Seismic data interpolation aims to reconstruct complete wavefields, with practical acquisition constraints often causing severe undersampling that compromises reconstruction reliability. Although implicit neural representations (INR) methods based on Fourier feature encoding can provide good interpolation results, their fixed frequency basis limits representational ability and may introduce noise. Multi-resolution hash encoding has been introduced in INR to enhance high-frequency detail representation, yet its potential for seismic data interpolation remains largely unexplored. In this paper, we propose a hybrid encoding strategy that integrates Fourier encoding and multi-resolution hash encoding into an INR framework. In the early stage, Fourier encoding dominates the reconstruction process by introducing global sinusoidal bases that preserve smooth structures and long-range continuity. As optimization progresses, the contribution of multi-resolution hash encoding gradually increases, enabling the model to capture high-frequency and spatially localized details through learnable local features. The Fourier and hash features are combined with progressive, epoch-dependent weights and then fed into a multilayer perceptron (MLP), which is optimized in a self-supervised manner using only the observed seismic traces. This hybrid design effectively balances global consistency and local detail recovery, thereby improving structural continuity while reducing aliasing artifacts. Experiments on both 2D and 3D seismic datasets demonstrate that the proposed method achieves more accurate and geologically coherent reconstructions under sparse sampling conditions.
Bathymetric mapping of ice-covered polar seas remains a major challenge. Persistent sea ice restricts systematic surveys, while icebreaking operations degrade multibeam echosounder (MBES) data quality through noise, hull vibrations, and ice interference. As a result, digital terrain models (DTMs) from bathymetric surveys in ice-covered regions are often fragmented, coarse, and incomplete, limiting interpretation of seafloor morphology, including glacial landforms used to infer former glacier extent and ice-sheet dynamics. Conventional interpolation for upscaling and gap filling often oversmooths morphology and suppresses geomorphologically important landforms. We developed a generative super-resolution approach to reconstruct bathymetry at 25 m resolution from degraded 100 m and 200 m inputs (4× and 8× super-resolution). We additionally examine whether fine-scale glacial bathymetric structure can be inferred from coarser geomorphological context. To address the issue of limited bathymetric training data from polar regions, we trained a conditional diffusion model on high-quality terrestrial DTMs containing landforms analogous to glaciated submarine settings. The model was evaluated on held-out Arctic MBES bathymetry from offshore northern Greenland using patch-based tests, full-grid evaluation, and qualitative assessment. Compared to bicubic interpolation, the model improves elevation-distribution agreement, feature-space similarity, and spectral consistency. However, gains in spatial cross-correlation remain limited, suggesting that coarse geomorphological context alone may provide insufficient information for reliable inference of fine-scale landforms. Qualitative assessment suggests clearer expression of glacial landforms, including lineations and retreat moraines, where only weak traces were visible in coarse bathymetry. These results suggest that proxy-trained generative super-resolution can serve as an interpretation-enhancing post-processing tool for degraded Arctic bathymetric grids.
