Statistical Benchmarking of Machine Learning Methods in Fisheries Morphological Analysis

Accurate prediction of fish morphological traits is critical for sustainable fisheries management and for optimizing aquaculture practices. The proposed approach systematically evaluates eleven machine learning models for predicting fish morphology from biometric measurements, using a dataset of 7159 samples across seven species. The dataset encompasses four key morphometric attributes such as weight, length, height, and width; with all attributes utilized as input features for comprehensive morphological analysis. The models assessed include both advanced probabilistic methods and traditional regression algorithms. The probabilistic models evaluated are Gaussian Process Regression (GPR), Variational Autoencoders (VAEs), and Bayesian Neural Networks (BNNs), while the traditional models comprise standard regression techniques. GPR achieved the highest predictive accuracy ( \(R^{2} = 0.9996\) , RMSE = 0.0170, MAE = 0.0117, MAPE = 2.44%), outperforming all other models. Within the VAE category, VQ-VAE ( \(R^{2} = 0.9977\) ) and Hierarchical VAE ( \(R^{2} = 0.9967\) ) also demonstrated strong performance. Among traditional approaches, Support Vector Regression ( \(R^{2} = 0.9775\) ) and Random Forest ( \(R^{2} = 0.9712\) ) were the most effective. Validation through residual, normality, and autocorrelation analyses confirms the robustness of these results. Overall, probabilistic models, particularly GPR, demonstrate superior capability in capturing nonlinear morphological relationships and provide a benchmark framework for automated fish morphology prediction in ecological and aquaculture applications.