LoRA in sonar: AUPRC more than doubles with 0.26% of weights

A frozen vision model, an adapter that trains 0.26 percent of the parameters, and the area under the precision recall curve going from 0.300 to 0.679. That is the headline result of LoRA Enhanced Contrastive Learning with SAS Vision Transformers, posted on arXiv on 22 September, a paper that adapts DINOv3 models to automatic target recognition (ATR) in synthetic aperture sonar (SAS) imagery.

The underlying problem is not the architecture, it is the domain. A Vision Transformer pretrained on photographs has seen cats, cars and faces, not the acoustic signature of a metal object sitting on sediment. On top of that come three conditions the paper lists without decoration: labelled target imagery is scarce, the seabed is full of rocks and formations that mimic man-made objects, and at the end of the chain there is a human operator reviewing every detection.

Three stages on a frozen backbone

The method splits into three steps, and the paper is honest about which one does the real work.

1. LoRA with the ViT backbone frozen, to bridge the gap between natural-image pretraining and underwater acoustic propagation.
2. Hard-negative mining, to harden the decision boundary against acoustic mimics: rocks and sediment formations with the wrong silhouette.
3. Supervised contrastive learning (SupCon), to separate target and clutter representations.

The primary effect belongs to the first stage. With the same frozen backbone, LoRA raises AUPRC from 0.300 to 0.679 ± 0.027, and it does so at rank 4, that is, training 0.26 percent of the model parameters.

Why AUPRC and not accuracy

The chosen metric says a lot about the problem. On a seabed, targets are extremely rare compared with clutter, so a classifier that always answers nothing here is right almost every time and useless. The area under the precision recall curve punishes exactly that. And since a human is reviewing, every false positive is paid in operator minutes, not in a cell of a confusion matrix.

The methodology is half the story

Three decisions in the evaluation deserve as much attention as the number.

The split is geographic and at mission level. Images are not shuffled at random between train and test, whole missions are separated. That is the difference between measuring whether the model recognises targets and measuring whether it has memorised a particular patch of seabed.

All comparisons are made at 85 percent test recall. Fixing the operating point avoids the usual trick of comparing two models in different regions of the curve and keeping whichever one flatters.

Each comparison is repeated across three random seeds and the deviation is reported. With small datasets, one lucky seed can account for a good share of an improvement.

None of this is exotic, but it is frequently skipped in applied vision. Having it done here makes the 0.679 readable with rather more confidence than a bare number.

What someone outside sonar can take away

The pattern transfers to any domain with scarce data far from natural images: industrial inspection, medical imaging, agricultural remote sensing, factory quality control. The recipe is concrete: freeze the backbone, adapt with low-rank LoRA before touching anything else, measure separately how much each stage contributes, and evaluate with a split that respects the real structure of the data.

It also leaves a warning. If the primary effect comes from stage 1, it is worth checking how much stages 2 and 3 genuinely add in your case before taking on the cost of implementing and maintaining them.

This is applied work, with a narrow domain and a defence context that not everyone will share, but the methodological part stands on its own. We prefer this kind of paper, with the operating point fixed and the seeds reported, over one with a bigger table and a random split.