A method, system, device and medium for predicting an asian elephant distribution area

Abstract

This invention relates to a method, system, device, and medium for predicting the distribution area of ​​Asian elephants, belonging to the field of Asian elephant trajectory prediction technology. The method includes: S1, performing spatiotemporal aggregation processing on the raw observation data of the target Asian elephant population to generate a daily representative coordinate sequence ordered by time; S2, converting the geographical coordinates in the daily representative coordinate sequence into local relative planar coordinates based on the reference origin corresponding to each population identifier; S3, constructing a model input sequence containing time-step feature vectors based on the daily representative coordinate sequence, local relative planar coordinates, and associated multidimensional environmental features; S4, inputting the model input sequence and population identifiers into a trained trajectory prediction model to obtain the predicted displacement for the next time step. This application achieves high-precision, automated prediction of the future location of Asian elephant populations through a complete technical process of data aggregation, coordinate transformation, feature fusion, and neural network prediction.

Description

Technical Field

[0001] This application belongs to the field of Asian elephant trajectory prediction technology, and in particular relates to a method, system, device and medium for predicting the distribution area of ​​Asian elephants. Background Technology

[0002] As an important protected species, the prediction of the Asian elephant's range is crucial for mitigating human-elephant conflict and developing scientific conservation strategies. Currently, predictions of the species' distribution area mainly rely on Species Distribution Models (SDMs). These models simulate habitat suitability and predict potential distribution ranges by analyzing the statistical relationships between species occurrence data (such as GPS tracking records and camera trap data) and environmental variables.

[0003] Common SDM (Site Detection and Management) technologies mainly include maximum entropy models (MaxEnt), ensemble learning models (such as those combining GLM, GAM, RF, MaxEnt, etc.), random forests (RF), Bayesian models (such as Bayesian logistic regression), and connectivity analysis models based on landscape ecology (such as minimum cost paths and circuit theory). These models typically couple remote sensing data with climate scenarios, outputting static habitat suitability indices or probability distribution maps under future scenarios, playing an important role in macro-level habitat assessment and protected area planning.

[0004] However, existing SDM (Specialized Dynamic Management) technologies have significant limitations when applied to the refined management and dynamic early warning of large, mobile animals such as Asian elephants: First, most models are based on static "presence-environment" relationships, making it difficult to capture and predict the actual movement trajectories and dynamic processes of animals over time, and failing to answer the crucial operational question of "where will the elephant herd go tomorrow?" Second, models typically ignore or simplify the heterogeneity within animal groups, treating different elephant herds as homogeneous units, and cannot distinguish and predict the behavior of groups with different habits and movement patterns. Third, their predictions heavily rely on complete and continuous species presence data, lacking robust mechanisms to handle the spatiotemporal discontinuities and GPS noise present in real-world observation data, resulting in decreased prediction accuracy in areas with sparse data or high noise levels. Fourth, most models are computationally complex, heavily reliant on multi-source, high-dimensional environmental data and high-performance computing resources, and their outputs are mostly fitness scores, requiring subjective thresholding to translate into specific geographical locations, which is not conducive to real-time, automated monitoring and early warning applications.

[0005] Therefore, existing technical solutions are significantly insufficient in achieving accurate, real-time, and interpretable spatiotemporal trajectory prediction for large-scale, multi-group, and dynamically changing wild animals. There is an urgent need for a method to predict the distribution area of ​​Asian elephants that can deeply integrate temporal behavioral patterns, group heterogeneity characteristics, and multi-source environmental factors, and has strong noise resistance. Summary of the Invention

[0006] This application aims to at least partially address one of the technical problems in related technologies. To this end, this application provides a method, system, device, and medium for predicting the distribution area of ​​Asian elephants. Through a complete technical process of data aggregation, coordinate transformation, feature fusion, and neural network prediction, it achieves high-precision and automated prediction of the future location of Asian elephant populations.

[0007] To achieve the above objectives, firstly, this application provides a method for predicting the distribution area of ​​Asian elephants, comprising: S1. Spatiotemporal aggregation processing is performed on the raw observation data of the target Asian elephant population to generate a daily representative coordinate sequence ordered by time; the raw observation data includes discrete spatiotemporal location points and their respective population identifiers; S2. Based on the reference origin corresponding to each group identifier, convert the geographical location coordinates in the daily representative coordinate sequence into local relative planar coordinates; S3. Based on the daily representative coordinate sequence, the local relative plane coordinates, and the associated multidimensional environmental features, construct a model input sequence containing time step feature vectors; S4. Input the model input sequence and the group identifier into the trained trajectory prediction model to obtain the predicted displacement of the next time step; wherein, the trajectory prediction model includes an embedding layer and a recurrent neural network; the embedding layer is used to map the group identifier into an embedding vector and fuse it with the time step feature vector to form the input of the recurrent neural network; S5. Obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

[0008] Preferably, the step of spatiotemporal aggregation processing in step S1 includes: calculating the average spatial dispersion of each group identifier in the daily set of observation points; If the average spatial dispersion is less than or equal to a preset threshold, the weighted average centroid is calculated based on the weight of each observation point and used as the representative coordinate for that day. If the average spatial dispersion is greater than the preset threshold, a weighted center point is selected from the set of observation points as the representative coordinate of the day. The weighted center point is the observation point that minimizes the sum of the weighted distances to all other observation points.

[0009] Preferably, the multidimensional environmental features in step S3 include at least one of vegetation index, topographic features, and meteorological data; The time step feature vector includes sine and cosine periodic encoded features calculated based on the date.

[0010] Preferably, in step S4, the embedding vector output by the embedding layer of the trajectory prediction model is copied in the time dimension and concatenated with the time step feature vector of each time step to form the input of the recurrent neural network.

[0011] Preferably, the time step feature vector constructed in step S3 includes a binary existence marker to indicate whether the data for that day is a true observation.

[0012] Preferably, the trajectory prediction model is trained using the Huber loss function, and before training, the displacement labels in the training data are truncated based on statistical quantiles. The displacement labels are the daily movement distances (Δx, Δy) of the target Asian elephant herd.

[0013] Preferably, the step of performing truncation based on statistical quantiles includes: Calculate the displacement modulus corresponding to all displacement labels in the training data, and use the predetermined quantile of the displacement modulus as the truncation threshold. For displacement labels whose displacement modulus exceeds the cutoff threshold, the displacement label is scaled down proportionally so that its modulus equals the cutoff threshold.

[0014] Preferably, the expression for the Huber loss function is: ; in, L δ This represents Huber's loss value. y This represents the actual displacement label value. f ( x ) represents the predicted displacement of the trajectory prediction model. d The preset threshold parameter is used to define the region of squared loss and the region of linear loss. y - f ( x )∣ represents the absolute value of the prediction error.

[0015] Preferably, during the training process, the trajectory prediction model dynamically weights the loss function according to the number of samples corresponding to each group identifier, with group identifiers with fewer samples having higher weights.

[0016] Preferably, step S5, which involves obtaining the local relative plane coordinates for the next time step based on the predicted displacement, includes: Based on the predicted displacement and recent historical displacement information, the local relative plane coordinates of the next time step are obtained by fusion and correction. The fusion correction is based on a gated fusion strategy, which is used to adaptively adjust the weight of the predicted displacement and the previous day's position in the final prediction result of the trajectory prediction model according to the recent movement status of the target Asian elephant population.

[0017] Preferably, the gating fusion strategy is configured as follows: Calculate the average displacement modulus of recent historical displacement information; The fusion coefficient is determined based on the ratio of the average displacement modulus to the preset static threshold. Based on the fusion coefficient, the predicted displacement of the trajectory prediction model is weighted and fused with the position of the previous day to obtain the corrected local relative planar coordinates of the next time step.

[0018] Secondly, this application provides a system for predicting the distribution area of ​​Asian elephants, comprising: The data preprocessing module is used to perform spatiotemporal aggregation and coordinate transformation on the raw observation data of the target Asian elephant population, and output the daily local relative planar coordinate sequence of each population identifier; the raw observation data includes discrete spatiotemporal location points and their respective population identifiers; The feature construction module is used to fuse multidimensional environmental features and time period features into the daily local relative planar coordinate sequence to construct a model input sequence containing time step feature vectors; The trajectory prediction module includes an embedding layer and a recurrent neural network; the embedding layer is used to map the input group identifiers into embedding vectors; the recurrent neural network is used to receive the model input sequence fused with the embedding vectors and output the predicted displacement for the next time step. The post-processing module is used to obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

[0019] Preferably, the post-processing module is further configured to perform gated fusion correction on the predicted displacement based on recent historical displacement information.

[0020] Thirdly, this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in any of the above-mentioned embodiments.

[0021] Fourthly, this application provides a computer-readable storage medium including a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described above.

[0022] Based on the above technical solution, it can be seen that the method for predicting the distribution area of ​​Asian elephants in this application has at least one of the following advantages compared with the prior art: 1. The method for predicting the distribution area of ​​Asian elephants in this application resolves the contradiction between "inefficient modeling for each group" and "the inability of a single model to fit heterogeneous groups" through the core technology of "cluster embedding". By using a recurrent neural network with shared parameters, combined with learnable group embedding vectors, a unified model is used to accurately model and predict the behavioral patterns of multiple heterogeneous animal groups, achieving personalized predictions for "a thousand groups, a thousand faces".

[0023] 2. This application, by deeply integrating multidimensional environmental features and periodic time-series coding, endows the trajectory prediction model with ecological perception capabilities, improving the interpretability of prediction results and the accuracy of medium- and long-term trends. By integrating multidimensional environmental features such as vegetation index, topography, and meteorology into the model input, and by applying sine and cosine periodic coding to the time variables, the ecological driving factors of Asian elephant migration are explicitly injected into the model. This allows for the capture of the impact of environmental changes and periodic patterns on animal behavior, resulting in more accurate and reliable inferences about medium- and long-term trends.

[0024] 3. In the data input stage, this application performs spatiotemporal aggregation processing on the raw observation data and introduces existence markers to distinguish between real data and imputed values. In the model training stage, it comprehensively utilizes the Huber loss function, a label truncation mechanism based on statistical distribution, and a dynamic weighting strategy tailored to different populations to effectively suppress outlier interference and balance the learning objective. In the result output stage, a gating fusion mechanism is used to adaptively smooth and calibrate the predicted trajectory. This consistent collaborative design significantly enhances the system's adaptability to complex field data, making its output trajectory not only smooth and stable but also consistent with the physical laws of animal behavior, thus possessing the reliability and practicality to support actual field monitoring and conservation management decisions. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating a method for predicting the distribution area of ​​Asian elephants provided in this application; Figure 2 This is a connection diagram of a system for predicting the distribution area of ​​Asian elephants provided in this application. Detailed Implementation

[0027] To facilitate understanding of this application, a more complete description will be provided below. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0028] 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 application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0029] In this application, terms such as "preferred," "better," "more suitable," and "ideal" are merely used to describe implementation methods or embodiments that achieve better results, and should be understood not to limit the scope of protection of this application.

[0030] In this application, terms such as "further," "even more," and "particularly" are used for descriptive purposes and to indicate differences in content, but should not be construed as limiting the scope of protection of this application.

[0031] In this application, "optionally," "optionally," and "optional" mean that something is optional, that is, it means that it is selected from either "with" or "without." If there are multiple "optional" entries in a technical solution, unless otherwise specified, and there are no contradictions or mutual constraints, each "optional" entry shall be independent.

[0032] In this application, the terms "first aspect," "second aspect," "third aspect," and "fourth aspect," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or quantity, nor should they be construed as implicitly indicating the importance or quantity of the indicated technical features. Moreover, "first," "second," "third," and "fourth," etc., serve only a non-exhaustive enumeration purpose and should be understood not to constitute a closed limitation on quantity.

[0033] In this application, the technical features described in an open-ended manner include both closed technical solutions consisting of the listed features and open technical solutions that include the listed features.

[0034] The technical terms used in this application are explained as follows: LSTM (Long Short-Term Memory) is a deep learning model used to capture temporal dependencies.

[0035] HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) is a clustering algorithm used to cluster raw observation points to identify different animal groups (elephant herds).

[0036] NDVI (Normalized Difference Vegetation Index) is a key environmental dynamic characteristic that reflects vegetation cover.

[0037] PyTorch is an open-source deep learning framework used to build and train LSTM models.

[0038] Cluster embedding is a technique used to map discrete group IDs (i.e., group identifiers) to continuous vectors to characterize group heterogeneity.

[0039] Entity Embedding is the specific technical principle used by Cluster Embedding.

[0040] Gated Fusion is a post-processing mechanism used to adjust prediction results based on recent motion momentum and suppress stationary drift.

[0041] like Figure 1 As shown in the embodiments of this application, a method for predicting the distribution area of ​​Asian elephants is provided, including: S1. Spatiotemporal aggregation processing is performed on the raw observation data of the target Asian elephant population to generate a daily representative coordinate sequence ordered by time; the raw observation data includes discrete spatiotemporal location points and their respective population identifiers; In one possible implementation, the step of spatiotemporal aggregation in step S1 includes: for each group identifier in the daily set of observation points, measuring the degree of spatial distribution concentration of a certain elephant group among all observation points within a single day by calculating its average spatial dispersion. If the average spatial dispersion is less than or equal to a preset threshold, the weighted average centroid is calculated based on the weight of each observation point as the representative coordinate of the day. The preset threshold can be set by combining historical data analysis of the daily activity range of Asian elephants, or it can be set by calculating the average spatial dispersion distribution of all groups on all dates and selecting a certain quantile as the preset threshold.

[0042] If the average spatial dispersion is greater than the preset threshold, a weighted center point is selected from the set of observation points as the representative coordinate of the day. The weighted center point is the observation point that minimizes the sum of the weighted distances to all other observation points.

[0043] In one possible implementation, to achieve accurate prediction of the spatiotemporal trajectory of Asian elephant herds, the raw observation data first needs to be cleaned and aggregated to construct a regular time-series input. The raw data comes from devices such as GPS collars and infrared cameras, and has typically been preprocessed using the HDBSCAN clustering algorithm to identify different elephant individuals within a herd, including information such as the latitude and longitude of each observation point, timestamp, herd identifier, and number of observations. Since the same elephant herd may generate multiple spatially discrete observation points within a single day, the core objective of this step is to aggregate these discrete points into a single representative coordinate for each day, forming a continuous time series.

[0044] The specific aggregation method employs an adaptive strategy. Under normal circumstances, for all observation points of a certain elephant herd on a certain day, the weighted centroid method is preferentially used to calculate its daily representative coordinates. This method fully considers the differences in the size of the elephant herd represented by different observation points, using the "number" of observation points as the weight to calculate the weighted average center. The calculation formula is as follows: ; in London i and Latin i The first i The longitude and latitude of each observation point w i The corresponding observation quantity weight, It is a weighted average longitude. It is a weighted average latitude.

[0045] This method can accurately reflect the "center of gravity" of elephant activity when the herd is relatively concentrated.

[0046] However, field observation data often contains outliers due to equipment errors or unusual animal behavior (such as temporary separation of groups in two locations). Directly using the weighted centroid method may cause representative points to deviate from the actual activity area. Therefore, this application introduces an adaptive noise-resistant backoff mechanism. This mechanism first calculates the average spatial dispersion among all observation points for the day to quantify the degree of group aggregation. If the average dispersion exceeds a preset reasonable threshold (e.g., 3000 meters), it is determined that there is significant noise or the group is extremely dispersed, and it will automatically backoff to a more robust weighted centroid algorithm. This algorithm selects a representative point from the actual observation points whose sum of weighted distances to all other points is the smallest, thus ensuring that the daily representative coordinates fall within the actual observation location range and effectively suppressing interference from outliers.

[0047] After completing the daily coordinate aggregation, the discontinuity in the time series needs to be addressed. Due to potential equipment failure or signal loss, observational data may be missing on certain dates. To address this, the system constructs a continuous calendar index for each elephant population. For dates with missing data, a combined "forward imputation" and "backward imputation" strategy is used: firstly, coordinates from the previous valid date are used for imputation; if none are available, coordinates from the next valid date are used. Crucially, while imputing data, the system generates a binary presence flag called "presence." For dates obtained from aggregated real observations, the presence flag is 1; for dates generated through imputation, the presence flag is 0. This design allows subsequent prediction models to clearly distinguish between "real observations" and "estimated imputations" in the input data, preventing the model from learning spurious patterns introduced by the imputation rules and ensuring the fidelity of the learning process.

[0048] S2. Based on the reference origin corresponding to each group identifier, convert the geographical location coordinates in the daily representative coordinate sequence into local relative planar coordinates; To construct spatiotemporal trajectory prediction inputs suitable for deep learning models, the problem of directly using geographic coordinates for numerical calculations must be addressed. Directly using latitude and longitude (WGS84 coordinates) has significant drawbacks: First, the Earth is a sphere, and latitude and longitude are angular quantities; the actual ground distances (in meters) represented by these coordinates are not equal at different latitudes, especially in high-latitude regions, where directly calculating the latitude and longitude difference between two points introduces severe distortion of ground distances. Second, latitude and longitude values ​​typically change only after several decimal places (e.g., 22.123456°), and such minute numerical differences can easily lead to gradient instability or vanishing gradients during neural network training. Therefore, this application employs a coordinate transformation strategy based on local projection to convert spherical coordinates into local planar coordinates conforming to Euclidean geometry.

[0049] The core of this strategy is the introduction of cluster anchor points. For each independent group of images (cluster k) identified by HDBSCAN clustering, its global weighted center is first calculated over the entire observation period. This center point... The calculation method is similar to the aforementioned daily aggregate weighted centroid method, but the data source covers all historical observation points of the cluster. This anchor point (i.e., the center point) will serve as the origin of the local planar coordinate system specific to this elephant group, thereby constraining all subsequent location calculations for this elephant group to the vicinity of its main activity area, effectively reducing the range of projection distortion.

[0050] The specific process of coordinate transformation employs a locally equidistant cylindrical projection model, which is an efficient method for linearly approximating spherical coordinates to planar coordinates within a small range. For the latitude and longitude coordinates of this elephant group at any time t (… L at t ,L on t The coordinates are transformed to local plane coordinates with the anchor point as the origin. ;in, London t and Latin t The first t The longitude and latitude of each observation point and 111320 and 110540 are the longitude and latitude of the center point of cluster k of the elephant group, respectively. 111320 and 110540 are the approximate ground distances corresponding to each degree of longitude and latitude near the Earth's equator. The term is used to correct the effect of latitude on the distance between meridians, ensuring that in the local area where the anchor point is located, the unit scale in the x and y coordinate axes is as close as possible to the actual metric distance.

[0051] After this transformation, the original latitude and longitude are converted into local Cartesian coordinates in meters. This transformation brings key benefits: First, it solves the distance distortion problem, ensuring that the coordinate difference (Δx, Δy) directly corresponds to the actual planar displacement (meters), giving the learned motion patterns a clear physical meaning. Second, the range of the converted coordinate values ​​is normalized, improving the numerical stability of the input data and facilitating gradient propagation and convergence of the deep learning model. Finally, the localization centered on the group's anchor points allows the model to focus more intently on learning the fine-grained motion patterns of the group within its habitual activity range, improving the specificity of the predictions. The coordinates will serve as the basis for the subsequent feature construction.

[0052] S3. Based on the daily representative coordinate sequence, the local relative plane coordinates, and the associated multidimensional environmental features, construct a model input sequence containing time step feature vectors; In one possible implementation, the multidimensional environmental features in step S3 include at least one of vegetation index, topographic features, and meteorological data. The time step feature vector includes sine and cosine periodic encoded features calculated based on the date.

[0053] In one possible implementation, the time step feature vector constructed in step S3 includes a binary existence marker to indicate whether the data for that day is a true observation.

[0054] After completing the spatiotemporal aggregation and local projection transformation of the coordinates, a comprehensive feature vector needs to be constructed for each time step as the direct input to the trajectory prediction model. The design of this feature vector aims to comprehensively encode the spatiotemporal, environmental, and periodic information that influences the Asian elephant's migration decisions, enabling the model to learn not only based on historical trajectories but also to understand the ecological logic driving migration behavior, thereby achieving interpretable and accurate predictions.

[0055] Feature vectors mainly contain the following three types of standardized information: 1. Basic Spatiotemporal State Characteristics: This part characterizes the core state of the elephant herd on a specific date. It first includes standardized local planar coordinates. The coordinates directly represent the precise location of the elephant herd in local Euclidean space. Secondly, the previously generated binary presence marker is introduced to clearly indicate whether the data for that day originates from actual observations or imputations, preventing model confusion. Finally, the herd count_sum (total population size) observed that day is included. This feature, standardized by Z-Score, reflects the dynamic changes in population size, providing the model with contextual information about population size.

[0056] 2. Multidimensional Ecological Environment Characteristics: To achieve ecology-driven prediction, this application deeply integrates static and dynamic environmental factors. Through a geographic information system interface, multiple habitat data for each coordinate point on the corresponding date are obtained, including: Topographical features, such as altitude and slope, affect the cost of animal movement and habitat selection.

[0057] Vegetation characteristics: The core indicator is the normalized vegetation index, which reflects the vegetation cover and lushness and is a key factor in determining the distribution of food resources.

[0058] Meteorological characteristics, such as daily cumulative precipitation and average daily temperature, directly affect the behavior and physiological needs of animals.

[0059] All environmental features are Z-score standardized before input using a Standardizer (a data processing module or tool that converts input data into a uniform format suitable for machine learning models). This standardization eliminates the influence of dimensions and ensures that the features are of comparable importance during model training.

[0060] 3. Time-periodic encoding features: To capture the periodic migration or activity patterns of Asian elephants that vary with seasons and years, the system continuously encodes the absolute time variable. Dates are converted to the day of the year (d), and sine and cosine functions are used to map them onto the unit circle, generating two periodic features: ; in, and These are the date sine and date cosine features, respectively. This encoding method ensures that the cycle boundaries (such as the 365th day and the 1st day) are continuous and smooth in the feature space, enabling the model to effectively learn and extrapolate seasonal behavioral patterns.

[0061] Finally, the input vector X at each time step t t It is composed of all the standardized basic features, environmental features, and periodic encoded features mentioned above. This method enables the model to simultaneously perceive spatial location, data confidence, population size, terrain constraints, vegetation resources, climate conditions, and seasonal cycles, thereby making trajectory predictions that conform to both historical movement trends and ecological principles, significantly improving the interpretability of the prediction results and the accuracy of medium- and long-term trends.

[0062] S4. Input the model input sequence and the group identifier into the trained trajectory prediction model to obtain the predicted displacement of the next time step; wherein, the trajectory prediction model includes an embedding layer and a recurrent neural network; the embedding layer is used to map the group identifier into an embedding vector and fuse it with the time step feature vector to form the input of the recurrent neural network; In one possible implementation, in step S4, the embedding vector output by the embedding layer of the trajectory prediction model is copied in the time dimension and concatenated with the time step feature vector of each time step to form the input of the recurrent neural network.

[0063] After constructing the comprehensive model input sequence, a trajectory prediction model is needed that can effectively handle long-term temporal dependencies and adapt to the heterogeneity of behavior among different elephant groups. To this end, this application constructs a deep learning model (i.e., a trajectory prediction model) called LSTMDeltaRegressor (a displacement regressor based on a long short-term memory network). The core innovation of this model lies in the organic combination of entity embedding technology and recurrent neural networks, achieving the goal of "one unified model accurately predicting multiple heterogeneous groups."

[0064] 1. Trajectory Prediction Model Input and Fusion Layer The input to the trajectory prediction model consists of two parts: First, there is the temporal feature tensor: its shape is [Batch, Window=14, Features]. Here, Batch represents the batch size, Window=14 represents the model observing data from the past 14 consecutive days, and Features are the time-step feature vectors obtained in the aforementioned feature construction steps. This tensor encodes the dynamic state and environmental information of the target elephant population over the past two weeks.

[0065] Second, the group identifier: an integer tensor of shape [Batch], where each integer represents a unique elephant group cluster number.

[0066] The trajectory prediction model processes group identification through an embedding layer. This layer is essentially a learnable lookup table that maps each discrete cluster number to a low-dimensional, continuous vector called the cluster embedding vector. The dimension of this vector can be set to 8 (those skilled in the art will understand that this dimension can be optimized and adjusted through conventional experimental verification based on the actual number of monitored groups and the complexity of their behavior, for example, selecting a suitable value within the range of 4 to 32). During training, the model automatically optimizes this vector through backpropagation, implicitly encoding the "personality" or long-term behavioral preferences of the corresponding elephant herd (such as migratory tendencies, settlement tendencies, dependence on water sources, etc.). To fuse this static group characteristic with dynamic temporal features, the system copies the cluster embedding vector of each sample 14 times along the time dimension, and then concatenates and fuses it with the temporal feature vectors of the corresponding 14 time steps to form the final input of the recurrent neural network (LSTM layer).

[0067] 2. Backbone network: Long Short-Term Memory network The fused feature sequence is fed into a Long Short-Term Memory (LSTM) layer. LSTM is a special type of recurrent neural network that effectively captures long-term dependencies in time series, avoiding the vanishing or exploding gradient problems of ordinary RNNs. In this embodiment, the input size of the LSTM layer is equal to the sum of the temporal feature dimension and the embedding vector dimension, and its hidden state dimension is 96 by default. The LSTM layer processes the input sequence step by step, finally extracting the hidden state of the last time step. As a condensed encoding of the entire past 14 days of sequence information, it includes movement trends, environmental responses, and group-specific behavioral patterns.

[0068] 3. Regression Output Head and Prediction Target The final part of the model is a regression head, consisting of fully connected layers, activation functions, and randomly deactivated layers. The specific structure is: Linear(96->96) → ReLU → Dropout → Linear(96->2). Its function is to process the high-level abstract features extracted by LSTM. This is mapped to specific predicted values. The key design of this trajectory prediction model lies in its prediction objective: the output layer does not directly predict the absolute coordinates of the next day, but rather predicts a two-dimensional displacement (Δx, Δy), that is, the offset of the next day relative to the current day in the local plane coordinate system (unit: meters). This "residual learning" strategy allows the model to focus on learning the patterns of changes in the direction and distance of motion, rather than memorizing absolute positions, effectively eliminating geographical bias and enhancing the model's generalization ability and accuracy in capturing motion trends.

[0069] In summary, this model architecture differentiates different groups through clustering embedding techniques, learns their temporal dynamics using LSTM, and focuses on the motion pattern itself through displacement prediction. This design enables a single model to adaptively serve multiple elephant groups with diverse behaviors, achieving efficient, accurate, and interpretable spatiotemporal trajectory prediction.

[0070] S5. Obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

[0071] In one possible implementation, step S5, which involves obtaining the local relative plane coordinates for the next time step based on the predicted displacement, includes: Based on the predicted displacement and recent historical displacement information, the local relative plane coordinates of the next time step are obtained by fusion and correction. The fusion correction is based on a gated fusion strategy, which is used to adaptively adjust the weight of the predicted displacement and the previous day's position in the final prediction result of the trajectory prediction model according to the recent movement status of the target Asian elephant population.

[0072] In one possible implementation, the gating fusion strategy is configured as follows: Calculate the average displacement modulus of recent historical displacement information; The fusion coefficient is determined based on the ratio of the average displacement modulus to the preset static threshold. Based on the fusion coefficient, the predicted displacement of the trajectory prediction model is weighted and fused with the position of the previous day to obtain the corrected local relative planar coordinates of the next time step.

[0073] After completing forward inference, the trajectory prediction model outputs the predicted displacement, which then undergoes a post-processing correction step to generate the final geolocation prediction result. This process aims to transform the model's direct output into more robust trajectory points that better conform to physical common sense.

[0074] First, regarding the initial position calculation: Given the temporal characteristics of the target elephant flock over a continuous time window (default 14 days), the trained trajectory prediction model (the aforementioned LTMDeltaRegressor model) will output a two-dimensional vector as the predicted displacement. The preliminary predicted location is expressed as: ; in,( , ) represents the local plane coordinates for that day, , () represents the predicted coordinates for the next day.

[0075] However, purely data-driven models often produce non-physical, minute displacement predictions due to input noise when dealing with prolonged periods of stillness or extremely low-speed movement in animals; this is known as "stillness prediction drift." To address this inherent deficiency, this application introduces a gated fusion mechanism. The core idea of ​​this mechanism is to dynamically adjust the trust weights of the model's prediction results based on the recent level of activity in the Asian region, achieving an adaptive and smooth transition. The specific implementation includes the following steps: 1. Calculate recent momentum: Analyze the past momentum of this elephant herd. N Day (in a preferred embodiment) N Calculate the average displacement modulus of the daily displacement of (=10). M recent This value quantifies the animal's overall recent level of physical activity.

[0076] 2. Determine the preset static threshold: During the model training phase, calculate the static threshold in advance from the displacement distribution of all training samples. Q Quantiles (in a preferred embodiment) Q =0.75) as the reference threshold t This threshold represents a relatively high-frequency statistical value of the species' daily displacement, used to define the boundary between "active movement" and "tendency to be still." This preset stillness threshold can also be set based on ecological common sense.

[0077] 3. Generate adaptive fusion coefficients: by comparing recent momentum M recent With static threshold t Calculate a fusion coefficient between 0 and 1. α The calculation formula is as follows: , where clip is the truncation function, ensuring that the coefficients fall within the [0,1] interval. When recent physical activity is intense (… M recent ≥ t )hour, α →1 indicates the prediction result of the trajectory prediction model with complete trust; when the trajectory tends to be stationary in the near future (M recent When →0), α →0 indicates a stronger belief that the animal's location has not changed.

[0078] 4. Perform weighted fusion correction: using fusion coefficients α The preliminary predicted displacement is weighted and fused with the previous day's position to obtain the final corrected predicted coordinates. : ; in, It's the position from the previous day. It predicts the displacement.

[0079] 5. Coordinate Inverse Calculation: Calculate the corrected local plane coordinates as described above. Using the clustering anchor points corresponding to this elephant group ( Latin 0k , London 0k The coordinates are converted back to standard latitude and longitude coordinates through the inverse operation of spatial projection transformation, which serve as the final prediction output of the system.

[0080] In summary, this post-processing workflow, through an intelligent gating fusion mechanism, effectively suppresses output noise in static or low-speed scenarios, resulting in smoother and more stable predicted trajectories that better align with the behavioral characteristics of large mammals—intermittent migration and prolonged residence—significantly improving the practicality and reliability of the prediction results. The parameters involved (such as the momentum calculation window) N quantiles Q It can be adjusted and optimized according to the activity characteristics and data of specific species.

[0081] In one possible implementation, the trajectory prediction model is trained using the Huber loss function, and the displacement labels in the training data are truncated based on statistical quantiles before training. The displacement labels are the daily movement distances (Δx, Δy) of the target Asian elephant herd.

[0082] In one possible implementation, the steps for performing statistical quantile-based truncation include: Calculate the displacement magnitude corresponding to all displacement labels in the training data, and use the predetermined quantile of the displacement magnitude as the truncation threshold. The predetermined quantile of the displacement magnitude can be selected according to the actual situation, or it can be selected through cross-validation to select the predetermined quantile that optimizes the performance of the trajectory prediction model validation set.

[0083] For displacement labels whose displacement modulus exceeds the cutoff threshold, the displacement label is scaled down proportionally so that its modulus equals the cutoff threshold. Preferably, the expression for the Huber loss function is: ; in, L δ This represents Huber's loss value. y This represents the actual displacement label value. f ( x ) represents the predicted displacement of the trajectory prediction model. d These are preset threshold parameters used to define the regions of squared loss and linear loss of the error. d The initial value is set to a multiple of the standard deviation of the label data, and can also be dynamically adjusted according to the gradient stability during training. y - f ( x )∣ represents the absolute value of the prediction error.

[0084] In one possible implementation, during the training process, the trajectory prediction model dynamically weights the loss function based on the number of samples corresponding to each group identifier, with group identifiers having a smaller number of samples receiving a higher weight.

[0085] To train a robust trajectory prediction model that can adapt to the characteristics of Asian elephant data, this application designs a comprehensive training and optimization strategy that runs through the entire process from data preparation to loss calculation, aiming to effectively address real-world challenges such as GPS noise and data imbalance.

[0086] First, before model training begins, the labels in the training data are preprocessed and augmented. This step mainly involves two key operations: 1. Label Truncation: To address the extremely large abnormal displacement values ​​that may be caused by GPS drift or data errors, this application truncates all real displacement values ​​before training. The process involves cleaning the data. Specifically, the modulus of all displacements in the training set is calculated, and their statistical quantile (e.g., the 99th percentile) is taken as a physically reasonable upper limit threshold for displacement. For any displacement label exceeding this threshold, it is scaled proportionally to that threshold, thereby physically limiting the model from learning unrealistic "teleportation" patterns at the source and providing the model with clean and reasonable supervision signals.

[0087] 2. Dynamic Sample Weighting: Due to the significant differences in the observation time and frequency among different elephant groups, the training data suffers from severe class imbalance. To prevent the model from being dominated by a few dominant groups with large datasets while neglecting rarer groups, the system weights each cluster... c Calculate a weight Toilet The weight is related to the number of samples. Nc It is inversely proportional to the square root, that is During training, the loss function is weighted according to the cluster to which the samples belong, forcing the trajectory prediction model to pay equal attention to all groups during optimization, ensuring fairness and broad applicability of the predictions. Those skilled in the art will understand that the specific form of the weighting function can be adjusted according to the severity of data imbalance; the core idea is to assign higher loss weights to clusters with smaller sample sizes.

[0088] Furthermore, to address the issue of extremely limited observational data for some elephant herds (long-tailed distribution), it is possible to... The loss function is dynamically weighted.

[0089] Secondly, a robust loss function is employed in the core stage of model training. To address the unavoidable large errors in animal trajectory data, this approach abandons the mean squared error loss, which is sensitive to outliers, and instead adopts Huber loss. The mathematical expression of this loss function is as described above; its core characteristic lies in defining a threshold parameter. d When the prediction error is less than d When the error is greater than the mean squared error, its behavior is the same, and fine-tuning is performed; when the prediction error is greater than the mean squared error, it is fine-tuned. d When the penalty term increases linearly, the training process becomes less sensitive to occasional, large GPS drift errors, effectively preventing gradient explosion or instability caused by penalizing abnormally large errors, and greatly improving the training stability and generalization performance of the model in noisy environments.

[0090] In summary, the training optimization process of this application constitutes a multi-layered collaborative mechanism through pre-process label cleaning and weighting, coupled with a robust Huber loss function. It not only teaches the trajectory prediction model to learn correct motion patterns (through cleaned labels) but also ensures that all monitored objects are fully learned (through dynamic weighting), while protecting the training process from the destructive effects of dirty data (through Huber loss). This combined strategy is a key guarantee for the model to successfully learn and achieve high-precision predictions on complex real-world wildlife observation data. The truncation quantiles and Huber loss thresholds involved in this process are also relevant. d Specific parameters can be optimized through cross-validation based on the actual data distribution.

[0091] like Figure 2 As shown, in one possible implementation, this application provides a system for predicting the distribution area of ​​Asian elephants, comprising: The data preprocessing module is used to perform spatiotemporal aggregation and coordinate transformation on the raw observation data of the target Asian elephant population, and output the daily local relative planar coordinate sequence of each population identifier; the raw observation data includes discrete spatiotemporal location points and their respective population identifiers; The feature construction module is used to fuse multidimensional environmental features and time period features into the daily local relative planar coordinate sequence to construct a model input sequence containing time step feature vectors; The trajectory prediction module includes an embedding layer and a recurrent neural network; the embedding layer is used to map the input group identifiers into embedding vectors; the recurrent neural network is used to receive the model input sequence fused with the embedding vectors and output the predicted displacement for the next time step. The post-processing module is used to obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

[0092] In one possible implementation, the post-processing module is further configured to perform gated fusion correction on the predicted displacement based on recent historical displacement information.

[0093] This embodiment constructs an end-to-end automated prediction system for the distribution area of ​​Asian elephants, realizing the full-process function from raw data input to final geographic coordinate output, which is convenient for deployment and application.

[0094] The Asian elephant distribution area prediction system in this embodiment resolves the contradiction between "inefficient individual modeling for each group" and "the inability of a single model to fit heterogeneous groups" through the core technology of "cluster embedding." By using a recurrent neural network with shared parameters, coupled with learnable group embedding vectors, a unified model can accurately model and predict the behavioral patterns of multiple heterogeneous animal groups, achieving personalized predictions for each group. Through a shared-parameter LSTM model, combined with learnable group embedding vectors, it can simultaneously serve hundreds or thousands of elephant groups with diverse behaviors, achieving "cost reduction and efficiency improvement" under large-scale monitoring. This lays the core algorithmic foundation for building large-scale intelligent wildlife monitoring and early warning platforms at the national and regional levels.

[0095] Furthermore, by deeply integrating multidimensional environmental features and periodic temporal coding, the trajectory prediction model acquires ecological awareness, improving the interpretability of prediction results and the accuracy of medium- and long-term trends. By integrating multidimensional environmental features such as vegetation index, topography, and meteorology into the model input, and using sine and cosine periodic coding for time variables, the ecological driving factors of Asian elephant migration are explicitly injected into the model. This allows for the capture of the impact of environmental changes and periodic patterns on animal behavior, resulting in more accurate and reliable inferences about medium- and long-term trends.

[0096] The Asian elephant distribution area prediction system in this embodiment performs spatiotemporal aggregation processing on the raw observation data during the data input stage and introduces existence markers to distinguish between real data and imputed values. During the model training stage, it comprehensively utilizes the Huber loss function, a statistical distribution-based label truncation mechanism, and dynamic weighting strategies tailored to different population groups to effectively suppress outlier interference and balance the learning objective. In the result output stage, a gating fusion mechanism is used to adaptively smooth and calibrate the predicted trajectory. This consistent collaborative design significantly enhances the system's adaptability to complex field data, ensuring that the output trajectory is not only smooth and stable but also conforms to the physical laws of animal behavior, thus possessing the reliability and practicality to support actual field monitoring and conservation management decisions.

[0097] In one possible implementation, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the steps of any of the methods described above. It also achieves the following functions: Through the core technology of "cluster embedding," it resolves the contradiction between "inefficient modeling for each group" and "the inability of a single model to fit heterogeneous groups." Through a recurrent neural network with shared parameters, combined with learnable group embedding vectors, it achieves accurate modeling and prediction of behavioral patterns of diverse animal groups using a unified model, realizing personalized predictions for "a thousand groups, a thousand faces." By deeply integrating multidimensional environmental features and periodic temporal coding, the trajectory prediction model possesses ecological perception capabilities, improving the interpretability of prediction results and the accuracy of medium- and long-term trends. By integrating multidimensional environmental features such as vegetation index, topography, and meteorology into the model input, and performing sine and cosine periodic coding on time variables, it explicitly injects the ecological driving factors of Asian elephant migration into the model, capturing the impact of environmental changes and periodic patterns on animal behavior, making medium- and long-term trend inferences more accurate and reliable. By performing spatiotemporal aggregation on the raw observation data during the data input stage and introducing existence markers to distinguish between real data and imputed values, and by comprehensively utilizing the Huber loss function, a label truncation mechanism based on statistical distribution, and dynamic weighting strategies tailored to different populations during the model training stage, outlier interference and the learning objective are effectively suppressed. In the output stage, a gating fusion mechanism is used to adaptively smooth and calibrate the predicted trajectory. This consistent collaborative design significantly enhances the system's adaptability to complex field data, ensuring that its output trajectory is not only smooth and stable but also conforms to the physical laws of animal behavior, thus possessing the reliability and practicality to support actual field monitoring and conservation management decisions.

[0098] The execution entity of the method described in this application embodiment can be an intelligent execution module composed of software, hardware, or a combination thereof. This execution module can receive relevant data input via wired communication, wireless communication, or a combination thereof, and can output control commands as needed. Furthermore, the execution module may also possess certain data processing and storage capabilities. The execution module can be used to control multiple target devices, including but not limited to physical servers and cloud servers deployed in remote locations and their running software systems, or host devices or servers deployed locally and their supporting software, for implementing operational control of devices at specific locations. In some application scenarios, the execution module can also be used to manage multiple data storage devices. These storage devices can be located in the same geographical location as the controlled devices or deployed in different regions to achieve flexible and efficient data access and management.

[0099] In one possible implementation, a computer-readable storage medium is provided, including a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described above. It achieves the following functions: Through the core technology of "cluster embedding," it resolves the contradiction between "inefficient modeling for each group" and "the inability of a single model to fit heterogeneous groups." Through a recurrent neural network with shared parameters, coupled with learnable group embedding vectors, it achieves accurate modeling and prediction of behavioral patterns in diverse animal groups using a unified model, realizing personalized predictions for each group. By deeply integrating multidimensional environmental features and periodic temporal coding, the trajectory prediction model possesses ecological awareness, improving the interpretability of prediction results and the accuracy of medium- and long-term trends. By integrating multidimensional environmental features such as vegetation index, topography, and meteorology into the model input, and performing sine and cosine periodic coding on time variables, it explicitly injects the ecological driving factors of Asian elephant migration into the model, capturing the impact of environmental changes and periodic patterns on animal behavior, making medium- and long-term trend inferences more accurate and reliable. By performing spatiotemporal aggregation on the raw observation data during the data input stage and introducing existence markers to distinguish between real data and imputed values, and by comprehensively utilizing the Huber loss function, a label truncation mechanism based on statistical distribution, and dynamic weighting strategies tailored to different populations during the model training stage, outlier interference and the learning objective are effectively suppressed. In the output stage, a gating fusion mechanism is used to adaptively smooth and calibrate the predicted trajectory. This consistent collaborative design significantly enhances the system's adaptability to complex field data, ensuring that its output trajectory is not only smooth and stable but also conforms to the physical laws of animal behavior, thus possessing the reliability and practicality to support actual field monitoring and conservation management decisions.

[0100] The computer program product can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0101] It should be noted that the above detailed description of this application in conjunction with the accompanying drawings and embodiments is intended to clearly and completely illustrate the technical solution and preferred implementation of this application, rather than to limit the scope of protection of this application. Under the guidance of the core concept of this application, those skilled in the art can make various modifications, simplifications or equivalent substitutions to the technical solution without departing from the spirit and scope of this application.

[0102] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0103] The foregoing has described specific embodiments of the present application. In some cases, the described actions or steps may be performed in a different order than those shown in the embodiments and the desired results may still be achieved. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0104] In the description of the embodiments of this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments of this application. In the embodiments of this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of this application, as well as the features of different embodiments or examples.

[0105] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0106] The above embodiments are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the distribution area of ​​Asian elephants, characterized in that, include: S1. Spatiotemporal aggregation processing is performed on the raw observation data of the target Asian elephant population to generate a daily representative coordinate sequence in time order; The original observation data includes discrete spatiotemporal location points and their respective group identifiers; S2. Based on the reference origin corresponding to each group identifier, convert the geographical location coordinates in the daily representative coordinate sequence into local relative planar coordinates; S3. Based on the daily representative coordinate sequence, the local relative plane coordinates, and the associated multidimensional environmental features, construct a model input sequence containing time step feature vectors; S4. Input the model input sequence and the group identifier into the trained trajectory prediction model to obtain the predicted displacement of the next time step; wherein, the trajectory prediction model includes an embedding layer and a recurrent neural network; the embedding layer is used to map the group identifier into an embedding vector and fuse it with the time step feature vector to form the input of the recurrent neural network; S5. Obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

2. The method according to claim 1, characterized in that, The steps for spatiotemporal aggregation processing in step S1 include: calculating the average spatial dispersion of each group identifier in the daily set of observation points; If the average spatial dispersion is less than or equal to a preset threshold, the weighted average centroid is calculated based on the weight of each observation point and used as the representative coordinate for that day. If the average spatial dispersion is greater than the preset threshold, a weighted center point is selected from the set of observation points as the representative coordinate of the day. The weighted center point is the observation point that minimizes the sum of the weighted distances to all other observation points.

3. The method according to claim 1, characterized in that, The multidimensional environmental features in step S3 include at least one of vegetation index, topographic features, and meteorological data; The time step feature vector includes sine and cosine periodic encoded features calculated based on the date.

4. The method according to claim 1, characterized in that, In step S4, the embedding vector output by the embedding layer of the trajectory prediction model is copied in the time dimension and concatenated with the time step feature vector of each time step to form the input of the recurrent neural network.

5. The method according to claim 4, characterized in that, The time step feature vector constructed in step S3 includes a binary existence marker, which is used to indicate whether the data for that day is a true observation.

6. The method according to claim 5, characterized in that, The trajectory prediction model is trained using the Huber loss function, and before training, the displacement labels in the training data are truncated based on statistical quantiles. The displacement labels are the daily movement distances (Δx, Δy) of the target Asian elephant herd.

7. The method according to claim 6, characterized in that, The steps for performing truncation based on statistical quantiles include: Calculate the displacement modulus corresponding to all displacement labels in the training data, and use the predetermined quantile of the displacement modulus as the truncation threshold. For displacement labels whose displacement modulus exceeds the cutoff threshold, the displacement label is scaled down proportionally so that its modulus equals the cutoff threshold.

8. The method according to claim 6, characterized in that, The expression for the Huber loss function is: ； in, L δ This represents Huber's loss value. y This represents the actual displacement label value. f ( x ) represents the predicted displacement of the trajectory prediction model. δ The preset threshold parameter is used to define the region of squared loss and the region of linear loss. y - f ( x )∣ represents the absolute value of the prediction error.

9. The method according to any one of claims 1-8, characterized in that, During the training process, the trajectory prediction model dynamically weights the loss function based on the number of samples corresponding to each group identifier, with group identifiers with fewer samples having higher weights.

10. The method according to any one of claims 1-8, characterized in that, Step S5, which involves obtaining the local relative plane coordinates for the next time step based on the predicted displacement, includes: Based on the predicted displacement and recent historical displacement information, the local relative plane coordinates of the next time step are obtained by fusion and correction. The fusion correction is based on a gated fusion strategy, which is used to adaptively adjust the weight of the predicted displacement and the previous day's position in the final prediction result of the trajectory prediction model according to the recent movement status of the target Asian elephant population.

11. The method according to claim 10, characterized in that, The gating fusion strategy is configured as follows: Calculate the average displacement modulus of recent historical displacement information; The fusion coefficient is determined based on the ratio of the average displacement modulus to the preset static threshold. Based on the fusion coefficient, the predicted displacement of the trajectory prediction model is weighted and fused with the position of the previous day to obtain the corrected local relative planar coordinates of the next time step.

12. A system for predicting the distribution area of ​​Asian elephants, characterized in that, include: The data preprocessing module is used to perform spatiotemporal aggregation and coordinate transformation on the raw observation data of the target Asian elephant population, and output the daily local relative planar coordinate sequence of each population identifier; the raw observation data includes discrete spatiotemporal location points and their respective population identifiers; The feature construction module is used to fuse multidimensional environmental features and time period features into the daily local relative planar coordinate sequence to construct a model input sequence containing time step feature vectors; The trajectory prediction module includes an embedding layer and a recurrent neural network; the embedding layer is used to map the input group identifiers into embedding vectors; the recurrent neural network is used to receive the model input sequence fused with the embedding vectors and output the predicted displacement for the next time step. The post-processing module is used to obtain the local relative planar coordinates of the next time step based on the predicted displacement, and then convert them back into geographical location coordinates for output.

13. The system according to claim 12, characterized in that, The post-processing module is also used to perform gated fusion correction on the predicted displacement based on recent historical displacement information.

14. A computer device comprising a memory and a processor, the memory being used to store computer programs, characterized in that, The processor is used to execute the computer program to implement the steps of the method according to any one of claims 1 to 11.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer or processor, causes the computer or processor to perform the steps of the method according to any one of claims 1 to 11.

Images