UC Davis

Bean screen size is a critical grading parameter in green coffee quality and trade, yet conventional sieving is time-consuming and labor-intensive. Here, we evaluate whether smartphone photographs of coffee beans placed on a calibrated reference sheet yield accurate bean size distributions. A Python-based image analysis algorithm measured the length and width of 2374 individual beans that were also each manually measured with digital calipers. We further used the image analysis algorithm to measure screen-size distributions from 19,509 beans across ten additional green coffee samples and compared them with distributions obtained via traditional mechanical sieving. Image-derived length and width measurements showed strong agreement with digital caliper measurements, with mean percent differences of 1.94% for length and 1.08% for width. Agreement was stronger for length (R² = 0.98) than for width (R² = 0.94), consistent with the greater geometric complexity of width estimation. Under controlled illumination, roast level and lighting had minimal influence on accuracy, whereas uncontrolled ambient lighting increased variability (p < 0.05). For sieve classification, mixed-effects modeling detected no statistically significant difference between photographic measurements and mechanical sieving (p = 0.976), and mean screen sizes strongly correlated with mechanical results (R² = 0.97; root mean square error = 0.22 screen units). We conclude that calibrated smartphone imaging closely aligns with conventional sieve grading and offers a portable, low-cost alternative for bean size classification across diverse coffee origins.