Use of Classification Trees to Apportion Acoustic Targets to Species: Application to the 2011 Lake-Wide Acoustic and Midwater Trawl Survey of Lake Superior

Acoustic (AC) methods are used to estimate the density of pelagic prey fish in large lakes, with midwater trawling (MT) used to apportion these estimates to species. Accurate apportionment of species composition is a challenge because only a small fraction of water sampled acoustically is sampled with MT gear, and catch composition can be biased because different fish species (and sizes) have different capture probabilities. We demonstrate a new method of apportionment that uses classification tree modeling to identify variables important for classifying AC targets to species.During the summer of 2011, we conducted a spatially-balance lake-wide AC and MT survey of Lake Superior that included 51 sites in four bathymetric strata (0-30 m, 30-100 m, 100-200 m and > 200 m). Using a data set of 83 MT samples, we developed tree-based models for each stratum. Fish length was the most important variable for classifying species in all strata. Using previously published data on AC target strengths (TSs) and lengths of individual fish, we developed a model to predict lengths that included stochastic variability. We verified our apportionment method using an independent data set (18 paired AC and MT samples) and found predicted proportions of important prey fish species did not vary significantly from observed MT catch proportions. Exotic rainbow smelt (Osmerus mordax) were the most common species at bathymetric depths < 100 m with their population estimated at 719 million (3,837 metric tons). Kiyi (Coregonus kiyi) were the most abundant species at depths > 100 m (396 million; 5,686 metric tons). Cisco (C. artedi) occupied all strata with their population estimated at 206 million (43,268 metric tons). Total prey fish abundance was estimated at 1.68 billion fish (209/ha), weighing an estimated 54,734 metric tons (6.80 kg/ha). Siscowet lake trout (Salvelinus namaycush siscowet) were the most abundant piscivore (24 million; 48,708 metric tons). Because the apportionment method we describe relies on AC sizing of fish, it is less sensitive to bias caused by species (and sizes) having different MT capture probabilities. Apportionment should be accurate provided catch composition of both small and large fish is similar to the true composition of these different size fish in a given lake.