Source-Free Domain Adaptation for Underwater Acoustic Target Recognition: When Batch Normalization Statistics Alone Suffice
Year: 2026
Type: preprint
Abstract: Deep classifiers for underwater acoustic target recognition (UATR) collapse when the deployment water area differs from the training area. Unsupervised domain adaptation (UDA) recovers much of the loss, but every UDA method requires joint access to the source data during adaptation. That assumption fails precisely where UATR matters most, because military and commercial sonar recordings are typically classified or proprietary. Source-free UDA (SFUDA), which adapts a frozen source model using only unlabeled target data, is therefore the practically relevant setting, yet it has never been benchmarked for underwater acoustics. We present the first SFUDA benchmark for UATR: three representative mechanisms — batch-normalization (BN) statistics re-estimation (bn_adapt), entropy minimization over BN affine parameters (TENT), and information maximization with pseudo-label self-training (SHOT) — evaluated on the two cross-water-area directions Oceanship→QiandaoEar22 and QiandaoEar22→Oceanship (2 directions × 3 methods × 3 seeds) under a strict label-free protocol, with every number directly comparable to previously reported source-data UDA results on the identical protocol. The central finding is a re-evaluation, not a new method: bn_adapt, the established AdaBN mechanism with zero gradients, zero model selection, and a per-configuration cost of 15–25 s, is the strongest SFUDA method in both directions (macro-F1 0.5023 ocean→lake, 0.5060 lake→ocean), statistically indistinguishable from the best source-data UDA method at n = 3 seeds in both directions (ocean→lake +0.038, p = 0.095; lake→ocean −0.087, p = 0.259; neither significant at n = 3), and it restores the collapsed passenger-ship class from 0/1,214 to up to 380/1,214 detections in the lake→ocean direction. Gradient-based SFUDA is weaker and structurally fragile: TENT reproducibly collapses into a single-class model that its own label-free selection criterion ranks as the best checkpoint, exposing label-free model selection rather than adaptation capacity as the primary bottleneck; SHOT is stable but drifts to a class-biased solution with below-chance accuracy. For channel-dominated shifts with small label spaces, BN statistics re-estimation should be the mandatory first step before any gradient-based adaptation.
Source: Preprints.org
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213306341