Four‐Dimensional Mapping of Sea Temperature and Salinity in the Coastal Waters Based on Generalized Spatio‐Temporal Autoregressive Neural Network

Abstract

Abstract The lack of sufficient in‐situ ocean observations significantly limits progress in studying the underwater marine environment. Fully analyzing the 4‐dimensional distribution of observation data, especially the changes over time and depth, allows for a broader and deeper application of marine data mining. This study introduces a tool named Generalized Spatio‐Temporal Autoregressive Neural Network (GSTARNN), which is designed to interpolate monthly gridded temperature and salinity data derived from underwater sampling in the coastal area of Zhejiang Province, China, from 2014 to 2018. Compared with five widely used geographical interpolation models, the results indicate that GSTARNN provides enhanced overall interpolation accuracy within the research area, with lower temperature and salinity errors relative to the other models. The ARMOR3D and IAP products were used to confirm the reliability of GSTARNN, with R2 of 0.81 for temperature and R2 of 0.935 for salinity, exhibiting that the reconstructed field retains the detailed spatio‐temporal data from on‐site observations. The exceptional performance of the proposed model in spatio‐temporal analysis successfully depicts the distribution patterns of temperature and salinity across both time and depth, with a resolution of 1/30° × 1/30°. The model reveals consistent seasonal patterns in ocean temperature and salinity from 2014 to 2018, with temperatures peaking in August and reaching their lowest in March, while salinity remains relatively stable but shows significant depth‐dependent variations, consistently reaching its minimum in October. The experimental results demonstrated the potential for robust and reliable application in marine environmental studies, enabling more meticulous insights into oceanic dynamics.

Keywords

• four‐dimensional spatio‐temporal interpolation
• subsurface ocean
• space‐time geostatistics
• autoregressive neural network