Automatic classification of acoustically detected krill aggregations: A case study from Southern Ocean

• Published in: Environmental Modelling & Software
• Main Authors: Fontana, Ignazio, Barra, Marco, Bonanno, Angelo, Giacalone, Giovanni, Rizzo, Riccardo, Mangoni, Olga, Genovese, Simona, Basilone, Gualtiero, Ferreri, Rosalia, Mazzola, Salvatore, Lo Bosco, Giosué, Aronica, Salvatore
• Format: Article in Journal/Newspaper
• Language: English
• Published: 2022
• Subjects: Hierarchical clustering; k-means; Krill; Ross Sea; Internal validation indices; Acoustic; Settore INF/01 - Informatica; Settore ING-INF/05 - Sistemi Di Elaborazione Delle Informazioni; Southern Ocean; Euphausia superba
• Online Access: http://hdl.handle.net/10447/537375; https://doi.org/10.1016/j.envsoft.2022.105357

Acoustic surveys represent the standard methodology to assess the spatial distribution and abundance of pelagic organisms characterized by aggregative behaviour. The species identification of acoustically observed aggregations is usually performed by taking into account the biological sampling and according to expert-based knowledge. The precision of survey estimates, such as total abundance and spatial distribution, strongly depends on the efficiency of acoustic and biological sampling as well as on the species identification. In this context, the automatic identification of specific groups based on energetic and morphological features could improve the species identification process, allowing to improve the precision of survey estimates or to overcome problems related to biases in biological sampling. In the present study, we test the use of well-known unsupervised clustering methods focusing on two important krill species namely Euphausia superba and Euphausia crystallorophias. In order to obtain a reference classification, the observed echoes were first classified according to specific criteria based on two parameters accounting for the acoustic response at 38 kHz and 120kHz. Different clustering methods combined with three distance metrics were then tested working on a wider set of parameters, accounting for the depth of insonified aggregation as well as for energetic and morphological features. The clustering performances were then evaluated by comparing the reference classification to the one obtained by clustering. Obtained results showed that the k-means performs better than the considered hierarchical methods. Our findings also evidenced that working on a specific set of variables rather than on all available ones highly impact k-means performances.