TL;DR
This paper introduces a novel 3D U-Net deep learning model that integrates multiple data sources (SST, SSHA, SSW) to accurately predict sea surface temperature (SST) and marine heat waves (MHWs) in the South China Sea. The model significantly outperforms traditional methods and ConvLSTM, offering improved accuracy and a new tool for disaster prevention and climate change research.
What Problem Does This Paper Address?
Accurate prediction of sea surface temperature (SST) is crucial for understanding and mitigating the impacts of climate change, such as extreme weather events and ocean circulation changes. Traditional physics-based numerical models are computationally expensive and slow. While data-driven methods, particularly deep learning, have shown promise, many existing models focus on single variables and overlook the complex interplay between different oceanic factors. This limits their physical relevance and predictive accuracy, especially in dynamic regions like the South China Sea (SCS), which is influenced by complex monsoon systems and ocean currents. This paper aims to address these limitations by developing a deep learning model that can effectively integrate multi-source data for improved SST and marine heat wave (MHW) prediction.
Methods
The core of this research is a novel 3D U-Net deep learning model. The authors adapted the standard 2D U-Net architecture, commonly used for image segmentation, by extending its convolutional operations to three dimensions. This allows the model to process not only spatial information but also temporal sequences across multiple input variables simultaneously. This 3D approach is designed to capture intricate spatiotemporal correlations between different sea surface variables.
For comparison, the study also employs a Convolutional Long Short-Term Memory (ConvLSTM) model, a well-established deep learning architecture for spatiotemporal forecasting. The 3D U-Net model was trained using a historical data window of 64 days to predict SST for the subsequent 30 days.
Datasets
The study utilizes the following datasets for the South China Sea region:
- Sea Surface Temperature (SST): NOAA Daily Optimum Interpolation SST (OISST) version 2.1 dataset (0.25° resolution), covering September 1, 1981, to the present.
- Sea Surface Height Anomaly (SSHA): Copernicus Marine and Environmental Monitoring Service (CMEMS) dataset (0.25° resolution), providing daily SSHA from January 1, 1993, to August 4, 2022.
- Sea Surface Wind (SSW): Cross-Calibrated Multi-Platform (CCMP) dataset v2.0 (0.25° resolution), providing eastward and northward SSW from July 10, 1987, to the present, with a temporal resolution of one-fourth of a day.
The study period for training and testing was from January 1, 1993, to December 31, 2021. Data was normalized before model training, and land areas were filled with zeros.
Key Insights and Ideas
- The power of multi-source data: Integrating SSHA and SSW alongside SST significantly improves prediction accuracy, highlighting that these variables are not just noise but crucial regulators of SST and MHW dynamics.
- Temporal dynamics of variable influence: The study reveals that SSW plays a more dominant role in early-stage predictions, while SSHA's influence grows as the forecast lead time increases, suggesting a complex, time-dependent interaction.
- 3D U-Net's spatiotemporal advantage: The 3D convolutional approach effectively captures both spatial patterns and temporal evolution across multiple variables, a key advantage over 2D models or those that treat variables independently.
- MHW detection as a novel application: Beyond SST forecasting, the 3D U-Net model demonstrates remarkable capability in accurately detecting and characterizing marine heat wave events, opening new avenues for early warning systems.
Key Results
- Superior SST Prediction: The 3D U-Net model consistently outperformed the ConvLSTM model across all lead times (1-30 days) and seasons.
- Accuracy Metrics: At a 30-day lead time, the 3D U-Net achieved MAE of 0.51 °C, RMSE of 0.69 °C, MAPE of 1.83%, and R of 0.93. This was better than ConvLSTM's MAE of 0.57 °C, RMSE of 0.77 °C, MAPE of 2.01%, and R of 0.92.
- Spatial Accuracy: Spatially, the 3D U-Net model exhibited low errors (RMSE < 0.5 °C) and high correlation (R > 0.9) across most of the SCS.
- Accurate MHW Detection: The model successfully captured the occurrence frequency, total duration, average duration, and average cumulative intensity of MHW events in the SCS for 2021, closely matching observed data.
- Robustness across Regions: The model demonstrated reliable performance in diverse regions of the SCS, with RMSE values generally below 0.54 °C and R values above 0.89 in selected test areas.
Limitations and Future Work
- Data-driven limitations: As a data-driven model, the 3D U-Net can still exhibit underestimation or overestimation, and MHW prediction is influenced by complex ocean dynamics and thermodynamics not fully captured by the input variables alone.
- Future Research Directions:
- Integrating Ocean Dynamics: The authors plan to incorporate more ocean dynamic mechanisms into the model to further improve SST and MHW prediction accuracy and forecasting skills.
- Expanding Parameter Prediction: The model's architecture could be extended to forecast other essential ocean parameters, such as sea surface wind, sea surface height, and thermohaline structure.
- Broader Applications: The model holds potential for applications in marine disaster prevention, marine ranching, and environmental protection.
Summary updated