Quantitative metrics to experimentally benchmark geoacoustic inversion methods

Abstract

Geoacoustic inversion schemes are complex methods that combine various experimental designs (e.g., arrays or single receivers, fixed or moving sources,…), different data processing methods (e.g., tracking ray arrivals, filtering modes, estimating spectral matrices,…) and different inversion methods (e.g., brute force grid search, stochastic optimization,…). Because of this complexity and variability, geoacoustic inversion methods are difficult to benchmark. Although simulated benchmarks are possible, experimental benchmarks are more difficult since (1) experimental ground truth is never fully available and (2) different methods may sense different features/scales of the seabed. Nonetheless, assessing the performance of geoacoustic inversion methods is of paramount importance, since the resulting geoacoustic models are used to predict acoustic field for various applications, including Navy operations or noise pollution monitoring. In this presentation, we propose quantitative metrics to assess the experimental performances of inversion methods. The proposed metrics allow inter-comparison of various geoacoustic models, but also comparison of predicted acoustic field with experimental (independently collected) data. The metrics are successfully applied on a subset of data collected during the Shallow Water 2006 experiment.