Asymmetric Dual-Branch Network for Gravity Anomaly Reference Map Super-Resolution in Underwater Gravity Matching Navigation

Gravity matching navigation is one of the most feasible ways to achieve long-duration and accurate underwater navigation. The positioning accuracy of the underwater gravity matching navigation is highly related to the resolution of the gravity anomaly reference map, which is commonly enhanced by interpolation techniques. However, the existing methods struggle to capture complex long-range dependencies of gravity anomalies and lack effective spatial feature modeling mechanisms, resulting in limited accuracy and poor adaptability across different regions. To address this issue, a novel asymmetric dual-branch network is proposed for gravity anomaly reference map super-resolution (SR). The proposed asymmetric dual-branch network employs a hybrid structure that consists of a convolutional neural network (CNN)-based local detail guidance branch and a Transformer-based long-range information interaction branch. A multiscale edge feature enhancement module is specially designed as the core of the local detail guidance branch to capture fine variations in local gravity anomalies, while the Swin Transformer is introduced for the long-range information interaction branch to model long-range spatial correlations of gravity anomalies. Additionally, a cross-scale feature reconstruction module is designed to fuse the complementary feature representations from both branches. Finally, a spatial–frequency domain joint-guided loss is constructed to further ensure the fine detail recovery while preserving the physical consistency of gravity anomalies. Experimental results demonstrate that the proposed method outperforms several existing methods in SR reconstruction of gravity anomaly reference map and effectively improves the positioning accuracy of the underwater gravity matching navigation.