Don’t work too hard: Subsampling leads to efficient analysis of large acoustic datasets

Abstract

Echo-integration measurements have been traditionally made from dedicated fisheries survey vessels, but extensive measurements from moorings, autonomous vehicles, and fishing vessels are increasingly available. Processing these data by traditional means developed for well-staffed fisheries surveys can be prohibitively time-consuming, which has limited their use. Automated processing methods exist to efficiently handle these large datasets; however, as compared to post-processing by trained analysts, these methods require substantial expertise and methodological development, and they often produce less certain results. Here, we evaluate the use of subsampling, which takes advantage of the spatial correlation common in many fish populations, to improve the efficiency of traditional processing methods while retaining a high level of precision. We subsampled data from an eastern Bering Sea walleye pollock (Gadus chalcogrammus) acoustic-trawl survey and compared estimates of pollock backscatter from subsamples to those from the full survey. Over a survey-wide scale, processing < 5% of the data resulted in estimates within 5% of those from processing the full survey. This suggests that in some applications there may be diminishing returns associated with exhaustively processing large spatially correlated datasets. We present an example that applies this simple approach by subsampling archived echosounder data from chartered fishing vessels to prioritize the areas surveyed in future surveys, which would not have been feasible without subsampling. When averaged values over large scales (e.g. over a survey domain) are required, precise echo integration estimates can be obtained with modest effort by processing relatively small subsamples of a dataset.