Decoding Diets: Applying Non-Linear Machine Learning Models to Geometric Morphometric Analysis of Bovid Dental Mesowear Signatures

Dental morphology and wear patterns provide insight into the dietary adaptations and ecological niches of living and extinct herbivores. Traditional classification statistics such as Linear Discriminant Analysis (LDA) are limited by assumptions of linearity, normality, and homoscedasticity. This study quantifies mesowear, the shape of molar cusps resulting from occlusal wear, and evaluates the performance of non-linear machine learning models in predicting herbivore diets based on geometric morphometric (GMM) data from adult mandibular second molars (M2) in bovids. We applied Generalized Procrustes Analysis and Principal Component Analysis (PCA) to digitized occlusal shape coordinates from 132 M2 specimens across 64 species. Using the resulting principal component scores, we compared the classification accuracy of LDA with three non-linear models: Random Forest, K-Nearest Neighbors, and Gradient Boosting. While LDA achieved a cross-validated accuracy of just 31%, all non-linear models achieved 99% cross-validation accuracy and 90% test accuracy, demonstrating substantially improved performance. Misclassification analyses revealed that non-linear models more effectively captured complex shape differences, particularly among species with overlapping wear patterns. Our findings support the integration of machine learning with geometric morphometrics to quantify mesowear and improve dietary classification, providing a framework for robust paleoecological inference.