Lightweight deep learning models for six-stage maturity classification of arabica coffee cherries using RGB imaging

This is the authors' abstract. We don't add key points for this paper.

Selective harvesting and postharvest grading for premium Arabica coffee production rely on the maturity of coffee cherries. Conventional visual assessment is subjective, labor-intensive, and difficult to standardize. This study developed an RGB image-based framework for classifying Arabica coffee cherries into six expert-defined visual maturity stages and compared four convolutional neural network architectures, namely Custom CNN, MobileNetV2, EfficientNet-B0, and ResNet-50. The dataset consisted of 1,010 original images of Coffea arabica L. ‘Catimor LC1662’ cherries collected from multiple orchard plots within a single production site in Northern Thailand. Model performance was measured using a held-out internal test set, a stratified five-fold cross-validation, bootstrap-based 95% confidence intervals, and paired statistical comparisons. The best held-out internal test performance was achieved by EfficientNet-B0 with 95.36% accuracy, 0.955 macro precision, 0.961 macro recall, 0.957 macro F1-score, and 0.953 weighted F1-score. MobileNetV2 and ResNet-50 had accuracies of 92.05% and 90.73% respectively, and the Custom CNN 80.13%. McNemar’s exact test showed that EfficientNet-B0 significantly outperformed the Custom CNN, whereas the difference with MobileNetV2 was not statistically significant. EfficientNet-B0 had 4.06 million parameters and an inference time of 0.007 s per image on CPU. The MobileNetV2 model showed a good balance between prediction performance, model size, and inference speed, whereas the Custom CNN was the smallest and fastest model but with a much lower classification performance. These results demonstrate the feasibility of lightweight deep learning for fine-grained visual maturity classification and its potential to contribute to more consistent harvest grading in premium coffee production. The models classify expert-assigned visual stages rather than directly measuring biochemical maturity, sensory attributes, or final coffee quality. Therefore, field and device-level validation is required before practical deployment.