YOLO-Lychee: A YOLOv8s-Based Detector for Lychee Growth-Stage Detection under Natural Orchard Conditions

Accurate lychee growth-stage detection in natural orchards is challenging because blossoms and fruits are often small, densely clustered, partially occluded, and visually similar to surrounding foliage. This study proposes YOLO-Lychee, a YOLOv8s-based detector for four developmental stages: blossom, young fruit, green fruit, and ripe fruit. The dataset contains 1,145 original orchard images and 19,422 annotated instances collected in Hai Duong and Bac Giang provinces, Vietnam. To reduce blossom-class imbalance, exposure-based augmentation was applied only to blossom samples in the training subset, resulting in 1,324 images and 20,261 annotations, while the validation and test sets remained unchanged. YOLO-Lychee replaces the original SPPF module with a Spatial and Channel Cross-Transformer module, incorporates a Context Augmentation Module in the neck, and uses EIoU instead of the default CIoU loss for bounding-box regression. Across six random seeds, YOLO-Lychee achieved a Precision of 83.02 ± 2.38%, Recall of 82.19 ± 1.87%, mAP50 of 88.42 ± 0.38%, and mAP50:95 of 72.76 ± 0.45%. Compared with recent YOLO-family detectors under the same protocol, the proposed model obtained the highest mAP50 and competitive mAP50:95 while maintaining real-time inference. Ablation results confirm the complementary contributions of SC3T, CAM, and EIoU, whereas qualitative analysis shows that blossom detection remains the most challenging case. These results demonstrate that YOLO-Lychee is a practical empirical baseline for lychee growth-stage detection and vision-assisted orchard monitoring.