Image-based machine learning reveals structural “fingerprints” of pelagic seascapes in shipboard echograms

Abstract Echograms generated using active shipboard acoustics data provide visually rich insights into the structure and behaviour of interior pelagic communities, yet current numerical approaches underuse this complexity. Here, we develop a novel approach for analysing diel echogram imagery using convolutional autoencoders (CAEs). We apply this approach to 1 596 daily echograms collected using uncalibrated Acoustic Doppler Current Profilers (ADCPs) from 306 cruises in the Pacific Ocean between 2001 and 2018 over nearly 500,000 km. By compressing each echogram into a low-dimensional feature space, we identify recurring structural patterns—such as diel vertical migrations and scattering layers—and use these features to cluster the echograms into six coherent regional modes across the basin. These modes, or acoustic “fingerprints”, provide a new, interpretable framework for classifying pelagic ecosystems based on structure rather than backscattering intensity. When mapped and predicted to the entire Pacific basin using environmental covariates, our classification aligns closely with previous biogeographies derived from calibrated scientific echosounder data, demonstrating the potential of unsupervised deep learning to extend ecosystem monitoring using widely available ADCP archives. The Pacific ADCP archive used in this study is nearly 40% larger than any of the 38 kHz archives used in previous global studies: by enabling analysis of uncalibrated acoustic data, this approach unlocks a vast global archive for ecological classification, expanding opportunities for ecosystem monitoring, biogeography, and change detection.