Feature extraction and classification of sonar echoes from rigid and elastic targets:Application to underwater mines detection and environmental risk mitigation in marine ecosystems

Details

Supervisors

Lambot, Sébastien ; Hurtado Erasso, Camilo

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences et technologies de l'environnement, à finalité spécialisée

Abstract

Reliable discrimination between rigid underwater targets, such as natural rocks, and elastic targets, including unexploded ordnance (UXO) and sea mines, is a fundamental challenge in active sonar-based survey operations. Elastic objects, such as the metallic shells typical of UXOs and mines, generate resonant acoustic echoes when insonified at acoustic frequencies whose wavelengths are comparable to the target dimensions (10–30 kHz in this study), producing backscatter signatures that differ from those of rigid scatterers. This work exploits these differences through a feature-based machine learning approach applied to a publicly available dataset of controlled in-air sonar measurements. The analysis is performed on backscattered acoustic time series after standard signal processing steps including matched filtering. Data from four targets — a solid polyurethane sphere and a hollow aluminum shell (elastic), and two medium-density fiberboard block letters (rigid) — acquired across four environments of increasing complexity (free-field, flat interface, rough interface with proud targets, and rough interface with partially buried targets) were used. Features were extracted from the real, imaginary, and envelope components of each echo window using the Python library tsfresh, and subsequently reduced via statistical feature selection. Classification was performed using Random Forest and XGBoost ensembles. In free-field and flat-interface conditions, both classifiers exceeded 0.97 accuracy. Performance degraded substantially in rough-interface scenarios, approximately 0.61 to 0.65, where surface clutter obscures the resonance signatures critical to classification. These results highlight the sensitivity of time-domain feature extraction to environmental conditions, motivating the development of more robust preprocessing strategies.