Deep Learning for Ocean Parameters Prediction: A Systematic Literature Review

Abstract Accurately predicting ocean temperature, salinity, and velocity is fundamental to capture the dynamics of subsurface thermohaline structure and understanding its impacts on marine ecosystems and the Earth’s climate system. With the rise of Deep Learning (DL), researchers have explored its potential to capture complex spatiotemporal dependencies in ocean parameter prediction. However, a comprehensive understanding of the application of DL in this field remains in its early stage. To address this gap, we conduct a systematic review of 122 peer-reviewed articles published between 2013 and 2024, aiming to answer four key Research Questions (RQs). In RQ1, we examine how marine datasets are categorized and represented for ocean parameter prediction. RQ2 explores the spatial, temporal, and input-level patterns of these datasets as utilized across various DL models. In RQ3, we analyze the DL architectures adopted for forecasting temperature, salinity, and velocity, highlighting trends across model types. Finally, RQ4 investigates the optimization strategies and evaluation metrics employed to assess predictive performance. Based on our findings, we highlight underexplored challenges and outline promising directions for future research.