A graph spatio-temporal sampling aggregation network for sea surface temperature prediction

Abstract

The application discloses a graph space-time sampling aggregation network for sea surface temperature prediction, comprising: a data preprocessing module configured to neighbor sample input sea surface temperature sequence data; a spatial information aggregation layer configured to aggregate information between adjacent nodes to obtain sequence data X (1) ; a full connection layer in communication connection with the sampling module, the full connection layer being configured to process information between adjacent nodes to obtain sequence data X (2) ; a time information convolution layer configured to obtain sequence data X (1) and sequence data X (2) and process them to capture time correlation node sequences and derive a prediction value. The graph space-time sampling aggregation network for sea surface temperature prediction disclosed by the application uses a neighbor sampling method for inductive learning, enhances the generalizability of the model, combines spatial information aggregation to capture spatial details, and combines time information convolution to master time dynamics.

Description

Technical Field

[0001] This invention belongs to the field of sea surface temperature prediction technology, specifically relating to a graph spatiotemporal sampling aggregation network for sea surface temperature prediction. Background Technology

[0002] Sea surface temperature (SST) is a key variable in marine environmental studies, playing a crucial role in understanding and predicting global climate and weather change. SST fluctuations not only influence climate models, such as causing El Niño and La Niña phenomena, but also have profound impacts on marine ecosystems and economic activities. Furthermore, SST is a critical factor in predicting marine hazards such as storm surges and red tides. Given the significant impact of SST on marine ecosystems and the atmospheric environment, accurate long-term and high-resolution forecasts are essential.

[0003] SST prediction methods are mainly divided into two categories: numerical methods and data-driven methods. Numerical methods are primarily based on physical models and ocean dynamics theories, often relying on complex air-sea interaction models combined with physical equations and oceanographic parameters to simulate and predict SST changes. However, building and running these models requires substantial computational resources and is highly sensitive to initial conditions and parameter settings, which can lead to model uncertainties.

[0004] In contrast, data-driven approaches primarily rely on historical data to predict future SST variations. Techniques such as machine learning and deep learning build predictive models by analyzing the relationships between historical SST data and other relevant variables, such as atmospheric pressure, wind direction, and wind speed. These methods are particularly well-suited for capturing and simulating complex nonlinear patterns in SST data. Starting with machine learning algorithms like random forests and support vector machines, classic deep learning algorithms, such as RNNs and their variant LSTM, have evolved to learn about variations in time series data. However, these algorithms do not consider spatial factors. Subsequently, Xiao, Changjiang, and others proposed ConvLSTM, an SST prediction algorithm that considers both spatiotemporal factors. Nevertheless, this algorithm does not account for spatial, structural, and connectivity irregularities. The presence of land or islands in areas lacking SST data can disrupt the information captured by the convolution.

[0005] Recently, graph learning methods have been widely applied in various fields, effectively considering the irregularities in space, structure, and connectivity. Traditional GNN algorithms cannot effectively capture and utilize structural information in graphs, nor can they effectively distinguish the relative importance between different nodes. The introduction of Graph Attention Networks (GATs) significantly solves this problem. However, the method of dynamically weighing the importance of neighboring nodes through the attention mechanism in GATs leads to increased computational costs. Therefore, Hamilton, Will, Zhitao Ying, and Jure Leskovec introduced GraphSAGE, which can use different aggregation functions. By learning how to aggregate the feature information of neighboring nodes, rather than learning specific graph feature representations, GraphSAGE improves the scalability and generalization ability of the model while saving computational costs. However, node-based aggregation methods do not fully consider the temporal aspect of node sequences; they focus more on the correlation between nodes. Summary of the Invention

[0006] This invention provides a graph spatiotemporal sampling aggregation network for sea surface temperature prediction. It leverages the advantages of graph neural networks in handling spatial structure, and based on GraphSAGE, utilizes a neighbor sampling method for inductive learning to enhance the model's generalization ability. It combines spatial information aggregation to capture spatial details and temporal information convolution to grasp temporal dynamics.

[0007] This invention is achieved through the following technical solution: This invention provides a graph spatiotemporal sampling aggregation network for sea surface temperature prediction, comprising: The data preprocessing module is configured to process the input sea surface temperature series data. Perform neighbor sampling to obtain information about neighboring nodes; The spatial information aggregation layer communicates with the data preprocessing module. This layer is configured to aggregate information between adjacent nodes to obtain sequence data. ; The fully connected layer communicates with the sampling module and is configured to process information between adjacent nodes to obtain sequence data. ; The temporal information convolutional layer communicates with the spatial information aggregation layer and the fully connected layer, respectively. The temporal information convolutional layer is configured to acquire sequence data. and sequence data The data is then processed to capture time-correlated node sequences and extract predicted values.

[0008] In some embodiments, the data preprocessing module is configured to sample the neighbors of a specified node and pass the embedding information of these neighbors to the target node; then aggregate the embedding information to update the embedding representation of the target node; and predict the label of the node based on the updated embedding.

[0009] In some embodiments, the spatial information aggregation layer includes a long short-term memory (LSTM) network module, which obtains sequence data based on information between adjacent nodes. The specific setup method for the Long Short-Term Memory (LSTM) network module is as follows: in, These are all learnable weight parameters from different layers. It involves concatenating x and y. It is a learnable bias parameter. It is the activation function sigmoid.

[0010] In some embodiments, the temporal information convolutional layer includes a convolutional layer, a flattened layer, a ReLU layer, and an FC-Layer layer arranged in sequence.

[0011] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention provides a graph spatiotemporal sampling aggregation network for sea surface temperature prediction. It employs neighbor sampling, capturing spatial details through spatial information aggregation and grasping temporal dynamics through temporal information convolution. This endows the model with significant inductive and generalization capabilities, effectively overcoming the limitations of SAGE aggregation in capturing the temporal correlation between node sequences. It also solves the problem of supervising temporal correlation, which is often neglected by traditional methods. By integrating the spatial information aggregation layer with the temporal convolutional layer, spatiotemporal dynamics can be effectively captured, and neighbor sampling aggregation technology is used to enhance the model's flexibility and generalization. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the structure of a graph spatiotemporal sampling aggregation network for sea surface temperature prediction provided in some embodiments of the present invention; Figure 2 A schematic diagram of a two-layer sampling neighbor sampling and aggregation process provided for some embodiments of the present invention, wherein K=2; Figure 3 This is a schematic diagram illustrating the working principle of the spatial information aggregation layer provided in some embodiments of the present invention; Figure 4 This is a schematic diagram illustrating the working principle of the temporal information convolutional layer provided in some embodiments of the present invention; Figure 5 Different South China Seas provided for some embodiments of the present invention The RMSE and MAE values ​​for the one-day forecast are shown below; where (a) is the South China Sea and (b) is the Bohai Sea. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0015] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., used to indicate the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship that the product of this invention is usually placed in during use. They are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0016] Furthermore, the use of terms such as "horizontal" and "vertical" in the description of this invention does not imply that the components are required to be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0017] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0018] The terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0019] To predict SST, we employed a graph-based approach to construct the prediction model. First, the SST data was divided into a grid format according to latitude, longitude, and resolution, forming N data points. These data points were then constructed into a graph model G=(V,E), where V represents the nodes in the graph, counted as N, and E represents the edges between nodes. The connections between these nodes can be represented using an adjacency matrix. To express.

[0020] For any given edge, we use the normalized Pearson correlation coefficient as the threshold for edge connectivity, defined as follows: The Pearson correlation coefficient is the Pearson correlation coefficient between the SST data of the two points. Distance is the distance between the two points on the grid.

[0021] By setting a threshold, we can remove items that are below the threshold. The edges, therefore any given edge All can be defined as: in This is the threshold for the correlation coefficient.

[0022] SST data is represented as a series of observations X = {X1, X2, ..., XT}, where X ∈ RN × C × T. Xn represents all observations of N nodes at time steps t, where t ranges from 1 to t, and C represents the feature dimension of each node. Our goal is to use past SST observations to predict future temperature changes. Specifically, by analyzing SST {Xtn−p+1, Xtn−p+2, ..., Xtn|n∈n}, our goal is to train a model to predict SST {Xt+1n, Xt+2n, ..., Xt+qn|n∈n} for the next q time steps. The observation value of the nth node at time step t.

[0023] Spatiotemporal sampling aggregation: The network framework of the graph spatiotemporal aggregation algorithm is as follows: Figure 1 As shown, it mainly consists of two parts: spatial information aggregation and temporal information convolution. The spatial information aggregation part is mainly composed of LSTM aggregation modules, and the temporal information convolution is mainly composed of one-dimensional temporal convolution. The input is SST sequence data. First, neighbor sampling is performed to obtain information about adjacent nodes. Then, a spatial information aggregation layer processes this information, aggregating relevant information between nodes to obtain... Simultaneously, X is processed through a fully connected layer to obtain... Then and The input is fed into a temporal information convolutional layer to capture the temporally relevant node sequence, which ultimately yields the predicted value.

[0024] Neighbor sampling and aggregation: SST data is processed using a neighbor sampling and aggregation strategy, such as Figure 2 As shown in the diagram, in this process, the model samples the neighbors of a specified node and passes the embedding information of these neighbors to the target node. Then, using a selected aggregation function, the model aggregates this information to update the embedding representation of the target node. The updated embedding is then used to predict the node's label.

[0025] By choosing an appropriate value for K, the neighbor sampling strategy allows the model to aggregate information from different levels, thus increasing the algorithm's flexibility. Unlike traditional GNNs, which rely on propagation training that requires information from the entire graph, the neighbor sampling strategy supports inductive learning. This means it can generate embedding representations for newly added nodes without retraining the entire model, significantly enhancing the model's scalability and practicality.

[0026] Spatial information aggregation layer: To enhance the expressive power of this method, a Long Short-Term Memory (LSTM) network is incorporated as part of the node aggregation strategy. After sampling neighboring nodes and acquiring their information, this information is sequentially input into the LSTM aggregation network to obtain comprehensive aggregated information, such as... Figure 3 As shown.

[0027] LSTM utilizes a series of carefully designed components—input gates, forget gates, output gates, and memory units—to store, update, and flow information. The cooperative operation of these components enables LSTM to efficiently handle and remember the dependencies between memory nodes during the aggregation process. The computational process involving these gate units is as follows: (3) (4) (5) (6) (7) (8) Temporal information convolutional layer: To accurately capture the evolving relationships between nodes in the graph over time, a temporal information convolutional layer based on one-dimensional temporal convolution is employed, such as... Figure 4 As shown. The purpose of this layer is to reveal the time series correlations in the SST data. The process is as follows: Initially, after processing through the spatial aggregation layer, we obtain... Subsequently, the input data is processed through a fully connected linear transformation to obtain... Then, through merging and We got : The data is then fed into a temporal convolutional layer to capture the correlation of its time series.

[0028] In some examples, the number of input channels is set to 2, and the number of output channels is expanded to 8. Finally, after sequence flattening and nonlinear transformation, the data is fed into a fully connected layer to generate prediction results.

[0029] Specific examples: Dataset and Experiment Setup The dataset used in the experiment came from the National Oceanic and Atmospheric Administration (NOAA) in Boulder, Colorado, specifically the best interpolated SST version 2.1 (OISSTV2.1), available at https: / / www.ncei.noaa.gov / . This dataset includes daily, weekly, and monthly average SST measurements globally from September 1981 to the present, ranging from 89.875°S to 89.875°N and 0.125°E to 359.875°E. The spatial resolution of the daily, weekly, and monthly average SST data is 0.25° x 0.25°.

[0030] For our analysis, we selected the South China Sea and Bohai Sea SST time series from January 1, 1982 to December 31, 2021. The South China Sea region ranges from 105.125°E to 124.125°E and from 0.125°N to 24.875°N. The Bohai Sea region ranges from 117°E to 122°E and from 37°N to 41°N. After obtaining these datasets, we designated the data from 1982 to 2002 as the training set, the data from 2002 to 2012 as the validation set, and the data from 2012 to 2022 as the test set.

[0031] This application compares the Graph Spatiotemporal Sampling Aggregation Method (GSSA) with ConvLSTM, GAT (Graph Attention Networks), and SAGE-LSTM (GraphSAGE with LSTM). To ensure a fair comparison of model performance, all models use the same sliding window length, specifically 7 for the average daily, weekly, and monthly SST predictions for the next day, three days, seven days, next week, three weeks later, and next month. We choose the Adam optimizer to train the models. For a valid comparison with GAT, both SAGE-LSTM and GSSA use only one sampling layer to collect all surrounding neighbor nodes. The hidden layer dimension between the GSSA aggregation layer and the convolutional layer is set to 16. The convolutional kernel size of ConvLSTM is set to 3, and GAT is implemented using an 8-head attention mechanism.

[0032] The experimental code was written in Python using the Py-Torch framework. The experiment was conducted on a platform equipped with a 5.4GHz Intel Core processor, an NVIDIA GeForce RTX 4060 graphics card, 32 GB of 5600 MHz DDR5 RAM, and a 1TB solid-state drive.

[0033] The root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the model's performance; the lower the value, the better the model's performance.

[0034] Results and Analysis Setting the edge connectivity threshold is critical to the GSSA network architecture. We have conducted a series of experiments with different thresholds to determine the network configuration that delivers superior performance metrics. Figure 5 The relationship between the RMSE and MAE of the one-day forecast and R0 is explained. The results show that the R0 value for predicting the SST in the South China Sea is 0.5, and the R0 value for predicting the SST in the Bohai Sea is 0.4.

[0035] Table 1. Forecast results for 1 day, 3 days, and 7 days from the SST dataset for the South China Sea and Bohai Sea Table 2. Weekly and monthly forecast results (RMSE and MAE) of SST data from the South China Sea and Bohai Sea Table 1 presents the forecast performance of four methods on the South China Sea and Bohai Sea SST datasets for 1, 3, and 7 days, with superior results shown in bold. Clearly, our GSSA method outperforms the other techniques across all evaluated timescales. The performance of the other three methods shows significant differences between the South China Sea and Bohai Sea datasets. For example, in the South China Sea dataset, SAGE-LSTM exhibits better accuracy on GAT due to smaller temperature variations, while in the Bohai Sea dataset, due to larger temperature fluctuations, GAT skillfully captures these fluctuations through its attention mechanism. However, GSSA effectively addresses this gap by introducing temporal convolutional layers to compensate for the shortcomings of SAGE-LSTM. On the Bohai Sea dataset, ConvLSTM's effectiveness is significantly lower than that of the South China Sea dataset, mainly due to the higher proportion of land and islands, which disrupts convolutional capabilities and weakens ConvLSTM's spatial information capture ability. Compared to the three alternative methods, our GSSA method demonstrates superior performance across all regions.

[0036] Table 2 shows the prediction performance of the four models for 1-week, 3-week, and 1-month SST. In long-term forecasts, the Bohai Sea SST fluctuates significantly, making prediction more difficult. However, GSSA consistently outperforms the other models across all time intervals. To provide a clear visualization of model performance and to facilitate discussion of prediction challenges in different regions, we chose to compare the error distributions of the SAGE-LSTM and GSSA algorithms across different times, locations, and terms. We selected specific temperature ranges to highlight areas that are difficult to predict. Overall, the GSSA algorithm improves upon the performance of SAGE-LSTM to some extent, reduces the scope of areas that are difficult to predict, and improves prediction accuracy.

[0037] This application proposes a graph spatiotemporal sampling aggregation algorithm, which mainly consists of two parts: spatial information aggregation and temporal information convolution. This algorithm aims to simultaneously consider spatial structure correlation and temporal variation, providing a comprehensive and integrated method for SST prediction. The graph spatiotemporal sampling aggregation algorithm uses adjacent sampling, capturing spatial details through spatial information aggregation and grasping temporal dynamics through temporal information convolution. This endows the model with significant inductive and generalization capabilities, effectively overcoming the limitations of SAGE aggregation in capturing the temporal correlation between node sequences. Experiments with SST data from the South China Sea and the Bohai Sea demonstrate that this algorithm outperforms traditional machine learning algorithms at various prediction levels.

[0038] In this application, we elaborate on the Graph Spatiotemporal Sampling Aggregation Method (GSSA), a novel approach that addresses the supervision problem of temporal relevance, which is often neglected by traditional methods. By integrating spatial information aggregation layers with temporal convolutional layers, spatiotemporal dynamics are effectively captured, and neighbor sampling aggregation techniques are employed to enhance the model's flexibility and generalization. Experiments determine the optimal parameters for constructing the adjacency matrix, and our method is compared with classical deep learning algorithms. Results show that GSSA outperforms other methods across various prediction ranges and regions.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A graph spatiotemporal sampling aggregation network for sea surface temperature prediction, characterized in that, First, the sea surface temperature data is divided into a grid format according to latitude, longitude, and resolution, forming N data points. These data points are then constructed into a graphical model G=(V, E), where V represents the nodes in the graph, with a count of N; E represents the edges between nodes; and the adjacency matrix A∈R... N×N To represent the connection between nodes; for any given edge, the normalized Pearson correlation coefficient is used as the threshold for edge connection, defined as follows: ; Where Pearson Correlation is the Pearson correlation coefficient of sea surface temperature data between two points; Distance is the distance between these two points on the grid; by setting a threshold, edges below the threshold R0 are removed, so any given edge... Defined as: ； The graph-based spatiotemporal sampling aggregation network for sea surface temperature prediction includes: A data preprocessing module is configured to process the input sea surface temperature sequence data. Perform neighbor sampling to obtain information about adjacent nodes; The spatial information aggregation layer, which is communicatively connected to the data preprocessing module, is configured to sequentially input information between adjacent nodes into the Long Short-Term Memory (LSTM) network aggregation module for aggregation, thereby obtaining sequence data X. (1) ; A fully connected layer, communicatively connected to the sampling module, is configured to perform a fully connected linear transformation on X to obtain sequence data X. (2) ; The temporal information convolutional layer is communicatively connected to both the spatial information aggregation layer and the fully connected layer. The temporal information convolutional layer is configured to acquire sequence data X. (1) and sequence data X (2) , obtain X (3) =concat(X (1) , X (2) ), X (3) The input temporal information convolutional layer captures the temporally relevant node sequence and extracts the predicted value; the temporal information convolutional layer includes a convld layer, a flatten layer, a ReLU layer and an FC-Layer layer set in sequence.

2. The graph spatiotemporal sampling aggregation network for sea surface temperature prediction according to claim 1, characterized in that, The data preprocessing module is configured to sample the neighbors of a specified node and pass the embedding information of these neighbors to the target node; then aggregate the embedding information to update the embedding representation of the target node; and predict the node's label based on the updated embedding.

3. The graph spatiotemporal sampling aggregation network for sea surface temperature prediction according to claim 1, characterized in that, The spatial information aggregation layer includes a long short-term memory (LSTM) network module, which obtains sequence data based on information between adjacent nodes. The specific configuration method of the Long Short-Term Memory (LSTM) network module is as follows: ； ； ； ； ； ； in, These are all learnable weight parameters from different layers. It involves concatenating x and y. It is a learnable bias parameter. It is the activation function sigmoid.

Images