Method for constructing spatio-temporal hypergraph conditional denoising diffusion probability model based on physical guidance

Abstract

The application discloses a physical guidance-based spatio-temporal hypergraph condition denoising diffusion probability model construction method, and relates to the field of ecological environment modeling and artificial intelligence prediction. The application innovatively combines the physical heat diffusion process and the spatio-temporal modeling of multi-modal environmental factors by introducing a temperature driving module based on heat conduction and a spatio-temporal perception fusion module. In addition, a lightweight convolution block attention module and a hierarchical hypergraph attention mechanism are embedded in the diffusion denoising network to realize high-order correlation modeling and dynamic fusion of multi-scale spatial features. The method can effectively capture the spatio-temporal evolution characteristics and ecological dependence of species distribution while maintaining reasonable computational complexity, thereby significantly improving the accuracy of distribution prediction. The experimental results on different types of data sets prove the superiority and robustness of the proposed PSTH-CDPM.

Description

Technical Field

[0001] This invention relates to the field of ecological environment modeling and artificial intelligence prediction, specifically to a method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model. Background Technology

[0002] With the increasing demand for modeling complex systems, how to characterize the dynamic evolution of target variables under multi-source spatiotemporal data conditions has become a key technical problem in the fields of artificial intelligence and spatiotemporal modeling. Existing methods usually rely on historical observation data and external condition information to predict or reconstruct the target distribution, but they still have significant shortcomings in complex spatiotemporal dynamic modeling, characterizing high-order spatial dependencies, and effectively utilizing condition information.

[0003] (1) Traditional model Early related methods often employed statistical modeling or shallow machine learning frameworks, achieving prediction by establishing a mapping relationship between conditional variables and target variables (Chu, Qingshuai, et al. 2024; Ramasamy, Maruthadurai, et al. 2022; Xian, Xiaoqing, et al. 2024). However, these methods typically assume that conditional features are independent of each other, making it difficult to express the nonlinear relationships between high-dimensional conditional variables. Furthermore, they have limited ability to characterize spatiotemporal dynamic changes, making it difficult to adapt to the evolutionary processes of complex systems that vary with time and space (de Oliveira, Lucas Rieger, et al. 2023; Liu, Yinghan, et al. 2024).

[0004] (2) Deep learning model In recent years, deep learning models have been gradually introduced into spatiotemporal modeling tasks, enhancing their ability to model spatial features and temporal dependencies through convolutional structures, recursive structures, or attention mechanisms (Davis, Courtney L, et al. 2023; Lake, Thomas A, et al. 2025; Smith, Austin M, et al. 2022). However, most existing deep models treat conditional information as static input or simply concatenate features, lacking a mechanism for continuously guiding conditional variables during the modeling process, resulting in insufficient responsiveness to changes in conditions. Furthermore, these methods typically focus on local neighborhood modeling, limiting their ability to express cross-regional and higher-order spatial dependencies, making it difficult to characterize long-range interaction structures in complex systems (Elias, Nathan. 2023; Guo, Yanhui, et al. 2022; Lake, Thomas A, et al. 2022).

[0005] (3) Diffusion model Diffusion models, as a novel generative modeling framework, possess strong expressive power for complex data distributions through forward noise diffusion and backward conditional denoising processes, and have demonstrated good performance in various generative tasks (Khachatryan, Levon, et al. 2023; Kumari, Nupur, et al. 2023; Saharia, Chitwan, et al. 2022). However, the application of existing diffusion models in spatiotemporal prediction scenarios remains relatively limited, especially under complex constraints. Mature modeling solutions for effectively injecting multi-source conditional information and characterizing higher-order spatial dependencies during the diffusion process are still lacking.

[0006] Furthermore, in real-world spatiotemporal systems, the evolution of some key conditional variables often follows explicit physical diffusion or propagation laws (Peng, Jiang-Zhou, et al. 2024; Tadeparti, Sidharth, and VishalVR Nandigana. 2022; Zhou, Zijie, et al. 2021). Existing data-driven models typically ignore such physical evolutionary mechanisms, relying solely on data fitting for modeling, which limits the model's ability to express real evolutionary processes and interpret the results. Therefore, how to integrate conditional diffusion modeling, physical evolutionary constraints, and higher-order spatial structure modeling within a unified modeling framework remains a key unresolved issue in current technology. Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model includes the following steps: Step 1: Data Construction Before proposing this model, we first acquired multi-temporal remote sensing data and multi-source environmental factor data of the target area, performed spatial alignment and temporal registration on them, and constructed an environmental cue dataset. }, and rasterize the species presence / absence records into a binary map of the target distribution { }

[0009] Step 2: Construct the network model: (1) Temperature-driven module based on heat conduction (TCTDM) Temperature is a crucial environmental factor influencing species distribution and spatial evolution. This module, based on the thermal diffusion equation, uses thermal diffusion convolution to simulate the temperature transfer process in space. Specifically, it assumes... The temperature field in two-dimensional space can be described by the discrete form of the thermal diffusion equation: (1) in The Laplace operator representing the temperature field, i.e., the second spatial derivative, It reflects the curvature of the temperature distribution in space. This operator is implemented through convolution: (2) in, This represents a discrete two-dimensional convolution operation with a fixed kernel, while This is a discrete Laplacian convolution kernel used to calculate the spatial gradient diffusion of the temperature field. Parameters The diffusion rate is used to control the intensity of heat diffusion. Through this iterative process, the temperature field gradually smooths out spatially, realizing the radiation of ecological potential energy from the high-temperature region to the surrounding low-temperature region. This heat conduction modeling not only conforms to the principle of thermodynamic equilibrium but also helps to construct a continuous and smooth temperature potential energy field, thereby enhancing the model's spatial perception of ecological suitability.

[0010] (2) Spatiotemporal Awareness Fusion Module (STAFM) This module uses a Convolutional Long Short-Term Memory (ConvLSTM) network to encode the spatiotemporal dependencies between historical distributions and environmental data, while adaptively introducing environmental cues for the current year through a gating mechanism. The specific process is as follows: First, the species distribution maps for the first two time steps. and Corresponding environmental prompts and The sequences are concatenated to form the historical input sequence: (3) It is worth noting that the temperature factors in all environmental cues were preprocessed as described in Part (1). The sequence was then progressively fed into a three-layer stacked ConvLSTM network to capture spatiotemporal dependencies. The update equations for each layer at each time step are shown below: (4) (5) (6) (7) (8) (9) in, This represents the standard two-dimensional convolution operation with a trainable kernel. This indicates element-wise multiplication. The activation function is Sigmoid. The hidden state of the third layer of the ConvLSTM at the final time step. Extracted as historical spatiotemporal semantic representation .

[0011] (10) (11) Subsequently, in order to provide environmental guidance for the current year Incorporate, and use a gating mechanism to connect with Dynamic fusion enables adaptive integration to adapt to the recent ecological context. Here, Represents a spatial gating mask and For convolution kernel parameters, This is a bias term. Features obtained through fusion. While preserving historical ecological patterns, it adaptively responds to environmental changes in the current year, capturing historical spatiotemporal patterns while adapting to current environmental differences, and serving as a guiding input for the subsequent PSTH-CDPM denoising process.

[0012] (3) Noise reduction network structure design and diffusion process The method adopts a diffusion model framework and uses U-Net as the denoising network. The overall operation process includes a forward noise diffusion stage and a reverse conditional denoising stage. In the forward stage, noise is gradually added to simulate the evolution of data from a clear state to a noisy state. In the reverse stage, guided by conditional information, noise is gradually removed and the target sample is reconstructed through the U-Net network to achieve generation and prediction that conforms to physical and spatiotemporal constraints.

[0013] ① Training phase 1) Forward diffusion stage exist Within each step, distribute the data to the original target. (i.e., the distribution of invasive alien species) Isotropic Gaussian noise is gradually added. As the noise accumulates, the data gradually degenerates into pure Gaussian noise, forming a forward diffusion trajectory. The process is shown in formula (12), where This represents the noise planning coefficient that follows a linear growth strategy. ,and The unit covariance matrix representing the standard Gaussian distribution.

[0014] (12) 2) Reverse denoising stage For each training sample, we randomly select the diffusion time step. Generate corresponding noise samples and link it with guiding conditions They are fed into a denoising network together to generate predicted noise. By employing a random time step sampling strategy, the model statistically covers the entire diffusion process during continuous training iterations, thereby achieving comprehensive noise prediction capabilities.

[0015] The denoising network U-Net employs an encoder-decoder structure, integrating residual modules, CBAM, HyperGAT, and positional coding. The input includes the noise map of the current diffusion step. Guiding condition variables and diffusion steps The noise graph Mapped to the hidden space representation via the projection layer and associated with the guiding conditions. Summation to obtain initial features: (13) The core unit of the network is the temporal conditional residual block, each containing two layers of 3×3 convolutions, group normalization, and the Swish activation function. Short-circuit connections are implemented through constant mapping to maintain residual learning. Diffusion step. First, a representation is generated using sinusoidal position encoding. The time embedding vector is then generated through two layers of fully connected networks. Finally, add it to the output of the first convolutional layer: (15) The following two attention mechanism modules are embedded in the main body of this U-Net to enhance feature representation capabilities and model higher-order spatial relationships: a. Convolutional Block Attention Module (CBAM) To enhance the network's representation capabilities, CBAM is embedded in the deep layers of the encoder, bottleneck blocks, and high-level layers of the decoder. CBAM consists of channel attention modules and spatial attention modules, effectively emphasizing key features and suppressing irrelevant features by sequentially focusing on the channel and spatial dimensions. For the input feature map... Channel attention is generated by applying global average pooling and max pooling operations to the feature map, and then fusing them through a shared multilayer perceptron (MLP) to produce channel attention weights. : (16) Finally, the channel-weighted features are obtained through element-wise multiplication: (17) Spatial attention is achieved by performing mean pooling and max pooling along the feature map channel dimensions, concatenating the resulting features, and then applying a convolution operation to generate spatial attention weights. : (18) The final fusion result is represented by the output of CBAM: (19) b. HyperGraph Attention Module (HyperGAT) Furthermore, we introduce HyperGAT to achieve cross-layer and multi-scale feature fusion, thereby enhancing high-order semantic modeling capabilities. Specifically, the feature maps output by each layer of the U-Net encoder are first aligned in spatial dimensions and then uniformly constructed into a set of feature nodes.

[0016] (20) vector This represents the first feature map extracted from different layers of the encoder. There are 10 feature nodes. Then, based on the Euclidean distance between feature vectors, the K-Nearest Neighbor (KNN) algorithm is used to construct a set of hyperedges.

[0017] (twenty one) Each hyperedge connects a group of semantically related nodes, forming a cross-layer semantic hypergraph structure. Building upon this structure, the HyperGAT architecture with a dual attention mechanism is employed to implement dynamic weight modeling between nodes and hyperedges. Specifically, HyperGAT's first... The layer update process is defined as follows: (twenty two) (twenty three) in, Represents a node In the Layer representation, Indicates the superedge In the The representation of layers, and It is with nodes A set of connected hyperedges. The corresponding hyperedge representation is generated by aggregating the features of the nodes within each hyperedge. This represents the process of aggregating hyperedge features back into node representations.

[0018] In the specific implementation, a dual attention mechanism is used to assign weights to the aggregation function, thereby enhancing the model's ability to focus on both key nodes and key hyperedges. During the edge-to-hyperedge attention phase (node ​​to hyperedge), the hyperedge is represented by a weighted summation of the nodes it connects to: (twenty four) in This represents a non-linear activation function (such as ReLU). For learnable weight matrix, Indicates the superedge internal nodes The attention weights are calculated as follows: (25) (26) in It is the context vector of the node-level attention mechanism.

[0019] During the edge-to-node attention phase (hyperedge to node), the node representation is obtained by weighted aggregation of its connected hyperedges: (27) in The hyperedge projection matrix, For super-edge For nodes The attention weights are calculated as follows: (28) (29) in The context vector represents the hyperedge-level attention mechanism, and the symbol "$||$" represents the feature concatenation operation. Through this dual attention mechanism, the HyperGAT layer not only captures the complex high-order semantic relationships between nodes and hyperedges but also dynamically adjusts the importance of different features, thereby enhancing the fusion effect of cross-scale features. Finally, after multiple HyperGAT iterations, the resulting fused cross-layer high-order semantic features are concatenated with the bottleneck layer output and input into the decoder. This process achieves more comprehensive multi-scale information fusion and representation.

[0020] ② Generation stage In the generation phase, initial noise is first obtained by sampling from a Gaussian distribution. The target distribution is then gradually generated through a reverse denoising process. To ensure fair performance, all models use the same random noise to maintain input consistency. This noise, along with the guiding conditions, is input into the denoising network. The model then iteratively predicts and denoises, progressively optimizing the output through a series of inverse diffusion steps. Refined step by step Specifically, as shown in the formula (i.e.) : (30) (4) Overall loss function The overall objective loss function of the physics-guided spatiotemporal hypergraph conditional denoising and diffusion probability model construction method can be summarized as follows: (31) The model is trained using a loss function (Equation (31)) to minimize prediction noise. With real noise The differences between them.

[0021] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model construction method, which unifies the modeling of physical constraint mechanisms, spatiotemporal dynamic modeling, and conditional diffusion generation processes. This effectively improves the modeling ability of conditional dependencies and dynamic evolution processes in complex spatiotemporal systems, overcoming the technical shortcomings of existing methods in terms of insufficient physical mechanism characterization, limited spatiotemporal feature expression, and weak high-order structure modeling ability. First, this invention introduces a temperature-driven module (TCTDM) based on the principle of heat conduction. Using the heat diffusion equation as a theoretical basis, it models the spatial propagation process of key conditional variables through convolution, enabling the model to explicitly incorporate physical evolution constraints during feature learning, thereby enhancing the model's ability to express real spatiotemporal evolution processes and the interpretability of results. Second, the proposed spatiotemporal awareness fusion module (STAFM) can adaptively fuse multi-temporal data and multi-source conditional information, realizing joint modeling of historical evolution information of target variables and conditional dynamic information, thus improving the model's ability to characterize complex spatiotemporal dynamic changes. Furthermore, the conditional denoising network of the diffusion model integrates a Convolutional Block Attention Module (CBAM) and a Hierarchical Hypergraph Attention Mechanism (HyperGAT). CBAM enhances the model's selective response to key channel features and salient spatial regions, while HyperGAT characterizes high-order structural dependencies across scales and regions, thus overcoming the limitations of traditional local neighborhood modeling and significantly enhancing feature representation and global structure awareness. Through the synergistic effect of these modeling mechanisms, the proposed method exhibits good modeling stability, robustness, and generalization ability in various complex spatiotemporal data scenarios, validating the effectiveness and applicability of the model in general spatiotemporal prediction and conditional generation tasks. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the architecture of the physical-guided spatiotemporal hypergraph conditional denoising diffusion probability model construction method in this invention. Detailed Implementation

[0023] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model includes the following steps: Step 1: Data Construction Before proposing this model, we first acquired multi-temporal remote sensing data and multi-source environmental factor data of the target area, performed spatial alignment and temporal registration on them, and constructed an environmental cue dataset. }, and rasterize the species presence / absence records into a binary map of the target distribution { }

[0025] Step 2: Construct the network model: (1) Temperature-driven module based on heat conduction (TCTDM) Temperature is a crucial environmental factor influencing species distribution and spatial evolution. This module, based on the thermal diffusion equation, uses thermal diffusion convolution to simulate the temperature transfer process in space. Specifically, it assumes... The temperature field in two-dimensional space can be described by the discrete form of the thermal diffusion equation: (1) in The Laplace operator representing the temperature field, i.e., the second spatial derivative, It reflects the curvature of the temperature distribution in space. This operator is implemented through convolution: (2) in, This represents a discrete two-dimensional convolution operation with a fixed kernel, while This is a discrete Laplacian convolution kernel used to calculate the spatial gradient diffusion of the temperature field. Parameters The diffusion rate is used to control the intensity of heat diffusion. Through this iterative process, the temperature field gradually smooths out spatially, realizing the radiation of ecological potential energy from the high-temperature region to the surrounding low-temperature region. This heat conduction modeling not only conforms to the principle of thermodynamic equilibrium but also helps to construct a continuous and smooth temperature potential energy field, thereby enhancing the model's spatial perception of ecological suitability.

[0026] (2) Spatiotemporal Awareness Fusion Module (STAFM) This module uses a Convolutional Long Short-Term Memory (ConvLSTM) network to encode the spatiotemporal dependencies between historical distributions and environmental data, while adaptively introducing environmental cues for the current year through a gating mechanism. The specific process is as follows: First, the species distribution maps for the first two time steps. and Corresponding environmental prompts and The sequences are concatenated to form the historical input sequence: (3) It is worth noting that the temperature factors in all environmental cues were preprocessed as described in Part (1). The sequence was then progressively fed into a three-layer stacked ConvLSTM network to capture spatiotemporal dependencies. The update equations for each layer at each time step are shown below: (4) (5) (6) (7) (8) (9) in, This represents the standard two-dimensional convolution operation with a trainable kernel. This indicates element-wise multiplication. The activation function is Sigmoid. The hidden state of the third layer of the ConvLSTM at the final time step. Extracted as historical spatiotemporal semantic representation .

[0027] (10) (11) Subsequently, in order to provide environmental guidance for the current year Incorporate, and use a gating mechanism to connect with Dynamic fusion enables adaptive integration to adapt to the recent ecological context. Here, Represents a spatial gating mask and For convolution kernel parameters, This is a bias term. Features obtained through fusion. While preserving historical ecological patterns, it adaptively responds to environmental changes in the current year, capturing historical spatiotemporal patterns while adapting to current environmental differences, and serving as a guiding input for the subsequent PSTH-CDPM denoising process.

[0028] (3) Noise reduction network structure design and diffusion process The method adopts a diffusion model framework and uses U-Net as the denoising network. The overall operation process includes a forward noise diffusion stage and a reverse conditional denoising stage. In the forward stage, noise is gradually added to simulate the evolution of data from a clear state to a noisy state. In the reverse stage, guided by conditional information, noise is gradually removed and the target sample is reconstructed through the U-Net network to achieve generation and prediction that conforms to physical and spatiotemporal constraints.

[0029] ① Training phase 1) Forward diffusion stage exist Within each step, distribute the data to the original target. (i.e., the distribution of invasive alien species) Isotropic Gaussian noise is gradually added. As the noise accumulates, the data gradually degenerates into pure Gaussian noise, forming a forward diffusion trajectory. The process is shown in formula (12), where This represents the noise planning coefficient that follows a linear growth strategy. ,and The unit covariance matrix representing the standard Gaussian distribution.

[0030] (12) 2) Reverse denoising stage For each training sample, we randomly select the diffusion time step. Generate corresponding noise samples and link it with guiding conditions They are fed into a denoising network together to generate predicted noise. By employing a random time step sampling strategy, the model statistically covers the entire diffusion process during continuous training iterations, thereby achieving comprehensive noise prediction capabilities.

[0031] The denoising network U-Net employs an encoder-decoder structure, integrating residual modules, CBAM, HyperGAT, and positional coding. The input includes the noise map of the current diffusion step. Guiding condition variables and diffusion steps The noise graph Mapped to the hidden space representation via the projection layer and associated with the guiding conditions. Summation to obtain initial features: (13) The core unit of the network is the temporal conditional residual block, each containing two layers of 3×3 convolutions, group normalization, and the Swish activation function. Short-circuit connections are implemented through constant mapping to maintain residual learning. Diffusion step. First, a representation is generated using sinusoidal position encoding. The time embedding vector is then generated through two layers of fully connected networks. Finally, add it to the output of the first convolutional layer: (15) The following two attention mechanism modules are embedded in the main body of this U-Net to enhance feature representation capabilities and model higher-order spatial relationships: a. Convolutional Block Attention Module (CBAM) To enhance the network's representation capabilities, CBAM is embedded in the deep layers of the encoder, bottleneck blocks, and high-level layers of the decoder. CBAM consists of channel attention modules and spatial attention modules, effectively emphasizing key features and suppressing irrelevant features by sequentially focusing on the channel and spatial dimensions. For the input feature map... Channel attention is generated by applying global average pooling and max pooling operations to the feature map, and then fusing them through a shared multilayer perceptron (MLP) to produce channel attention weights. : (16) Finally, the channel-weighted features are obtained through element-wise multiplication: (17) Spatial attention is achieved by performing mean pooling and max pooling along the feature map channel dimensions, concatenating the resulting features, and then applying a convolution operation to generate spatial attention weights. : (18) The final fusion result is represented by the output of CBAM: (19) b. HyperGraph Attention Module (HyperGAT) Furthermore, we introduce HyperGAT to achieve cross-layer and multi-scale feature fusion, thereby enhancing high-order semantic modeling capabilities. Specifically, the feature maps output by each layer of the U-Net encoder are first aligned in spatial dimensions and then uniformly constructed into a set of feature nodes.

[0032] (20) vector This represents the first feature map extracted from different layers of the encoder. There are 10 feature nodes. Then, based on the Euclidean distance between feature vectors, the K-Nearest Neighbor (KNN) algorithm is used to construct a set of hyperedges.

[0033] (twenty one) Each hyperedge connects a group of semantically related nodes, forming a cross-layer semantic hypergraph structure. Building upon this structure, the HyperGAT architecture with a dual attention mechanism is employed to implement dynamic weight modeling between nodes and hyperedges. Specifically, HyperGAT's first... The layer update process is defined as follows: (twenty two) (twenty three) in, Represents a node In the Layer representation, Indicates the superedge In the The representation of layers, and It is with nodes A set of connected hyperedges. The corresponding hyperedge representation is generated by aggregating the features of the nodes within each hyperedge. This represents the process of aggregating hyperedge features back into node representations.

[0034] In the specific implementation, a dual attention mechanism is used to assign weights to the aggregation function, thereby enhancing the model's ability to focus on both key nodes and key hyperedges. During the edge-to-hyperedge attention phase (node ​​to hyperedge), the hyperedge is represented by a weighted summation of the nodes it connects to: (twenty four) in This represents a non-linear activation function (such as ReLU). For learnable weight matrix, Indicates the superedge internal nodes The attention weights are calculated as follows: (25) (26) in It is the context vector of the node-level attention mechanism.

[0035] During the edge-to-node attention phase (hyperedge to node), the node representation is obtained by weighted aggregation of its connected hyperedges: (27) in The hyperedge projection matrix, For super-edge For nodes The attention weights are calculated as follows: (28) (29) in The context vector represents the hyperedge-level attention mechanism, and the symbol "$||$" represents the feature concatenation operation. Through this dual attention mechanism, the HyperGAT layer not only captures the complex high-order semantic relationships between nodes and hyperedges but also dynamically adjusts the importance of different features, thereby enhancing the fusion effect of cross-scale features. Finally, after multiple HyperGAT iterations, the resulting fused cross-layer high-order semantic features are concatenated with the bottleneck layer output and input into the decoder. This process achieves more comprehensive multi-scale information fusion and representation.

[0036] ② Generation stage In the generation phase, initial noise is first obtained by sampling from a Gaussian distribution. The target distribution is then gradually generated through a reverse denoising process. To ensure fair performance, all models use the same random noise to maintain input consistency. This noise, along with the guiding conditions, is input into the denoising network. The model then iteratively predicts and denoises, progressively optimizing the output through a series of inverse diffusion steps. Refined step by step Specifically, as shown in the formula (i.e.) : (30) (4) Overall loss function The overall objective loss function of the physics-guided spatiotemporal hypergraph conditional denoising and diffusion probability model construction method can be summarized as follows: (31) The model is trained using a loss function (Equation (31)) to minimize prediction noise. With real noise The differences between them.

[0037] The present invention has experimentally verified the above method and achieved significant results, as described below: 1. Dataset We used two high-quality, multi-temporal datasets for pine wilt nematode (dataset A) and fall webworm (dataset B). These datasets cover species distribution records and corresponding environmental factors from 2016 to 2020. Data on the occurrence of these invasive alien species were obtained from the annual "Announcement of Pine Wilt Nematode / Fall Webworm Epidemic Areas in China" (http: / / www.forestry.gov.cn) published by the National Forestry and Grassland Administration from 2017 to 2021. These records reflect the actual distribution of this species in county-level and above administrative regions of China during the study period. Latitude and longitude coordinates of the affected areas were extracted using Google Maps (https: / / mapsplatform.google.com / ) and processed using ArcGIS (version 10.8) to generate species distribution maps that meet the research requirements. The distribution of invasive alien species is influenced by a combination of climate, topography, human disturbance, and vegetation-related environmental factors. Relevant data were sourced from the WorldClim database (http: / / worldclim.org / ) and the Global Resources Data Cloud Platform (http: / / gis5g.com / ), covering 2016 to 2020. This dataset contains a rich set of variables, including climate factors such as average / minimum / maximum temperature, precipitation, relative humidity, sunshine duration, wind speed, and soil moisture; topographic factors are represented by altitude, and human disturbance factors are quantified using GDP, population density, and nighttime light intensity. Vegetation factors are characterized by vegetation cover. After preprocessing and screening, 13 environmental factors were ultimately retained.

[0038] This invention employs real annual species distribution records and corresponding environmental data for all model training, comparative experiments, and in-depth analysis to ensure scientific rigor and practical relevance. However, invasive alien species distribution data are typically recorded annually, resulting in a relatively limited sample size. Although environmental data offers finer temporal resolution (e.g., monthly), the overall dataset remains sparse. This sparsity makes comprehensively capturing the complex spatiotemporal dynamics of species distribution patterns challenging. To alleviate this limitation, we propose a data simulation strategy using dataset A as a case study. This method integrates existing annual species distribution records with monthly environmental data to construct a synthetic daily resolution dataset. Combining ecologically sound interpolation methods with random perturbation techniques, this technique significantly expands the temporal sample size, enriches the time series structure, and enhances the model's ability to learn dynamic evolutionary features in high-temporal-resolution sequences. It is important to note that the simulated dataset is only used to assist in validating the model's generalization performance in large sample sizes and high-resolution time series, and is not intended as the primary training data or a source of comparative experimental data.

[0039] 2. Evaluation Indicators To evaluate all methods from multiple perspectives, we used five widely used evaluation metrics: Area Under the Receiver Operating Characteristic Curve (AUC), True Skill Statistic (TSS), Kappa coefficient, Intersection over Union (IoU), and F1 score. For all metrics, higher scores indicate better performance.

[0040] 3. Implementation details All deep learning models (such as ConvLSTM, PredRNN++, etc.) were reimplemented with the original parameter settings. In addition to the deep learning models, other baseline models were implemented using the BIOMOD_EnsembleModeling and BIOMOD_Modeling functions in the BIOMOD2 R package.

[0041] For PSTH-CDPM, the total number of diffusion steps Set as (Right now The noise variance scheduling scheme is defined as follows: Its value is from linearly increasing to The model was trained using the Adam optimizer, with an initial learning rate set to... Furthermore, a cosine annealing scheduling scheme is incorporated to promote stable convergence. During the generation phase, the optimal threshold is determined through verification, and the prediction results are binarized: pixels above the threshold are assigned a value of 1, and pixels below the threshold are mapped to 0.

[0042] 4. Comparison Methods Our model was benchmarked against sixteen established models, including Convolutional Long Short-Term Memory (ConvLSTM), Predictive Recurrent Neural Network (PredRNN), Predictive Recurrent Neural Network++ (PredRNN++), Memory Network (MIM), Motion Aware Unit (MAU), Predictive Recurrent Neural Network V2 (PredRNN.V2), Video Prediction Model (SimVP), Temporal Attention Unit (TAU), Wavelet-based Spatiotemporal Model (WaST), and several traditional species distribution models, such as Artificial Neural Network (ANN), Generalized Additive Model (GAM), Progressive Regression Tree Model (GBM), Multivariate Adaptive Regression Spline Function (MARS), Maximum Entropy Model (MaxEnt), Random Forest Model (RF), Extreme Gradient Boosting Algorithm (XGBOOST), and Ensemble Model (EM).

[0043] Table 1: Experimental results for dataset A Table 2: Experimental results for dataset B Table 3: Experimental Results of Simulation Data Experimental results were presented on all three datasets, with the best performance highlighted in bold. While the proposed PSTH-CDPM model performed well but not at the highest in AUC and TSS metrics, it achieved the best results in the three core metrics of Kappa, IoU, and F1-score. This is mainly because AUC and TSS focus more on global binary classification ability, which may be more prominent on negative samples and fail to reflect spatial accuracy and pattern consistency in ecological distribution prediction; while Kappa, IoU, and F1-score can more comprehensively measure the model's performance in terms of spatial consistency, class balance, and ecological rationality. Therefore, this study emphasizes the latter to demonstrate the model's comprehensive advantages in predicting the spatial distribution of invasive species. In summary, PSTH-CDPM exhibits superior predictive performance across different species and geographic regions. This performance improvement stems from its generative modeling framework based on diffusion processes, which generates high-quality species distribution maps. Furthermore, the model incorporates TCTDM to simulate the physical diffusion process of environmental factors, thereby constructing an ecological potential field and enhancing the spatial continuity representation of environmental factors. By capturing the spatiotemporal correlations of historical data through STAFM, the generation process is guided to better align with ecological evolution patterns. In the denoising network, CBAM enhances the model's ability to focus on key regions, while HyperGAT enables high-order semantic modeling of cross-level, multi-scale features—overcoming the limitations of the traditional U-Net architecture in cross-scale information interaction. The deep integration of these components significantly enhances the model's ability to simulate ecological processes and the spatial distribution of generation, thereby improving the model's generalization ability and ecological adaptability.

[0044] This invention proposes a species distribution prediction method based on a Physically Guided Spatiotemporal Hypergraph Conditional Denoising Diffusion Probability Model (PSTH-CDPM). It innovatively combines physical mechanism modeling, spatiotemporal feature fusion, multimodal conditional guidance, and diffusion generation learning, effectively improving the accuracy and generalization ability of invasive alien species distribution prediction. First, this invention introduces a Temperature-Driven Model (TCTDM) based on thermal conduction, simulating the physical evolution of environmental factors through thermal diffusion convolution. This achieves dynamic constraints on the spatial diffusion effect of environmental temperature on species distribution, enhancing the model's physical consistency and interpretability. Second, this invention designs a Spatiotemporal Awareness Fusion Model (STAFM) to fuse multi-temporal remote sensing information and multi-source environmental factors, achieving joint modeling of historical changes in species distribution and environmental dynamics, significantly improving the model's ability to capture complex spatiotemporal dependencies. Third, this invention introduces a diffusion probability model framework into the field of species distribution prediction for the first time, and uses a U-Net-based denoising network structure to progressively reverse-reconstruct noisy data, thereby achieving conditional generation and high-precision distribution recovery. Within the denoising network, a Convolutional Block Attention (CBAM) module is incorporated to enhance the feature responses of salient channels and spatial regions. Simultaneously, a Hierarchical Hypergraph Attention (HyperGAT) mechanism is introduced to achieve cross-level and multi-scale feature fusion, thereby enhancing high-order semantic modeling capabilities and effectively improving the model's representational power and robustness. Experimental results on different datasets demonstrate the superiority and robustness of our proposed PSTH-CDPM in ecosystem spatial prediction tasks.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0046] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0047] References Chu, Qingshuai, et al. "Invasive alien plants in the Qinghai-XizangPlateau (China): Current state and future predictions." Ecological Indicators166 (2024): 112488. Ramasamy, Maruthadurai, Bappa Das, and R. Ramesh. "Predicting climatechange impacts on potential worldwide distribution of fall armyworm based onCMIP6 projections." Journal of Pest Science 95.2 (2022): 841-854. Xian, Xiaoqing, et al. "Temperature extremes nip invasive macrophyteCabomba caroliniana A. Gray in the bud: Potential geographic distributionsand risk assessment based on future climate change and anthropogenicinfluences." Frontiers in Plant Science 15 (2024): 1393663. de Oliveira, Lucas Rieger, et al. "Non-native decapods in SouthAmerica: Risk assessment and potential impacts." Diversity 15.7 (2023): 841. Liu, Yinghan, Xuemei Zhang, and Shixiang Zong. "Prediction of thePotential Distribution of Teinopalpus aureus Mell, 1923 (Lepidoptera,Papilionidae) in China Using Habitat Suitability Models." Forests 15.5(2024): 828. Davis, Courtney L., et al. "Deep learning with citizen science dataenables estimation of species diversity and composition at continentalextents." Ecology 104.12 (2023): e4175. Lake, Thomas A., Ryan D. Briscoe Runquist, and David A. Moeller. "Twodecades of satellite images reveal the spatial and temporal dynamics of leafyspurge invasion and improve species distribution models." bioRxiv (2025):2025-05. Smith, Austin M., César Capinha, and Andrew M. Kramer. "Predictingspecies distributions with environmental time series data and deep learning."bioRxiv (2022): 2022-10. Elias, Nathan. "Deep learning methodology for early detection andoutbreak prediction of invasive species growth." Proceedings of the IEEE / CVFWinter Conference on Applications of Computer Vision. 2023. Guo, Yanhui, et al. "A novel invasive plant detection approach usingtime series images from unmanned aerial systems based on convolutional andrecurrent neural networks." Neural Computing and Applications 34.22 (2022):20135-20147. Lake, Thomas A., Ryan D. Briscoe Runquist, and David A. Moeller. "Deep learning detects invasive plant species across complex landscapes usingWorldview‐2 and Planetscope satellite imagery." Remote Sensing in Ecology andConservation 8.6 (2022): 875-889. Khachatryan, Levon, et al. "Text2video-zero: Text-to-image diffusionmodels are zero-shot video generators." Proceedings of the IEEE / CVFInternational Conference on Computer Vision. 2023. Kumari, Nupur, et al. "Multi-concept customization of text-to-imagediffusion." Proceedings of the IEEE / CVF conference on computer vision andpattern recognition. 2023. Saharia, Chitwan, et al. "Photorealistic text-to-image diffusionmodels with deep language understanding." Advances in neural informationprocessing systems 35 (2022): 36479-36494. Peng, Jiang-Zhou, et al. "HCP-PIGN: Efficient heat conductionprediction by physics-informed graph convolutional neural network."International Journal of Heat and Fluid Flow 109 (2024): 109552. Tadeparti, Sidharth, and Vishal VR Nandigana. "Convolutional neuralnetworks for heat conduction." Case Studies in Thermal Engineering 38 (2022):102089. Zhou, Zijie, Baofeng Zhang, and Xiao Yu. "Infrared handprintclassification using deep convolution neural network." Neural ProcessingLetters 53.2 (2021): 1065-1079.

Claims

1. A method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model, characterized in that, Includes the following steps: Step 1: Data Construction Acquire multi-temporal remote sensing data and multi-source environmental factor data of the target area, perform spatial alignment and temporal registration, and construct an environmental alert dataset. }, and rasterize the species presence / absence records into a binary map of the target distribution { }; Step 2: Construct the network model: A temperature-driven module TCTDM based on heat conduction, a spatiotemporal awareness fusion module STAFM, and a denoising network model are constructed. The temperature-driven module TCTDM is used to process temperature factors, the spatiotemporal awareness fusion module STAFM is used to fuse historical spatiotemporal information with current environmental cues to generate guiding conditions, and finally, a denoising network model with U-Net as its core is used to generate a target species distribution prediction map from noise through a reverse denoising process under the guiding conditions.

2. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 1, characterized in that, The temperature-driven module TCTDM based on heat conduction uses the heat diffusion equation as its basis and employs heat diffusion convolution to simulate the temperature conduction process in space; assuming... Representing the temperature field in two-dimensional space, the diffusion process is described by the discrete form of the thermal diffusion equation: (1) in The Laplace operator, representing the temperature field, i.e., the second spatial derivative, reflects the curvature of the temperature distribution in space. This Laplace operator is implemented through convolution. (2) in, This represents a discrete two-dimensional convolution operation with a fixed kernel. For the discrete Laplacian convolution kernel used to calculate the spatial gradient diffusion of the temperature field, the parameters are... This indicates the diffusion rate, used to control the intensity of thermal diffusion.

3. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 1, characterized in that, The spatiotemporal awareness fusion module STAFM uses a Convolutional Long Short-Term Memory (ConvLSTM) network to encode the spatiotemporal dependence of historical distribution and environmental data. Simultaneously, it adaptively introduces environmental cues for the current year through a gating mechanism. The specific process is as follows: First, the species distribution maps for the first two time steps. and Corresponding environmental prompts and The sequences are concatenated to form the historical input sequence: (3) The sequence is then fed stepwise into a three-layer stacked ConvLSTM network to capture spatiotemporal dependencies. The update equation for each layer at each time step is shown below: (4) (5) (6) (7) (8) (9) in, This represents the standard two-dimensional convolution operation with a trainable kernel. This indicates element-wise multiplication. Using the Sigmoid activation function, the hidden state of the third layer of ConvLSTM at the final time step. Extracted as historical spatiotemporal semantic representation : (10) (11) Subsequently, in order to provide environmental guidance for the current year Incorporate and use a gating mechanism to connect with Dynamic integration, adaptively integrating with the recent ecological context. Represents a spatial gating mask and For convolution kernel parameters, As a bias term, the features obtained by fusion This serves as a guiding condition input for the subsequent PSTH-CDPM denoising process.

4. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 1, characterized in that, The training phase of the denoising network model includes a forward noise diffusion phase and a backward conditional denoising phase: 1) Forward diffusion stage exist Within each step, distribute the data to the original target. Isotropic Gaussian noise is gradually added. As the noise accumulates, the data gradually degenerates into pure Gaussian noise, forming a forward diffusion trajectory. The process is shown in formula (12), where This represents the noise planning coefficient that follows a linear growth strategy. ,and The unit covariance matrix representing the standard Gaussian distribution; (12) 2) Reverse denoising stage For each training sample, a diffusion time step is randomly selected. Generate corresponding noise samples and link it with guiding conditions They are fed into a denoising network together to generate predicted noise. The denoising network U-Net adopts an encoder-decoder structure, integrating residual modules, CBAM, HyperGAT, and position coding elements; The input includes the noise map of the current diffusion step. Guiding condition variables and diffusion steps ; Among them, the noise graph Mapped to the hidden space representation via the projection layer and associated with the guiding conditions. Summation to obtain initial features: (13) The core unit of the network is the temporal conditional residual block. Each unit contains two layers of 3×3 convolutions, group normalization, and the Swish activation function. Short-circuit connections are achieved through constant mapping to maintain residual learning, and the diffusion step... First, a representation is generated using sinusoidal position encoding. Subsequently, the time embedding vector is generated through two layers of fully connected networks. Finally, add it to the output of the first convolutional layer: (15)。 5. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 4, characterized in that, The following two attention mechanism modules are embedded in the main body of the denoising network U-Net to enhance feature representation ability and high-order spatial correlation modeling: A. Convolutional Block Attention Module (CBAM) CBAM consists of channel attention modules and spatial attention modules, which are embedded in the deep layers of the encoder, bottleneck blocks, and high-level layers of the decoder for the input feature map. The channel attention module generates weight coefficients based on the global average pooling and max pooling results: (16) Finally, the channel-weighted features are obtained through element-wise multiplication: (17) The spatial attention module generates a spatial mask based on the response of features in the spatial dimension: (18) The final fusion result is: (19) B. HyperGAT (HyperGraph Attention Module) The feature maps output from each layer of the U-Net encoder in the denoising network are first aligned in spatial dimensions and then uniformly constructed into a set of feature nodes: (20) vector This represents the first feature map extracted from different layers of the encoder. For each feature node, a set of hyperedges is constructed using the K-Nearest Neighbors (KNN) algorithm based on the Euclidean distance between feature vectors. (21) Each hyperedge connects a group of semantically related nodes, forming a cross-layer semantic hypergraph structure. .

6. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 5, characterized in that, In the cross-layer semantic hypergraph structure Based on this, the HyperGAT architecture with a dual attention mechanism is adopted to realize dynamic weight modeling between nodes and hyperedges.

7. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 6, characterized in that, The HyperGAT architecture is the first The layer update process is defined as follows: (22) (23) in, Represents a node In the Layer representation, Indicates the superedge In the The representation of layers, and It is with nodes A set of connected hyperedges; The corresponding hyperedge representation is generated by aggregating the features of the nodes within each hyperedge. This represents the process of aggregating hyperedge features back into node representations.

8. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 7, characterized in that, The HyperGAT architecture employs a dual attention mechanism to assign weights to aggregation functions; During the edge-in-attention phase, the hyperedge representation is calculated through a weighted sum of its connected nodes: (24) in Represents a non-linear activation function. For learnable weight matrix, Indicates the superedge internal nodes The attention weights are calculated as follows: (25) (26) in It is the context vector of the node-level attention mechanism; During the edge-to-edge attention phase, node representations are obtained through weighted aggregation of their connected hyperedges: (27) in The hyperedge projection matrix, For super-edge For nodes The attention weights are calculated as follows: (28) (29) in This represents the context vector of the hyperedge-level attention mechanism, and the symbol "$||$" represents the feature concatenation operation.

9. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 1, characterized in that, The generation stage of the denoising network model includes: First, pure Gaussian noise is sampled as the initial generation target. The pure Gaussian noise and guiding conditions are input into the denoising network model. The denoising network model then iteratively predicts and denoises, gradually optimizing the output result through a series of inverse diffusion steps. Refined step by step Specifically, as shown in the formula: (30) 。 10. The method for constructing a physics-guided spatiotemporal hypergraph conditional denoising diffusion probability model according to claim 9, characterized in that, The overall loss function of the denoising network model is: (31) The denoising network model is trained using equation (31) to minimize the predicted noise. With real noise The differences between them.

Images