Optimizing automated feeder algorithms for feed intake prediction in barramundi (Lates calcarifer)

Abstract

Efficient aquaculture management requires optimizing feed intake to balance fish productivity and environmental waste. Automated feeders with daily feed intake (DFI) prediction capabilities are a popular tool for achieving this balance. The aim of this study was to optimize DFI prediction using various learning algorithms and nonlinear functional forms, while also assessing the utility of adaptive algorithms for the development of automated feeders. We selected 11 commercial feeds that represent common farm uses for barramundi (Lates calcarifer) in Southeast Asia and conducted three 8-week trials feeding 1872 juveniles to satiation. We collected 30 variables, including environmental factors (E), fish characteristics (F), feed chemical composition (C), and pellet physical characteristics (P). We compared three statistical methodologies: stepwise regression (SWR), Bayesian model averaging (BMA), and multiple factorial polynomial (MFP). The selection of commercial feed had a significant impact on fish growth, feed conversion ratio, total feed intake and nutrient retentions (P < 0.001). All three methodologies consistently identified the same 10 variables as significant (P < 0.05) for accurately describing DFI, yielding high accuracy levels (R2 = 0.85–0.88). Three variables, namely fish length, fish batch, and water temperature, accounted for the vast majority (77%) of the DFI variance. Beyond feed chemical composition, the preference for sinking pellets and the regular increase in pellet size emerged as the predominant factors enhancing DFI. SWR modelling was robust, but MFP model demonstrated the highest accuracy (R2 = 0.88), indicating that nonlinear functional relationships between variables was beneficial. BMA showed that a multimodel approach was effective in developing adaptive DFI prediction, with the most accurate models incorporating at least one variable from each category (E, F, C, P). This study proposes an optimization of DFI prediction for automated feeders, which can reduce the risk of overfeeding in barramundi farms. The use of adaptive algorithms was shown to facilitate the prioritization of sensor integration, lower development costs, and enable technology level adjustments from low-tech to high-tech. These findings provide valuable insights for the development of highly efficient and cost-effective feeding systems that align more effectively with the diverse demands of the aquaculture industry.