Feng Nan, Zhuolin Li, Jie Yu, Suixiang Shi, Xinrong Wu, Lingyu Xu. Prediction of three-dimensional ocean temperature in the South China Sea based on time series gridded data and a dynamic spatiotemporal graph neural network[J]. Acta Oceanologica Sinica.
Citation:
Feng Nan, Zhuolin Li, Jie Yu, Suixiang Shi, Xinrong Wu, Lingyu Xu. Prediction of three-dimensional ocean temperature in the South China Sea based on time series gridded data and a dynamic spatiotemporal graph neural network[J]. Acta Oceanologica Sinica.
Feng Nan, Zhuolin Li, Jie Yu, Suixiang Shi, Xinrong Wu, Lingyu Xu. Prediction of three-dimensional ocean temperature in the South China Sea based on time series gridded data and a dynamic spatiotemporal graph neural network[J]. Acta Oceanologica Sinica.
Citation:
Feng Nan, Zhuolin Li, Jie Yu, Suixiang Shi, Xinrong Wu, Lingyu Xu. Prediction of three-dimensional ocean temperature in the South China Sea based on time series gridded data and a dynamic spatiotemporal graph neural network[J]. Acta Oceanologica Sinica.
Prediction of three-dimensional ocean temperature in the South China Sea based on time series gridded data and a dynamic spatiotemporal graph neural network
Ocean temperature is an important physical variable in marine ecosystems, and ocean temperature prediction is an important research objective in ocean-related fields. Currently, one of the commonly used methods for ocean temperature prediction is based on data-driven, but research on this method is mostly limited to the sea surface, with few studies on the prediction of internal ocean temperature. Existing graph neural network-based methods usually use predefined graphs or learned static graphs, which cannot capture the dynamic associations among data. In this study, we propose a novel dynamic spatiotemporal graph neural network (DSTGN) to predict three-dimensional ocean temperature (3D-OT), which combines static graph learning and dynamic graph learning to automatically mine two unknown dependencies between sequences based on the original 3D-OT data without prior knowledge. Temporal and spatial dependencies in the time series were then captured using temporal and graph convolutions. We also integrated dynamic graph learning, static graph learning, graph convolution, and temporal convolution into an end-to-end framework for 3D-OT prediction using time-series grid data. In this study, we conducted prediction experiments using high-resolution 3D-OT from the Copernicus global ocean physical reanalysis, with data covering the vertical variation of temperature from the sea surface to 1,000 m below the sea surface. We compared five mainstream models that are commonly used for ocean temperature prediction, and the results showed that the method achieved the best prediction results at all prediction scales.
Figure 1. (a) The blue box represents the experimental area. (b) A 3D representation of the experimental area, displaying average temperatures from 2000 to 2019, with the color bar above indicating sea surface temperature and the color bar below indicating internal ocean temperature.
Figure 2. Real seawater temperature profiles at three points in the experimental area. Points A (10°N, 111.25°E) and B (10°N, 112°E) are located at the sea surface, and point C (10°N, 111.25°E) is located 30 m below the sea surface.
Figure 3. The overall framework of DSTGN. A: The blue cube represents the raw input $ \in {R}^{T\times D\times {L}_{1}\times {L}_{2}} $ , which is first converted to a 2D tensor $ \chi \in {R}^{T\times N} $ and then sampled using a sliding window with input window size $ {W}_{in} $ and output window size $ {W}_{out} $ to obtain the input $ \chi \in {R}^{B\times {W}_{in}\times N} $ and the true label $ {Y}_{true}\in {R}^{B\times {W}_{out}\times N} $ . B: The input data and node embedding matrix are dynamically learned to obtain the dynamic graph matrices, while the node embedding matrix is statically learned to obtain the static matrix. C: Details of the dynamic graph learning. D: The original data and the two types of graph adjacency matrices are transformed by K-layers of temporal convolution module and graph convolution module to extract features, and the features extracted by the temporal convolution module in each layer are linked to the output module through skip connections. Finally, the predicted result $ {Y}_{pre}\in {R}^{B\times {W}_{out}\times N} $ is obtained.
Figure 4. Seawater temperature trends at different depths at the same location (10°N, 111.25°E).
Figure 5. Comparison of the prediction results of different models.
Figure 6. Comparison of the prediction results of different models at different depths.
Figure 7. Fit plot of average temperature predictions at different depths. The blue curve represents the average actual temperature in the experimental area, and the orange curve represents the average predicted temperature in the corresponding area.
Figure 8. Visualization of the 3D-OT distribution and the predicted results of DSTGN for different depth layers. "Tem_real" represents the true temperature, "Tem_DSTGN" represents the predicted temperature by DSTGN, and "Abs_error" represents the absolute error, which is defined as |Tem_DSTGN – Tem_real|.
Figure 9. Prediction error distributions of different models in different ocean temperature layers. The error plots of SVR, FC-LSTM, TPA-LSTM, DSANet, Graph WaveNet, and DSTGN are 0 m, 30 m, 100 m, 500 m, and 1,000 m under the sea from top to bottom. The color bar on the right represents the MAE value.
Figure 10. Prediction Error Curves in Different Input Windows.
Figure 11. Boxplots of Prediction Errors with Different Output Windows and Input Windows.
Figure 12. Pearson correlation coefficients for points A, B and C (Fig. 1). The time range is from 1 January 2016 to 31 December 2019, the green curve represents the Pearson correlation coefficient for the true temperature and the red curve represents the Pearson correlation coefficient for the predicted values of the DSTGN.