Exploring matrix and tensor factorization for discovering latent structures in large echosounder datasets

Moored autonomous echosounders are increasingly popular as an integrated component in ocean observing systems for measuring biological response to environmental changes. Different from ship-based observations, these large echo datasets often lack concurrent ground truth information from net or optical samples and reliable calibration, both of which are required for conventional analysis routines. However, even though accurate inference of organism abundance is not possible in such scenarios, rich spatio-temporal structures exist in the data and may be exploited to capture variation of migration patterns of different animal groups in the water column. Here, we explore the use of unsupervised matrix and tensor factorization approaches to analyze the long-term echo data sets from the Ocean Observatories Initiative (OOI) network. The data stretch in multiple dimensions, including time, space, and frequency. We restructure the echo data to explore latent structures at different temporal scales using different factorization formulations. Outputs from different methods are compared among one another and with those from conventional echo analysis routines. The results show the importance of augmenting generic factorization formulations with temporal and spatial continuity constraints for biologically meaningful analysis of large echosounder datasets.