Methodological study demonstrates unsupervised multifrequency radar fusion in Arctic sea ice, indicating improved structural imaging without labeled training data.
Key Points
To develop and evaluate an unsupervised deep learning framework that fuses multifrequency ground-penetrating radar data for Arctic sea ice exploration without requiring paired ground-truth labels.
Constructed UDBA-Net featuring dual-branch frequency-specific encoders (400 MHz and 900 MHz) and a convolutional block attention module for adaptive channel and spatial recalibration.
Optimized the network using an unsupervised hybrid loss integrating local intensity, structural similarity, and Laplacian-gradient constraints.
Evaluated performance using synthetic radar simulations and field-acquired Arctic multiyear sea ice profiles.
Enhanced relative structural contrast and preserved reflection-event continuity across multiyear sea ice profiles.
Merged shallow high-resolution textures with deep low-frequency stratigraphic interfaces under label-scarce conditions without relying on handcrafted fusion heuristics.