ATRNet-STAR: Depression Angle Diversity, Few-Shot Recognition, and Cross-Dataset Generalization in SAR Automatic Target Recognition
Mousa Shaya · Shaya Software and Systems LLC | Arizona State University
arXiv preprint, May 2026

Paper (arXiv) → Code →


Overview

Synthetic Aperture Radar (SAR) enables all-weather, day/night target recognition — a critical capability for defense and intelligence applications. This study systematically evaluates SAR ATR generalization across three axes that matter operationally but are rarely studied together: depression angle variation, data scarcity (few-shot), and cross-dataset transfer.

The work uses two benchmarks: ATRNet-STAR (40 vehicle classes across 4 depression angles: 15°, 30°, 45°, 60°) and MSTAR (10 vehicle classes, standard operating conditions and extended operating conditions).


Key Results

Experiment Accuracy
SOC baseline (ConvNeXt-Small) 97.2%
EOC — 15° only model, tested at other angles 43.7%
EOC — adding 60° to training 74.1% (+30.4 pp)
Prototypical (5 shots per class) 98.2%
MSTAR (cross-dataset transfer) 99.1%
Cross-dataset transfer (ATRNet→MSTAR) 61.1%
DINOv2 optical features on SAR 6.1% (near-chance)
Test-time augmentation (TTA) −24.3 pp vs. baseline

Key finding: Depression angle is the dominant challenge in operational SAR ATR. A model trained on a single angle collapses to 43.7% at other angles. Adding multi-angle training data is more effective than any architectural improvement. Optical foundation model features (DINOv2) are near-useless on SAR — SAR requires SAR-trained features.


Feature Space Analysis (UMAP)

The figures below show the 768-dimensional ConvNeXt feature space projected to 2D via UMAP, color-coded by vehicle class.

(UMAP figures will appear here — images in /assets/images/projects/sar-atr/)


Methods

Backbone: ConvNeXt-Small pretrained on ImageNet-1k, fine-tuned on single-channel SAR chips (64×64 px, log-amplitude normalized).

Few-shot: Prototypical networks — class prototypes computed as mean of support embeddings; test samples classified by nearest prototype in embedding space.

Evaluation: Standard Operating Conditions (SOC) follow the MSTAR convention. Extended Operating Conditions (EOC) test across depression angles not seen during training.

Cross-dataset: Model trained on ATRNet-STAR, evaluated zero-shot on MSTAR (no MSTAR training data used).


Datasets

Dataset Classes Angles Resolution Size
ATRNet-STAR 40 15°, 30°, 45°, 60° ~0.3m ~100K chips
MSTAR 10 15°, 17° ~0.3m ~5K chips

Citation

@article{shaya2026atrnetstar,
  title={ATRNet-STAR: Depression Angle Diversity, Few-Shot Recognition,
         and Cross-Dataset Generalization in SAR Automatic Target Recognition},
  author={Shaya, Mousa},
  journal={arXiv preprint arXiv:2506.XXXXX},
  year={2026}
}