Unsupervised dual-branch attention network for multifrequency ground-penetrating radar data fusion: an application to Arctic sea ice exploration

Abstract

Abstract Ground-penetrating radar (GPR) is a nondestructive electromagnetic (EM) method used for subsurface target detection and stratigraphic imaging. The use of a single-frequency GPR profile is inherently limited by the trade-off between penetration depth and vertical resolution. For example, low-frequency antennas provide deeper penetration with lower resolution, whereas high-frequency antennas perform better in shallow areas but attenuate rapidly. Hence, multifrequency GPR data fusion can have complementary advantages. Conventional fusion methods usually depend on handcrafted feature extraction and fusion rules, whereas supervised deep learning-based methods are difficult to apply in polar field surveys due to the unavailability of absolute paired ground-truth labels. To address these limitations, this study proposes an unsupervised dual-branch attention network (UDBA-Net) for the fusion of 400/900 MHz multifrequency GPR data and applies it to Arctic multiyear sea ice exploration. The proposed network combines frequency-specific dual-branch encoders, a convolutional block attention module (CBAM)-like channel- or spatial-attention recalibration module, and an unsupervised hybrid loss comprising local intensity, structural similarity, and Laplacian-gradient constraints. The low- and high-frequency profiles are encoded separately to reduce early cross-band feature aliasing, whereas the attention module adaptively recalibrates the concatenated features based on channel contribution and spatial reflection-event salience. The main objective of the method is the structural and morphological enhancement, rather than preservation, of absolute electromagnetic reflection coefficients. Experimental analyses using controlled synthetic data and Arctic field GPR profiles confirm that UDBA-Net enhances relative structural contrast, improves reflection-event continuity, and integrates shallow, high-resolution textures with deeper, low-frequency interfaces under label-scarce conditions.