LoRA Fine-Tuning DINOv3 for Sonar Target Recognition: One Efficient Adaptation Stage Is All You Need

The Deep Learning Challenge in Underwater Target Recognition

Automatic Target Recognition (ATR) for Synthetic Aperture Sonar (SAS) is a critical capability in modern naval reconnaissance, but adapting deep learning to this domain is far from straightforward. In this arXiv paper, researchers identify three major pain points: scarce target image samples, severe background clutter, and evaluation pipelines that still require human-in-the-loop intervention. These constraints make conventional end-to-end deep learning training highly impractical.

Compounding the challenge is the unique nature of underwater acoustic propagation — there is a substantial domain gap between vision models pre-trained on natural images and sonar imagery. Natural formations like rocks and sediment often "disguise" themselves as man-made targets, creating hard-to-distinguish acoustic mimics. How to teach a model to differentiate real targets from imposters under data-scarce conditions is the central question this work aims to answer.

A Three-Stage Parameter-Efficient Adaptation Framework

The research team transferred a DINOv3 Vision Transformer (ViT) model to the underwater SAS ATR task using a three-stage parameter-efficient framework:

Stage 1: Frozen Backbone Fine-Tuning via LoRA

The first stage freezes the entire ViT backbone and fine-tunes only through Low-Rank Adaptation (LoRA), bridging the gap between natural image pre-training and underwater acoustic propagation. This is by far the most impactful step in the entire pipeline.

LoRA (Low-Rank Adaptation), proposed by Hu et al. in 2021, is a parameter-efficient fine-tuning method built on a key insight: the weight update matrices in pre-trained models have low intrinsic rank during downstream adaptation and can therefore be approximated by the product of two small matrices. In practice, a pair of low-rank matrices $A$ ($d\times r$) and $B$ ($r\times d$) are inserted in parallel with the Transformer's attention weight matrices, where the rank $r$ is much smaller than the original dimension $d$. During training, the original weights are frozen and only $A$ and $B$ are optimized. A rank of 4 means this matrix pair introduces very few additional parameters, yet captures task-relevant feature shifts in a low-dimensional subspace. In the vision domain, LoRA has been widely used to adapt large-scale pre-trained ViTs to small-sample specialized domains such as medical imaging and remote sensing. Its key advantage is significantly reducing overfitting risk while preserving the general feature representations learned by the pre-trained backbone.

Stage 2: Hard-Negative Mining

The second stage introduces hard-negative mining, targeting rocks and sediment structures that visually resemble man-made targets to sharpen decision boundaries — with the goal of making the model more sensitive to "acoustic mimics."

Stage 3: Supervised Contrastive Learning

The third stage applies Supervised Contrastive Learning (SupCon) to further separate targets from clutter in the representation space, theoretically producing cleaner class geometry.

Supervised Contrastive Learning (SupCon), proposed by Khosla et al. in 2020, is a supervised extension of self-supervised contrastive learning methods like SimCLR. Its loss function pulls encoded representations of same-class samples closer together while pushing different-class samples apart, forming compact and well-separated class clusters in the feature space rather than relying on a single classification hyperplane. Compared to cross-entropy loss, SupCon encourages the model to learn more generalizable geometric structure, and is theoretically advantageous for tasks with high intra-class variance or high inter-class similarity. However, its effectiveness is highly dependent on the feature space not yet being well-organized before this stage — if prior stages have already aligned the features, there is little room left for SupCon to make additional geometric adjustments. This is precisely the underlying reason for the null results observed in this paper.

Experimental Design: Rigorous Controls and Statistics

Evaluation is based on real-world SAS data collected at sea, using mission-level geographic splits to prevent data from the same geographic region leaking across training and test sets. All experimental branches are compared at 85% test recall, with results repeated across three random seeds to ensure statistical reliability.

This design deserves recognition — rather than relying on single-point metric comparisons, it uses multi-seed repetition and mean ± standard deviation reporting, allowing the true contribution of each stage to be quantitatively verified.

Mission-level geographic split is an important methodological choice in sonar ATR evaluation. Because SAS image collection repeatedly covers the same geographic areas across missions, a random image-level split could allow the training and test sets to share the same seafloor terrain, sediment textures, or even different viewpoints of the same target — leading to inflated performance estimates. Geographic splitting enforces that the test set comes from sea areas never seen during training, more faithfully simulating the domain shift pressure encountered in real deployment. By contrast, natural image benchmarks (e.g., ImageNet classification) typically do not have this spatial correlation problem, so sonar research that directly adopts random splitting risks drawing conclusions with limited practical value. The use of this splitting strategy in this paper makes the experimental findings significantly more relevant to real-world deployment.

Key Finding: LoRA Is the Main Effect; Stacked Refinement Is an Illusion

The experimental data tells a remarkably instructive story. LoRA carries nearly all of the performance gain: under the same frozen backbone, the Area Under the Precision-Recall Curve (AUPRC) rises from 0.300 to 0.679±0.027. Even more striking, the Rank 4 configuration achieves this while training only 0.26% of the weights — an extraordinarily high level of parameter efficiency.

The two subsequent refinement stages, however, yield almost no additional benefit. Hard-negative mining shows an AUPRC change of just -0.0045±0.0119 compared to a random curriculum of equivalent size; SupCon produces an AUPRC change of +0.0002±0.0096 relative to the preceding stage. Both fall well within the margin of error — statistically null results.

The researchers offer a coherent explanation: hard-negative mining operates on data the encoder has already been fitted to, and therefore cannot provide new learning signal; and the supervised stage has in practice already established most of the target-clutter geometric structure, leaving SupCon with nothing meaningful left to do.

Implications for Engineering Practice

The value of this work lies not only in achieving a SOTA-style numerical improvement, but also in its honest reporting of what does not work. The conclusion is clear and measured: one efficient adaptation stage is sufficient; stacked refinement is not worth the cost.

For resource-constrained real-world deployments, this is a pragmatic signal — rather than blindly stacking complex training techniques, it pays to use parameter-efficient methods like LoRA well. Achieving more than a 2× improvement in AUPRC with only 0.26% trainable weights is highly attractive for edge devices and rapidly iterated military or civilian sonar systems.

This also reinforces an emerging consensus in the transfer learning field: in data-scarce specialized domains, choosing the right adaptation paradigm often matters more than stacking training pipelines.