Robust multitask SAR ATR with outlier detection for persistent false alarm discrimination

Operational synthetic aperture radar (SAR) automatic target recognition (ATR) requires rejecting out-ofdistribution non-target objects while maintaining classification accuracy on known targets, but standard models trained on the MSTAR and SAMPLE datasets are prone to learning background texture shortcuts that undermine both goals. We establish this causally through target–background chimera decomposition and dual saliency analysis: baseline domain classifiers achieve high accuracy by reading background texture (background dominance index = 0.97), not target signatures, a learning shortcut invisible to accuracy metrics. To correct this, we develop an integrated multi-task ATR pipeline combining chimera augmentation, base triplet loss, and chimera paired triplet mining, validated through a component ablation where each addition is shown to be load-bearing. The multitask ATR pipeline redirects the domain head from background to target (occlusion interior/exterior ratio: 0.9× → 63.9×), and critically, 99.9% domain accuracy survives even as gross domain clustering vanishes in the embedding space, indicating the domain head has shifted from background texture to subtle target-level scattering differences. This decoupling enables multiple scoring methods that naturally specialize by non-target object type: domain entropy and maximum softmax probability for structureless clutter; P(measured) and Mahalanobis distance for false alarm objects that exhibit similar scattering characteristics. Baseline models also achieve perfect false alarm AUROC on standard evaluation, but via background texture. A chimera stress test (synthetic targets on measured backgrounds) breaks this ceiling: the scratch baseline’s P(measured) collapses to 0.622 while the full pipeline holds at 0.947, confirming that the pipeline’s detection mechanism is grounded in target signatures rather than background cues.