A large-scale proof-of-concept benchmark for commercial coffee-bean classification using external phenotypic features and advanced machine learning algorithms.

Coffee beans are economically and culturally important agricultural commodities traded globally across cultivation, processing and consumer supply-chain stages. Faced with challenges in distinguishing coffee beans of different geographical origins caused by high visual similarity and uncertainties in feature extraction, this study compared 12 advanced machine-learning algorithms (TabPFN-3, XGBoost, CatBoost, Random Forest, LightGBM, Logistic Regression, Support vector machine, K-nearest neighbor, FSC-GA-DT, ROPAC-L, ROPAC-M, GATree) for lot-level classification of 10 sampled single commercial coffee lots originating from 10 producing countries, using 23,283 green coffee-bean samples. This study constitutes a large-scale proof-of-concept phenotyping benchmark under standardized controlled imaging conditions. All samples originate from one single commercial lot per country; classification was evaluated via individual-bean-level cross-validation, and no external validation was performed on fully independent coffee lots. Experimental results show that TabPFN-3 yielded the best overall within-dataset metrics for the sampled single commercial lots (accuracy 0.9627, precision 0.9617, recall 0.9606, F1 score 0.9610). Therefore, TabPFN-3 exhibits promising discriminative performance under this proof-of-concept experimental setup. This study provides preliminary insights for phenotype-driven benchmarking of commercial coffee-lot classification; generalizable geographical-origin authentication remains to be established using multi-lot datasets in future study. All reported results represent within-dataset bean-level classification under controlled imaging conditions, performance on unseen multi-lot independent samples remains unconfirmed.