Deep learning-based method for predicting influence of climate change on biodiversity

Abstract

The invention belongs to the technical field of environment prediction, and relates to a method for predicting the influence of climate change on biodiversity based on deep learning, and the method comprises the steps: converting climate data into a 4D tensor, employing depth separable convolution to extract multi-scale spatial-temporal features, and capturing a spatial-temporal dependency relationship; a kernel density estimation method is adopted to convert the biodiversity data into a density field, a dynamic heterogeneous graph is obtained through a heterogeneous graph attention network, effective alignment of the climate data and the biodiversity data is achieved by constructing a transmembrane hypergraph and carrying out space-time projection, a multi-modal uncertainty propagation mechanism is introduced, and the dynamic heterogeneous graph is obtained. According to the method, a three-dimensional thermodynamic diagram of joint uncertainty is output based on a multi-modal uncertainty propagation mechanism, and finally, a prediction result of species distribution is generated based on climate and biological data of space-time alignment and the thermodynamic diagram of joint uncertainty, so that the prediction precision of the influence of climate change on biological diversity is improved, and the prediction accuracy of species distribution is improved. And the interpretability of the prediction result is enhanced.

Description

Technical Field

[0001] This application belongs to the technical field of environmental prediction, and more specifically, relates to a method for predicting the impact of climate change on biodiversity based on deep learning. Background Art

[0002] The impact of climate change is one of the key topics in global environmental research. Biodiversity is an important indicator of ecosystem health, and its changes are closely related to climate conditions. Therefore, accurately predicting the impact of climate change on biodiversity is of great significance for environmental protection, ecological restoration, and species conservation.

[0003] However, the impact of climate change on biodiversity is a multi-factor and dynamically changing process; climate data is usually time-series and has a complex distribution pattern in space; while biodiversity data is usually observed at specific spatial location points and has strong locality and irregularity; this difference makes it a major challenge to effectively align the two for joint analysis. Summary of the Invention

[0004] The present invention provides a method for predicting the impact of climate change on biodiversity based on deep learning, aiming to solve the technical problem of the difficulty in effectively aligning climate data and biodiversity data.

[0005] The method for predicting the impact of climate change on biodiversity based on deep learning includes the following steps: Climate data collection and processing: Collect climate data, convert the collected climate data into a 4D tensor, and use depthwise separable convolution to extract multi-scale spatio-temporal features to obtain a climate data tensor; Biodiversity data collection and processing: Obtain biodiversity data based on species observation points, convert it into a dynamic spatial density field using kernel density estimation, and then model it through constructing a heterogeneous graph attention network to obtain a dynamic heterogeneous graph of species; Spatio-temporal alignment: Based on the climate data tensor and the dynamic heterogeneous graph, construct a transmembrane state hypergraph and perform spatio-temporal projection to obtain spatio-temporally aligned climate and biological data; Multi-modal uncertainty propagation: Model the climate prediction error based on a neural network differentiator, and at the same time separate the noise of biological observation data through a double-branch feature extractor to obtain observation noise from different sources, and then fuse the covariance of the climate prediction error and the biological observation noise to output a three-dimensional heat map of joint uncertainty; Dynamic feedback prediction architecture: Based on the spatio-temporally aligned climate and biological data and the three-dimensional heat map, construct a multi-layer spatio-temporal graph structure, and process it based on the spatio-temporal graph attention network to generate a prediction result of species distribution.

[0006] The present invention converts climate data into a 4D tensor and uses depthwise separable convolution to extract multi-scale spatio-temporal features and capture spatio-temporal dependencies. Then, the kernel density estimation method is used to convert biodiversity data into a density field, and a dynamic heterogeneous graph is obtained through a heterogeneous graph attention network. Then, by constructing a transmembrane state hypergraph and performing spatio-temporal projection, effective alignment of climate data and biodiversity data is achieved, enabling the two to be jointly modeled and predicted within the same spatio-temporal framework. In addition, the present invention also introduces a multi-modal uncertainty propagation mechanism, and outputs a three-dimensional heat map of joint uncertainty based on the multi-modal uncertainty propagation mechanism, further improving the credibility and accuracy of the prediction. Finally, based on the spatio-temporally aligned climate and biological data and the heat map of joint uncertainty, a multi-layer spatio-temporal graph structure is constructed and processed using a spatio-temporal graph attention network to generate a prediction result of species distribution, which not only improves the prediction accuracy of the impact of climate change on biodiversity, but also enhances the interpretability of the prediction result.

[0007] Preferably, obtaining the climate data tensor includes the following steps: Depth convolution: Using 3D convolution based on the 4D tensor to extract local spatio-temporal features in the climate data, where each channel is convolved separately; Point convolution: Performing channel fusion on the features extracted by the depth convolution through 1D convolution to obtain a spatio-temporal feature map; Multi-scale convolution: Using a first convolution kernel to extract a first feature; using a second convolution kernel to extract a second feature, where the scale of the second convolution kernel is larger than that of the first convolution kernel; then performing a concatenation operation on the first feature and the second feature to obtain a multi-scale feature map; Vertical correlation modeling: Based on the multi-scale feature map, calculating the attention weights between each pair of height layers, and performing weighted fusion on the features of different height layers based on the calculated attention weights to obtain a new height-correlated feature map; Spatio-temporal feature encoding: Inputting the multi-scale feature map and the height-correlated feature map into a GRU, using a gating mechanism to control the information flow to capture the dynamic pattern in time, and finally outputting the climate data tensor.

[0008] Preferably, the steps for processing the biodiversity data include: Dynamic density conversion and gradient calculation: According to the species occurrence point data, through the kernel density estimation algorithm, converting discrete observation points into a continuous species density field, where the kernel density estimation considers the influence of space and time during calculation, using a Gaussian kernel function and a time decay function to obtain a dynamically transformed species density value, that is, the species density field; based on the calculated species density field, using a spatial gradient calculation method to extract the spatial gradient of the species density field; Spatial graph structure construction: Define each spatial grid point as a node of the graph. The features of the node include species density value, magnitude and direction of the gradient, terrain features, accuracy and coverage radius of the observation device. If the gradient directions of the species density fields of two nodes are similar, an edge connection is established between the two nodes. For each adjacent node, the weight of the edge is determined based on gradient similarity, geographical distance, and device accuracy, based on which the constructed spatial graph structure is obtained; Heterogeneous graph attention network: Map the features of each node through a trainable linear layer, and then use the multi-head heterogeneous graph attention mechanism to calculate the attention coefficients for each node and its adjacent nodes. Then, based on the adjacent nodes of each node and the calculated attention coefficients, weighted fusion of information is performed to update the node features.

[0009] Preferably, the spatio-temporal alignment includes the following steps: Construct a climate hypergraph: For each climate grid, extract all climate grid points within its spatio-temporal neighborhood to form a hyperedge, and then aggregate the neighborhood climate features through 3D convolution to obtain the hyperedge feature representation of the climate; Construct a biological hypergraph: For each biological node, select its Top-k neighbor nodes according to the diffusion probability to form a hyperedge, and aggregate the features of the node and its neighbor nodes through the multi-head graph attention mechanism to obtain the hyperedge feature representation of the biology; Transmembrane state hypergraph connection: Connect the climate grid points and biological nodes. When both the spatial distance and the temporal difference meet the predetermined conditions, the climate grid points and biological nodes share a transmembrane state hyperedge; Projection of climate data to biological data: For each biological node, find the climate grid points within its spatial neighborhood, calculate the projection weight of each climate grid point on the biological node, where the projection weight considers the similarity of climate features and biological features, and introduce geographical distance constraints for calculation; Based on the calculated projection weight combined with the climate grid point features, obtain the climate feature representation on the biological node; Projection of biological data to climate data: For each climate grid point, find the biological nodes within its spatial neighborhood, and perform weighted aggregation of the biological node features through the diffusion probability to obtain the biological feature representation on the climate grid point.

[0010] Preferably, modeling the climate prediction error based on a neural network differentiator includes: 3D convolution: Use 3D convolution based on the climate data tensor to extract local spatio-temporal features, and add an activation function after the convolutional layer to increase non-linearity; Fully connected layer: Based on the features obtained after the 3D convolution operation, use a fully connected layer to process the features and output the error prediction result; Error covariance calculation: Based on the error prediction results of the output, calculate the covariance matrix of the errors to describe the error correlation at different spatial points.

[0011] Preferably, the dual-branch feature extractor includes two branches; The first branch is used to extract the true biological information of the observed data; the second branch is used to extract features related to noise; Calculate the covariance matrix of the biological observation noise based on the extracted noise-related features to represent the correlation of the noise.

[0012] Preferably, the dynamic feedback prediction architecture is as follows: Construct a multi-layer spatio-temporal graph structure: Construct it based on the spatio-temporally aligned climate and biological data. The spatio-temporal graph structure includes a spatial graph and a temporal graph. Each spatial point in the spatial graph is regarded as a node in the graph, and the edges between the nodes represent the spatial adjacency relationship; each time point in the temporal graph is regarded as a node in the graph, and the edges between the nodes represent the evolutionary relationship in time; obtain spatio-temporal graph data based on the constructed multi-layer spatio-temporal graph structure; Spatio-temporal graph attention network processing: Based on the spatio-temporal graph data and the three-dimensional heat map of uncertainty as the input of the spatio-temporal attention network, assign different importance weights to each spatio-temporal node through the spatio-temporal attention mechanism, and update the features of each spatio-temporal node based on the calculated importance weights; Species distribution prediction: Based on the features of the updated out-of-control nodes, use the species distribution prediction network to predict the distribution of species.

[0013] The beneficial effects of the present invention include: The present invention converts climate data into a 4D tensor, and uses depthwise separable convolution to extract multi-scale spatio-temporal features to capture spatio-temporal dependencies; then uses the kernel density estimation method to convert biodiversity data into a density field, and then obtains a dynamic heterogeneous graph through the heterogeneous graph attention network, and then realizes the effective alignment of climate data and biodiversity data by constructing a transmembrane state hypergraph and performing spatio-temporal projection, enabling the two to be jointly modeled and predicted in the same spatio-temporal framework. In addition, the present invention also introduces a multi-modal uncertainty propagation mechanism, and outputs a three-dimensional heat map of joint uncertainty based on the multi-modal uncertainty propagation mechanism to further improve the credibility and accuracy of the prediction. Finally, based on the spatio-temporally aligned climate and biological data and the heat map of joint uncertainty, a multi-layer spatio-temporal graph structure is constructed, and the spatio-temporal graph attention network is used for processing to generate the prediction results of species distribution, which not only improves the prediction accuracy of the impact of climate change on biodiversity, but also enhances the interpretability of the prediction results. Description of the Drawings

[0014] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0015] Figure 1 It is the overall step block diagram provided by the embodiment of the present invention. Detailed implementation manners

[0016] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the following further details the present application in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0017] See Figure 1 As shown, an energy efficiency optimization method for diving equipment based on environmental data prediction includes the following steps: A prediction method for the impact of climate change on biodiversity based on deep learning includes the following steps: Climate data collection and processing: Collect climate data, convert the collected climate data into a 4D tensor, and use depthwise separable convolution to extract multi-scale spatio-temporal features to obtain a climate data tensor; Exemplarily, obtain climate data with spatial and temporal dimensions through means such as climate sensors, satellite remote sensing, or meteorological stations, including meteorological variables such as temperature, humidity, precipitation, wind speed, and air pressure. Assume the obtained climate data matrix is C, with dimensions [T, H, W, F], where T represents the time step (e.g., monthly or annually), H represents the number of spatial height layers; W represents the spatial width (longitude and latitude dimensions); F represents the type of climate variable (e.g., temperature, humidity, precipitation, etc.); Depth convolution: Based on the obtained climate data matrix, use 3D convolution for local spatio-temporal feature extraction, where the 3D convolution operation is expressed as: ; In the formula: represents the convolution kernel, and the convolution operation extracts spatio-temporal features by applying the convolution kernel, and each channel is convolved separately, that is, the convolution operation is applied to each climate variable (such as temperature, humidity, etc.) respectively to obtain multiple output feature maps; Point convolution: Use 1D convolution to perform channel fusion on the features obtained by depth convolution. Let the output of each channel be , and the 1D convolution operation is expressed as: ; In the formula: Denote the 1D convolution kernel. The convolution operation is fused along the time or spatial dimension to generate a spatio-temporal feature map; Multi-scale convolution: To further capture features at different scales, two convolution kernels with different scales are used to extract features respectively: Based on the output results of point convolution, the first convolution kernel and the second convolution kernel are used to extract the first feature and the second feature respectively, where the scale of the second convolution kernel is larger than that of the first convolution kernel. Then, the first feature and the second feature are concatenated to obtain a multi-scale feature map ; Vertical correlation modeling: Calculate the attention weights between each pair of height layers on the multi-scale feature map and perform weighted fusion on the features of different height layers based on these weights. Assume the attention weight matrix between height layers is A. Multiply the weight matrix by the multi-scale feature map to obtain a highly correlated feature map ; Among them, the steps to calculate the attention weight matrix are as follows: Perform a linear transformation on the multi-scale feature map to obtain the query, key, and value representations of each height layer. Use a learnable weight matrix to project each height layer in the multi-scale feature map respectively to obtain the query, key, and value; Calculate the similarity between the query vector of each pair of height layers and all key vectors. The dot product can be used to measure the correlation between the query and the key; then normalize the calculated correlation values to obtain the attention weight corresponding to each query; based on this, obtain the attention weight matrix; Spatio-temporal feature encoding: Input the multi-scale feature map and the highly correlated feature map into a GRU (Gated Recurrent Unit) for spatio-temporal feature encoding to obtain a 4D climate data tensor.

[0018] Biodiversity data collection and processing: Obtain biodiversity data based on species observation points, convert it into a dynamic spatial density field using kernel density estimation, and then model it through constructing a heterogeneous graph attention network to obtain a dynamic heterogeneous graph of species; In this embodiment, the collection of biodiversity data mainly relies on the observation point data of species. The observation point data mainly includes the occurrence location of species and the corresponding observation time. The observation points of each species include: spatial coordinates (longitude, latitude), timestamp (observation time), and species identifier (species information corresponding to each observation point); Since the collected biodiversity data is discrete data and cannot be directly used for modeling, further processing is required. In this embodiment, the collected discrete data is converted into a species density field. The specific steps are as follows: In this embodiment, the discrete data is converted into a species density field through the kernel density estimation method. By applying a kernel function (such as a Gaussian kernel) around each observation point, the species density at each location is estimated smoothly. Exemplarily: Given a set of species observation point data , where represents the spatial coordinates of observation point i, i.e., longitude and latitude; the observation time of observation point i; represents the species identifier corresponding to observation point i, and the kernel density estimation function is as follows: ; ; In the formula: represents the species density at location x and time t; N represents the total number of observation points; h represents the bandwidth parameter, which controls the size of the kernel; represents the Gaussian kernel function; represents the time decay function; represents the decay factor; where x and t are the specific spatial and temporal points of the species density to be estimated; The gradient operator is used to differentiate the species density field to obtain the spatial gradient of the species density field: ; In the formula: represents the species density field in the x and y directions of the spatial coordinates. The gradient is used to describe the distribution change of the species in space, and further reveals the response pattern of the species to climate change; Spatial graph structure construction: Each grid point is used as a node of the graph, and the features of the node include: The species density value, the species density calculated by KDE ; The modulus and direction extracted in space, the modulus and direction of calculated by the spatial gradient; Terrain data, as an additional feature of the node; The accuracy and coverage radius of the observation equipment; The edge connection rule of the graph is based on the following criteria: If the gradient directions of two nodes are similar, an edge connection is established; For each adjacent node, the weight of the edge is calculated based on the following factors: First, calculate the similarity of the gradient directions of the two nodes : ; In the formula: represents the species density field of node i in the x and y directions of the spatial coordinates; Represents the species density field of node j Gradients in the x and y directions of the spatial coordinates; Geographical distance calculation: The Haversine formula is used to calculate the distance between two points on the Earth's surface; Based on the calculated similarity and geographical distance, a weighted summation method is used to calculate the comprehensive weight.

[0019] Heterogeneous graph attention network: Based on the constructed spatial graph structure, a heterogeneous graph attention network is used to process node features. The specific steps are as follows: To enable the neural network to better process the feature information of nodes, the original features of each node are mapped to a new feature space through a linear transformation by node feature mapping. The specific process is as follows: Given the feature vector of each node i , it is mapped to a new feature space through a linear transformation (such as a trainable weight matrix W) to obtain a transformed node feature : ; In the formula: W represents the trainable weight matrix; represents the original feature of node i; b represents the bias term; Based on this, the multi-head attention mechanism is further applied for processing. The interaction between each node and its adjacent nodes is realized by calculating the attention coefficient . The attention coefficient represents the correlation between node i and adjacent node j, and based on this, the features of adjacent nodes are weighted and fused. How to calculate the attention coefficient belongs to the conventional technical means in this field, so it will not be elaborated in this embodiment; When the attention coefficient is calculated, the attention coefficient will be used to weightedly fuse the information of each adjacent node. By weightedly summing the features of adjacent nodes, the feature of the current node i is updated.

[0020] Based on the updated node features, we will obtain the final updated output dynamic heterogeneous graph of the species.

[0021] Spatiotemporal alignment: Based on the climate data tensor and the dynamic heterogeneous graph, a transmembrane state hypergraph is constructed and spatiotemporal projection is performed to obtain spatiotemporally aligned climate and biological data; As a possible implementation manner of this embodiment, the spatiotemporal alignment includes the following steps: In this step, in order to convert the climate data and biodiversity data into a unified spatiotemporal framework so that the climate data (4D tensor) and biological data (dynamic heterogeneous graph) can correspond to each other and be analyzed under the same spatiotemporal pattern, the specific steps are as follows: Construct a climate hypergraph: For each climate grid point , extract all climate grid points in the spatio-temporal neighborhood around the climate grid point . Assume the time interval of the grid is , and the spatial interval is . Select a domain range R, where R is the spatio-temporal radius; all grid points within this neighborhood form a hyperedge, connecting the grid points within the domain range R; 3D convolutional aggregation: Aggregate the climate features within the spatio-temporal neighborhood of the climate grid point to obtain the feature representation of the climate hyperedge. Specifically, aggregate the climate features of the spatio-temporal neighborhood through a 3D convolutional operation, and the expression form is as follows: ; In the formula: represents the climate data of the spatio-temporal neighborhood around the climate grid point ; represents the 3D convolutional operation, which performs convolution in the three dimensions of time, latitude, and longitude to aggregate the climate features within the neighborhood; represents the feature representation of the climate hyperedge; Construction of the climate hypergraph: Each climate grid point is connected to other climate grid points within its domain through hyperedges to obtain the climate hypergraph.

[0022] Construct a biological hypergraph: The biological hypergraph is used to model biodiversity data, focusing on the relationships between biological nodes, and aggregates the features of neighbors through an attention mechanism. The specific steps are as follows: For each biological node , select the Top-k neighbor nodes according to its diffusion probability to form a hyperedge, where the diffusion probability is expressed as follows: ; In the formula: represents the diffusion probability between the climate grid and the biological node ; represents the geographical distance between the climate grid point and the biological node ; represents a hyperparameter that controls the diffusion range; represents the maximum number of neighbors; represents the geographical distance between the climate grid point and the biological node ; Aggregate the features of the biological node and its neighbor nodes through a multi-head attention mechanism: ; In the formula: and respectively represent the feature representations of climate grid points and biological neighbor nodes ; represents the multi-head attention mechanism, which is used to aggregate multiple neighbor features; Based on this, the biological node and its Top-k neighbors are connected by a hyperedge to obtain a biological hypergraph.

[0023] In spatio-temporal alignment, the climate grid and biological nodes need to be connected in the same spatio-temporal framework. The transmembrane state hypergraph connection realizes the alignment of climate grid points and biological nodes. The specific steps are as follows: Spatially, if the distance between two nodes is less than the set threshold , and temporally, the time difference between two nodes is less than the set time threshold , then a transmembrane state hyperedge is shared between the climate grid point and the biological node. The spatial distance is calculated by the Haversine formula, and the time difference can be directly calculated as the difference between two time points. If both the spatial distance and the time difference meet the above conditions, then a transmembrane state hyperedge will be formed between the climate grid and the biological node .

[0024] Projection of climate data onto biological data: By calculating the projection weights, the climate data is mapped into the biological data as follows: For each biological node , calculate the projection weight of the climate grid points within its spatial neighborhood on this biological node. The formula for calculating the weight is as follows: ; In the formula: represents the geographical distance between the climate grid point and the biological node ; represents the geographical distance between the climate grid point and the biological node ; Based on the calculated projection weights, combined with the features of the climate grid points , the climate feature representation on the biological node is obtained: ; In the formula: represents the climate feature representation on the biological node.

[0025] For each climate grid point , find the biological nodes within its spatial neighborhood, and perform weighted aggregation on the features of the biological nodes through the diffusion probability: ; In the formula: represents the biological feature representation at the climate grid point.

[0026] In this embodiment, by constructing a climate hypergraph and a biological hypergraph, as well as the connection and projection of the transmembrane state hypergraph, the alignment of climate data and biodiversity data is achieved. This not only helps to integrate two different modalities of data into the same spatio-temporal framework, but also improves the accuracy of alignment through weighted aggregation and the attention mechanism, providing a reliable data basis for subsequent predictions.

[0027] Multi-modal uncertainty propagation: Model the climate prediction error based on a neural network differentiator, and at the same time separate the noise of biological observation data through a double-branch feature extractor to obtain observation noise from different sources. Then, fuse the covariance of the climate prediction error and the biological observation noise to output a three-dimensional heat map of joint uncertainty; As a possible implementation manner of this embodiment, the multi-modal uncertainty propagation includes the following steps: Use a 3D convolutional layer to extract local spatio-temporal features. Assume that the neural network uses standard convolution operations, and the convolution kernel size is ( ), and the output feature map is , where , , represents the spatial and temporal dimensions after convolution operation, , represents the high dimension of space and the width dimension of space, such as longitude and latitude; represents the number of output channels; Further process the convolution output through a fully connected layer to predict the error, where the calculation formula of the fully connected layer is as follows: ; In the formula: and represent the weights and biases of the fully connected layer; represents the operation of expanding the convolution output; represents the predicted error; According to the output error calculate the covariance matrix of the error, which is used to describe the error correlation of different spatial points, and the calculation formula is as follows: ; In the formula: represents the number of samples; represents the mean value of the error; Represents the i-th error value; To handle the noise in biodiversity data, a dual-branch feature extractor is adopted to extract real biological information and noise-related features respectively; Among them, the first branch extracts the real biological information in the observed data through a convolutional layer or other neural network layers to obtain the features representing the real information of biodiversity , exemplarily, the processing using a convolutional layer is as follows: ; In the formula: Represents the activation function (such as ReLU, Sigmoid activation function); Represents the weight matrix in the first branch; Represents the biological observation data; Represents the bias term of the first branch; The second branch extracts the features related to noise , and calculates the covariance matrix of the biological observation noise. Similarly, the noise feature extraction process can also be carried out using a convolutional layer to obtain the noise features ; Then, based on this, the covariance matrix of the noise is calculated: ; In the formula: Represents the covariance matrix of the noise features; M represents the number of samples; Represents the noise feature of the i-th sample; Represents the mean value of the noise features; Represents the transpose operation.

[0028] Covariance fusion: The covariance matrix of the climate error and the covariance matrix of the biological observation noise are fused by weighted summation to obtain the joint covariance matrix ; Based on the joint covariance matrix a three-dimensional heat map is generated as the output of uncertainty ; This three-dimensional heat map represents the joint uncertainty between climate change and biodiversity in the spatial and temporal dimensions.

[0029] Dynamic feedback prediction architecture: Based on the spatio-temporal aligned climate and biological data and the three-dimensional heat map, a multi-layer spatio-temporal graph structure is constructed, and processed based on the spatio-temporal graph attention network to generate the species distribution prediction result; As a possible implementation manner of this embodiment, the dynamic feedback prediction architecture specifically includes the following steps: Construct a multi-layer spatio-temporal graph structure based on the spatio-temporal aligned data, specifically including a spatial graph and a temporal graph. Each spatial point in the spatial graph serves as a node, and the edges between the nodes represent the adjacency relationship in space. For example, neighboring spatial points (such as adjacent pixels) can be connected by edges. Each time point in the temporal graph serves as a node, and the edges between the nodes represent the evolutionary relationship in time. For example, the edges connect consecutive time points. The spatio-temporal graph structure combines the spatial graph and the temporal graph into a multi-layer spatio-temporal graph. The features of each node include both the spatial information and the temporal information of the node. The spatio-temporal data is represented by the following tensor: ; where: N represents the number of spatial points (i.e., the number of nodes in the graph); T represents the number of time dimensions; D represents the feature dimension of each spatio-temporal node; Processing of the spatio-temporal graph attention network: The input of the spatio-temporal graph attention network includes spatio-temporal graph data , and the three-dimensional heat map of joint uncertainty ; The goal of the spatio-temporal graph attention mechanism is to assign different importance weights to each node according to the context information of each spatio-temporal node, use the attention mechanism to calculate the weights of each spatio-temporal node, and update the features of the nodes based on the calculated weights. Assume that the input matrix of the spatio-temporal graph network is , then the output of the spatio-temporal attention network is expressed as: ; where: represents the spatio-temporal attention operation, which calculates the weights and updates the node features based on the context information (spatial neighborhood and temporal neighborhood) of each node and the uncertainty heat map. The specific calculation formula is as follows: ; In the formula: and are the feature vectors of spatio-temporal nodes i and j respectively; represents the feature transformation matrix; represents the feature concatenation operation; represents the transpose of the weight vector of the attention mechanism; represents the attention weight of node i to node j; The species distribution prediction network processes the time node features after passing through the time attention network and an uncertainty three-dimensional heat map to predict species distribution; among them, the species distribution prediction network can adopt a hybrid model of CNN and LSTM, which can process information in space and time simultaneously; based on the given model and input data, how to use this model to achieve species distribution prediction belongs to the conventional technical means of the technical solution in this field, so it will not be described in detail and no application will be made.

[0030] In the present invention, climate data is converted into a 4D tensor, and depthwise separable convolution is used to extract multi-scale spatio-temporal features and capture spatio-temporal dependence relationships; then the kernel density estimation method is used to convert biodiversity data into a density field, and a dynamic heterogeneous graph is obtained through a heterogeneous graph attention network, and then the effective alignment of climate data and biodiversity data is achieved by constructing a transmembrane state hypergraph and performing spatio-temporal projection, so that the two can be jointly modeled and predicted in the same spatio-temporal framework. In addition, the present invention also introduces a multi-modal uncertainty propagation mechanism, and based on the multi-modal uncertainty propagation mechanism, a three-dimensional heat map of joint uncertainty is output, further improving the credibility and accuracy of the prediction. Finally, based on the spatio-temporal aligned climate and biological data and the heat map of joint uncertainty, a multi-layer spatio-temporal graph structure is constructed, and the spatio-temporal graph attention network is used for processing to generate the prediction result of species distribution, which not only improves the prediction accuracy of the impact of climate change on biodiversity, but also enhances the interpretability of the prediction result.

[0031] The above are only the preferred embodiments of the present application, and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the impact of climate change on biodiversity based on deep learning, characterized in that: The following steps are involved: Climate data collection and processing: Collect climate data, convert the collected climate data into 4D tensors, and use deep separable convolution to extract multi-scale spatiotemporal features to obtain climate data tensors; Biodiversity data collection and processing: Biodiversity data is obtained based on species observation points, and converted into a dynamic spatial density field using kernel density estimation. Then, a heterogeneous graph attention network is constructed for modeling to obtain a dynamic heterogeneous graph of species. Spatiotemporal alignment: Based on the climate data tensor and dynamic heterogeneous graph, a transmembrane hypergraph is constructed and spatiotemporally projected to obtain spatiotemporally aligned climate and biological data; Multimodal uncertainty propagation: climate prediction errors are modeled based on a neural network differentiator. The noise of biological observation data is separated by a dual-branch feature extractor to obtain observation noise from different sources. The covariance of climate prediction errors and biological observation noise is then fused to output a three-dimensional heat map of joint uncertainty. Dynamic feedback prediction architecture: Based on spatiotemporal aligned climate and biological data and three-dimensional heat maps, a multi-layer spatiotemporal graph structure is constructed, and species distribution prediction results are generated based on spatiotemporal graph attention network processing.

2. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: Obtaining the climate data tensor comprises the following steps: Deep convolution: 3D convolution is used based on 4D tensors to extract local spatiotemporal features in climate data, where each channel is convolved separately; Point convolution: The features extracted by deep convolution are fused through 1D convolution to obtain spatiotemporal feature maps; Multi-scale convolution: Use the first convolution kernel to extract the first feature; use the second convolution kernel to extract the second feature, where the scale of the second convolution kernel is larger than the scale of the first convolution kernel; then concatenate the first feature and the second feature to obtain a multi-scale feature map; Vertical correlation modeling: Based on the multi-scale feature map, the attention weights between each pair of height layers are calculated, and the features of different height layers are weightedly fused based on the calculated attention weights to obtain a new height correlation feature map; Spatiotemporal feature encoding: Multi-scale feature maps and highly correlated feature maps are input into GRU, and the gating mechanism is used to control the information flow to capture dynamic patterns in time, and finally the climate data tensor is output.

3. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: The steps of processing the biodiversity data include: Dynamic density conversion and gradient calculation: Based on the species occurrence point data, the discrete observation points are converted into a continuous species density field through the kernel density estimation algorithm. The kernel density estimation takes into account the influence of space and time during calculation, and uses the Gaussian kernel function and time decay function to obtain a dynamically transformed species density value, namely the species density field. Based on the calculated species density field, the spatial gradient of the species density field is extracted using the spatial gradient calculation method. Spatial graph structure construction: define each spatial grid point as a node of the graph. The characteristics of the node include species density value, gradient modulus and direction, terrain characteristics, accuracy of observation equipment and coverage radius. If the gradient direction of the species density field of two nodes is similar, an edge connection between the two nodes is established. For each adjacent node, the weight of the edge is determined based on the gradient similarity, geographical distance and equipment accuracy, and the constructed spatial graph structure is obtained. Heterogeneous Graph Attention Network: The features of each node are mapped through a trainable linear layer, and then the multi-head heterogeneous graph attention mechanism is used to calculate the attention coefficient for each node and its adjacent nodes. Then, the information is weightedly fused based on the adjacent nodes of each node and the calculated attention coefficient to update the node features.

4. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: The spatiotemporal alignment comprises the following steps: Constructing a climate hypergraph: For each climate grid, extract all climate grid points in its spatiotemporal neighborhood to form a hyperedge, and then aggregate the domain climate features through 3D convolution to obtain the hyperedge feature representation of the climate; Constructing biological hypergraph: For each biological node, select its top-k neighbor nodes according to the diffusion probability to form a hyperedge. Through the multi-head graph attention mechanism, aggregate the features of the node and its neighbor nodes to obtain the biological hyperedge feature representation; Transmembrane hypergraph connection: connect the climate grid point and the biological node. When the spatial distance and the time difference meet the predetermined conditions, the climate grid point and the biological node share a transmembrane hyperedge. Projection of climate data to biological data: For each biological node, find the climate grid points in its spatial neighborhood, calculate the projection weight of each climate grid point on the biological node, where the projection weight takes into account the similarity of climate characteristics and biological characteristics, and introduces geographic distance constraints for calculation; based on the calculated projection weight combined with the climate grid point characteristics, the climate feature representation on the biological node is obtained; Projection of biological data to climate data: For each climate grid point, find the biological nodes in its spatial neighborhood, perform weighted aggregation on the biological node features through diffusion probability, and obtain the biological feature representation on the climate grid point.

5. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: Modeling climate prediction errors based on neural network differentiators includes: 3D convolution: 3D convolution is used to extract local spatiotemporal features based on the climate data tensor, and an activation function is added after the convolution layer to increase the linearity; Fully connected layer: Based on the features obtained after the 3D convolution operation, a fully connected layer is used to process the features and output the error prediction results; Error covariance calculation: Based on the output error prediction results, the error covariance matrix is ​​calculated to describe the error correlation of different spatial points.

6. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: The dual-branch feature extractor includes two branches; The first branch is used to extract the real biological information of the observed data; the second branch is used to extract the features related to the noise; The covariance matrix of the biological observation noise is calculated based on the extracted noise-related features, which represents the correlation of the noise.

7. The method for predicting the impact of climate change on biodiversity based on deep learning according to claim 1, characterized in that: The dynamic feedback prediction architecture is as follows: Construct a multi-layer spatiotemporal graph structure: Based on the spatiotemporal aligned climate and biological data, the spatiotemporal graph structure includes a spatial graph and a temporal graph. Each spatial point in the spatial graph is regarded as a node in the graph, and the edges of the node represent the adjacency relationship in space; each time point in the temporal graph is regarded as a node in the graph, and the edges of the node represent the evolutionary relationship in time; based on constructing a multi-layer spatiotemporal graph structure, the spatiotemporal graph data is obtained; Spatiotemporal graph attention network processing: Spatiotemporal graph data and a three-dimensional heat map of uncertainty are used as the input of the spatiotemporal attention network. Different importance weights are assigned to each spatiotemporal node through the spatiotemporal attention mechanism, and the features of each spatiotemporal node are updated based on the calculated importance weights. Species distribution prediction: Based on the updated features of out-of-control nodes, the species distribution prediction network is used to predict the distribution of species.

Images