<![CDATA[This study evaluates YOLOv8 performance for detecting Robusta coffee ripeness in unstructured Bondowoso plantations. Models trained on idealized data often degrade in chaotic real-world visual fields. To address this, this research establishes a realistic evaluation framework using a dual-source dataset strategy, leveraging public datasets for training and compiling a specialized primary field dataset from Bondowoso for independent testing under severe occlusion and dynamic illumination. Systematic tuning established an optimal configuration: 640x640 resolution, batch size of 8, and the AdamW optimizer. Internal testing achieved a robust mAP50 of 0.8144. However, in real-world deployment, the mAP50 dropped to 0.5254 due to environmental complexities. Despite this, the system maintained high precision (0.8390) and an exceptional standalone inference speed of 121.51 FPS. While high precision ensures reliable harvest assistance, the significant performance gap, characterized by a severe drop in recall, highlights architectural vulnerabilities in anchor-free detectors when handling dense clustering and foliage obstructions. These findings define the operational boundaries of lightweight detectors, providing a vital baseline for future domain adaptation and optimization in precision agriculture. Keywords: Bondowoso; Ripeness detection; Robusta coffee; Unstructured environment; YOLOv8.]]>
Copyrights © 2026
