SARPCNet: Spaceborne SAR Feature Extraction Network Coupled With Phase Characteristics

The feature extraction task for spaceborne synthetics aperture radar (SAR) image matching faces the challenge of maintaining stable feature performance under different imaging conditions. The existing feature extraction networks mainly rely on amplitude domain features, which are easily affected by scattering and speckle noise in complex terrain. However, the phase characteristics of spaceborne SAR images have structural invariance and strong noise resistance, which can more reliably retain the geometric details that are often lost in the amplitude domain. To address this issue, a feature extraction network coupled with SAR phase characteristics is proposed, which combines the amplitude information and phase information of SAR images. The amplitude feature maps of SAR images are first extracted through a shallow feature extraction module. Subsequently, a phase fusion module is employed to couple the shallow features with phase information, thereby enriching the spatial semantic representation and enhancing structural sensitivity. The fused multidimensional SAR features are then further encoded and represented via a deep feature extraction module. In addition, phase consistency is incorporated in the form of a loss function to improve the stability of phase feature extraction. The loss function is computed based on the residual matrix between the feature maps before and after registration. The experimental results demonstrate that the proposed method achieves a matching accuracy better than 1.5 pixels, outperforming current state-of-the-art SAR image matching algorithms, while yielding several-fold improvements in the number of correctly matched points compared with the competing methods. The robustness and dependability of the suggested method for SAR image feature extraction are confirmed by these results.