Classification of broadband target spectra in the mesopelagic using physics-informed machine learning.

Abstract

Broadband echosounders measure the scattering response of an organism over a range of frequencies. When compared with acoustic scattering models, this response can provide insight into the type of organism measured. Here, we train the k-Nearest Neighbors algorithm using scattering models and use it to group target spectra (25-40 kHz) measured in the mesopelagic near the New England continental shelf break. Compared to an unsupervised approach, this creates groupings defined by their scattering physics and does not require significant tuning. The model classifies human-annotated target spectra as gas-bearing organisms (at, below, or above resonance) or fluid-like organisms with a weighted F1-score of 0.90. Class-specific F1-scores varied-the F1-score exceeded 0.89 for all gas-bearing organisms, while fluid-like organisms were classified with an F1-score of 0.73. Analysis of classified target spectra provides insight into the size and distribution of organisms in the mesopelagic and allows for the assessment of assumptions used to calculate organism abundance. Organisms with resonance peaks between 25 and 40 kHz account for 43% of detections, but a disproportionately high fraction of volume backscatter. Results suggest gas bearing organisms account for 98.9% of volume backscattering concurrently measured using a 38 kHz shipboard echosounder between 200 and 800 m depth.