Classification of Southern Ocean krill and icefish echoes using random forests

Abstract

Listen

Translate article

Abstract Target identification remains a challenge for acoustic surveys of marine fauna. Antarctic krill, Euphausia superba, are typically identified through a combination of expert scrutiny of echograms and analysis of differences in mean volume backscattering strengths (SV; dB re 1 m−1) measured at two or more echosounder frequencies. For commonly used frequencies, however, the differences for krill are similar to those for many co-occurring fish species that do not possess swimbladders. At South Georgia, South Atlantic, one species in particular, mackerel icefish, Champsocephalus gunnari, forms pelagic aggregations, which can be difficult to distinguish acoustically from large krill layers. Mackerel icefish are currently surveyed using bottom-trawls, but the resultant estimates of abundance may be biased because of the species' semi-pelagic distribution. An acoustic estimate of the pelagic component of the population could indicate the magnitude of this bias, but first a reliable target identification method is required. To address this, random forests (RFs) were generated using acoustic and net sample data collected during surveys. The final RF classified as krill, icefish, and mixed aggregations of weak scattering fish species with an overall estimated accuracy of 95%. Minimum SV, mean aggregation depth (m), mean distance from the seabed (m), and geographic positional data were most important to the accuracy of the RF. Time-of-day and the difference between SV at 120 kHz (SV 120) and that at 38 kHz (SV 38) were also important. The RF classification resulted in significantly higher estimates of backscatter apportioned to krill when compared with widely applied identification methods based on fixed and variable ranges of SV 120–SV 38. These results suggest that krill density is underestimated when those SV-differencing methods are used for target identification. RFs are an objective means for target identification and could enhance the utility of incidentally collected acoustic data.