Year: 2026
Type: conference-paper
Abstract: False detections in Internet of Underwater Things (IoUT) monitoring create unnecessary alerts and communication burden. This paper presents Physics-Guided Dynamic STR-SubNet, a compact anchor-free detector that conditions structural-textural feature sharing on an imagederived degradation proxy. A controlled five-fold ablation separates hard routing (A0), learned routing with a trainingonly constant gate (A1), physics gating with fixed mixing and A0-matched capacity (A2), and the full mechanism (A3). Micro- F1 changes from 0.6067 (A0) to 0.6306, 0.6324, and 0.6301 for A1-A3. A2 improves all five folds without added parameters, whereas A3 gives the strongest false-positive control (205 to 150). Across 17 proxy-weight settings, one-at-a-time ±20% perturbations keep micro-F1 within 0.6285-0.6307. A validationselected branch-preservation fallback leaves clean F1 unchanged after rounding and improves pooled dark, lowcontrast, and severe F1 from 0.3149 to 0.3511, although severe collapse remains poor (0.0913). A public LIACI test does not transfer the controlled gain (0.1087 to 0.1054). Actual held-out detections replayed as QoS-1 MQTT events reduce false-event payload by 26.294%; latency remains a co-located softwarepipeline measurement rather than underwater-link evidence.
Language: English
Subfield: Water Science and Technology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221241651