A Novel Method for Sea Surface Temperature Prediction using a Featural Granularity-Based and Data-Knowledge-Driven ConvLSTM Model

Abstract. Many data-driven methods predict sea surface temperature (SST) at specific locations using only the previous period SST values as predictors, which ignores the spatiotemporal dependencies of SST variability and the influence of multiple variables on SST patterns. Additionally, these methods have difficulty in capturing the feature dependencies involved in SST fluctuations, limiting the accuracy and horizon of SST predictions. This study proposes a new medium- and long-term forecasting model to address these issues, which includes two sub-models: a featural granularity model and a data-knowledge-driven ConvLSTM prediction model. The former restacks the one-dimensional time-series of each variable into multidimensional feature variables using an adaptive granulation-based method. The latter integrates parameters that affect ocean dynamics and thermodynamics, along with pixel-to-pixel similarity, to achieve partition predictions. The multidimensional feature variables were fed into the ConvLSTM model to exploit the feature- spatiotemporal patterns for predictions. Experiments conducted in three sea areas of the western Pacific and Indian Oceans indicate that the use of featural granularity can enhance the ability of the prediction model to capture dynamic characteristics in the time domain and internal dependencies of the features and extend prediction horizons. The combination of knowledge-driven and study area segmentation concepts can help the prediction model better capture the unique features and dynamics of the local area, further improving prediction accuracy. Validation against observations and cross-comparisons with baseline models in three different sea areas for the prediction of monthly SST with lead times ranging from 1 to 120 months demonstrate that the proposed model can generate consistent and more accurate regional SST predictions. The differences between predicted and observed values range from -0.7 to 0.7 K, with an RMSE of approximately 0.3 to 0.57 K for SST predictions. This developed model provides a promising approach for medium- and long-term sea surface temperature forecasting, which can be easily adapted to other ocean parameter prediction tasks.