ReviewAnalysis of backscatter properties and application of classification procedures for the identification of small pelagic fish species in the Central Mediterranean

•We investigate discrimination amongst schooling fish using ΔdB and school descriptors.•Horse mackerel, anchovy and sardine cannot be distinguished using ΔdB alone.•School depth plays an important role in the identification of these species.•Energetic or morphometric parameters alone classify sardine reasonably well.•All the available school descriptors classify 85% of the known species.

The pelagic realm of the Central Mediterranean Sea is populated by four main species of fish: sardine, anchovy, horse mackerel and a mix of other pelagic fish species. In this study we employed a multifrequency acoustics approach to detect and classify fish schools of these groups. Monospecific trawl catches were selected from eight acoustic surveys and examined in relation to the coincident acoustic data. The backscattering properties of the three main species were determined using the decibel difference (Sv120 − Sv38) and the frequency response (NASC120/NASC38). The results indicate that schools of these species cannot be distinguished on the basis of energetic properties alone, because they are very similar in physiology and scattering is dominated by the swimbladder, which is similar in shape and size. However, the use of classification models (classification tree, random forest), using energetic features, as well as bathymetric and morphometric parameters, allowed for some discrimination among the groups. According to the classification tree, school depth was found to play an important role in the identification of these fish groups, especially for anchovy and horse mackerel, for which the contribution to the overall performance of the tree was about 20%. The tree models, with only energetic or morphometric parameters, were able to classify sardine schools reasonably well, but not so well for anchovy and horse mackerel. Using a random forest method, which accounted for the variability in the learning sample, an accuracy of 85% in the overall classification rate was reached with a greater power of discrimination for sardine and anchovy schools.