Underwater geolocalization based on attention-enhanced dual-focal-length polarization imaging

Underwater geolocalization is challenging due to electromagnetic attenuation and limited visibility. We propose an attention-enhanced dual-focal-length polarization imaging method for underwater geolocalization, in which synchronized 25 mm wide-FoV and 50 mm narrow-FoV cameras are used to capture complementary polarization patterns. A dual-FoV cross-attention ResNet18 network is developed to fuse cross-FoV features and predict solar azimuth and zenith angles from the angle of polarization, degree of polarization, and intensity images. Geographic coordinates are then estimated through nonlinear least-squares optimization based on the solar position algorithm. Experiments show angle prediction MAEs of 0.091° and 0.039° for solar azimuth and zenith, respectively, with geolocalization errors of 4.4–7.6 km, demonstrating improved accuracy over single-focal-length baselines.