A team from UC San Diego's Scripps Institution of Oceanography has published a new method for deriving high-resolution ocean current maps from existing satellite data. Their system, GOFlow, applies a trained neural network to thermal imagery from geostationary weather satellites, producing velocity fields at a granularity and frequency that surpass conventional satellite altimetry.

The engineering approach centered on a supervised learning model. Researchers first trained a convolutional neural network on a corpus of simulated ocean current data, establishing a baseline for the relationship between thermal patterns and subsurface flow. This model was then applied to real-world thermal images from the GOES satellite series. By tracking the displacement of surface temperature features across sequential images, the network infers the velocity vectors of the underlying currents.

Validation presented a key system design challenge. The team benchmarked GOFlow's outputs against in-situ measurements from research vessels in the Gulf Stream and against standard satellite-derived products. While results aligned with these established sources, GOFlow provided spatial and temporal detail previously only achievable in high-fidelity computational fluid dynamics models, not from direct observation.

Luc Lenain, the project's lead, framed the work as a data extraction problem. "The physics are already present in the satellite signal," he noted. "Our model helps recover that information, which traditional methods have struggled to isolate efficiently." The architecture acknowledges limitations, primarily data gaps from cloud cover, with plans to integrate multi-spectral satellite sources to improve coverage.

In a move for practical adoption, the team is releasing the full codebase publicly. This positions GOFlow not as a black-box solution, but as a reproducible component for processing large-scale remote sensing datasets, offering a template for merging physical observation with machine learning inference.

Source: CNET