Compact representation of temporal processes in echosounder time series via matrix decomposition

The recent explosion in the availability of echosounder data from diverse ocean platforms has created unprecedented opportunities to observe the marine ecosystems at broad scales. However, the critical lack of methods capable of automatically discovering and summarizing prominent spatio-temporal echogram structures has limited the effective and wider use of these rich datasets. To address this challenge, a data-driven methodology is developed based on matrix decomposition that builds compact representation of long-term echosounder time series using intrinsic features in the data. In a two-stage approach, noisy outliers are first removed from the data by principal component pursuit, then a temporally smooth nonnegative matrix factorization is employed to automatically discover a small number of distinct daily echogram patterns, whose time-varying linear combination (activation) reconstructs the dominant echogram structures. This low-rank representation provides biological information that is more tractable and interpretable than the original data, and is suitable for visualization and systematic analysis with other ocean variables. Unlike existing methods that rely on fixed, handcrafted rules, this unsupervised machine learning approach is well-suited for extracting information from data collected from unfamiliar or rapidly changing ecosystems. This work forms the basis for constructing robust time series analytics for large-scale, acoustics-based biological observation in the ocean.