Morphological and Physicochemical Variability of Passion Fruit (Passiflora edulis) Across Production Sites in Veracruz, Mexico: Multivariate Analysis and Machine-Learning Assessment

This study integrated multivariate statistical analysis and machine learning to characterize fruit quality variability and prediction of soluble solids content across nine passion fruit production sites in Veracruz, Mexico. A total of 143 valid observations were analyzed, including fruit weight, pulp weight, fruit height, fruit diameter, rind thickness, number of seeds, peel color, location (site code), and soluble solids (°Brix). Fruit weight ranged from 40.7 to 204.5 g, pulp weight from 8.3 to 115 g, and site-level mean soluble solids ranged from 10.31 to 23.39 °Brix. Canonical Variate Analysis (CVA) showed that the first two canonical axes accounted for 75.97% of the total discriminant variation among production sites (CV1 = 45.42%; CV2 = 30.55%). CV1 was primarily associated with soluble solids content, whereas CV2 was associated with fruit dimensions. Under nested leave-one-site-out cross-validation, Random Forest achieved the highest performance among the evaluated machine-learning models (R2 = −0.408; RMSE = 6.051 °Brix), but did not outperform the training-mean baseline (R2 = −0.190; RMSE = 5.562 °Brix). Permutation importance identified fruit weight and diameter as the most influential predictors in the fitted Random Forest model, although their predictive contributions were insufficient for reliable generalization to previously unseen production sites. These findings provide insights into fruit quality variability across nine passion fruit production sites in Veracruz, Mexico, during the 2025 sampling period, while highlighting the complementary value of multivariate analysis and machine learning for characterizing fruit traits and evaluating predictive performance across locations.