Segmentation of Sonar Spectrograms Using Physically Informed Deep Learning
Year: 2026
Type: article
Abstract: Detecting and classifying ship-radiated noise is challenging due to overlapping acoustic phenomena and complex interference patterns. We develop a simulation-driven approach to segment LOFARgrams, low-frequency spectrograms used in passive sound navigation and ranging analysis, into harmonic machinery tones, nonharmonic signals, and Lloyd’s mirror interference fringes. Training data are generated using a deterministic and physically grounded simulation based on additive synthesis for harmonic and nonharmonic tonal-like signals, while fractional delay-based modeling is used to simulate Lloyd’s interference patterns. We formulate the problem as a multilabel segmentation task and compare it against class-specialized models trained independently for each signal type. Although the multilabel model is intended to leverage shared structure across acoustic phenomena, our ablation study shows that this assumption does not hold for the considered data set: class-specialized models consistently outperform the joint model across all signal types. We further analyze how spectrogram resolution and encoder receptive field affect the detection of low-frequency fundamentals and higher order harmonics, showing that global context is essential for distinguishing harmonic sets from locally similar nonharmonic features. Frequency-preserving encoder architectures, including shallow convolutional neural networks and transformer models with global attention, balance context and resolution. A human annotation study quantifies inter-annotator agreement, providing a benchmark for model–human comparison.
Authors: Nils Müller, Jens Reermann, Tobias Meisen
Institutions: Atlas Elektronik (Germany), University of Wuppertal
Language: English
Subfield: Oceanography
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7221668041