Long-lead daily sea surface salinity prediction using time-series CCI L4 SSS satellite data and a temporal convolutional deep learning model

Abstract

Accurate long-lead daily Sea Surface Salinity prediction remains a significant challenge, primarily attributed to complex oceanic dynamics and the cumulative propagation of prediction errors over extended time frames. Existing methodologies, encompassing classical machine learning approaches and recurrent deep learning architectures, struggle to balance computational efficiency with the modeling of long-range temporal dependencies in SSS time series. This study introduces a temporal convolutional network (TCN)-based model to address these challenges, leveraging dilated causal convolutions to model multi-scale SSS dynamics while mitigating error accumulation in long-lead forecasting. The proposed deep learning model enables the automatic capture and modeling of SSS temporal dependencies using only historical time-series SSS data from satellite remote sensing. Comprehensive experiments conducted at eight geographically dispersed sites in the Indian Ocean, utilizing European Space Agency (ESA) Climate Change Initiative (CCI) Level 4 (L4) SSS satellite data, demonstrate that the proposed model outperforms both baseline machine learning and deep learning models, demonstrating its superior capability for long-lead daily SSS prediction.