A carregar...
Publicação
Analysis of Empirical Ocean Sound‑Speed Divergence: An Explainable Machine Learning Approach Using ARMOR3D
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.31 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Sound-velocity profiles (SVPs) are essential for studying sonar signals, establishing underwater
communications, and understanding how marine-mammal calls travel through the ocean.
However, widely used empirical equations for calculating sound speed in water often produce
inconsistent results when applied beyond their original calibration regions, which can create
uncertainty for acoustic operations. This thesis investigates the extent and causes of these
discrepancies by applying five standard sound-speed equations to 747 million records from
the Copernicus ARMOR3D dataset. The analysis produces a global, four-dimensional data cube
(latitude, longitude, depth, and time) illustrating where, when and by how much the equations
diverge. Our results revealed that differences can exceed 0.6 m/s near the surface and
gradually decrease with depth. A Random Forest model using only temperature, salinity, and
depth reproduces 95% of the mapped spread, achieving a root-mean-square error of 0.05 m/s.
Importantly, explainable machine learning analysis (through SHAP) indicates that temperature
accounts for about 65% of the variance in predicted sound-speed divergence across the test
dataset, hydrostatic pressure for 26%, and salinity for 8%. Adding mixed-layer depth and
surface current speed as input variables improves local accuracy by up to 10% in dynamic
regions such as the Gulf Stream and Kuroshio during winter, the Antarctic Circumpolar Current
in spring, and the Agulhas Return Current in autumn. These findings support a practical twolevel approach: a simple core model for most of the ocean, and an enhanced version that
activates region-specific variables in dynamic zones (like western boundary currents) where
they significantly reduce prediction error. This work provides a benchmark for SVP modelling,
supporting future improvements in sonar planning, offshore operations, and marine
ecosystem research that rely on precise sound-speed predictions.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Business Intelligence
Palavras-chave
Ocean Acoustics Sound-Velocity Profiles ARMOR3D Machine Learning SHAP Explainable Artificial Intelligence SDG 14 - Life below water