Transfer Learning With Fuzzy Elevation-Aware Experts for Bathymetric DEM Super-Resolution

Abstract

High-resolution land–ocean integrated digital elevation models (DEMs) are critical for navigation, coastal management, and marine engineering in strait environments. However, existing DEM products are too coarse, while bathymetric observations are sparse, heterogeneous, and unevenly distributed, making high-resolution reconstruction difficult. To address this problem, this study proposes a transfer learning-based fuzzy elevation-aware mixture of experts (TFEA-MoE) framework for DEM super-resolution in the Taiwan Strait. The core methodological contribution is an elevation-aware fuzzy expert routing mechanism, which uses learnable Gaussian membership functions to softly assign pixels to shallow, transition, and deep-water experts. Embedded in a Bass-Strait-to-Taiwan-Strait transfer-learning framework, this routing strategy enables smooth and physically interpretable residual correction under sparse ATL24 and shipborne supervision. Experimental results show that TFEA-MoE achieves the lowest errors among the compared methods under the evaluated protocols. Based on held-out sparse bathymetric observations, compared with the average performance of all baseline methods, TFEA-MoE reduces root mean square error (RMSE), mean absolute error (MAE), slope RMSE, and slope MAE by 23.87%, 34.01%, 22.02%, and 21.48%, respectively. Ablation studies confirm the contributions of transfer learning and fuzzy expert routing, with fuzzy refinement reducing the four metrics by 12.36%, 34.89%, 12.26%, and 16.05%, respectively. Because these observation-level improvements may be partially influenced by spatial autocorrelation, spatially independent testing is used as the main evidence for target-domain generalization, with uncertainty and seam-continuity analyses supporting reconstruction stability. These results indicate that TFEA-MoE provides a promising and interpretable approach for DEM super-resolution under sparse and heterogeneous bathymetric observations.