PA-HydroPreNet: Multi-Task Spectral Enhancement and Bioacoustic Detection for Hydrophone-Array Monitoring
Year: 2026
Type: article
Abstract: Hydrophone-array recordings support passive acoustic monitoring, but overlapping biological, anthropogenic, and geophysical sources can complicate enhancement and event detection. This study presents PA-HydroPreNet, a multi-task spectral enhancement and bioacoustic detection framework for four-channel observations. Evaluation used 8,200 controlled mixtures constructed from real biological and passive underwater recordings, with partition control at the biological source-folder and background-deployment levels. Across three independent seeds, the model achieved a mean-centred SI–SNRmag of $12.9298 \pm 0.1156$ dB, compared with 7.8960 dB for the raw mixtures. Waveforms synthesized using the observed mixture phase improved SI–SNR from 5.0372 to 9.6020 dB. However, the model does not estimate phase. Biological detection achieved an AUROC of $0.8382 \pm 0.0306$ and an F1-score of $0.8253 \pm 0.0242$ . Environmental-label perturbation and probe-estimated metadata produced marginal changes under the evaluated conditions. A matched seed-42 no-FiLM conditioning diagnostic also produced comparable performance to the full seed-42 configuration, with SI–SNRmag of 13.0610 versus 12.8701 dB and AUROC of 0.8674 versus 0.8664. Source-aware branches showed source-related information but negative source-specific waveform SI–SNR and are interpreted as auxiliary representations rather than physical separators. The exploratory delay-index head did not outperform constant references. A seed-42 diagnostic associated the multi-scale detection head with higher enhancement than its single-scale counterpart; this single-seed result is descriptive rather than cross-seed evidence. The findings support spectral enhancement and biological acoustic detection under controlled four-channel conditions, while continuous field-array, unseen-geometry, variable-channel-count, and downstream localization performance remain unvalidated.
Source: IEEE Access
Authors: Vaka Suresh, Mohan Allam
Institution: VIT-AP University
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7204151273