Grid-Encoded Neural Networks for Precise Interpolation of Oceanic Temperature and Salinity

The spatial-temporal interpolation of ocean temperature and salinity is a fundamental and critical task in oceanography. Traditional interpolation method can cause a great loss of accuracy. This study introduces a novel Grid-Encoded Neural Network (GENN) approach to enhance the precision of ocean temperature and salinity interpolation. By discretizing geographic coordinates into a grid system, GENN emphasizes temporal correlations, surpassing the performance of conventional Deep Neural Networks (DNNs) in efficiency and accuracy. Specifically, case studies in the North Atlantic and South China Sea regions validate the method's effectiveness and highlight its superiority in various marine environments. Also, GENN demonstrates quicker convergence and lower loss function values, indicative of more effective model parameter optimization and a reduced risk of overfitting. This research emphasizes the potential of GENN for three-dimensional ocean temperature and salinity reconstruction, expands the application of machine learning in oceanography.