A Reliable Radar Clutter Classification Using Dual-Path Complex-Real U-Net with Feature Fusion

Radar clutter classification plays an important role in the field of target detection, tracking and recognition. Modern deep learning methods achieve highly reliable clutter classification through means such as data mining and model training, but they suffer from performance degradation in small-sample scenarios and exhibit insufficient adaptability to multi-classification tasks in complex scenes. This paper proposes a Dual-Path Complex-Real U-Net (DPC-U-Net) method, aiming to effectively handle the complex information in radar echoes and address the class imbalance issue. First, in terms of model framework design, by decoupling the explicit and implicit correlations of complex information in radar echoes, parallel real-feature and complex-feature branches are constructed. The former branch extracts multi-scale explicit features based on the classic U-Net architecture, enabling it to capture intuitive features at different scales, while the latter branch nests complex convolution modules within the U-Net structure to mine implicit features in the complex domain, thereby obtaining more abundant feature information through the two branches. Second, cross-channel feature fusion is performed on the features extracted by the two branches, then the fused features are processed through convolutional operations to finally output multi-class clutter classification results, effectively addressing the multi-class clutter classification problem. Third, to address the instability of model performance caused by class imbalance in small-sample environments, a Weighted Focal Loss function is introduced. Experiments on a dataset which is constituted of background noise, meteorological clutter, ground clutter, sea clutter, and folded clutter show that the fusion of real-valued and complex-valued features enhances the richness of feature representation, thereby significantly improving classification performance. Compared with MLP, CNN, FCN, and U-Net algorithms, DPCU-Net has achieved higher classification accuracy.