DeepVision for Precision Farming: Automated Ripeness Grading of Coffee Cherries Using Deep Learning Algorithms

Resumen

Objective: To develop an automated computer vision system using the YOLO framework for classifying coffee cherry ripeness stages, specifically addressing selective harvesting challenges faced by small-scale producers through accessible and cost-effective technology.

Design/methodology/approach: YOLOv8, YOLOv9 and YOLOv11 models were implemented and evaluated using transfer learning on a dataset of 7,778 images in four maturity classes. The methodology included comprehensive preprocessing, data augmentation, and rigorous validation with a 70/20/10 train/validation/test split. Statistical validation employed paired t-tests and Wilcoxon signed-rank tests to compare model performance.

Results: YOLOv11m demonstrated superior performance with overall mAP@50 of 0.864, precision of 0.806, and recall of 0.813. Class-wise analysis revealed highest performance for unripe cherries (AP: 0.905) and most challenging detection for overripe class (AP: 0.817). The model showed consistent 7.5% improvement in mAP@50 over YOLOv8m, with statistical significance confirmed through rigorous testing.

Study limitations/implications: The natural imbalance of classes, particularly in overripe, represents the main limitation. The results imply the feasibility of implementation in affordable mobile applications.

Findings/conclusions: The effectiveness of YOLOv11 for maturity classification in real conditions was demonstrated. The system represents an accessible solution that can optimize harvesting and improve grain quality.