New S-DEIM Method Estimates Sea Surface Temperatures 40% More Accurately

Researchers at North Carolina State University have developed a computational method, called Sparse Discrete Empirical Interpolation Method (S-DEIM), that estimates global sea surface temperatures from sparse data more accurately than existing techniques and faster than the best AI model tested against it. The team was led by Mohammad Farazmand, associate professor of mathematics at NC State.

Sea surface temperature (SST) data underpins weather forecasting, climate predictions and marine-ecosystem research, but real-world data is often sparse. Buoys measure temperatures accurately but are limited in number, while satellites cover more area, though atmospheric conditions can interfere with their readings. Federal agencies such as the National Oceanic and Atmospheric Administration (NOAA) have historically relied on complex differential equations to fill these gaps, Farazmand said.

S-DEIM builds on an existing technique, the Discrete Empirical Interpolation Method (DEIM), which specifies a basis, or combination of patterns, describing the SST field being estimated. DEIM does not perform well with sparse data. To compensate, S-DEIM uses historical data to estimate a "kernel vector," a value with no closed-form formula, that helps fill in missing measurements.

The team tested S-DEIM against DEIM and a convolutional neural network (CNN), the best-performing AI model in the comparison, using 30 years of NOAA sea surface temperature data. They withheld the final year, had each method predict it, then compared the predictions to the historical record. S-DEIM was 40% more accurate than DEIM and 2% more accurate than the CNN, while taking only one minute to train compared with one and a half hours for the CNN.

The work, part of a research experience for undergraduates, appears in the Journal of Geophysical Research: Machine Learning and Computation. Co-authors include Cassidy All of the University of Colorado Boulder, Kevin Ho of Mississippi State University, Maya Magnuski of Bard College, Christopher Nicolaides of Indiana University, and NC State graduate student Louisa Ebby. Farazmand said the researchers hope to keep improving S-DEIM's accuracy.