We leverage SWOT satellite data see e.g.,for some background. We use the latest release (v3.0) of Level-3 SWOT data, namely the SWOT_L3_SSH “Basic” (2 km resolution) product, derived from the L3 SWOT Ka-band Radar Interferometer (KaRIn) Low rate ocean data products provided by NASA/JPL and CNES. The Level-3 processing removes SWOT's systematic errors, and has been extensively validated using other altimeters, numerical models, and in situ data, in the global ocean . These datasets are produced and freely distributed by the AVISO and DUACS teams as part of the DESMOS Science Team project .

Here, we use the “filtered” Sea Level Anomaly (SLA) variable. The filtering procedure is described in . It is based on a U-Net-based convolutional neural network, trained on simulated North Atlantic data, and shown to outperform classical filtering methods. The method reduces significantly the noise (by a factor of 2), while preserving the balanced part of the signal at spatial scales down to 10 km . This method allows extraction of most of the balanced signal from the full Sea Surface Height (SSH) field, but it also reduces the overall energy, which can complicate interpretation at small scales where signal levels approach the noise floor see e.g.,. This effect primarily impacts the high-wavenumber part of the spectrum, where reduced energy can become comparable to instrumental noise. However, this limitation does not affect our analysis, as we focus on mesoscale structures whose signal amplitude remains well above the noise level . Recent studies have indeed demonstrated SWOT's ability to robustly resolve mesoscale variability that was previously undetected in gridded products .

We consider data during the “science phase” (with a 21 d repeat orbit), from cycle #6 to cycle #41, covering the period from 11 November 2023 to 17 November 2025, thus two whole seasonal cycles. We extract SWOT measurement in the Labrador Sea area, which represents 74 SWOT passes in total (see one cycle of measurement in the area of interest in Fig. ).

We complement the SWOT satellite observations with ship-based velocity measurements obtained from Shipboard Acoustic Doppler Current Profilers (SADCPs). SADCPs provide high-resolution profiles of horizontal current velocity along the ship track by measuring the Doppler shift of acoustic signals reflected by waterborne scatterers. The instruments operated at frequencies between 38 and 150 kHz, providing near-surface velocity estimates at an along-track resolution of roughly 200–300 m. SADCP data collected during MSM129 in June–July 2024 aboard the RV Maria S. Merian is used, with the instrument operating at 75 kHz, and providing near-surface velocity estimates at an along-track resolution of roughly 200–300 m. The data were processed following the same procedure as for MSM74, as detailed in . The processed ship-based velocity measurements were then used to fit idealized eddy solutions along the SADCP transects, allowing reconstructions of eddy centers and key dynamical properties (see for methodological details). These reconstructions provide a robust in situ “truth” against which we can directly assess the performance and reliability of the SWOT-based eddy detection, enabling a direct comparison of mesoscale eddy structures between SWOT and MSM129 (Fig. ). In addition, we compare the statistical results of , which are based on SADCP observations from the MSM40, MSM54, and MSM74 cruises and comprise 40 well-resolved eddies with radii of 3–39 km and azimuthal velocities of 7–58 cm s⁻¹, with the SWOT-based statistics, as shown in the histograms in Fig. .

Eddy detection statistics from a classical 1/4° satellite altimetry are also used for comparison purpose. The detection was performed by using the AMEDA algorithm Angular Momentum Eddy Detection and Tracking Algorithm;, and is referred to as “CMEMS” for Copernicus Marine Environment Monitoring Service in e.g., Fig. b and d. Note that these eddy statistics are only representative of eddies along or close to the ship tracks of MSM40, MSM54, and MSM74 campaigns. Also, note that higher-resolution products such as the DUACS DT2024 1/8° fields are now available and could improve the representation of smaller-scale structures (see e.g., Figs. c and d). Nevertheless, we do not expect that increasing the resolution from 1/4 to 1/8° would qualitatively alter the main conclusions presented in Fig. b and d.

We use the sea ice concentration from the OSI SAF dataset , and the ETOPO Global Relief Model dataset to assess respectively the sea ice concentration, and the depth at each eddy location.

Each SWOT pass consists of two adjacent swaths, each approximately 50 km wide, resulting in a combined swath width of about 120 km. These swaths are separated by a nadir gap of approximately 20 km, where no KaRIn measurements are collected. See for example the “raw” pass #563 of cycle #10 in the top left of Fig. . The swaths are oriented along the satellite's ground track, providing detailed across-track coverage. Within each pass, the measurement grid is composed of 69 across-track pixels, and N along-track pixels, with each pixel representing a 2 km by 2 km area on the Earth's surface. Measurement gaps can occur within the swaths due to various factors, including rain cells which can attenuate the radar signal and lead to missing data in certain regions, or sea ice for example.

We apply an eddy-detection procedure directly on the 2 km resolution swath product of SWOT, without performing any subsequent gridded assimilation, in contrast to classical gridded altimetric product based eddy detection approaches, that now include SWOT measurements see e.g.,who used MIOST, the Experimental multimission gridded L4 sea level heights and velocities with SWOT KaRIn data. Working in the native swath geometry maximizes our ability to identify small‐scale features that might otherwise be lost in the smoothing and gridding stages of standard products. Additionally, the mesoscale eddies in our high-latitude study region are smaller than this swath width, hence an eddy can typically wholly be contained within a single swath pass – enabling reliable detection and sizing. The procedure is as follows (see also Fig.  for an example):

We use the native 2 km resolution SWOT swaths, which are 69 pixels wide in the across-track direction. To enable efficient processing, each swath is first extended by padding its lateral edges with NaNs values (step #1 in Fig. a), leading to 128-pixel wide swaths. The padded swath is then subdivided into overlapping 128×128 pixel windows, with an overlap of 64 pixels between adjacent tiles (step #2 in Fig. a). Each tile therefore contains the original data along with regions of missing values (NaNs), including the nadir gap. These gaps are then filled independently within each tile using a biharmonic inpainting method, resulting in complete 128×128 SLA images (step #3 in Fig. a). Specifically, we use the inpaint_biharmonic function from the skimage.restoration Python module, which reconstructs missing values by solving the biharmonic equation (∇4u=0) inside the gaps. Note that the tiling procedure is needed to make the inpainting method computationally tractable.

This inpainting method ensures a smooth interpolation by minimizing the Laplacian of the interpolated field, producing a physically consistent and continuous surface that preserves the mesoscale eddies structure and allows closed contours to form even across the gap between the two swaths. Note that the gap-filling methodology may lead to misdetections, particularly for larger eddies, as these structures more frequently intersect gap regions; however, eddies with radii ≲15 km remain largely constrained by observations, so the inpainting does not significantly bias the statistics in the size range considered here (see Appendix A for details).

Eddies are then identified in each sub-image using the py-eddy-tracker package . We relied on this eddy detection method with standard parameter choices, ensuring robustness, reproducibility, and consistency with existing studies. Note that sensitivity to the shape error parameter used in py-eddy-tracker was conducted, and it was shown to not bias the statistical properties of the eddy population, we therefore used the value of 85 % to maximize the number of detections.

Eddies are assumed to be predominantly geostrophic, and although ageostrophic (cyclostrophic) corrections may become relevant for the most intense structures e.g.,, they are neglected here as a first-order approximation, following standard practice for mesoscale structures in altimetric data.

We sort detected eddies using a set of predefined criteria to ensure robustness (step #4 in Fig. a). First, a maximum of 40 % of the eddy contour (defined by the region of maximum azimuthal velocity) is allowed to overlap with gaps in the SLA field, preventing spurious features generated by the inpainting procedure. This condition is suppressed when the eddy center lies near the middle of the pass (in the gap between swaths) and a tolerance of 60 % of gaps is then allowed. Finally, only eddies exceeding a minimum radius of 5 km and a minimum amplitude of 0.01 m are retained. We chose relatively conservative thresholds to ensure robust detection of balanced structures in the filtered SWOT SSH fields, minimizing spurious detections due to noise or residual unbalanced motions; while this choice may exclude the weakest eddies, it does not significantly affect the statistical distributions and ensures the reliability of the analysis. Note that sensitivity tests to all these parameters were performed to choose their values – low variability of probability distributions was obtained from these sensitivity tests.

The remaining eddy arrays from all sub-images are merged, with duplicates in the overlapping regions removed, providing a complete set of eddy detections for each SWOT pass (step #5 in Fig. a).

Finally, for each cycle, duplicates are removed by identifying eddies whose centers are separated by a distance shorter than their radius of maximum velocity (hereafter radius or Rmax for conciseness) and a distance threshold of 10 km. This threshold corresponds to the typical displacement of an eddy advected by Rossby wave propagation over one SWOT repeat cycle. When duplicates are identified, one of the two eddies is retained without preferential selection – specifically, the first occurrence in the sequence of passes is kept. This ensures that the procedure does not introduce systematic biases in eddy characteristics. As an example, the anticyclonic eddy located at 60° N, 54° W (left panel in Fig. b) is detected in multiple passes during the cycle but retained only once after merging.

Eddies detected at a position where sea ice concentration is larger than 15 % are removed. Further, for the comparison with ship-derived eddy characteristics (see Sects.  and ), we only consider eddy detections at depths larger than 2000 m to solely focus on eddies in the central Labrador Sea, similarly to .

This methodology therefore gives for each SWOT cycle all detected eddies in the area of interest (see Fig. b). It is important to note that an eddy may be re-detected in successive cycles; therefore, this methodology does not provide the absolute number of eddies, but rather a detection probability per cycle. The method may also detect a few submesoscale vortices where RD is large, but these cases fall near the mesoscale–submesoscale transition (R/RD∼1; see Fig. ). We therefore treat all detections as mesoscale eddies in this study.

Detection consistency using this methodology was qualitatively assessed using SWOT Cal/Val data in the Labrador Sea (notably pass #20). The analysis of consecutive daily detections shows overall good temporal consistency despite occasional, expected misdetections due to evolving structures and sampling variability. No quantitative metric was applied at this stage.

Detection was performed using both Sea Level Anomaly (SLA) and Absolute Dynamic Topography (ADT) fields. In the Labrador Sea, the resulting eddy statistics and validation against in situ observations are very similar for the two approaches. In the main manuscript, we nevertheless present the SLA-based detections, as most recent studies investigating eddies from SWOT observations have relied on SLA fields e.g.,. For comparison purposes, ADT-based results are additionally shown in Fig.  and in Appendix B. A detailed assessment of the impact of the Mean Dynamic Topography (MDT) provided in SWOT product and of the relative performance of SLA versus ADT for SWOT-based eddy detection is beyond the scope of the present study and is left for future work.

Validating the eddy detection method requires comparing the eddy characteristics obtained from SWOT data with independent in situ observations of eddies in the same region: this comparison must rely on another source of measurement, with eddy statistics derived from an alternative detection approach. However, establishing a suitable ground truth is not straightforward. On one hand, classical gridded altimetry products cannot serve as a reference, since their coarse resolution and interpolation algorithm prevent them from resolving mesoscale eddies in high-latitude regions . On the other hand, one could consider using eddies detected from high-resolution numerical simulations or reanalyses, either directly on their native grid or by simulating SWOT-like tracks and applying the same detection algorithm. Yet, such an approach would implicitly assume that the model perfectly represents the mesoscale dynamics of the region – an assumption that has not itself been validated, leading to a circular argument.

During the MSM129 cruise conducted in spring 2024 in the Labrador Sea, four anticyclonic mesoscale eddies (hereafter A1–4) were sampled with the SADCP. The specific tracks crossing the eddies, as well as their associated surface velocity fields (averaged between 22 and 102 m depth), are shown in Fig. a, e and f. We take advantage of the method developed by and applied it to this specific dataset, from which the main dynamical properties of the eddies are retrieved to validate the outcome of our SWOT-eddy detection method. The findings are summarized in the table in Fig. g.

All four eddies were successfully captured in coincident SWOT passes and detected by our SWOT-based eddy detection method. The eddies identified in SWOT exhibit radii and maximum azimuthal velocities that are in good agreement with those measured during the MSM129 cruise. It is important to note that the SWOT detections and the in situ observations were not made at the same time. For instance, A1 was observed in situ on 24 June 2024, while its SWOT counterpart was detected on 27 June 2024 (it was also identified on 25 June and 6 July but duplicate detections across cycles were automatically removed; see Fig. a–c). The slight reduction in velocity amplitude seen in the SWOT data can be attributed to the instrument's spatial smoothing, which tends to dampen kinetic energy. Also, SWOT altimetry measures an integrated quantity, which tends to be weaker than directly measured surface velocity from SADCP directly. Nevertheless, the close spatial coincidence between eddy centers and contours identified by the two independent approaches, together with their comparable dynamical characteristics, confirms the robustness and reliability of our detection algorithm when applied to the native 2 km SWOT swath data. Note that given the limited number of colocated cases, this agreement should be interpreted qualitatively. Also, note that this in situ validation is based solely on anticyclonic cases; cyclonic eddy properties are validated statistically in the following.

For comparison with standard gridded altimetry, we extracted SLA from the CMEMS 1/4 and 1/8° products on the same date as the SWOT observation of A1, and applied the same py-eddy-tracker detection parameters (see Fig. c and d). This yields velocity for A1 of Vmax=0.29 m s⁻¹ in both datasets, and radii of Rmax=33.4 and 25.6 km for the 1/8 and 1/4° products, respectively. This example clearly illustrates the improved effective resolution of SWOT compared to conventional altimetry.

To assess the robustness of our detection method at the statistical level, we use the comprehensive in situ dataset published by , which provides a unique reference for eddy statistics in the Labrador Sea. Using their in situ methodology, a large number of mesoscale and submesoscale eddies were identified and characterized, allowing the construction of histograms of the main eddy dynamical properties – including radius, maximum azimuthal velocity, and Rossby number. This dataset thus represents an unprecedented “ground truth” for the eddy field in this high-latitude region, against which we can directly compare the SWOT-based results. The distributions obtained from SWOT (SLA- or ADT-based) detections show remarkable agreement with the in situ statistics, both in terms of typical eddy scales and dynamical intensity, see Fig. . When including eddies detected above depth < 2000 m, we find an increased number of small-radius eddies, consistent with the reduced Rossby deformation radius on the continental shelf, but reaching the limitations of SWOT capabilities. Quantitatively, SADCP eddies have a radius between 3 an 39 km (mean 15 km, std 9 km), and azimuthal velocity between 7 and 58 cm s⁻¹ (mean 26 cm s⁻¹, std 13 cm s⁻¹); this compares favorably with SWOT eddies, with radius between 5 and 39 km (mean 14 km, std 6 km), and azimuthal velocity between 5 and 89 cm s⁻¹ (mean 20 cm s⁻¹, std 10 cm s⁻¹). In contrast, the eddy statistics derived from CMEMS gridded altimetry display much larger radii and weaker velocities, reflecting the well-known smoothing and scale limitations of mapped altimetry products at these latitudes; CMEMS eddies have a radius between 25 an 68 km (mean 42 km, std 11 km), and azimuthal velocity between 3 and 12 cm s⁻¹ (mean 8 cm s-1, std 3 cm s⁻¹)

Note that using the recently released v3.0 of SWOT data instead of v2.0.1 improves eddy detection, in particular by (1) reducing the number of spurious small eddies and (2) rebalancing the number of cyclonic and anticyclonic detections (with more cyclones than anticyclones in v2 and approximately equal proportions in v3). This improvement likely results from enhanced noise filtering, i.e., an improved version of the filtered SLA product, and the adoption of the new Mean Sea Surface which significantly reduces the imprint of geodetic structures in the resulting SSH data. A detailed assessment of differences between versions is beyond the scope of this study.

With the in situ validation confirming the reliability and robustness of our detection approach, we can confidently apply it and analyze eddies detected using SWOT in the Labrador Sea. In the following section, we examine the eddy field observed in the Labrador Sea throughout 2024–2025, focusing on the spatial distribution and dynamical characteristics of the detected eddies. This provides a spatial statistical description of the regional mesoscale activity at such high latitude at an unprecedented resolution, significantly extending previous estimates derived from lower-resolution altimetry products.