Spatiotemporal Deep Learning for Target Classification in High-Resolution 3D-ISAR Radar Images

Inverse Synthetic Aperture Radar (ISAR) has recently advanced to volumetric 3D-ISAR imaging, creating new opportunities and challenges for automatic target recognition (ATR). This work proposes a spatiotemporal deep learning framework that jointly learns target structure and motion dynamics from high-resolution 3D-ISAR sequences. A CNN backbone (ResNet) extracts per-frame spatial features, which are fed to temporal models Bidirectional LSTM and/or ConvLSTM to capture micro-Doppler cues and aspect-dependent scattering over time; the pipeline is supported by physics-aware formation and backprojection-style 3D reconstruction. We evaluate on a four-class dataset (aircraft, helicopter, drone, tank) comprising 400 labeled samples drawn from MSTAR and simulated 3D-ISAR sequences, with standard train/validation/test partitions and targeted denoising, normalization, and augmentation to enhance robustness. The proposed model achieves strong performance across metrics: an overall accuracy of 95% on the final evaluation set with near-ideal class separability (AUC ≈ 0.98–1.00), and a best accuracy of 96.7% when all preprocessing and geometric/data-level augmentations are enabled. Ablation and robustness studies show consistent gains from motion-aware temporal modeling and the preprocessing stack under low-SNR and distortion conditions, while confusion is largely confined to visually and dynamically similar aerial classes. These results demonstrate that coupling modern spatiotemporal architectures with principled ISAR signal processing yields reliable, accurate, and deployment-oriented ATR for 3D-ISAR systems.