Non-intrusively estimating the live body biomass of Pintado Real® fingerlings: A feature selection approach

• Published in: Ecological informatics 2022-05, Vol.68, p.101509, Article 101509
• Main Authors: Pache, Marcio Carneiro Brito, Sant'Ana, Diego André, Rezende, Fábio Prestes Cesar, de Andrade Porto, João Vitor, Rozales, João Victor Araújo, de Moraes Weber, Vanessa Aparecida, da Silva Oliveira Junior, Adair, Garcia, Vanir, Naka, Marco Hiroshi, Pistori, Hemerson
• Format: Article
• Language: English
• Subjects: Computer vision; Live fish in movement; Machine learning; Morphometric measurements; Predictive model
• ISSN: 1574-9541
• DOI: 10.1016/j.ecoinf.2021.101509

Aquaculture has undergone significant technological advances in recent decades, which has enabled the expansion of fish protein production worldwide. However, some conventional processes in fish farming facilities, such as the weighing of fingerlings, usually occur manually and laboriously, which can cause physical damage to the fingerlings, precision balance errors, and financial losses for fish farmers. The main contribution of this research is the creation of an automatic weighing application for fingerlings and juveniles, in contrast to the laborious manual weighing that occurs particularly in small and medium aquaculture facilities. This paper presents results for the prediction of biomass of moving live fingerlings using supervised learning algorithms. It is applied in two new datasets, the first with illumination and the second without illumination. For both datasets, the images are pre-processed and segmented to extract the characteristic vectors systematically trained with cross-validation by four regression algorithms. So, the selection of attributes was performed based on correlation and relative importance which allowed the removal of some attributes which implied null significance for the model. The best result was for the experiment with attribute and frame selection applied to the lighting dataset and the Linear Regressor obtained a R2 = 0.76 and MAE = 0.83 g. The proposed model shows as promising in comparison with other approaches in the literature. •Automatic weighing fingerlings in contrast to the laborious manual•Availability of a new dataset of the hybrid species called Pintado Real® fingerlings•Training and tests conducted by four supervised learning regression algorithms•The combination of thirty-one extractors to compose the characteristic vector•Attributes selection based on correlation and relative importance