FISH 203

Webinar Link: https://attendee.gotowebinar.com/register/4413110263879871319

FISH 203

Webinar Link: https://attendee.gotowebinar.com/register/5487004370240294486

FISH 203

Webinar Link: https://attendee.gotowebinar.com/register/6020211434618619488

FISH 113

Optimizing traditional and model-predicted fish ageing methods: a study of best practices
for quantifying ageing error and mitigating its various sources

Derek Chamberlin 1 and Jason Conner 1

1 National Marine Fisheries Service, Alaska Fisheries Science Center, 7600 Sand Point Way N.E.,
Building 4, Seattle, WA 98115

Age data are fundamental to population ecology and critical to age structured stock assessments.
The misspecification of ageing error can result in overly optimistic estimates of stock status
leading to inadvertent overfishing. The implementation of model-predicted age estimation
derived from Fourier transform near-infrared spectroscopy (FT-NIR) observations coupled with
multimodal convolution neural network (MMCNN) models introduces new sources of error in
the age estimation process. Given the impact of ageing error on stock assessment and new
sources of uncertainty in model predicted ages, the accurate quantification of ageing error is
critical to accurate stock assessment and sustainable fisheries management. We examined both
the quantification of ageing error and its effects on MMCNN predictions via two projects 1) a
simulation test of the optimal number of readers and proportion of a collection that is read by
those readers to reliably quantify ageing error and 2) a bootstrap resampling of age data based on
estimated ageing error to propagate ageing error through machine learning models. Initial results
suggest the proportion of collections that are tested by multiple readers can be reduced from the
current 20% but that test samples should be selected using a stratified random design.
Additionally, estimated ageing error resulted in a minimal decrease in FT-NIR model
performance (R 2 = 0.87 with ageing error and R 2 = 0.92 without), suggesting the FT-NIR
MMCNN is robust to noise in calibration age data. Finally, we explored application of the Punt
et al. 2008 ageing error matrix model to estimate total uncertainty in model predicted ages.

Webinar Link: https://attendee.gotowebinar.com/register/2275558504413507935