Thu, Aug 05, 2021:On Demand
Background/Question/Methods
Estimating the age and growth of fishes is essential for understanding fisheries ecology and management. Traditional microscopic age estimation methods of fish otoliths are labor-intensive and subject to poor repeatability particularly for long-lived species. Northern rockfish can live to more than 80 years and is a valuable component of the demersal rockfish complex in the Gulf of Alaska (GOA) and Aleutian Islands for which management relies on age composition information. Here we explored advanced technologies using Fourier transform near infrared (FT-NIR) spectroscopy of northern rockfish otoliths coupled with deep machine learning to estimate fish age more rapidly and with greater efficiency than traditional approaches. FT-NIR spectroscopy measured the absorption of light in the near infrared region and recorded an otolith spectrum consisting of overtones and combination vibrations of molecules that contain CH, NH or OH groups. Deep machine learning explored the underlying relationships between otolith spectra with fish age and other biological and geospatial data that have an effect on fish growth. We employed TensorFlow to train and test a 4-layer deep neural network (DNN) models using Keras with hyperband (HB) and Bayesian optimization (BO) tuning algorithms with the fused data set combining spectral, biological, and geospatial data from northern rockfish collected in the GOA in 2015-2019.
Results/Conclusions Otolith spectra in the 4,000-6000 cm-1 wavenumber region, which is associated with C-H and N-H functional groups, had the highest impact predicting fish age. Otolith weight, for which both fish age and somatic growth are determinants, had greater impact than fish length. The latitude had an impact due to fish ontogenetic habitat shifts. Coefficients of determination (R2) for the DNN model with BO tuner were 0.89 and 0.86 for the training and test data, respectively. Root mean square errors (RMSE) of training and testing data sets were 3.84 and 4.36, respectively. R2 for the DNN model with HB tuner were 0.90 and 0.88 for the training and test data, respectively. RMSE of training and testing data sets were 3.75 and 4.15, respectively. All were comparable to RMSE of 3.80 for traditional microscopic methods. Age estimation outcomes between the traditional and FT-NIR DNN methods were largely indistinguishable based on Bland-Altman plots. Since FT-NIR spectroscopy method is about ten times faster than traditional age estimation methods, these results suggest that FT-NIR spectroscopy of otoliths coupled with deep machine learning can predict fish age more rapidly, with greater efficiency, and with comparable precision.
Results/Conclusions Otolith spectra in the 4,000-6000 cm-1 wavenumber region, which is associated with C-H and N-H functional groups, had the highest impact predicting fish age. Otolith weight, for which both fish age and somatic growth are determinants, had greater impact than fish length. The latitude had an impact due to fish ontogenetic habitat shifts. Coefficients of determination (R2) for the DNN model with BO tuner were 0.89 and 0.86 for the training and test data, respectively. Root mean square errors (RMSE) of training and testing data sets were 3.84 and 4.36, respectively. R2 for the DNN model with HB tuner were 0.90 and 0.88 for the training and test data, respectively. RMSE of training and testing data sets were 3.75 and 4.15, respectively. All were comparable to RMSE of 3.80 for traditional microscopic methods. Age estimation outcomes between the traditional and FT-NIR DNN methods were largely indistinguishable based on Bland-Altman plots. Since FT-NIR spectroscopy method is about ten times faster than traditional age estimation methods, these results suggest that FT-NIR spectroscopy of otoliths coupled with deep machine learning can predict fish age more rapidly, with greater efficiency, and with comparable precision.