Decoding marine ecosystem complexity through 3D acoustic sensing and Data Science: an integrated approach to environmental status assessment

Marine ecosystem monitoring faces critical limitations: traditional approaches rely on costly and in- termittent observations that fail to capture continuous ecological dynamics or provide early warning signals of system transitions. While passive acoustic monitoring offers a scalable and continuous al- ternative, existing methodologies often lack a robust theoretical and analytical framework to reliably interpret soundscape variability in terms of environmental dynamics. This thesis introduces UW-AcousticML, a physically grounded data science framework that interprets underwater acoustic observations as indicators of environmental dynamics and air–sea interaction regimes. The approach integrates metrological validation, physically interpretable spec- tral descriptors, and explainable machine learning to bridge multi-scale acoustic measurements with underlying physical processes, enabling their interpretation as proxies of system-level dynamics. Using multi-scale acoustic data collected at the Acqua Alta Oceanographic Tower across sea- sonal conditions, the study extracts scale-invariant spectral features and energy-based indicators linked to environmental forcing. These descriptors are complemented by deep learning models ap- plied to spectrogram sequences, enabling the detection of regime-dependent acoustic patterns. A convolutional neural network with temporal attention achieves a classification accuracy of 60.9% under realistic data constraints, with interpretability analysis demonstrating that model decisions are driven by physically meaningful time–frequency structures. The results show that underwater acoustic data encode signatures of air–sea interaction regimes and can serve as sensitive indicators of environmental variability. Although direct validation of ecological regime shifts remains limited by data availability, the combined use of physically grounded descriptors and interpretable machine learning provides a robust framework for analyzing complex acoustic datasets. This work advances the use of underwater acoustics as an integrative, system- level environmental observable, demonstrating how data science methods can be embedded within a physically consistent framework to support scalable ecosystem monitoring and assessment. This work contributes to complexity science by developing an integrative monitoring methodol- ogy that bridges physically grounded analysis with data-driven approaches, highlighting the poten- tial of lightweight models to extract meaningful patterns from complex acoustic datasets. Future work will extend this framework across diverse environmental conditions and integrate additional observations to further advance ecosystem-scale indicators.