Radar Classification Method for Small Targets on the Sea Surface Based on Multi-Domain Standardization and Class-Conditional Distribution Alignment

Accurate classification of small sea-surface targets remains a critical technological challenge for advancing marine monitoring and defense security capabilities. In sea clutter environments， radar echoes from small surface targets are characterized by low signal strength and pronounced non-stationarity. Furthermore， distributional discrepancies across different radar systems limit the effectiveness of classification approaches based on single-domain features， which are unable to simultaneously capture radar cross-section variability， motion dynamics， and cross-platform generalization. To address these limitations， this study proposes a radar-based classification framework for small sea-surface targets that integrates multi-domain standardization with class-conditional distribution alignment. First， a multi-domain feature standardization preprocessing pipeline was developed to achieve consistent normalization of the target amplitude sequence， Doppler amplitude spectrum， and time-frequency representation. Subsequently， a three-branch parallel feature extraction network incorporating a Multi-Scale Dual Attention （MSDA） module was developed. The time-frequency branch employs a two-dimensional feature extraction architecture based on self-calibrated convolutions. In contrast， the amplitude sequence branch and the Doppler branch leverage the multi-scale dual attention module to capture local variations and long-range temporal dependencies， respectively. During the feature fusion stage， a class-conditional maximum mean discrepancy （CC-MMD） constraint was incorporated to align the cross-domain feature distributions of samples belonging to the same class within a shared latent space. This process effectively mitigates inter-class confusion， particularly for challenging targets such as floating spheres and boats. Experiments are conducted on the IPIX， CSIR， and measured UAV datasets. The results show that， compared with a single-branch network based exclusively on time-frequency features， the proposed method improves overall classification accuracy from 0.7665 to 0.8155. In addition， relative to the three-domain fusion network without class-conditional alignment， the accuracy was further increased by approximately 2.4%. These findings demonstrate that the proposed method significantly improves the classification performance of small sea-surface targets and provides robust discrimination for visually and statistically similar classes.