Using a Multi-Output Neural Network Model to Standardize Heterogeneous Fisheries Data

Biological data in fishery ecology have complex structures and are highly heterogeneous. Catch per unit effort (CPUE) estimated from fishery-dependent data are often used to characterize abundance indices (AI) of fish species, which is critical in fish stock assessment. However, additional considerations need to be undertaken to ensure robust estimation because of the latently complicated structures in fishery-dependent data. Here, we elaborated the process of constructing multi-output artificial neural network models to standardize CPUE for heterogeneous fishing operations and applied it to the skipjack tuna (Katsuwonus pelamis) in the western and central Pacific Ocean (WCPO). Seasonal, spatial, and environmental factors were input variables, and the CPUE of four types of skipjack tuna fisheries were set as output variables. The optimal structure for multi-output neural network was evaluated by systematic comparison in 100 runs hold-out cross-validation. The results showed that the final multi-output neural network model with high accuracy can predict the spatial and temporal trends of skipjack tuna abundance.