Abstract
Underwater acoustic target recognition (UATR) models trained in one body of water may fail to generalize to another; when the target water is unlabeled, unsupervised domain adaptation (UDA) is the standard remedy. On a frozen binary protocol spanning four corpora—the ocean corpus Oceanship from the Strait of Georgia, the freshwater-lake corpus QiandaoEar22, an AIS-auto-labeled VTUAD reconstruction, and the nearshore corpus ShipsEar—we benchmark three UDA families (correlation alignment (CORAL), multi-kernel MMD, and adversarial DANN/CDAN) against source-only and target-only supervised references, across 40 method–direction combinations and 344 frozen-matrix runs. Adaptation can work, but is strongly direction-dependent. Lake-to-sea, MMD raises macro-F1 from 0.359 (near-total majority-class collapse) to 0.524 ± 0.090—74% of the supervised reference—and restores passenger-ship recall to 662/1,214; all four UDA methods improve significantly over source-only, while pairwise differences among them are unresolved. Sea-to-lake, the supervised reference is low (macro-F1 0.551; minority test class: 89 segments), so the measurable value of UDA there is variance compression (variance ratios 0.044–0.376) and class-bias correction. The most stable method in each direction is always a moment-matching method; CDAN shows direction-dependent early-stopping collapse. CORAL leads the VTUAD directions (+0.129, p = 0.0008), whereas MMD produces significant negative transfer under the frozen split (−0.147, p = 0.005)—a cell whose sign reverses under a hull-disjoint MMSI re-split (128 additional runs; +0.084, p < 0.001). On the hardest ShipsEar direction (QiandaoEar22→ShipsEar) no evaluated method is usable (macro-F1 ≤ 0.194 against a 0.8356 majority baseline). Transfer direction and target-side constraints dominate the choice of UDA objective.