HMG-Former: a hierarchical meta-graph transformer framework for robust multi-class SAR target recognition

Automatic Target Recognition (ATR) using Synthetic Aperture Radar (SAR) imagery plays a vital role in defence surveillance and remote sensing due to its capability to operate reliably under all-weather and day-and-night conditions. However, robust SAR ATR remains challenging because of variations in imaging geometries, target appearance, speckle noise, and complex background clutter. Moreover, existing deep learning approaches typically process SAR images as regular grids or token sequences, overlooking the structural relationships among target components. To overcome these issues, we proposed a novel Hierarchical Meta Graph Transformer (HMG-Former) for SAR ATR. This framework uses a Convolutional Neural Network (CNN) to eliminate discriminative feature maps. Further, it transformed into a hierarchical graph containing local patch nodes, intermediate part-level nodes, and a comprehensive target node. The multi-head self-attention graph transformer model facilitates effective learning of global structural dependencies and fine-grained scattering characteristics among these nodes. The final global graph representation is being used for target classification. To evaluate the effectiveness of the framework, the experiments have been performed on the benchmark eight-class MSTAR dataset with the 128 × 128 greyscale SAR images. The proposed framework outperformed achieving 99.42% test accuracy and an average 5-fold cross-validation accuracy of 99.39% (±0.06) over the available state-of-the-art baseline model. The framework also demonstrates strong robustness to speckle noise by maintaining 99.31% accuracy while achieving an inference speed of 17.26 FPS. These results confirm that hierarchical graph-based structural reasoning significantly enhances SAR target recognition accuracy, robustness, and computational efficiency.