SA-SRDC-Loc: Single-AUV-assisted three-dimensional underwater localization via selective residual distance learning and depth-constrained robust optimization

Accurate three-dimensional localization in underwater acoustic sensor networks remains challenging because irregular acoustic propagation and multi-hop connectivity weaken the relationship between network topology and physical distance. This paper proposes SA-SRDC-Loc, a single-AUV-assisted localization framework that treats successive AUV reports as temporally ordered observations of one continuously moving platform rather than as independent anchors. A global DNN estimates Euclidean node-to-AUV distances from observable topology, depth, reported AUV positions, and acoustic context, while validation-gated regional residual branches selectively correct hop-dependent errors. The learned ranges are then used for depth-constrained all-reference initialization, followed by holdout-safe robust refinement in which candidate updates are accepted only when both fit and unused-observation objectives improve. Across 30 paired deployments, SA-SRDC-Loc achieves a mean localization error of 23.63 ± 0.84 m, corresponding to a 15.3% reduction relative to the strongest competing method for this metric. Ablation results further show that holdout-safe refinement provides the larger localization gain, while selective residual correction offers a complementary improvement.