Deep learning for temporal reconstruction of FESOM-derived sea surface temperature and 3D ocean variables

In oceanography, the growing volume of data collection raises the challenge of efficient data storage and computational efficiency for ocean simulations. To address this, we adopted an interpolation-based approach for reducing data storage requirements while ensuring accurate reconstruction of sea surface temperature (SST) and 3D ocean variables, including temperature and horizontal velocities (u and v) derived from the Finite-volumE Sea ice–Ocean Model (FESOM2).
Our research explores three distinct strategies: single-step interpolation, temporal-window interpolation, and multi-step interpolation. To achieve these goals within a supervised learning framework, we propose a hybrid CNN-BiLSTM deep learning model for both accurate interpolation and data storage optimization. This model extracts features from time-series data by combining the strengths of convolutional neural networks (CNNs) for capturing spatial patterns and bidirectional long short-term memory (BiLSTM) networks for capturing temporal dependencies in time-series data. Additionally, our approach uses a multi-input multi-output (MIMO) strategy to improve computational efficiency and prevent error accumulation, resulting in high prediction accuracy. By incorporating neighborhood information and preprocessing data to consider spatial and temporal dependencies, our approach effectively prepares the data for training. We use subsampling to reduce memory usage and introduce diverse, uncorrelated samples.
The results show that our model outperforms traditional methods, including linear interpolation (LI) and linear regression (LR), achieving lower mean squared error (MSE) and higher correlation coefficients when reconstructing ocean variables. Single-step interpolation achieved a 75% improvement, temporal-window interpolation demonstrated an overall 46.80% improvement, and multi-step interpolation achieved an overall 59.51% improvement in predictive accuracy. Remarkably, the model predicts multiple time steps for 3D ocean fields without introducing artifacts, achieving this within a single model framework.
After training on diverse datasets, the model generalizes well to unseen data without requiring additional computational resources, making it a scalable and sustainable solution for addressing the growing need for data storage efficiency in oceanographic research. To improve model performance, key techniques were applied, including overfitting prevention, balanced batching, multi-GPU training, adaptive learning rate schedules such as polynomial decay and cyclic learning, and fine-tuning.
We also demonstrate the practical application of our model by computing ocean heat transport in zonal and meridional directions. The results confirm that our model predicts ocean heat content more accurately than traditional methods. However, the computation of ocean heat transport using mixed-sign velocity values presents challenges, despite individual velocity predictions being highly accurate. This work highlights the potential of advanced neural networks in reconstructing temperature and velocity data, accurately matching observed values, and enhancing our understanding of ocean heat dynamics.