Enhancing intelligent identification of internal solitary waves via SWOT multimodal feature fusion

Internal solitary waves (ISWs) are prevalent nonlinear hydrodynamic phenomena in topographically complex ocean regions. Automated identification of internal solitary waves mainly relies on surface-texture imagery, leaving the incremental value of sea-surface height information uncertain. We evaluated a framework using Surface Water and Ocean Topography (SWOT) observations for ISW segmentation in the East China Sea. It combines normalized radar cross section (σ 0 ), background-removed sea surface height anomaly (SSHA), and its gradient (∇SSHA). A TransUNet model was evaluated across five training runs on one fixed scene-level partition. Radar backscatter provided the dominant predictive information. With a prescribed σ 0 :SSHA:∇SSHA input-amplitude scaling of 95:4:1, W1 achieved a mean foreground-class intersection over union (IoU) of 0.790. The architecture-matched σ 0 reference (A1) achieved 0.766. The mean full-test-set gains in Precision, Recall, F1-score, and IoU were modest and not statistically significant (all p > 0.05). On 346 challenging-positive test patches, W1 increased mean Recall from 83.57% to 85.40% and foreground-class IoU from 69.06% to 71.00%. On 346 challenging-negative test patches, mean pixel-level false-positive rate decreased from 1.60% to 1.52%. These subset comparisons are descriptive. Shuffled, random-field, and σ 0 -gradient controls had lower mean performance than W1, suggesting that co-registered height-related inputs may provide incremental predictive value beyond generic channel-count or edge effects. The predicted masks also supported crest-line extraction and inter-crest spacing analysis. The framework extends texture-based ISW segmentation by incorporating height-related observations, with modest observed gains under the evaluated conditions.