TFEA-MoE Taiwan Strait 1 Arc-Second Integrated DEMTFEA-MoE 台湾海峡 1 弧秒陆海一体 DEM

Dataset overview

This record describes a 1 arc-second land-ocean integrated DEM for the Taiwan Strait and its surrounding marine and terrestrial areas. The final product combines a NASADEM land surface with a learned marine surface generated by the transfer-learning-based Fuzzy Elevation-Aware Mixture of Experts framework, abbreviated TFEA-MoE.

The target region is difficult to model with one uniform reconstruction rule. Shallow water is supported mainly by ICESat-2 ATL24 photons, the transition zone has sparse supervision, and deeper water relies more heavily on shipborne observations. TFEA-MoE uses a Bass Strait source domain to learn transferable terrain priors before adapting to Taiwan Strait observations.

• GEBCO 2025 15 arc-second input is reconstructed to a 1 arc-second output with a 15x scale factor
• The 32 x 32 low-resolution input corresponds to a 480 x 480 high-resolution prediction patch
• The marine field is divided into shallow water, transition water and deep water regimes at approximately 0 to -40 m and -40 to -200 m
• The Taiwan Strait target model is fine-tuned with sparse ATL24 and shipborne bathymetric observations
• The final map preserves NASADEM over land and uses the reconstructed marine DEM over water
• Spatially independent testing, nearshore seam checks and ablation studies are included with the research release

Data sources and preprocessing

All records are harmonized in WGS84 geographic coordinates and quality controlled before sample construction. Duplicate points, invalid coordinates, land contamination, obvious outliers and records outside the study area are removed. The Taiwan Strait GEBCO grid is split into overlapping 32 x 32 patches with a stride of 16, and sparse observations are projected onto the matching 480 x 480 high-resolution grid. Patches with fewer than five valid observations are excluded. Development and validation samples use an 80/20 split, while a reserved 20% of the bathymetric points is held out for testing.

The Bass Strait source pairs use GEBCO 2025 as the low-resolution input and a 1 arc-second AusSeabed DEM as the high-resolution reference. The paired products are co-registered, divided into 480 x 480 and 32 x 32 patches, and normalized from approximately -7000 to 4000 m into the range -1 to 1. The input and reference are different DEM products, so the model learns both spatial super-resolution and cross-product correction rather than a simple downsampling inverse.

Method

TFEA-MoE has three stages: source-domain pre-training on the Bass Strait, target-domain fine-tuning in the Taiwan Strait, and inference of the final integrated land-ocean DEM. A Swin Transformer backbone first generates a coarse 480 x 480 high-resolution DEM from each 32 x 32 input patch. A fuzzy elevation-aware refinement module then predicts depth-specific residual corrections.

The fuzzy router uses learnable Gaussian membership functions rather than hard depth thresholds. Each pixel receives a soft weight for a shallow expert, a transition expert and a deep-water expert. The three lightweight convolutional experts estimate residual corrections, which are normalized, fused and added to the coarse Swin output. This lets neighboring depth regimes overlap smoothly while retaining a physical interpretation of the routing signal.

The source model is optimized with dense DEM and terrain-gradient supervision. During transfer learning, the first two residual Swin Transformer blocks are frozen to preserve general terrain priors, while the remaining backbone layers and the fuzzy mixture-of-experts module adapt to sparse ATL24 and shipborne observations. The main settings are four residual Swin Transformer blocks with depths [4, 4, 4, 4], window size 8 x 8, embedding dimension 180, 6 attention heads per block, Adam optimization, cosine annealing, batch size 32, 100 pre-training epochs, 100 fine-tuning epochs and early stopping patience 10.

Validation and results

The paper evaluates six metrics: RMSE, MAE, PSNR, SSIM, Slope RMSE and Slope MAE. Five baselines are compared with TFEA-MoE: Bicubic, EDSR, ESRGAN, SwinIR and DDPM. The target-domain evaluation uses held-out ATL24 photons and shipborne points, spatially independent test regions and nearshore land-sea seam indicators.

Across the main target-domain comparison, TFEA-MoE reduces RMSE, MAE, Slope RMSE and Slope MAE relative to the average baseline performance by 23.87%, 34.01%, 22.02% and 21.48%. The paper also reports that fuzzy refinement improves the four metrics by 12.36%, 34.89%, 12.26% and 16.05% in the corresponding ablation comparison.

Output and access

The output is a 1 arc-second GeoTIFF land-ocean integrated DEM for the Taiwan Strait. The marine area is reconstructed by TFEA-MoE and the land area is filled with NASADEM. Dataset files are supplied through the access request form on this page; the paper and model details are recorded in the project documentation for reference.