Accurate mapping of oceanic internal waves (IWs) is essential for understanding ocean mixing and energy transfer, yet remains challenging due to weak and transient sea-surface signatures and heterogeneous sensor limitations. In this study, we propose a geometry-aware deep learning framework, MI-IWNet, which incorporates a direction-aware encoder design (including Direction-Aware Strip Convolution, DASC) to represent the anisotropic and curvilinear morphology of IW crests in multi-sensor satellite imagery; we further develop a graph-based Structure-Aware Wave Packet Aggregation algorithm to assemble pixel-level predictions into object-level Oriented Bounding Boxes (OBBs) for robust geophysical analysis. Experiments on large-scale Sentinel-1 SAR and optical observations demonstrate that the proposed framework enables geometry-consistent mapping. By mapping these extracted OBBs onto regional bathymetry in a decoupled post-processing analysis, we reveal distinct topographic modulation effects in IW evolution across marginal seas: in the South China Sea, IW packets exhibit coherent trans-basin propagation with local obstacle-induced modulation and shelf-break accumulation; in the semi-enclosed Sulu Sea, wave packets show source-controlled radiation confined by basin boundaries; and in the Andaman Sea, the Andaman–Nicobar Ridge induces strong directional anisotropy and complex interference patterns. These results highlight how key topographic units systematically modulate packet density, pathway continuity, and directional organization, and indicate that MI-IWNet provides an effective and transferable tool for automated IW segmentation, characterization, and comparative regional analysis.
Xiao et al. (Mon,) studied this question.