Answers are grounded in this paper and cite their sources.
Egypt ranks among the world’s leading citrus producers and exporters; however, orchard operations remain constrained by subjective, labor-intensive manual quality grading and time-consuming yield estimation that require specialized expertise. This study presents an AI-driven orchard-intelligence framework for automated fruit quality grading and dynamic yield estimation under the unstructured conditions of Egyptian orange farms. To address the lack of in-field public datasets, a domain-specific instance-segmentation dataset of 618 original pixel-level annotated images, expanded to 1,545 augmented training images, and 15 videos was constructed. This dataset was used to benchmark single-stage segmentation models (YOLOv8, YOLOv11) against a two-stage (Mask R-CNN), alongside a transformer-based bounding-box detection (RT-DETR), under dense canopy occlusion, clustered fruit instances, and variable natural illumination. Each fine-tuned model was evaluated across two stages: static quality grading and dynamic yield estimation. For static grading (Good vs. Defective), YOLOv11l ranked among the top tier of evaluated segmenters, reaching mean Average Precision (mAP) of 0.932 with balanced precision-recall behavior, while maintaining real-time feasibility at 26.1 ms per image. For dynamic yield estimation, a multi-object tracking pipeline was implemented using BoT-SORT with camera-motion compensation and a virtual line-crossing counting mechanism. The evaluation reveals a three-way dissociation: the highest-mAP static segmenter (YOLOv11l), the best dynamic total counter (RT-DETR, WMAPE = 5.66%, despite the lowest box-level static accuracy), and the best quality-resolved dynamic grader (YOLOv11s, WMAPE = 6.97% at 14.2 ms per frame) are three different systems; RT-DETR did not resolve fruit quality reliably under motion. Counting accuracy and grading capability are thus separable properties requiring separate evaluation. Overall, the framework enables objective fruit quality assessment and actionable in-field fruit counting to support harvest logistics and labor planning, providing a practical foundation for smart agriculture in the Egyptian citrus sector.
No takes yet. Share an insight, caveat, or question.
Farhoud et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: