A Framework to Investigate the Effects of Observation Error on Neural Network Predictions of Fish Age

Abstract

Technological innovations for predicting fish age represent a paradigm shift from conventional age estimation methods used in fisheries science. Recently developed secondary methods rely on models trained on conventional age estimates, derived from subjective interpretation of growth patterns and a biological property of the fish to predict age. Hence, quantifying the error that propagates from these models and the conventional age estimates on which they are trained is critical to fully account for uncertainty in predicted ages used in fisheries stock assessments. We review the development of three secondary age prediction methods: image analysis, epigenetics and Fourier transform near‐infrared spectroscopy (FT‐NIRS); the use of artificial intelligence (AI) within each method; and present an approach to examine the effects of ageing error on AI age prediction models. As a case study, we conducted an empirical study, coupled with simulation, to quantify the effects of ageing error on the performance of a multimodal convolutional neural network (MMCNN) model used to predict eastern Bering Sea walleye pollock ( Gadus chalcogrammus ) ages from otoliths analysed using FT‐NIRS. Results indicated repeatability of predicted ages was high between instrument operators, while adding ageing error resulted in a slight decrease in model performance from R 2 = 0.92 and CV = 7.6% to R 2 = 0.87 and CV = 10.2% on test datasets. Our results also suggest the MMCNN model is robust to noise in calibration age data and model performance may be better than performance metrics indicate when the model is evaluated against data with increased error.

Bibliographic Information