As a remote sensing tool, moored echosounders have played an important role in observing temporal changes of animal distributions in the water column over large temporal scales, ranging from days, months to seasons and even years. In this work we take advantage of the power of matrix decomposition techniques in exploiting regularity in the data to automatically discover low-dimensional structures in large data sets, and develop a methodology that can effectively remove noise and extract dominant daily echogram patterns from long-term echosounder series collected by moored echosounders deployed by the U.S. Ocean Observatories Initiative (OOI). These echosounders are located on the continental shelf and the shelf break in the rich Northern California Current System that is strongly influenced by seasonal upwelling. The echosounders are collocated with a suite of oceanographic sensors, allowing systematic analysis of multi-modal data streams. Our analysis results in an array of daily echogram patterns (components) whose time-varying linear combination (activations) captures major structures in these time series. Together, these components and variations provide a compact representation that allows intuitive interpretation of such a long-term observational dataset.