Automatic Identification of RD Spectrum Components Implemented Using LAU-Net

The accurate identification of components in the range-Doppler (RD) spectrum image is deemed crucial for radar interference suppression and parameter inversion in high-frequency hybrid sky-surface wave radar (HFSSWR) data processing. Previous studies have primarily focused on the manual selection of components in the RD spectrum image using regular boxes, with limited ability to simultaneously identify each component automatically. To address this issue, a small dataset comprising 364 images and 1190 annotations using real radar returns was constructed. Building upon the U-Net architecture, the lightweight aligned U-Net (LAU-Net) structure is introduced, where ResNet50 serves as the encoder, an alignment module is integrated to standardize image sizes, and focal loss is employed to mitigate the effects of class imbalance within the dataset on model performance. The test results demonstrate that the LAU-Net model outperforms the U-Net model, achieving improvements in frames per second (FPS) by 77.10%, mean pixel accuracy (mPA) by 13.48%, and mean IoU (mIoU) by 10.92%.