Robust reconstruction of sea level fields from sparse observations using shallow recurrent decoders

This research introduces a novel data-driven approach for reconstructing high-dimensional sea level (SL) fields using sparse observational data, employing the Shallow Recurrent Decoder (SHRED) architecture. The methodology was evaluated using an extensive dataset of daily SL records collected over twelve years from 16 monitoring stations along Turkey's coastline. The SHRED framework effectively learns dominant spatiotemporal patterns, enabling accurate field reconstruction from limited sensor inputs. A key finding of this study is that the model maintains high reconstruction accuracy when using data from only a single monitoring station. SHRED successfully reconstructs spatially distributed SL fields with low reconstruction error and high efficiency values. This highlights the model's ability to generalize complex physical dynamics from minimal information. The model also shows consistent performance across a wide range of randomized sensor configurations, which indicates robustness to different sensor configurations. Most importantly, the results indicate that reconstruction accuracy is relatively insensitive to the specific locations of the input stations for the dataset considered in this study, suggesting that reliable reconstruction can be achieved using randomly selected monitoring stations. Visual analyses, including scatter plots and three-dimensional reconstructions, further confirm that SHRED produces physically coherent and realistic representations of SL behavior. These results show that the SHRED framework can be used as a data-driven tool for reconstructing SL fields from sparse observations. This approach holds significant promise for improving the representation and reconstruction of coastal hydrodynamic processes and other physical dynamical systems in data-scarce environments.