calculate the straight-line distance from the center point of each hexagonal node to the nearest road, obtaining the corresponding Euclidean distance value. The system retrieves forestry resource survey data of the study area and obtains the canopy closure value corresponding to the center point of each hexagonal node through spatial location matching. The system integrates the obtained elevation values, slope values, Euclidean distance values ​​to the nearest water source, NDVI values, Euclidean distance values ​​to the nearest road, and canopy closure values ​​to generate an initial environmental feature set for each hexagonal node. The initial environmental feature set is stored in the form of a numerical sequence, completely recording the various environmental attribute information corresponding to each hexagonal node.

[0046] S13: Normalize and scale the initial set of environmental features to construct an environmental state space consisting of all hexagonal nodes and their corresponding normalized multidimensional environmental features.

[0047] Specifically, the system reads the initial environmental feature set corresponding to all hexagonal nodes and performs normalization scaling on the values ​​of each type of environmental feature in the initial environmental feature set. The system first iterates through and statistically analyzes all values ​​for each type of environmental feature to determine the value range of that type of environmental feature and to identify the maximum and minimum values. The system then uses a normalization scaling algorithm to transform all values ​​corresponding to each type of environmental feature into a unified numerical range based on the value range of each type of environmental feature, maintaining the relative magnitude of the values ​​of each type of environmental feature during the transformation process. The core purpose of normalization scaling is to eliminate the dimensional differences between different environmental features, avoiding feature weight imbalances in subsequent model calculations due to different value ranges of various features, and ensuring that all environmental features participate in subsequent calculations with equal weight.

[0048] Furthermore, during the normalization and scaling process, the system detects and processes outlier values. It identifies outliers exceeding reasonable ranges through numerical distribution analysis and corrects them using interpolation methods, ensuring the accuracy and reasonableness of the normalized data. After completing the normalization and scaling of all environmental feature values, the system associates and binds each hexagonal node with its corresponding normalized multidimensional environmental feature, establishing a correspondence between node identifiers and normalized environmental features. This ensures that each hexagonal node accurately matches all its corresponding normalized environmental features. The system iterates through all hexagonal nodes, completing the association operation between each node and its corresponding normalized multidimensional environmental feature, integrating all associated data to construct an environmental state space. This environmental state space uses hexagonal nodes as basic spatial units, with normalized multidimensional environmental features as the core attributes of each unit, forming a standardized, discretized spatial modeling carrier.

[0049] In one embodiment, S2 of the sparse scene forestry wildlife trajectory prediction method provided by the present invention specifically includes the following steps:

[0050] S21: For each observation point in the historical appearance data of wild animals in the forestry area, perform timestamp and spatial location analysis, and filter continuous observation point pairs with clear movement direction indications based on field record information to generate a set of candidate start and end point pairs.

[0051] Specifically, the system acquires historical sighting data of wild animals in the forestry area under study. This data includes information related to all wildlife observation points recorded in field surveys. The system performs timestamp parsing and spatial location parsing operations on each observation point in the historical sighting data. During timestamp parsing, the system reads the time record information corresponding to each observation point, converts the time record into a standardized timestamp format, and clarifies the specific observation time of each observation point, ensuring that the time information of all observation points is under a unified standard, facilitating subsequent time series analysis. During spatial location parsing, the system reads the original spatial location record of each observation point, parses it to obtain the geographic coordinate information of each observation point, and simultaneously checks the validity of the parsed spatial location coordinates, discarding observation points whose coordinates are outside the scope of the forestry area under study or whose coordinates contain obvious errors, ensuring that the spatial location information of each observation point is accurate and within the study area.

[0052] Furthermore, the system retrieves field records, which include the observation sequence of observation points, individual wildlife identifiers, and field tracking notes. Based on these field records, the system filters the analyzed observation points, selecting consecutive observation point pairs with clear directional movement indications. During the filtering process, the system combines the timestamp sequence of the observation points with the observation logic of the field records to determine the order of consecutive observation point pairs, ensuring that each pair reflects the movement trend of the wildlife and has a clear directional movement indication. The system integrates all the filtered consecutive observation point pairs to generate a candidate start-end point pair set. Each pair in the candidate start-end point pair set is assigned a unique identifier, and the system records the original observation point information and field record basis for each pair.

[0053] S22: Perform spatial coordinate transformation on each observation point in the candidate start and end point pair set, calculate the Euclidean distance between each observation point and all hexagonal nodes in the hexagonal grid, and match each observation point to the nearest hexagonal node to generate a mapping point pair consisting of the start hexagonal node and the end hexagonal node.

[0054] Specifically, the system retrieves preset coordinate transformation parameters from the system parameter configuration library. These parameters are used to convert the original geographic coordinates of the observation points into planar geographic coordinates consistent with the hexagonal grid layer, eliminating deviations caused by different coordinate systems and ensuring that the coordinates of the observation points and the coordinates of the hexagonal nodes are within the same spatial reference frame. After completing the spatial coordinate transformation for each observation point, the system reads the planar geographic coordinates of all hexagonal nodes in the hexagonal grid layer. For each observation point, the system calculates the Euclidean distance between the observation point and all hexagonal nodes, strictly adhering to the Euclidean distance calculation logic to ensure the accuracy of the distance calculation results. The system sorts all Euclidean distances corresponding to each observation point, selects the hexagonal node corresponding to the smallest Euclidean distance, and identifies this hexagonal node as the matching node corresponding to the observation point, completing the matching operation between a single observation point and a hexagonal node.

[0055] Following the above process, the system sequentially completes the matching operation for each observation point in the candidate start-end point pair set. It matches the start observation point in each candidate start-end point pair to the corresponding start hexagonal node and the end observation point to the corresponding end hexagonal node, forming a mapping point pair for each candidate start-end point pair. Each mapping point pair contains the identifiers of the start and end hexagonal nodes, as well as relevant association information of the corresponding original observation points. The system integrates all generated mapping point pairs, establishing a correspondence between the mapping point pairs and the original candidate start-end point pairs. This ensures the traceability of subsequent filtering operations and provides a data foundation for deduplication and filtering of mapping point pairs.

[0056] S23: Perform deduplication and filtering on the mapped point pairs, and remove invalid point pairs with the same starting point and ending point or a distance less than the preset minimum effective migration distance threshold to generate a sparse movement task set.

[0057] Specifically, the system extracts the start-point and end-point hexagonal node identifiers for each mapping point pair. Using the combination of these identifiers as the deduplication criterion, the system iterates through all mapping point pairs, eliminating duplicate pairs with identical start-point and end-point identifier combinations, retaining only one pair with the same combination. This ensures the uniqueness of the mapping point pairs and prevents duplicate data from interfering with subsequent calculations. After deduplication, the system filters the deduplicated mapping point pairs. During this filtering process, the system first determines whether the start-point and end-point hexagonal nodes of each mapping point pair are the same node. If they are the same node, the mapping point pair is deemed invalid and discarded. For mapping point pairs with different start and end points, the system calculates the Euclidean distance between the center points of the start and end hexagonal nodes and retrieves a preset minimum effective migration distance threshold. This threshold, obtained from the system parameter configuration library, is used to determine whether the movement corresponding to the mapping point pair is a valid movement. The system compares the calculated Euclidean distance with a preset minimum effective migration distance threshold. If the distance is less than the threshold, the mapped point pair is deemed invalid and discarded; if the distance is not less than the threshold, the mapped point pair is deemed valid and retained. After completing the deduplication and filtering operations for all mapped point pairs, the system integrates all retained valid mapped point pairs to generate a sparse movement task set.

[0058] In one embodiment, S3 of the sparse scene forestry wildlife trajectory prediction method provided by the present invention specifically includes the following steps:

[0059] S31: Count the frequency of wild animal occurrences for each hexagonal node in the environmental state space, label the positive and negative samples of each node based on historical occurrence data, and generate a training sample set for the random forest model.

[0060] Specifically, the system reads all hexagonal nodes in the environmental state space and retrieves the matched historical wildlife occurrence mapping data, which includes the hexagonal node identifiers corresponding to all historical occurrence points. The system establishes the association between environmental state space nodes and historical occurrence mapping data through the node identifiers, and performs wildlife occurrence frequency statistics for each hexagonal node. During the statistics process, the system can use frequency calculation logic, with the following formula:

[0061]

[0062] in, The frequency of wild animals appearing at the i-th hexagonal node. Let represent the hexagonal node identifier after matching the k-th historical occurrence point, i represent the currently counted hexagonal node identifier, M represent the total number of historical occurrence points, and I be the indicator function. When i matches, I takes the value 1; otherwise, it takes the value 0. The system calculates the frequency of occurrence for each node based on this formula, verifies the validity of the calculation results, and removes abnormal frequencies caused by data association errors. The system labels each hexagonal node with positive and negative samples based on historical occurrence data; nodes with a frequency greater than zero are labeled with positive samples, and nodes with a frequency of zero are labeled with negative samples. The system associates and binds the normalized multidimensional environmental features of each node with the corresponding sample label to form a single training sample. It iterates through all nodes to complete the construction of the training samples and integrates them to generate the training sample set for the random forest model. The system cleans the training sample set, removing invalid samples with missing features or incorrect labels, and stores the cleaned training sample set in the model training database.

[0063] S32: Using the multidimensional environmental features of each sample in the training sample set as input and the label of whether wild animals appear in the sample as output, the random forest model is trained through multiple rounds of decision tree growth to construct a resource selection probability model that represents the nonlinear relationship between environmental features and the appearance of wild animals.

[0064] Specifically, the system retrieves the cleaned training sample set, defines the input and output parameters for model training, with the input parameters being the normalized multidimensional environmental features of each training sample and the output parameters being the positive and negative sample labels of each training sample. The system initializes the random forest model, retrieves the preset initial model parameters, and completes the basic structure construction of the model. Based on the training sample set, the system performs multiple rounds of decision tree growth training on the random forest model. In each round of training, samples are randomly extracted from the training sample set to form a training subset, and some environmental features are randomly selected as decision tree splitting features. Based on the correlation between sample labels and feature data, a single decision tree is grown, and this process is repeated to generate multiple independent decision trees.

[0065] During training, the system monitors model performance through cross-validation, adjusts training parameters to avoid overfitting or underfitting, and stops training and saves model parameters once the model's prediction performance reaches a stable state, thus completing the resource selection probability model. The core calculation logic formula of this model is:

[0066]

[0067] in, X represents the probability of wild animals appearing corresponding to the input environmental feature X, and T represents the total number of decision trees in the random forest. Let represent the prediction result of the t-th decision tree for input X, where is the prediction of a positive sample. The value is 1 when the prediction is a negative sample and 0 when the prediction is a negative sample. This formula calculates the final resource selection probability through the voting results of multiple decision trees, which represents the nonlinear relationship between environmental characteristics and the occurrence of wild animals.

[0068] S33: Input the multidimensional environmental features of all hexagonal nodes in the environmental state space into the trained resource selection probability model, perform forward propagation prediction processing on each node, and generate the resource selection probability value corresponding to each hexagonal node.

[0069] Specifically, the system retrieves the trained resource selection probability model, reads all hexagonal nodes in the environment state space and their corresponding normalized multidimensional environment features, ensuring that the data format input to the model is consistent with that used during training. The system preprocesses the input environment feature data, checking its completeness and validity, supplementing missing features using interpolation methods, and correcting abnormal data to ensure the input data meets the model's prediction requirements. The system sequentially inputs the preprocessed multidimensional environment features of each hexagonal node into the resource selection probability model for forward propagation prediction. During forward propagation, the model performs layer-by-layer calculations on the input environment feature data based on preset parameters and decision logic, completing the resource selection probability value calculation in the same manner as during model training. The system records the resource selection probability value corresponding to each hexagonal node, associating and binding the probability value with the node identifier to ensure that each node accurately matches its prediction result. After completing the forward propagation prediction for all nodes, the system integrates the resource selection probability values ​​of all nodes, generating an association table containing node identifiers, environment features, and probability values.

[0070] S34: Perform raster interpolation on the resource selection probability value of each hexagonal node, convert the discrete node probability values ​​into a continuously covered raster layer, and generate an environmental resource suitability distribution map covering the entire forestry area under study.

[0071] Specifically, the system reads the resource selection probability values ​​and corresponding spatial location information of all hexagonal nodes, sets the basic parameters for raster interpolation, and determines the raster size based on the node distribution density and the spatial scale of the study area to ensure that the interpolated raster layer completely covers the study area and meets the spatial resolution requirements of trajectory prediction. The system uses a spatial interpolation algorithm to perform raster interpolation processing on discrete node probability values. The core interpolation formula is:

[0072]

[0073] in, Represents grid cell The resource selection probability value, where n represents the number of adjacent hexagonal nodes participating in the interpolation calculation. This represents the resource selection probability value of the i-th adjacent hexagonal node. Represents the i-th adjacent node pair of grid cells The interpolation weights are determined based on the spatial distance between nodes and raster cells. Using this formula, and based on the spatial location of hexagonal nodes, the system calculates the probability value of each raster cell through interpolation logic, converting discrete node probability values ​​into continuously distributed raster probability values. During interpolation, the system verifies the validity of the interpolation results, removing outliers to ensure the results accurately reflect the regional probability distribution. After interpolation, the system generates a raster layer, which is then spatially corrected to ensure its spatial location accurately matches the geographical boundaries of the study area, eliminating spatial offset errors. The system integrates the corrected raster layer to generate an environmental resource suitability distribution map covering the entire forestry area, storing this map in a spatial database to provide environmental suitability data for subsequent individual path preference learning and path search operations.

[0074] In one embodiment, S4 of the sparse scene forestry wildlife trajectory prediction method provided by the present invention specifically includes the following steps:

[0075] S41: Based on the preset number of virtual individuals, a high-dimensional real number vector is randomly initialized for each virtual individual as the initial value of the individual memory vector, and the memory vectors of all virtual individuals are combined and stored to generate an initial experience memory bank.

[0076] Specifically, the system retrieves a preset parameter for the number of virtual individuals, obtained from the system parameter configuration library, to determine the total number of virtual individuals participating in individual path preference learning. Based on this parameter, the system generates a corresponding number of virtual individuals, each corresponding to a different type of real wild animal individual in the population. The system randomly initializes a high-dimensional real-valued vector for each virtual individual, using this vector as the initial value for the individual's memory vector. The dimension of the individual's memory vector is consistent with the dimension of the multi-dimensional environmental features of the hexagonal nodes in the environmental state space, ensuring subsequent coupled computation with environmental features. The initialization formula for the virtual individual's memory vector is as follows: .in This represents the initial memory vector of the k-th virtual individual. The expression represents a normal distribution, where μ represents the mean of the normal distribution. The variance, representing the normal distribution, is used by the system to randomly initialize the initial memory vectors of all virtual individuals. The system then integrates the initial memory vectors of all virtual individuals, establishing a correspondence between virtual individual identifiers and initial memory vectors, thus generating an initial experience memory bank.

[0077] S42: Perform path generation processing for each task in the sparse movement task set, couple the memory vector of each virtual individual with the environmental resource suitability distribution map, and use a reinforcement learning framework to allow each virtual individual to explore the movement path from the task starting point to the end point, and record the cumulative reward value of each exploration path.

[0078] Specifically, the system reads a sparse set of movement tasks, retrieves all of them, and performs path generation for each task. Simultaneously, the system retrieves the memory vectors of all virtual individuals from the initial experience memory bank, as well as the previously generated environmental resource suitability distribution map. It couples each virtual individual's memory vector with the environmental resource suitability distribution map, using vector dot product operations to associate the path preferences encoded by the individual's memory vector with the environmental suitability of each hexagonal node, thus determining the virtual individual's preference weights for different nodes. The system constructs a reinforcement learning framework, using the starting hexagonal node of each movement task as the initial exploration position of the virtual individual and the ending hexagonal node as the exploration target position. Each virtual individual starts from the starting point and, based on the coupled preference weights and environmental suitability information, gradually explores the movement path until reaching the destination.

[0079] Furthermore, the system records all hexagonal nodes traversed by each virtual individual during exploration, as well as the environmental suitability score corresponding to each node. The cumulative reward value for each exploration path is calculated by accumulating these values. The formula for calculating the cumulative reward value is as follows:

[0080]

[0081] in, This represents the cumulative reward value for the exploration path of the k-th virtual individual. This represents the set of hexagonal nodes traversed by the exploration path of the k-th virtual individual. This represents the environmental resource suitability score of the i-th hexagonal node. The system uses this formula to calculate and record the cumulative reward value for all exploration paths.

[0082] S43: Sort and normalize the cumulative reward values ​​of each virtual individual under each task, calculate the attention weight of each virtual individual when completing the task, and generate the individual experience contribution distribution corresponding to each task.

[0083] Specifically, the system reads the cumulative reward value of all virtual individuals under each movement task, sorts the cumulative reward values ​​under the same movement task in descending order of value, and records the position of each virtual individual in the sort. The system then performs normalization processing on the sorted cumulative reward values ​​to eliminate differences in the numerical range of cumulative reward values ​​between different movement tasks, ensuring the rationality of subsequent attention weight calculations. Preferably, the normalization formula is:

[0084]

[0085] in, This represents the normalized cumulative reward value for the k-th virtual individual completing the m-th movement task. This represents the original cumulative reward value for the k-th virtual individual completing the m-th movement task. Let m be the minimum cumulative reward value for all virtual individuals completing the m-th movement task. This represents the maximum cumulative reward value for all virtual individuals completing the m-th movement task. Based on the normalized cumulative reward value, the system calculates the attention weight for each virtual individual when completing the movement task. The weight is positively correlated with the normalized cumulative reward value; the higher the cumulative reward value, the greater the attention weight. The system integrates the attention weights of all virtual individuals under each movement task to generate an individual experience contribution distribution for each movement task. This distribution records the degree to which each virtual individual's experience contributes to the path selection for that movement task.

[0086] S44: Based on attention weights, the exploration paths of each virtual individual are weighted and fused to generate an integrated path as the optimal path template for the current iteration round.

[0087] Specifically, the system retrieves the exploration paths of all virtual individuals under each movement task, along with their corresponding attention weights, and performs a weighted fusion operation on these exploration paths based on this data. The core logic of the weighted fusion is to assign weights to the exploration paths of different virtual individuals according to their attention weights; the larger the attention weight, the higher the proportion of the corresponding path in the fusion process. The formula for path weighted fusion is:

[0088]

[0089] in, This represents the integration path for the m-th mobile task. This represents the attention weight for the k-th virtual individual completing the m-th movement task. This represents the exploration path for the k-th virtual individual to complete the m-th movement task, where K represents the total number of virtual individuals. The system uses this formula to perform a weighted fusion of all movement task exploration paths. The system determines the fused integrated path as the optimal path template for the current iteration. This template integrates the high-quality path experience of all virtual individuals and accurately reflects the optimal path selection logic for this movement task. The system records the optimal path template corresponding to each movement task, associates and binds it with the start and end nodes of the movement task, and stores it in the iterative training database. This provides a template reference for subsequent updates to the virtual individual memory vectors, ensuring that the memory vector updates align with the optimal path preferences.

[0090] S45: Perform gradient calculation on the individual memory vector of each virtual individual to maximize the cumulative reward and combine it with the integrated path template for behavioral cloning constraints. Update the memory vector of each virtual individual through the gradient ascent method to generate an updated experience memory bank.

[0091] Specifically, the system performs gradient calculation on the individual memory vector of each virtual individual. The goal of gradient calculation is to maximize the cumulative reward value of the virtual individual for completing all movement tasks. The update direction and magnitude of the memory vector are determined through gradient calculation. The system retrieves the optimal path template for the current iteration and uses it as a behavior cloning constraint to constrain the update direction of the virtual individual's memory vector, ensuring that the updated memory vector can learn the path preference of the optimal path template and avoiding deviations in memory vector updates. The system uses gradient ascent, combined with a preset learning rate and regularization coefficient, to iteratively update the memory vector of each virtual individual. The update formula for the memory vector of each virtual individual is:

[0092]

[0093] in, Let be the memory vector of the k-th virtual individual after the (t+1)-th iteration. Let be the current memory vector of the k-th virtual individual in the t-th iteration. The learning rate hyperparameter for the memory vector. The total number of tasks in the sparse mobile task set. The cumulative reward value obtained for the k-th virtual individual to complete the m-th movement task. The regularization coefficient hyperparameter for behavior cloning constraints. The path selection strategy distribution for the k-th virtual individual Distribution with integrated path strategy KL divergence between them The regularization coefficient hyperparameter is used to represent the experience collaboration among individuals. This refers to the set of virtual individuals whose memory similarity to the k-th virtual individual is higher than a preset threshold. The empirical collaboration weight coefficient between the k-th virtual individual and the j-th virtual individual is... For two virtual individual memory vectors and Cosine similarity between them.

[0094] S46: Repeat S42 to S45 until the preset iteration rounds are completed or the cumulative reward value converges, generating an individual experience memory bank containing multiple differentiated individual memory vectors.

[0095] Specifically, the system retrieves preset iteration round parameters, monitors the current number of iterations, and tracks the cumulative reward value of all virtual individuals completing all movement tasks, then executes an iteration stop judgment operation. The formula for determining iteration stop is:

[0096]

[0097] in, This represents the average cumulative reward value of all virtual individuals in the (t+1)th iteration. This represents the average cumulative reward value of all virtual individuals in the t-th iteration. This represents the preset convergence threshold. When this formula is satisfied, it indicates that the cumulative reward value has reached convergence. The system repeats steps S42 to S45. After each iteration, the experience memory is updated, recording the current iteration number, average cumulative reward value, and changes in memory vectors. If the current iteration number reaches the preset iteration number, or the average cumulative reward value of all virtual individuals meets the convergence condition, the system stops the iterative training operation. The system retrieves the finally updated experience memory, which contains multiple differentiated individual memory vectors. Each memory vector encodes the unique path preferences and experiences of the corresponding virtual individual.

[0098] In one embodiment, S5 of the sparse scene forestry wildlife trajectory prediction method provided by the present invention specifically includes the following steps:

[0099] S51: Obtain the current start and end positions of the target wild animal, perform spatial coordinate transformation on the current start and end positions, and map the current start and end positions to the nearest hexagonal node in the environmental state space based on the principle of minimum Euclidean distance, thereby generating the start and end nodes of the task to be predicted.

[0100] Specifically, the system acquires the current start and end positions of the target wild animal. The location data originates from geographic coordinate information collected by field monitoring equipment or recorded through manual surveys. The system performs spatial coordinate transformation operations on the acquired start and end positions, retrieving preset coordinate transformation parameters. These parameters are consistent with the coordinate system of the environmental state space and the hexagonal grid layer, converting the original geographic coordinates of the start and end points into unified planar geographic coordinates to eliminate deviations caused by different coordinate systems. The system reads the planar geographic coordinates of all hexagonal nodes in the environmental state space, calculates the Euclidean distance between the current start coordinates and the coordinates of all hexagonal nodes, and simultaneously calculates the Euclidean distance between the current end coordinates and the coordinates of all hexagonal nodes. The formula for calculating the Euclidean distance is:

[0101]

[0102] in, represents the Euclidean distance between the current location and the i-th hexagonal node, and x and y represent the planar geographic coordinates of the current location. and This represents the planar geographic coordinates of the center point of the i-th hexagonal node. Based on the principle of minimum Euclidean distance, the system selects the hexagonal node with the smallest distance to the current starting point and designates it as the starting point node for the task to be predicted; similarly, it selects the hexagonal node with the smallest distance to the current ending point and designates it as the ending point node for the task to be predicted. The system records the identification information of the starting and ending nodes, associates and binds them as the discretized start and end nodes of the current task to be predicted, and stores them in the prediction database.

[0103] S52: Perform path cost evaluation for each virtual individual in the individual experience memory bank. With the starting node as the start point and the ending node as the target, generate candidate paths for each virtual individual based on the environmental resource suitability distribution map and calculate the path cost value based on attention weight.

[0104] Specifically, the system uses the starting node as the starting position and the ending node as the target position. Combining this with an environmental resource suitability distribution map, and based on the path preferences encoded by the memory vector of each virtual individual, it generates candidate paths for that virtual individual. Each candidate path consists of a series of consecutive hexagonal nodes, ensuring that the path starts from the starting node and gradually transitions to the ending node, and that adjacent nodes on the path have spatial connectivity. The system analyzes the candidate paths for each virtual individual, determining the total number of steps. Simultaneously, it extracts the resource selection probability value of each hexagonal node at each step on the path, retrieves the virtual individual's access records for all nodes during historical training, and counts the number of visits to each node and the maximum number of visits. Based on a preset path cost calculation formula, the system calculates a weighted path cost value for each virtual individual's candidate paths. The formula for calculating the path cost value is:

[0105]

[0106] in, The weighted path cost value for the k-th virtual individual to complete the current predicted task. The total number of steps in the candidate paths generated for the k-th virtual individual. For discount factor hyperparameters, The hexagonal node at step t of the path obtained from the environmental resource suitability distribution map. The probability value of resource selection. The memory decay coefficient hyperparameter, For the k-th virtual individual to access nodes during the historical training process Total number of times Let $k$ be the maximum number of visits the k-th virtual individual makes to all nodes. Explore the hyperparameters of the excitation coefficient for the new node. For indicator functions, when node The value is 1 if the virtual individual has never been visited during its historical training, and 0 otherwise. After the system calculates the path cost for all virtual individuals, it associates and binds the cost value with the corresponding virtual individual identifier and stores it in the cost assessment database.

[0107] S53: Perform softmax normalization on the path cost value, calculate the attention weight of each virtual individual for completing the current task to be predicted, and generate the individual experience call weight distribution.

[0108] Specifically, the system reads the weighted path cost values ​​corresponding to all virtual individuals and performs softmax normalization on all cost values. The core purpose of normalization is to eliminate the differences in the numerical range of path cost values ​​among different virtual individuals, converting the cost values ​​into attention weights that conform to a probability distribution, thus ensuring the rationality and standardization of weight calculation. Preferably, the softmax normalization calculation formula is as follows:

[0109]

[0110] in, This represents the attention weight of the k-th virtual individual in completing the current task to be predicted. Let K represent the weighted path cost of the k-th virtual individual, where K represents the total number of virtual individuals. The formula is an exponential function, with the denominator being the sum of the exponentialized cost values ​​of all virtual individuals. This formula converts cost values ​​into weight values; the lower the cost value, the greater the corresponding attention weight, indicating a higher degree of matching between the virtual individual's path experience and the current task to be predicted. The system calculates the attention weight for each virtual individual sequentially based on this formula, verifies the validity of the calculated weight values ​​to ensure that the sum of all weight values ​​is 1, and removes abnormal weight values ​​caused by calculation errors. The system integrates the attention weights of all virtual individuals to generate an individual experience call weight distribution. This distribution records the call priority of each virtual individual's experience in the current task to be predicted, and the system stores this weight distribution in a weight database.

[0111] S54: Based on the weight distribution of individual experience, the resource selection probability of each node on the environmental resource suitability distribution map is weighted and fused to generate an integrated path search guidance map that integrates group experience.

[0112] Specifically, the system retrieves the weight distribution of individual experience calls, and simultaneously reads the environmental resource suitability distribution map and the node resource selection probability evaluation results corresponding to each virtual individual. Based on this data, a weighted fusion operation of resource selection probabilities is performed. The core logic of weighted fusion is to sum the evaluated node resource selection probabilities of each virtual individual according to its attention weight, and integrate the group experience to form an integrated resource selection probability that is more suitable for the current task to be predicted. Preferably, the calculation formula for weighted fusion is:

[0113]

[0114] in, The probability value of node s in selecting integrated resources after integrating the group's experience. The attention weight represents the k-th virtual individual. Represents the k-th virtual individual's pair of nodes The system calculates the integrated resource selection probability value for each hexagonal node in the environmental state space, based on the formula, and associates the calculated integrated probability value with the corresponding node identifier. The system then rasterizes the integrated probability values ​​of all nodes, converting discrete node probability values ​​into continuously covered raster data to generate an integrated path search guidance map that integrates group experience. This guidance map presents the integrated resource selection probability of each node within the study area in raster form; the higher the probability value, the higher the path guidance priority of the node.

[0115] S55: Start a greedy search on the integrated path search guidance graph from the starting node, and successively select the adjacent node with the highest guidance value to perform state transitions until the end node is reached. The sequence of nodes passed through is used as the main predicted path output.

[0116] Specifically, the system retrieves the integrated path search guidance graph and reads the start and end nodes of the task to be predicted. A greedy search operation is then performed on the integrated path search guidance graph to generate the main prediction path. The core logic of the greedy search is to successively select the adjacent node with the highest guidance value for state transition, ensuring that each move aligns with the optimal path guidance derived from the fusion of group experience. Preferably, the node selection formula for the greedy search is:

[0117]

[0118] in, This represents the node selected in step t+1. Represents the node at step t. represent The set of all adjacent hexagonal nodes, where argmax represents the set of nodes that make argmax the set of adjacent hexagonal nodes. The largest node s, The guidance value represents the probability value of selecting integrated resources for node s in the integrated path search guidance graph. The system uses the starting node as the initial search position, selects the optimal neighboring node for the first step according to this formula, and records the node identifier. Then, using this node as the current position, it continues to select the next neighboring node with the highest guidance value, and so on, performing state transition operations in a cyclical manner. During the search process, the system monitors in real time whether the current node is the destination node. If the current node is the destination node, the search operation stops; if the destination node has not been reached, node selection and state transition continue. The system records all nodes traversed during the search process, integrates them sequentially to form a node sequence, outputs this node sequence as the main predicted path, and stores the main predicted path in the prediction result database.

[0119] S56: Sort each virtual individual in descending order according to the weight distribution based on individual experience, select the candidate path corresponding to the virtual individual in the preset ranking, and output the selected candidate path as the set of alternative prediction paths.

[0120] Specifically, the system retrieves the individual experience call weight distribution, reads the attention weights and corresponding candidate paths of all virtual individuals, and sorts all virtual individuals in descending order based on their attention weights. During the sorting process, the virtual individual identifiers and corresponding candidate paths are arranged sequentially according to the order of attention weights from largest to smallest, ensuring that virtual individuals with higher weights are ranked higher, representing a higher reliability and matching degree of their path experience. The system retrieves a preset ranking threshold, which is obtained from the system parameter configuration library, to determine the number of candidate prediction paths to be screened. Based on the descending sorting results, the system selects virtual individuals ranked within the preset ranking threshold, extracts the candidate paths corresponding to these virtual individuals, and performs deduplication processing on the extracted candidate paths, eliminating duplicate candidate paths with identical path node sequences to ensure the diversity of the candidate prediction path set. The system integrates the deduplicated candidate paths to generate a candidate prediction path set, which contains multiple different candidate paths, covering potential movement trajectories under different individual experiences. The system verifies the validity of the set of candidate prediction paths, removes invalid paths that cannot reach the end point from the starting node or whose path nodes are discontinuous, stores the verified set of candidate prediction paths in association with the main prediction path, and outputs the set of candidate prediction paths to complete the entire trajectory prediction process, providing trajectory reference for the protection and management of forestry wildlife.

[0121] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0122] Based on the same inventive concept, this application also provides a sparse scene forestry wildlife trajectory prediction system for implementing the sparse scene forestry wildlife trajectory prediction method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more sparse scene forestry wildlife trajectory prediction system embodiments provided below can be found in the limitations of the sparse scene forestry wildlife trajectory prediction method described above, and will not be repeated here.

[0123] Preferably, such as Figure 2As shown, the present invention provides a sparse scene forestry wildlife trajectory prediction system 600, which is configured with the following modules:

[0124] The hexagonal mesh construction module 610 is used to divide the forestry area under study into hexagonal meshes, dividing the forestry area under study into a mosaic mesh composed of multiple hexagonal nodes, and extracting the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features.

[0125] The mobile task pairing and mapping module 620 is used to pair the start and end points of historical appearance point data of wild animals in the forestry area. It selects observation point pairs with clear start and end relationships in the historical appearance point data, maps the start and end points of each observation point pair to the nearest hexagonal node, and generates a sparse mobile task set composed of the start hexagonal node and the end hexagonal node.

[0126] The resource suitability assessment module 630 is used to estimate the probability of resource selection based on the environmental state space. It calls the random forest model to learn and fit the nonlinear relationship between multidimensional environmental characteristics and the historical occurrence points of wild animals, and generates an environmental resource suitability distribution map covering the entire study area.

[0127] The individual path preference learning module 640 is used to learn individual path preferences based on a sparse mobile task set and an environmental resource suitability distribution map. It constructs a trainable individual memory vector for each virtual individual and iteratively trains the individual memory vector based on a reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors.

[0128] The attention path search module 650 is used to obtain the current start and end positions of the target wild animal, map the current start and end positions to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and the individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within the preset ranking as alternative prediction paths.

[0129] Preferably, the hexagonal mesh construction module 610 provided in this application is configured with the following units:

[0130] Hexagonal grid segmentation units are used to analyze the geographic spatial boundaries of the forestry area under study. Based on the preset hexagonal unit size parameters, the forestry area under study is segmented in a mosaic manner to generate a hexagonal grid layer covering the entire study area.

[0131] The multidimensional environmental feature extraction unit is used to perform geographic coordinate positioning processing on each hexagonal node in the hexagonal grid layer, extract the elevation value, slope value, Euclidean distance value to the nearest water source, normalized vegetation index value, Euclidean distance value to the nearest road and canopy closure value of the center point of each hexagonal node, and generate the initial environmental feature set of each hexagonal node.

[0132] The environment state space construction unit is used to normalize and scale the initial set of environment features to construct an environment state space consisting of all hexagonal nodes and their corresponding normalized multidimensional environment features.

[0133] Preferably, the mobile task pairing and mapping module 620 provided in this application is configured with the following units:

[0134] The observation point screening and pairing unit is used to analyze the timestamps and spatial locations of each observation point in the historical appearance data of wild animals in the forestry area, and to filter continuous observation point pairs with clear movement direction indications based on field record information, generating a set of candidate start and end point pairs;

[0135] The start and end point grid mapping unit is used to perform spatial coordinate transformation on each observation point in the candidate start and end point pair set, calculate the Euclidean distance between each observation point and all hexagonal nodes in the hexagonal grid, match each observation point to the nearest hexagonal node, and generate a mapping point pair consisting of the start hexagonal node and the end hexagonal node.

[0136] The invalid point pair removal unit is used to deduplicatize and filter mapped point pairs, removing invalid point pairs whose start and end points are the same or whose distance is less than the preset minimum effective migration distance threshold, and generating a sparse movement task set.

[0137] Preferably, the resource suitability assessment module 630 provided in this application is configured with the following units:

[0138] The training sample construction unit is used to count the frequency of wild animal occurrences for each hexagonal node in the environmental state space, and to label the positive and negative sample labels of each node based on historical occurrence data to generate the training sample set of the random forest model.

[0139] The random forest model training unit is used to train the random forest model through multiple rounds of decision tree growth, taking the multidimensional environmental features of each sample in the training sample set as input and the label of whether wild animals appear in the sample as output, to construct a resource selection probability model that represents the nonlinear relationship between environmental features and the appearance of wild animals.

[0140] The resource selection probability prediction unit is used to input the multidimensional environmental features of all hexagonal nodes in the environmental state space into the trained resource selection probability model, perform forward propagation prediction processing on each node, and generate the resource selection probability value corresponding to each hexagonal node.

[0141] The suitability distribution map generation unit is used to perform raster interpolation on the resource selection probability value of each hexagonal node, converting the discrete node probability values ​​into a continuously covered raster layer, and generating an environmental resource suitability distribution map covering the entire forestry area under study.

[0142] Preferably, the individual path preference learning module 640 provided in this application is configured with the following units:

[0143] The individual memory initialization unit is used to randomly initialize a high-dimensional real number vector for each virtual individual as the initial value of the individual memory vector based on a preset virtual individual quantity parameter, and to combine and store the memory vectors of all virtual individuals to generate an initial experience memory bank.

[0144] The virtual individual path exploration unit is used to generate paths for each task in the sparse movement task set. It couples the memory vector of each virtual individual with the environmental resource suitability distribution map. Through a reinforcement learning framework, each virtual individual explores the movement path from the task start point to the end point, and records the cumulative reward value of each exploration path.

[0145] The attention weight calculation unit is used to sort and normalize the cumulative reward values ​​of each virtual individual under each task, calculate the attention weight of each virtual individual when completing the task, and generate the individual experience contribution distribution corresponding to each task.

[0146] The optimal path template generation unit is used to perform weighted fusion of the exploration paths of each virtual individual based on attention weights, and generate an integrated path as the optimal path template for the current iteration round.

[0147] The individual memory vector update unit is used to perform gradient calculation on the individual memory vector of each virtual individual. With the goal of maximizing the cumulative reward and combining the integrated path template for behavioral cloning constraints, it updates the memory vector of each virtual individual through the gradient ascent method to generate an updated experience memory bank.

[0148] The iterative convergence memory unit is used to repeatedly execute path exploration, weight calculation, template generation and memory update steps until the preset iteration rounds are completed or the cumulative reward value converges, generating an individual experience memory bank containing multiple differentiated individual memory vectors.

[0149] Preferably, the attention path search module 650 provided in this application is configured with the following units:

[0150] The start and end point grid mapping unit is used to obtain the current start and end positions of the target wild animal, perform spatial coordinate transformation on the current start and end positions, and map the current start and end positions to the nearest hexagonal node in the environmental state space based on the principle of minimum Euclidean distance, thereby generating the start and end nodes of the task to be predicted.

[0151] The candidate path cost evaluation unit is used to evaluate the path cost of each virtual individual in the individual experience memory bank. Taking the starting node as the start point and the ending node as the target, it generates candidate paths for each virtual individual based on the environmental resource suitability distribution map and calculates the path cost value based on attention weight.

[0152] The experience call weight calculation unit is used to perform softmax normalization on the path cost value, calculate the attention weight of each virtual individual for completing the current prediction task, and generate the individual experience call weight distribution.

[0153] The integrated path search guidance graph construction unit is used to weight and fuse the resource selection probabilities of each node on the environmental resource suitability distribution map based on individual experience and the weight distribution, so as to generate an integrated path search guidance graph that integrates group experience.

[0154] The greedy master path generation unit is used to perform a greedy search on the integrated path search guidance graph starting from the starting node, successively selecting the adjacent node with the highest guidance value for state transition until the end node is reached, and outputting the sequence of nodes passed as the master predicted path.

[0155] The alternative path filtering output unit is used to sort each virtual individual in descending order based on the weight distribution of individual experience, select the candidate path corresponding to the virtual individual in the preset ranking, and output the selected candidate path as the set of alternative prediction paths.

[0156] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for predicting the trajectory of forestry wildlife in sparse scenes.

[0157] In one embodiment, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the trajectory of forestry wildlife in sparse scenes.

[0158] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0159] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0160] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the trajectory of forestry wildlife in sparse scenes, characterized in that, Includes the following steps: S1: Divide the forestry area under study into hexagonal grids, dividing the forestry area into a mosaic grid composed of multiple hexagonal nodes, and extract the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features. S2: Pair start and end points of historical occurrence data of wild animals in the forestry area. Select observation point pairs with clear start and end relationships of movement direction from the historical occurrence data. Map the start and end points of each observation point pair to the nearest hexagonal node to generate a sparse movement task set composed of the start hexagonal node and the end hexagonal node. S3: Based on the environmental state space, resource selection probability is estimated, and a random forest model is called to learn and fit the nonlinear relationship between the multidimensional environmental features and the historical occurrence points of wild animals, generating an environmental resource suitability distribution map covering the entire study area. S4: Based on the sparse mobile task set and the environmental resource suitability distribution map, perform individual path preference learning, construct a trainable individual memory vector for each virtual individual, and iteratively train the individual memory vector based on the reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors. S5: Obtain the current starting point and ending point of the target wild animal, map the current starting point and ending point to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and the individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within the preset ranking as alternative prediction paths.

2. The method according to claim 1, characterized in that, S1 includes: S11: Analyze the geographic spatial boundary of the forestry area under study, and perform mosaic subdivision of the forestry area under study according to the preset hexagonal unit size parameters to generate a hexagonal grid layer covering the entire study area. S12: Perform geographic coordinate positioning processing on each hexagonal node in the hexagonal grid layer, extract the elevation value, slope value, Euclidean distance value to the nearest water source, normalized vegetation index value, Euclidean distance value to the nearest road and canopy closure value of the center point of each hexagonal node, and generate the initial environmental feature set of each hexagonal node. S13: Normalize and scale the initial set of environmental features to construct an environmental state space consisting of all hexagonal nodes and their corresponding normalized multidimensional environmental features.

3. The method according to claim 1, characterized in that, S2 includes: S21: For each observation point in the historical appearance data of wild animals in the forestry area, the timestamp and spatial location are analyzed. Based on the field record information, continuous observation point pairs with clear movement direction indications are selected to generate a set of candidate start and end point pairs. S22: Perform spatial coordinate transformation on each observation point in the candidate start and end point pair set, calculate the Euclidean distance between each observation point and all hexagonal nodes in the hexagonal grid, and match each observation point to the nearest hexagonal node to generate a mapping point pair consisting of the start hexagonal node and the end hexagonal node; S23: Perform deduplication and filtering on the mapped point pairs, and remove invalid point pairs with the same starting point and ending point or a distance less than the preset minimum effective migration distance threshold to generate a sparse mobile task set.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform wildlife occurrence frequency statistics on each hexagonal node in the environmental state space, label the positive and negative samples of each node based on the historical occurrence point data, and generate a training sample set for the random forest model. S32: Using the multidimensional environmental features of each sample in the training sample set as input and the label of whether wild animals appear in the sample as output, perform multiple rounds of decision tree growth training on the random forest model to construct a resource selection probability model that represents the nonlinear relationship between environmental features and the appearance of wild animals. S33: Input the multidimensional environmental features of all hexagonal nodes in the environmental state space into the trained resource selection probability model, perform forward propagation prediction processing on each node, and generate the resource selection probability value corresponding to each hexagonal node; S34: Perform raster interpolation on the resource selection probability value of each hexagonal node to convert the discrete node probability values ​​into a continuously covered raster layer, generating an environmental resource suitability distribution map covering the entire forestry area under study.

5. The method according to claim 1, characterized in that, S4 includes: S41: Based on the preset number of virtual individuals, a high-dimensional real number vector is randomly initialized for each virtual individual as the initial value of the individual memory vector, and the memory vectors of all virtual individuals are combined and stored to generate an initial experience memory bank. S42: Perform path generation processing on each task in the sparse movement task set, couple the memory vector of each virtual individual with the environmental resource suitability distribution map, and use a reinforcement learning framework to allow each virtual individual to explore the movement path from the task starting point to the end point, and record the cumulative reward value of each exploration path; S43: Sort and normalize the cumulative reward values ​​of each virtual individual under each task, calculate the attention weight of each virtual individual when completing the task, and generate the individual experience contribution distribution corresponding to each task; S44: Based on the attention weights, the exploration paths of each virtual individual are weighted and fused to generate an integrated path as the optimal path template for the current iteration round; S45: Perform gradient calculation on the individual memory vector of each virtual individual, with the goal of maximizing the cumulative reward and combining it with the integrated path template to perform behavioral cloning constraints, update the memory vector of each virtual individual through the gradient ascent method, and generate an updated experience memory bank. S46: Repeat S42 to S45 until the preset iteration rounds are completed or the cumulative reward value converges, generating an individual experience memory bank containing multiple differentiated individual memory vectors.

6. The method according to claim 5, characterized in that, The update formula for the memory vector of each virtual individual is: in, Let be the memory vector of the k-th virtual individual after the (t+1)-th iteration. Let be the current memory vector of the k-th virtual individual in the t-th iteration. The learning rate hyperparameter for the memory vector. The total number of tasks in the sparse mobile task set. The cumulative reward value obtained for the k-th virtual individual to complete the m-th movement task. The regularization coefficient hyperparameter for behavior cloning constraints. The path selection strategy distribution for the k-th virtual individual Distribution with integrated path strategy KL divergence between them The regularization coefficient hyperparameter is used to represent the experience collaboration among individuals. This refers to the set of virtual individuals whose memory similarity to the k-th virtual individual is higher than a preset threshold. The empirical collaboration weight coefficient between the k-th virtual individual and the j-th virtual individual is... For two virtual individual memory vectors and Cosine similarity between them.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: Obtain the current start and end positions of the target wild animal, perform spatial coordinate transformation on the current start and end positions, and map the current start and end positions to the nearest hexagonal node in the environmental state space based on the principle of minimum Euclidean distance, thereby generating the start and end nodes of the task to be predicted. S52: For each virtual individual in the individual experience memory bank, perform path cost evaluation. Taking the starting node as the origin and the ending node as the destination, generate candidate paths for each virtual individual based on the environmental resource suitability distribution map and calculate the path cost value based on attention weights. The formula for calculating the path cost value is as follows: in, The weighted path cost value for the k-th virtual individual to complete the current predicted task. The total number of steps in the candidate paths generated for the k-th virtual individual. For discount factor hyperparameters, The hexagonal node at step t of the path obtained from the environmental resource suitability distribution map. The probability value of resource selection. The memory decay coefficient hyperparameter, For the k-th virtual individual to access nodes during the historical training process Total number of times Let $k$ be the maximum number of visits the k-th virtual individual makes to all nodes. Explore the hyperparameters of the excitation coefficient for the new node. For indicator functions, when node The value is 1 if the virtual individual has never been accessed in the training history, and 0 otherwise. S53: Perform softmax normalization on the path cost value, calculate the attention weight of each virtual individual for completing the current task to be predicted, and generate the individual experience call weight distribution; S54: Based on the individual experience call weight distribution, the resource selection probability of each node on the environmental resource suitability distribution map is weighted and fused to generate an integrated path search guidance map that integrates group experience; S55: On the integrated path search guidance graph, start a greedy search from the starting node, and successively select the adjacent node with the highest guidance value to perform state transition until the end node is reached. Output the sequence of nodes passed through as the main predicted path. S56: Based on the individual experience call weight distribution, sort each virtual individual in descending order, select the candidate path corresponding to the virtual individual in the preset ranking, and output the selected candidate path as the set of candidate prediction paths.

8. A sparse scene forestry wildlife trajectory prediction system, characterized in that, The system includes: The hexagonal mesh construction module is used to divide the research forestry area into hexagonal meshes, dividing the research forestry area into a mosaic mesh composed of multiple hexagonal nodes, and extracting the multidimensional environmental features corresponding to each hexagonal node to construct an environmental state space composed of all hexagonal nodes and corresponding multidimensional environmental features. The mobile task pairing and mapping module is used to pair start and end points of historical occurrence data of wild animals in the forestry area. It selects observation point pairs with clear start and end relationships of movement direction from the historical occurrence data, maps the start and end points of each observation point pair to the nearest hexagonal node, and generates a sparse mobile task set composed of the start hexagonal node and the end hexagonal node. The resource suitability assessment module is used to estimate the probability of resource selection based on the environmental state space, call the random forest model to learn and fit the nonlinear relationship between the multidimensional environmental characteristics and the historical occurrence points of wild animals, and generate an environmental resource suitability distribution map covering the entire study area. The individual path preference learning module is used to learn individual path preferences based on the sparse mobile task set and the environmental resource suitability distribution map, construct a trainable individual memory vector for each virtual individual, and iteratively train the individual memory vector based on a reinforcement learning framework to generate an individual experience memory bank containing multiple differentiated individual memory vectors. The attention path search module is used to obtain the current start and end positions of the target wild animal, map the current start and end positions to the corresponding hexagonal nodes in the environmental state space, perform weighted path search based on the environmental resource suitability distribution map and the individual experience memory bank, assign attention weights to the memory vectors of each virtual individual in the individual experience memory bank, take the hexagonal node sequence with the highest score after weighted integration as the main prediction path, and take the individual paths generated by virtual individuals whose attention weights are ranked within a preset range as alternative prediction paths.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images