Wildlife habitat prediction method and system based on mask representation and density model
By using masked representation and density model, and leveraging the JEPA self-supervised learning framework and normalized flow model, the problem of insufficient generalization ability and prediction accuracy of species distribution models in existing technologies is solved, thus achieving efficient habitat suitability assessment and conservation decision support.
Patent Information
- Application Number
- CN202610327293.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-16
AI Technical Summary
Existing species distribution models have low generalization ability and prediction accuracy when dealing with complex bioclimatic characteristics that are high-dimensional and nonlinear. Furthermore, their reliance on spurious negative samples for modeling leads to significant biases in the prediction results, making it difficult to meet the needs of refined management.
We employ masked representation and density model, train the context encoder and preference environment density model through the JEPA self-supervised learning framework, and use maximum likelihood estimation and normalized flow model to learn the stable combination law between environmental variables, thus avoiding the use of spurious negative samples.
It improves the accuracy of habitat suitability assessment in the absence of precise non-distribution points, reduces sample label uncertainty, achieves unified comparability of scores across the entire region and continuous spatial distribution of suitability, and supports conservation decisions in large-scale areas.
Smart Images
Figure CN122220982A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wildlife habitat prediction technology, and in particular to a method and system for wildlife habitat prediction based on mask representation and density model. Background Technology
[0002] With the intensification of global climate change and human activities, biodiversity loss has become a major ecological and environmental problem urgently needing to be addressed worldwide, posing a severe challenge to regional ecological security and sustainable social development. Wildlife, as an important component of ecosystems, directly reflects the health of the regional environment through its population status. However, due to the highly concealed and mobile nature of wildlife activities, coupled with the high time costs and complex geographical conditions limiting field surveys, there remains considerable uncertainty regarding the spatiotemporal distribution patterns of rare and endangered species and their potential suitable habitat ranges. Therefore, accurately identifying and assessing potential wildlife habitats has become a crucial and challenging task in national park construction and biodiversity conservation management.
[0003] Species distribution models, as a quantitative assessment tool, play a crucial role in evaluating wildlife habitat suitability. In recent years, with the development of ecoinformatics, statistical methods, represented by the maximum entropy model, have been widely applied in habitat simulation. However, existing technologies still have significant limitations in practical applications: On the one hand, wildlife survey data typically consists of distribution point data, lacking precise non-distribution points. Traditional methods often simply remove existing data points from the study area and randomly select points as spurious negative samples, ignoring potential undiscovered suitable areas hidden within unrecorded species distribution areas, leading to bias in model training samples. On the other hand, single traditional statistical models often struggle to meet the needs of refined management in terms of generalization ability and prediction accuracy when dealing with complex, high-dimensional, nonlinear bioclimatic characteristics. Therefore, it is urgent to introduce algorithms with stronger data mining capabilities and develop optimization strategies for unlabeled samples to improve the scientific rigor and robustness of habitat assessment.
[0004] Therefore, this application proposes a method and system for predicting wildlife habitats based on mask representation and density model to solve the technical problems existing in the prior art. Summary of the Invention
[0005] This invention proposes a method and system for predicting wildlife habitats based on mask representation and density model, in order to solve the problems that existing species distribution models have large prediction biases due to their reliance on pseudo-negative samples for modeling, and that their generalization ability and prediction accuracy are low when dealing with complex bioclimatic features of high dimension and nonlinearity, making it difficult to meet the requirements of refined management.
[0006] The present invention achieves the above objectives through the following technical solutions:
[0007] This invention relates to a wildlife habitat prediction method based on mask representation and density model, comprising:
[0008] Obtain field survey data of the target species, and obtain corresponding multidimensional environmental variable data based on the field survey data of the target species;
[0009] The multidimensional environmental variable data is input into the trained environmental representation extraction model, and the output is an environmental representation vector. The environmental representation extraction model is a context encoder trained based on the multidimensional environmental variable training dataset and the JEPA self-supervised learning framework.
[0010] The environmental representation vector is input into the trained preference environment density model, and the output is the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples.
[0011] The probability score or logarithmic probability score is graded according to a preset threshold, and the suitability level is output.
[0012] Furthermore, based on the multidimensional environmental variable training dataset, a context encoder is trained using the JEPA self-supervised learning framework, including:
[0013] Based on the field survey data of the target species, the corresponding multidimensional environmental variable training dataset is extracted from the preset environmental variable layer or raster data.
[0014] Missing value processing and scaling uniformity processing are performed on the multidimensional environmental variable training dataset to obtain the environmental variable input vector;
[0015] Construct the environment variable input vector into context input and target input;
[0016] The JEPA self-supervised learning framework is constructed, which includes a context encoder, an object encoder, and a predictor.
[0017] According to the preset masking strategy, some variable dimensions in the context input are masked and mask markers are generated to obtain the masked context input and mask marker vectors. The masked context input and mask marker vectors are input into the context encoder to output the context environment representation. The context environment representation is input into the predictor to output the environment representation.
[0018] The target input is fed to the target encoder, which outputs a representation of the target environment.
[0019] The alignment loss between the environment representation and the target environment representation is calculated, and the context encoder is updated based on the backpropagation of the alignment loss to obtain the trained context encoder.
[0020] Furthermore, both the context encoder and the target encoder include an input processing layer, a feature fusion layer, and a representation output layer. The input processing layer is used to organize the environmental variable input vector and process the mask label. The feature fusion layer is used to fuse multivariate relationships. The representation output layer is used to output a fixed-length environmental representation vector.
[0021] Furthermore, using the environmental representation vector training dataset as positive samples, a normalized flow model is trained based on maximum likelihood estimation, including:
[0022] Input the multidimensional environment variable training dataset into the trained context encoder, and output the environment representation vector training dataset.
[0023] Construct a normalized flux density model;
[0024] The training dataset of environmental representation vectors is used as positive samples to input into the normalized flow density model. The model parameters of the normalized flow density model are updated by maximizing the log-likelihood of the positive samples under the normalized flow density model, thus obtaining the trained normalized flow model.
[0025] Furthermore, the normalized flux density model includes an input layer, a transformation block stack layer, a basic distribution layer, and a scoring output layer. The transformation block stack layer is composed of multiple transformation blocks connected in series. Each transformation block includes a dimension rearrangement unit, a parameter network, and a transformation calculation unit. The dimension rearrangement unit is used to change the order of vector dimensions, the parameter network is used to generate transformation parameters, and the transformation calculation unit is used to complete the reversible transformation of this layer based on the parameters and output the result.
[0026] Furthermore, the probability score or logarithmic probability score is graded according to a preset threshold, and the suitability level is output, including:
[0027] The probability score or log probability score is backfilled into the spatial grid of the target area to form a spatial distribution map of habitat suitability;
[0028] The scores are normalized and mapped to form habitat suitability scores in the [0,1] interval;
[0029] The habitat suitability score is classified according to a preset threshold, and the classification results are used as the habitat suitability output.
[0030] Furthermore, data on the occurrence points of the target species in the field survey are obtained, and corresponding multidimensional environmental variable data are obtained based on the data on the occurrence points of the target species in the field survey, including:
[0031] Collect occurrence point data of the target species, including field transect survey data, infrared camera monitoring data, location tracking data and / or historical data;
[0032] Perform spatial coordinate unification, duplicate point removal, and outlier point removal on the point data;
[0033] Based on the spatial location of the point data, the corresponding multidimensional environmental variable data is extracted from the preset environmental variable layer or raster data. The multidimensional environmental variable data includes topographic variables, climate variables, vegetation variables and / or human disturbance variables.
[0034] The multidimensional environmental variable data is processed for missing values and scale unification to form an environmental variable input vector.
[0035] The present invention also provides a system for the aforementioned wildlife habitat prediction method based on mask representation and density model, comprising:
[0036] The acquisition module is used to acquire field survey data of the target species and acquire corresponding multidimensional environmental variable data based on the field survey data of the target species.
[0037] The extraction module is used to input multidimensional environmental variable data into a trained environmental representation extraction model and output an environmental representation vector. The environmental representation extraction model is a context encoder trained based on the JEPA self-supervised learning framework using a multidimensional environmental variable training dataset.
[0038] The scoring module is used to input the environmental representation vector into the trained preference environment density model and output the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples.
[0039] The scoring output module is used to classify probability scores or logarithmic probability scores according to preset thresholds and output the suitability level.
[0040] The beneficial effects of this invention are as follows:
[0041] 1) In the absence of exact non-distribution points, habitat suitability can be inferred without constructing pseudo-negative samples, thereby reducing the systematic bias caused by sample label uncertainty;
[0042] 2) Through predictive mask representation learning, the model can learn stable combination patterns between environmental variables, obtain environmental representations that are consistent with missing and perturbations, and improve robustness under complex nonlinear environmental conditions.
[0043] 3) By modeling with normalized flow density, the probability distribution of the target species’ preferred environment is learned only from the occurrence points, so that the global score has a uniform and comparable scale, which is convenient for generating a continuous spatial distribution of suitability and supporting hierarchical management.
[0044] 4) This method can be used for habitat suitability assessment and conservation decision-making in large-scale areas, providing technical support for habitat identification, conservation planning and management optimization for key species. Attached Figure Description
[0045] Figure 1 This is a flowchart of the wildlife habitat prediction method in the embodiments of this application;
[0046] Figure 2 This is a schematic diagram of the process of training a context encoder based on the JEPA self-supervised learning framework in the embodiments of this application;
[0047] Figure 3 This is a schematic diagram illustrating the process of the preferred environment density model processing the environment representation vector in an embodiment of this application. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0049] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. Here, JEPA stands for Joint Embedding Predictive Architecture, and is abbreviated as JEPA.
[0050] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0051] like Figure 1 As shown, the method and system for predicting wildlife habitats based on mask representation and density model are characterized by comprising:
[0052] S1: Obtain field survey data of the target species, and obtain corresponding multidimensional environmental variable data based on the field survey data of the target species;
[0053] S2: Input the multidimensional environmental variable data into the trained environmental representation extraction model and output the environmental representation vector. The environmental representation extraction model is a context encoder trained based on the multidimensional environmental variable training dataset and the JEPA self-supervised learning framework.
[0054] S3: Input the environmental representation vector into the trained preference environment density model, and output the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples.
[0055] S4: Classify the probability score or logarithmic probability score according to the preset threshold and output the suitability level.
[0056] In one embodiment, step S1 includes:
[0057] Step A1: Collect occurrence point data of the target species, including field transect survey data, infrared camera monitoring data, location tracking data and / or historical data;
[0058] The occurrence point data consists of two parts: the first part is the acquisition of global environmental variables and spatial rasterization processing, which divides the study area into a 1km×1km spatial grid and extracts the corresponding multidimensional environmental variable data for each grid unit to form a grid-level environmental feature record; the second part is the acquisition of occurrence point data of the target species, which is used to obtain the exact occurrence point location of the target species and complete coordinate unification and quality control.
[0059] Step A2: Perform spatial coordinate unification, duplicate point removal, and outlier point removal on the occurrence point data. Use the Pearson correlation coefficient to evaluate the correlation between environmental variables, remove highly correlated redundant variables, and obtain a set of effective environmental variables for modeling.
[0060] Step A3: Extract the corresponding multidimensional environmental variable data from the preset environmental variable layer or raster data according to the spatial location of the point of occurrence. The multidimensional environmental variable data includes topographic variables, climate variables, vegetation variables and / or human disturbance variables.
[0061] Step A4: Perform missing value processing and scaling on the multidimensional environmental variable data to form an environmental variable input vector.
[0062] In one embodiment, training the context encoder based on the JEPA self-supervised learning framework in step S2 includes:
[0063] Step B1: Construct the context input and target input based on the environment variable input vector obtained in Step A;
[0064] Step B2: Mask some variable dimensions of the context input according to a preset masking strategy and generate mask markers to obtain the masked context input;
[0065] Step B3: The masked context input is fed to the context encoder to obtain the latent context representation;
[0066] Step B4: The target input is fed to the target encoder to obtain the target latent representation, and the target latent representation is predicted by the predictor based on the context latent representation to achieve latent space alignment and obtain the environment representation vector.
[0067] In one embodiment, both the context encoder and the target encoder include an input processing layer, a feature fusion layer, and a representation output layer. The input processing layer is used to organize the input of environmental variables and process the mask markers. The feature fusion layer is used to fuse multivariate relationships. The representation output layer is used to output a fixed-length environmental representation vector.
[0068] In one embodiment, the parameters of the target encoder are updated by the parameters of the context encoder using a slow follow-up update mechanism to keep the target representation output stable and reduce training fluctuations.
[0069] In one embodiment, step S3, training the normalized flow model based on maximum likelihood estimation, includes:
[0070] Step C1: Construct a normalized flux density model, using the standard normal distribution as the base distribution;
[0071] Step C2: Input the environmental representation vector corresponding to the occurrence point as a training sample into the normalized flow density model for training, so as to learn the probability distribution of the target species’ preferred environment in the environmental representation space;
[0072] Step C3: After training, obtain the target species' preferred environment density model, and use the density model as the preferred environment density model for subsequent score inference.
[0073] In one embodiment, the normalized flow density model includes an input layer, a transform block stack layer, a base distribution layer, and a scoring output layer. The transform block stack layer is composed of multiple transform blocks connected in series and is used to progressively map the input environment representation vector to the base distribution space. The scoring output layer is used to output a probability score or a log probability score.
[0074] In one embodiment, each transformation block includes a dimension rearrangement unit, a parameter network, and a transformation calculation unit. The parameter network is used to generate parameters for transforming the remaining dimensions based on a portion of the dimensions of the input vector. The transformation calculation unit is used to complete the transformation of this layer based on the parameters and output the result. The dimension rearrangement unit is used to change the order of vector dimensions between adjacent transformation blocks to enhance the fitting ability to complex distribution patterns.
[0075] In one embodiment, step S4 includes mapping the probability score or log probability score back to a spatial grid to form a habitat suitability distribution map, and classifying the suitability score according to a preset threshold to output high suitability, medium suitability and low suitability levels.
[0076] In one embodiment, the present invention also provides a method and system for predicting wildlife habitats based on mask representation and density models, comprising:
[0077] The acquisition module is used to acquire field survey data of the target species and acquire corresponding multidimensional environmental variable data based on the field survey data of the target species.
[0078] The extraction module is used to input multidimensional environmental variable data into a trained environmental representation extraction model and output an environmental representation vector. The environmental representation extraction model is a context encoder trained based on the JEPA self-supervised learning framework using a multidimensional environmental variable training dataset.
[0079] The scoring module is used to input the environmental representation vector into the trained preference environment density model and output the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples.
[0080] The scoring output module is used to classify probability scores or logarithmic probability scores according to preset thresholds and output the suitability level.
[0081] In one embodiment, step S2, training the context encoder based on the JEPA self-supervised learning framework, specifically includes:
[0082] Step B1: Construct representation learning training samples. Establish a training sample set for the environmental variable input vectors formed in Step A, with each sample corresponding to an environmental variable input vector at a spatial location. For each sample, construct a context input and a target input, where the target input uses the complete environmental variable input vector, and the context input uses an input view obtained by applying a preset perturbation to the target input. The perturbation includes at least one of adding small noise to some variable dimensions, randomly discarding a small number of dimensions, and / or randomly scaling the variables, to simulate environmental variable measurement errors and missing data.
[0083] Step B2: Generate a predictive masking task. Randomly mask the context input along the variable dimension according to a preset masking ratio to obtain the masked context input; simultaneously, generate a masking marker vector, which is consistent with the environment variable dimension and is used to indicate the position of the masked dimension; the masked variable values can be replaced with preset mask values or learnable masking vectors to avoid mistaking missing values for true values.
[0084] Step B3: Establish context representation and target representation. The masked context input and the masked marker vector are input into the context encoder to obtain the context latent representation; the target input is fed into the target encoder to obtain the target latent representation. The context encoder and target encoder use the same encoding structure, both including an input processing layer, a feature fusion layer, and a representation output layer. The input processing layer organizes the environmental variable input and processes the masked markers; the feature fusion layer fuses multivariate relationships and extracts combined features; and the representation output layer outputs a fixed-length latent representation.
[0085] Step B4: Achieve latent space alignment and obtain the environment representation extraction model. Input the context latent representation into the predictor and output the predicted representation. Use the consistency between the predicted representation and the target latent representation in a unified representation space as a training constraint to complete the latent space alignment. After training, fix the context encoder as the environment representation extraction model, and use the environment representation extraction model to output the environment representation vector for any spatial location as an input vector of environmental variables.
[0086] Preferably, the training stabilization strategy in step B includes: the parameters of the target encoder are updated by the parameters of the context encoder using a slow follow-up update mechanism to keep the target latent representation output stable and reduce training fluctuations, thereby providing a smooth alignment reference for the predictor.
[0087] In one embodiment, step S3, training the normalized flow model based on maximum likelihood estimation, specifically includes:
[0088] Step C1: Construct preference distribution modeling samples. Extract the set of occurrence points from Step A, and use the environmental representation extraction model trained in Step B to infer the environmental variable input vector corresponding to each occurrence point, obtaining the set of environmental representation vectors corresponding to the occurrence points, which serves as the positive sample set for normalized flow density modeling.
[0089] Step C2: Construct a normalized flux density model. A normalized flux density model is constructed, comprising an input layer, a stacked transform block layer, a base distribution layer, and a scoring output layer, with a standard normal distribution set as the base distribution. The stacked transform block layer consists of multiple transform blocks connected in series. Each transform block includes a dimension rearrangement unit, a parameter network, and a transform calculation unit. The dimension rearrangement unit changes the order of vector dimensions, the parameter network generates transform parameters, and the transform calculation unit performs the reversible transform of this layer based on the parameters and outputs the result.
[0090] Step C3: Train the preference environment density model. Input the positive sample set into the normalized flow density model, and update the model parameters by maximizing the log-likelihood of the positive samples under the normalized flow density model to learn the probability distribution of the target species' preferred environment in the environmental representation space; after training, obtain the target species' preference environment density model, and use the density model as the preference environment density model for subsequent rating inference.
[0091] In one embodiment, step D specifically includes:
[0092] Step D1: Generate global environmental representation vectors. Extract corresponding multidimensional environmental variable data for each spatial location within the target area and form environmental variable input vectors; input the environmental variable input vectors into the environmental representation extraction model trained in step B, and output the environmental representation vectors for each spatial location.
[0093] Step D2: Perform suitability score inference. Input the environmental representation vector of each spatial location into the preferred environment density model, and output a probability score or a log probability score; wherein, the probability score or log probability score is used to characterize the degree of matching between the environmental representation of the spatial location and the target species preference distribution.
[0094] Step D3: Output spatial distribution results. The probability score or log probability score is backfilled into the spatial grid of the target area to form a spatial distribution map of habitat suitability; further, the score can be normalized and mapped to form a habitat suitability score in the [0,1] interval.
[0095] Step D4: Graded output of habitat suitability levels. Based on preset thresholds, the habitat suitability scores are graded into high suitability, medium suitability, and low suitability levels, and the grading results are used as the habitat suitability output results.
[0096] The above processes will be explained in detail below with reference to the accompanying drawings.
[0097] This invention provides a wildlife habitat suitability prediction system based on predictive mask representation learning and normalized flux density modeling, comprising:
[0098] Figure 2 This is a schematic diagram of the environment representation learning process based on the JEPA joint embedding architecture provided by the present invention. It is used to accurately illustrate how the input environment variables are masked, encoded, and predicted and aligned to finally form the internal data flow and hierarchical structure of the environment representation vector.
[0099] Input and mask construction: such as Figure 2 As shown, the system uses a certain area within the study region. Grid cells For example, multidimensional environmental variables are extracted from the grid and concatenated in a fixed order to form the input vector x. x can be composed of topographic variables (elevation, slope, etc.), climate variables (average annual temperature, annual precipitation, etc.), vegetation variables (NDVI, etc.), and anthropogenic interference variables (road density, nighttime light intensity, etc.). To construct the predictive masking task, the system uses the complete input x as the target input and simultaneously performs occlusion on a partial set of dimensions m of x to obtain the context input. The occlusion method can be as follows: replacing the occluded dimension with a fixed mask value or a learnable mask vector. Simultaneously, a mask marker vector m is generated, such that m has the same dimension as x, and satisfies the following condition: if the k-th dimension is occluded... ,otherwise In this way, the context path simultaneously obtains explicit indications at the input end of both "observable variable values" and "which dimensions are missing".
[0100] The hierarchical structure and data flow of a context encoder: A context encoder is used to extract data from... The environmental characterization that remains stable under missing conditions is extracted and can be summarized as a three-layer sequential process.
[0101] (1) Input processing layer: The model is formed by taking m as input, for example, by using a splicing method. Or map m to... Same-dimensional tag embedding and They are added together to enable the model to distinguish between "true low values" and "occluded placeholder values".
[0102] (2) Feature fusion layer: The multivariate relationship is fused and modeled to learn the combined features between variables.
[0103] (3) Representation output layer: Maps the fused hidden features to a fixed-length latent representation. And can be used for Normalization is performed to stabilize the training scale. The final result is... It represents the "environmental representation vector inferred under occlusion conditions".
[0104] Hierarchical structure and output of the target encoder: The target encoder receives the complete target input x and outputs the latent representation of the target. Its network structure is consistent with that of the context encoder, but its input does not include missing information caused by occlusion, therefore It serves as a more complete reference representation to guide contextual path learning.
[0105] The role of the predictor and the alignment path: The predictor receives... It outputs a predicted representation p, such that p is as close as possible to... During the training phase, alignment constraints are used to make p and... Consistency within the same representation space allows the ability to recover occluded information to be pushed into the representations learned by the context encoder.
[0106] Target-side stabilization mechanism and training loop: To avoid alignment instability caused by frequent target reference drift during training, the target encoder parameters are updated slowly by the context encoder parameters during the training phase; simultaneously, an alignment loss is constructed based on the similarity between the predicted representation and the target representation, for example, using...
[0107] The parameters of the context encoder and predictor are updated through backpropagation, thereby forming a stable alignment training loop.
[0108] Output and use of environment representation: After training, the context encoder is fixed as the environment representation extraction model. During the inference phase, the environment variable vector (organized as x according to the same variable order) is input to any grid cell and set... (Indicates no mask), the context encoder outputs the environmental representation vector of the mesh, which is then used as input to the subsequent density modeling module and suitability score inference module.
[0109] To address the inherent challenges in wildlife habitat suitability prediction tasks, this invention, after completing environmental representation learning, also needs to solve the problems of how to characterize preference distribution using only occurrence points, and how to transform preference distribution into a globally interpretable suitability score, as shown in the appendix. Figure 3 The principles of normalized flow density modeling and suitability output are explained.
[0110] Density modeling layer: as attached Figure 3As shown, the normalized flow density model is used to learn the probability density of the occurrence point environment representation in the representation space. It consists of an input layer, a reversible transformation stacked layer, a basic distribution layer, and a scoring output layer. The input layer receives the environment representation vector output by the environment representation learning module; the basic distribution layer is set as the standard normal distribution as the reference distribution; the reversible transformation stacked layer is composed of multiple transformation blocks connected in series, which is used to gradually transform the complex preference distribution to the basic distribution space, so that the model can not only express complex distribution patterns, but also provide a computable evaluation of density.
[0111] Reversible Transformation Layer: Appendix Figure 3 The internal structure of the transform block is further illustrated. Each transform block includes a dimension rearrangement unit, a parameter network, and a transform computation unit.
[0112] (1) Dimension rearrangement unit: The vector dimension order is rearranged between adjacent transformation blocks to fully mix different dimensions in multi-level transformation;
[0113] (2) Parameter network: Based on a portion of the dimensions of the input vector, it generates the parameters required to transform the remaining dimensions;
[0114] (3) Transformation calculation unit: Performs an invertible transformation on the input vector according to the parameters and outputs the transformation result.
[0115] By cascading multiple transform blocks, the model can progressively characterize the nonlinear form of the target species' preferred environment in the representation space. For example, it can simultaneously express joint preferences for variables such as temperature, precipitation, vegetation cover, and human disturbance, rather than simply making a threshold judgment on a single variable.
[0116] The scoring output layer of the normalized flow density model outputs either a probability score or a log-probability score. Its computational logic comprises two parts: one part is the probability mapped to the base distribution, and the other part is a correction term for volume changes during the invertible transformation. Through the combined effect of these two parts, the model can provide a consistent density score for any input environment representation. A higher score indicates that the environment representation is closer to the center of the preference distribution formed by the point samples, while a lower score indicates a greater deviation from the preference distribution.
[0117] During the training phase, the system first fixes the environment representation extraction model, then uses the environment representation vectors corresponding to the occurrence points as training samples to input the normalized flow density model. The density model parameters are updated by maximizing the log-likelihood of these training samples under the model. After training, a preferred environment density model is obtained, which serves as the core scorer for subsequent suitability inference.
[0118] During the inference phase, for each 1km×1km grid cell within the study area, the system first extracts the corresponding environmental variable data and generates an environmental variable input vector. Then, the environmental representation extraction model outputs the environmental representation vector for that grid. Subsequently, this environmental representation vector is input into the preferred environmental density model to obtain the probability score or log probability score for that grid. The system backfills the scores of all grids into the spatial grid, forming a continuous habitat suitability distribution map. Furthermore, a unified scale mapping can be applied to the scores to make the scores from different regions or species comparable. Suitability scores are also graded according to preset thresholds, outputting high suitability, moderate suitability, and low suitability levels, facilitating direct use in scenarios such as patrol deployment, habitat restoration, and protected area boundary optimization.
[0119] This invention provides a method and system for predicting wildlife habitat suitability based on predictive mask representation learning and normalized flow density modeling. Unlike traditional species distribution models that rely on binary classification and pseudo-negative sample construction, this invention overcomes the problems of high noise from pseudo-negative samples and significant boundary learning bias when definite non-distribution points are lacking. This invention does not directly treat unlabeled locations within the study area as negative samples for training. Instead, through the core mechanisms of "predictive mask representation learning" and "preference distribution density modeling," the model learns stable environmental representations from environmental variables across the entire region and learns its preferred environmental distribution based solely on the environmental representations of the exact locations where the target species appears. This allows for continuous scoring and spatial mapping of potential suitable habitats within the study area without the need for pseudo-negative sample labeling.
[0120] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, semiconductor memory, optical memory, etc.) containing computer-usable program code.
[0121] This invention is described with reference to flowchart illustrations and / or system architecture diagrams of a method, system, and computer program product for predicting wildlife habitat suitability based on predictive mask representation learning and normalized flow density modeling, according to embodiments of the invention. It will be understood that each process and / or module in the flowchart illustrations and / or architecture diagrams, as well as combinations of processes and / or modules in the flowchart illustrations and / or architecture diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, server, or artificial intelligence computing platform, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process... Figure 1 One or more processes and / or architectures Figure 1 A device for a function specified in one or more modules.
[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer, server, or dedicated artificial intelligence computing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or architectures Figure 1 The specified functions are defined in one or more modules. Specifically, the instruction device can realize functions such as environmental variable data acquisition and processing, environmental characterization extraction, normalized flux density modeling, and habitat suitability score inference.
[0123] These computer program instructions can also be loaded onto a computer, server cluster, or programmable data processing device in the cloud, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or architectures Figure 1 The steps of the functions specified in one or more modules. These steps include, but are not limited to: spatially rasterizing the study area and extracting environmental variable data, performing predictive mask representation learning on the environmental variable input vector to generate environmental representation vectors, training a normalized flow density model based only on the environmental representation vectors of occurrence points to obtain a preferred environmental density model, and probabilistically scoring the global grid environmental representation vectors and outputting the spatial distribution results of habitat suitability.
[0124] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method and system for predicting wildlife habitats based on mask representation and density model, characterized in that, include: Obtain field survey data of the target species, and obtain corresponding multidimensional environmental variable data based on the field survey data of the target species; The multidimensional environmental variable data is input into the trained environmental representation extraction model, and the output is an environmental representation vector. The environmental representation extraction model is a context encoder trained based on the multidimensional environmental variable training dataset and the JEPA self-supervised learning framework. The environmental representation vector is input into the trained preference environment density model, and the output is the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples. The probability score or logarithmic probability score is graded according to a preset threshold, and the suitability level is output.
2. The method and system for predicting wildlife habitat based on mask representation and density model according to claim 1, characterized in that, Based on the multidimensional environmental variable training dataset, a context encoder is trained using the JEPA self-supervised learning framework, including: Based on the field survey data of the target species, the corresponding multidimensional environmental variable training dataset is extracted from the preset environmental variable layer or raster data; Missing values and scaling are processed on the multidimensional environmental variable training dataset to obtain the environmental variable input vector; Construct the environment variable input vector into context input and target input; The JEPA self-supervised learning framework is constructed, which includes a context encoder, a target encoder, and a predictor. According to the preset masking strategy, some variable dimensions in the context input are masked and mask markers are generated to obtain the masked context input and mask marker vectors. The masked context input and mask marker vectors are input into the context encoder to output the context environment representation. The context environment representation is input into the predictor to output the environment representation. The target input is fed to the target encoder, which outputs a representation of the target environment. The alignment loss between the environment representation and the target environment representation is calculated, and the context encoder is updated based on the backpropagation of the alignment loss to obtain the trained context encoder.
3. The method and system for predicting wildlife habitat based on mask representation and density model according to claim 2, characterized in that, Both the context encoder and the target encoder include an input processing layer, a feature fusion layer, and a representation output layer. The input processing layer is used to organize the environmental variable input vector and process the mask label. The feature fusion layer is used to fuse multivariate relationships. The representation output layer is used to output a fixed-length environmental representation vector.
4. The method and system for predicting wildlife habitat based on mask representation and density model according to claim 2, characterized in that, Using the environmental representation vector training dataset as positive samples, a normalized flow model is trained based on maximum likelihood estimation, including: Input the multidimensional environment variable training dataset into the trained context encoder, and output the environment representation vector training dataset. Construct a normalized flux density model; The training dataset of environmental representation vectors is used as positive samples to input into the normalized flow density model. The model parameters of the normalized flow density model are updated by maximizing the log-likelihood of the positive samples under the normalized flow density model, thus obtaining the trained normalized flow model.
5. The method and system for predicting wildlife habitats based on mask representation and density model according to claim 4, characterized in that, The normalized flux density model includes an input layer, a transformation block stack layer, a basic distribution layer, and a scoring output layer. The transformation block stack layer is composed of multiple transformation blocks connected in series. Each transformation block includes a dimension rearrangement unit, a parameter network, and a transformation calculation unit. The dimension rearrangement unit is used to change the order of vector dimensions, the parameter network is used to generate transformation parameters, and the transformation calculation unit is used to complete the reversible transformation of this layer based on the parameters and output the result.
6. The method and system for predicting wildlife habitat based on mask representation and density model according to claim 1, characterized in that, The probability score or logarithmic probability score is graded according to a preset threshold, and the suitability level is output, including: The probability score or log probability score is backfilled into the spatial grid of the target area to form a spatial distribution map of habitat suitability; The scores are normalized and mapped to form habitat suitability scores in the [0,1] interval; The habitat suitability score is classified according to a preset threshold, and the classification results are used as the habitat suitability output.
7. The method and system for predicting wildlife habitat based on mask representation and density model according to claim 1, characterized in that, Obtain field survey data of the target species, and based on this data, obtain corresponding multidimensional environmental variable data, including: Collect occurrence point data of the target species, including field transect survey data, infrared camera monitoring data, location tracking data and / or historical data; Perform spatial coordinate unification, duplicate point removal, and outlier point removal on the point data; Based on the spatial location of the point data, the corresponding multidimensional environmental variable data is extracted from the preset environmental variable layer or raster data. The multidimensional environmental variable data includes topographic variables, climate variables, vegetation variables and / or human disturbance variables. The multidimensional environmental variable data is processed for missing values and scale unification to form an environmental variable input vector.
8. A system for the wildlife habitat prediction method based on mask representation and density model as described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire field survey occurrence data of the target species and acquire corresponding multidimensional environmental variable data based on the field survey occurrence data of the target species. The extraction module is used to input multidimensional environmental variable data into a trained environmental representation extraction model and output an environmental representation vector. The environmental representation extraction model is a context encoder trained based on the JEPA self-supervised learning framework using a multidimensional environmental variable training dataset. The scoring module is used to input the environmental representation vector into the trained preference environment density model and output the matching degree of the environmental representation vector under the target species preference distribution. The matching degree is a probability score or a log probability score. The preference environment density model is a normalized flow model trained based on the maximum likelihood estimation method, using the environmental representation vector training dataset as positive samples. The scoring output module is used to classify probability scores or logarithmic probability scores according to preset thresholds and output the suitability level.