Automatic target recognition (ATR) for synthetic aperture sonar (SAS) enables advanced naval capabilities, yet deep‑learning methods are hampered by limited target imagery, background clutter, and the need for human‑in‑the‑loop assessment. We transfer DINOv3 Vision Transformer (ViT) models to underwater SAS ATR using a three‑stage, parameter‑efficient adaptation framework.\ \ Stage 1 freezes the ViT backbone and trains only Low‑Rank Adaptation (LoRA) modules, bridging the gap between natural‑image pre‑training and underwater acoustic propagation. Stage 2 applies hard‑negative mining to sharpen the decision boundary against acoustic mimics such as rocks and sediment formations. Stage 3 introduces Supervised Contrastive Learning (SupCon) to further separate target and clutter representations.\ \ Evaluation on at‑sea SAS data uses a mission‑level geographic split; all arms are compared at 85 % recall and each comparison is repeated over three random seeds. LoRA alone raises the area under the precision‑recall curve (AUPRC) from 0.300 to 0.679 ± 0.027 while training only 0.26 % of the weights. Subsequent hard‑negative mining and SupCon stages yield negligible changes (‑0.0045 ± 0.0119 and +0.0002 ± 0.0096 respectively) versus matched controls, indicating that the encoder had already captured most target‑clutter geometry.\ \ These findings suggest that a single efficient adaptation stage suffices; stacked refinements do not provide additional benefit.\ \ Review