ResNet-18 Citrus Classification with Augmentation: Peel Color Associations with pH and Vitamin C

Dewi, Titin Afrina Priani (Department of Mathematics, Ahmad Dahlan University, Yogyakarta 55191, Indonesia)
Surono, Sugiyarto (Department of Mathematics, Ahmad Dahlan University, Yogyakarta 55191, Indonesia)
Thobirin, Aris (Department of Mathematics, Ahmad Dahlan University, Yogyakarta 55191, Indonesia)

Zero : Jurnal Sains, Matematika, dan Terapan
Vol 9, No 3 (2025): Zero: Jurnal Sains Matematika dan Terapan

Limes and lemons are widely consumed, but field quality assessment is often subjective. This study built a ResNet-18 citrus classifier using transfer learning and data augmentation, and evaluated it with stratified 5-fold cross-validation on 80 images (40 limes, 40 lemons). All analyses were conducted per fold to reduce optimistic bias. The model reached 98.75% mean validation accuracy, misclassifying one lime image while correctly recognizing all lemons. For interpretation, peel regions were quantified using NDYI and CIELab (L*, a*, b*), and related to pH and vitamin C using Spearman correlation. Uncertainty was quantified with bootstrap-based 95% confidence intervals for each correlation. Peel color features were more consistently associated with pH (especially in limes), whereas correlations with vitamin C were weak and non-significant for both fruits. Results indicate strong performance under controlled imaging, but using peel color as a vitamin C proxy requires broader data and external validation across cameras and lighting.