SWO: A Lightweight Window Spatiotemporal Attention Network Reconstructs Subsurface Temperature Structure

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 19274 - 19287
• Main Authors: Jiang, Jiawei, Wang, Jun, Liu, Yiping, Feng, Liqiang, Jiang, Qiufu, Huang, Chao, Xiang, Liang, Zhang, Xiangguang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial intelligence; Computational modeling; Computer applications; Computer architecture; Computing costs; Dynamic topography; High resolution; Image reconstruction; Image resolution; Lightweight; Ocean temperature; Oceans; Remote sensing; Satellite imagery; Satellite observation; satellite observations; Satellites; Sea surface temperature; spatiotemporal attention mechanism; Spatiotemporal data; Spatiotemporal phenomena; Spatiotemporal Window Ocean (SWO); Surface salinity; Temperature requirements; temperature structure reconstruction
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3427845

Satellite remote sensing enables the extensive, long-term observation of oceanic changes. To achieve transparent ocean observation, artificial intelligence is innovatively used to reconstruct the three-dimensional ocean structure from remote sensing data with spatiotemporal attention. However, high-resolution satellite imagery imposes high computational demands. This study proposes Spatiotemporal Window Ocean (SWO), which is an efficient network that uses a special calculation strategy and spatiotemporal window attention to reduce complexity. The model integrates multiple satellite data sources, including sea surface temperature, absolute dynamic topography, and sea surface salinity. Our experimental results demonstrate that SWO achieves lower computational costs and superior performance compared with recent commonly used spatiotemporal sequence models. Specifically, SWO requires a training time equivalent to 1/3 of SimVP, 1/4 of PredRNN, and 1/5 of SA-ConvLSTM while achieving a root mean square error index that is 13.8%, 20.3%, and 13.2% better, respectively. The computational advantages of SWO offer an important technical means for the high-resolution reconstruction of ocean phenomena in the future.