AI performs on high-resolution three-dimensional ocean temperature forecasting: remote sensing data-driven becomes a new possibility

Accurate prediction of three-dimensional (3D) ocean structures is essential for understanding oceanic processes. While AI-based ocean models demonstrate superior forecasting performance, they typically depend on numerically simulated 3D background structures as input, leading to operational limitations and significant computational expenses. This study explores a method for directly forecasting high-spatiotemporal resolution 3D ocean temperature structures using multisource remote sensing data in ‘2D-to-3D’ mode, comparing it with predictions from 3D numerical simulation background structure profiles in ‘3D-to-3D’ mode. We propose a multiscale residual spatiotemporal window attention model (MSWO) for 1/12° resolution forecasting. Extensive experiments are conducted using the world-leading ocean prediction intercomparison and validation task team (IV-TT) Class4 intercomparison framework to evaluate the model's performance. Benchmarked against mainstream forecasting systems, MSWO achieves comparable accuracy to operational models in 2D-to-3D mode and superior accuracy in 3D-to-3D mode. Furthermore, the MSWO model outperforms other data-driven artificial intelligence models in terms of training cost and accuracy. This study demonstrates the feasibility of deriving high-spatiotemporal-resolution 3D ocean forecasts from satellite remote sensing.