Comparative analysis of statistical and machine learning approaches for predicting fish length from otoliths

Accurate estimation of fish length from otoliths is essential for stock assessment, bycatch monitoring, and automated length reconstruction workflows. However, the strength and consistency of otolith-somatic scaling can vary across species and environments. This study examined the predictive performance of otolith morphometric variables for estimating fish length across six pelagic and demersal species from major Philippine fishing grounds. We evaluated 11 otolith morphometric and shape metrics using linear and nonlinear regressions, generalized additive models (GAMs), and machine learning (ML) algorithms. Model performance was compared across species to identify both the most informative predictors and ecological factors influencing otolith-length relationships. Otolith length (OL) and otolith area (OA) consistently produced the highest predictive power, whereas otolith perimeter (OP) showed the weakest performance. Demersal species and the midwater schooling Decapterus kurroides exhibited highly predictable otolith-length relationships ( R 2 > 0.95), reflecting relatively stable habitats and uniform growth dynamics. In contrast, Selar crumenophthalmus and Thunnus albacares displayed lower predictability ( R 2 ≤ 0.70), likely due to exposure to dynamic thermal regimes, variable prey fields, and ontogenetic shifts that increase plasticity in otolith accretion. ML models, particularly Random Forest, outperformed classical approaches for species with heterogeneous growth patterns by capturing nonlinearities and interactions among morphometric variables. Our findings demonstrate that OL and OA are robust and broadly transferable predictors of fish length in tropical multispecies fisheries, while species inhabiting variable pelagic environments benefit from more flexible ML frameworks. Integrating ecological context with advanced modeling tools can significantly improve otolith-based size estimation. The study highlights the value of expanding sampling coverage and refining ML approaches to enhance future applications in fisheries monitoring and assessment.