Beschreibung:
Recent advances in technology have produced a deluge of acoustic data that offer opportunities to study ecological processes at scales that are not possible previously. A prominent example is the continuous data flow from the fleet of upward-looking echosounders installed by the Ocean Observatories Initiatives (OOI) at diverse global locations. However, it is unclear if conventional echosounder analysis routines are effective in analyzing these data sets, due to the generally scarce ground-truth resources and limited empirical knowledge at many locations. In this study, we explored the use of signal decomposition methods in discovering daily patterns of marine organism activities in long-term echosounder time series. Using non-negative matrix factorization (NMF), we show that the echograms can be decomposed into a weighted combination of discrete components, each with acoustic features that can be exploited for inferring the underlying biological assemblage. In addition, the component weights provide an avenue to capture echogram variations in a significantly reduced dimensional space, which can be utilized to describe changes in the ecosystem. Building on these results, we are developing new formulations to incorporate continuity in both time and space to construct a fully scalable framework for data-driven discovery using long-term echosounder data.