Morphological Database and Quantification of Liberica Coffee Beans Using a Hybrid CNN Model

Abstract

Traditional evaluation of Liberica coffee (Kapeng Barako) in Batangas is commonly performed through manual visual inspection and sensory judgment, which are slow, subjective, and difficult to reproduce across farms, buyers, and grading personnel. These limitations can lead to inconsistent quality classification, weak traceability, and reduced bargaining power for local producers. This study develops and validates a web-based morphological analysis system for Liberica coffee beans using a hybrid pipeline that combines U-Net convolutional neural network segmentation with deterministic feature extraction. A database of 4,000 high-resolution images collected from farms in Batangas was curated and divided into training, validation, and test subsets using a 70:15:15 ratio. To evaluate generalization, an independent 500-image dataset from previously unseen farms was also used. The model was trained with Dice plus binary cross-entropy loss, Adam optimization, and augmentation strategies designed to simulate field variations. The 96.0% accuracy reported in this study is explicitly defined as feature-level agreement: the percentage of test images in which the primary extracted traits fall within accepted tolerances relative to a manually verified reference. On the internal test set, the proposed method achieved a Dice coefficient of 0.971, intersection-over-union of 0.944, and high agreement for key morphological traits including area (R² = 0.976), eccentricity (R² = 0.972), solidity (R² = 0.973), and centroid coordinates (R² > 0.962). Compared with classical image processing and CNN baselines, the hybrid approach provided a favorable balance of accuracy, interpretability, and deployment speed. The resulting platform provides farmers, cooperatives, and researchers with a practical, low-cost, and reproducible tool for data-driven Liberica coffee quality assessment and morphology database development.