Do otolith contour descriptors improve age classification? A test with non-contour baselines and grouped validation in Japanese jack mackerel (Trachurus japonicus)

Otolith shape is often proposed as an aid to age classification, but two problems limit the evidence: performance is rarely compared with a non-contour baseline, and validation often allows otoliths from the same collection event into both the training and the test set. We examined both in Japanese jack mackerel Trachurus japonicus , using 1444 sagittal otoliths from fish aged 0–4 years. Thirty-one combinations of eleven feature sets and six classifiers were evaluated under two main validation designs. When whole purchase batches were kept within folds, the best model used eight non-contour variables alone (fork length, four otolith measurements and three dimensionless indices; macro F1 0.790, 95% interval 0.782–0.799), and no set containing Fourier or wavelet coefficients performed better. The effect of adding contour descriptors depended on the classifier: the six comparisons using tree ensembles were positive and the twelve using other classifiers negative, and after correction for dependence among folds only the negative differences remained significant. Permuting all 63 wavelet coefficients as a block changed macro F1 by 0.003, against 0.448 for the four otolith measurements. Age explained 11.0% of contour variation before standardisation to otolith length and 1.3% after. On 300 otoliths read twice, the model agreed with the reference ages in 80.3% of cases and the two readings agreed with each other in 79.7%, whereas the model agreed with the blind repeat reading in only 67.0%. Otolith size, rather than contour shape, carries most of the age information in this species.