Skip to main content
Back to timeline
arXivSource publication:

SWOTnet carries wide-swath resolution into conventional altimetry, finding about 30% more eddies with the same weights

Synopsis

Detecting mesoscale ocean eddies has long relied on hard-coded thresholds and a set of physical constraints that is not uniquely defined, making results sensitive to small changes in the input fields. A multimodal, multi-task model based on a shallow U-Net segments eddies while reconstructing the co-located sea surface temperature anomaly, yielding two variants: NADIRnet trained on conventional nadir altimetry and SWOTnet trained through a curriculum on SWOT-enriched maps. Applied to conventional nadir altimetry, SWOTnet detects about 30% more eddies than the dynamical algorithm and generalises without retraining to altimetry products and processing levels unseen in training.

Source-provided article image: Improving Eddy Detection from Wide-Swath Altimetry with Multimodal Data-Driven Models
arXiv

Interpretation

Eddy detection no longer rests only on a set of hard-coded thresholds and constraints; a multi-task network outputs both an eddy mask and a sea surface temperature anomaly. Earlier dynamical detectors rely on a set of necessary physical conditions that is not uniquely defined and that merges several nearby vorticity structures into one. On the validation split, NADIRnet reaches a pixel-wise F1 of 0.81 and SWOTnet 0.78; strict structure-wise F1 ranges are 0.77 to 0.88 and 0.73 to 0.86, with residual disagreement concentrated at boundaries rather than in missed or spurious detections.

The network takes sea surface height, sea surface temperature, geostrophic velocity components and phase velocities as input and processes the full global grid in a single forward pass. A shallow U-Net encoder downsamples and doubles channel width at each stage, while a decoder with transposed convolutions restores resolution and skip connections preserve fine spatial detail. Training uses a 1/10-degree global grid with an 80/20 training-validation split; the full single-day detection cycle runs in under a minute on a laptop CPU.

The sea surface temperature regression branch is not a side output; through boundary-coupling and gradient-matching terms it constrains the geometric quality of the detected eddy boundaries. Trained with the detection term alone, boundaries can be irregular and noisy and fail to follow the sea surface height anomaly extremum. Adding the regression and boundary-coupling terms visibly smooths the boundary, and adding the gradient term makes the contour track the sea surface height gradient, shown for a representative eddy.

Perspective

The demonstrated scope is mesoscale eddy segmentation and sea surface temperature anomaly reconstruction on a global 1/10-degree grid, evaluated against a dynamical reference eddy field rather than the true eddy field; model weights and training and evaluation code are public.

The optimal thresholds selected on the validation products are not guaranteed to remain optimal for products with different resolution, noise or processing chains; for L3 swaths a higher threshold was used to avoid noisy detections at swath borders.