A Depth-Aware Ensemble Deep Learning Framework With Layer-Wise Recursive Prediction for Multisource Subsurface Ocean Temperature and Salinity Forecasting

Abstract

Accurate prediction of upper ocean temperature (UOT) and salinity (SS) within the critical 0–50-m depth range is essential for understanding climate variability and supporting sustainable marine ecosystem management. However, existing approaches often struggle to capture complex spatiotemporal dependencies and explicitly model vertical interactions across ocean depths, particularly when integrating heterogeneous data sources. This study proposes a depth-resolved upper ocean forecasting framework with three main contributions. First, a structured multisource data fusion strategy integrates satellite remote sensing products with in situ Argo float observations to enhance the spatial and vertical representation of upper ocean conditions. Second, an ensemble multimodel (EMM) architecture combines complementary deep learning models, including a transformer-based spatiotemporal transformer (TST) model, a standard transformer, long short-term memory (LSTM), and gated recurrent unit (GRU) networks, to learn oceanographic dynamics from heterogeneous observations. Third, a layer-wise depth recursive prediction (LDRP) strategy conditions deeper layer predictions on predicted states from shallower layers, enabling explicit modeling of vertical dependencies without relying on deeper layer ground truth during inference. Experiments conducted on a multisource Western Pacific dataset demonstrate consistent improvements in depth-resolved forecasting accuracy across multiple subsurface layers compared with standalone baseline models, with stronger gains observed for temperature forecasting and task-dependent improvements for SS forecasting. The proposed method provides a robust and depth-resolved forecasting framework for upper ocean monitoring, climate analysis, and marine resource management.