Accurate detection of mesoscale oceanic eddies from satellite altimetry data is crucial for understanding ocean dynamics and climate processes. Eddies’ variable morphology and dense distribution in ocean flow fields present significant challenges for automated identification systems. While multi-scale segmentation networks have advanced eddy detection capabilities, conventional approaches using fixed sampling patterns inadequately capture irregular eddy geometries and produce boundary artifacts in regions of closely-packed eddies. This paper introduces a novel dual-adaptive feature extraction framework that addresses these limitations through two complementary mechanisms. First, our Geometry-Responsive Sampling Module (GRSM) implements bounded deformable convolutions that dynamically adjust sampling positions by learning optimal displacement fields constrained within physically meaningful ranges. This enables adaptive matching to irregular eddy boundaries across different scales. Second, our Spatial Weighting Module (SWM) employs dedicated convolution branches to generate context-aware sampling weights that enhance eddy boundary features while suppressing redundant information, effectively resolving boundary delineation uncertainties in closely-packed eddy fields. Experiments across oceanographically distinct regions validate the framework’s robustness and transferability. Compared to established methods such as SegNet, FCN-8s, EddyNet, PSPNet, and DeepLabv3+, our approach demonstrates superior performance, with improvements of 1.3–2.9% in mean Intersection over Union (mIoU) and 0.8–2.4% in mean Dice Coefficient (mDice) over the leading baselines. The method achieves 89.52% mIoU and 92.78% mDice on the South China Sea-Northwest Pacific dataset, and maintains 87.05% mIoU and 92.61% mDice on the South Atlantic dataset. The consistent performance across regions with contrasting eddy dynamics, scale distributions, and flow regimes demonstrates the framework’s capability to accurately segment mesoscale eddies under diverse oceanographic conditions. • Geometry-aware approach for mesoscale eddy segmentation based on SSH. • Deformable sampling and spatially weighting for adaptive feature extraction. • Robust across diverse oceanic regions and eddy types.
Song et al. (2026) studied this question.