CFNet: Advancing automated pollination with YOLOv8-based cantaloupe floral detection

Abstract

The escalating crisis of pollinator decline, projected to culminate in 60–70% losses of commercial honeybee colonies in the United States by 2025, coupled with pervasive labor shortages, imperils global agricultural productivity, particularly for pollination-dependent crops such as cantaloupe (Cucumis melo). This study presents CFNet, an advanced YOLOv8-derived (You Only Look Once version 8) object detection paradigm engineered for real-time detection of cantaloupe flowers and buds, to enable robotic pollination. Unlike conventional architectures, CFNet integrates GSConv for parameter-efficient convolutions, FDCA for context-aware channel recalibration, EMA-BiFPN for bidirectional multi-scale feature fusion, and SPPF for robust scale aggregation, optimizing inference on resource-constrained devices. Evaluated on a custom dataset exceeding 743 annotated greenhouse images, CFNet achieves mAP@0.5 of 94.2% and mAP@0.5:0.95 of 74.2%, surpassing YOLOv8 (93.0%, 72.5%), RT-DETR (92.8%, 71.4%), DINO (91.5%, 71.5%), RTMDet (89.5%, 69.2%), DETR (79.5%, 59.6%), YOLOv5 (77.1%, 57.8%), and SSD (74.5%, 52.3%). CFNet demonstrates particular superiority over recent transformer-based detectors including DINO and RT-DETR, validating its architectural innovations. With 2.7 M parameters and 22 ms latency, it demonstrates a 15% precision improvement for buds under occlusions while maintaining real-time performance. Deployed in a pollination prototype, CFNet engenders 20–30% yield enhancements in empirical trials. This endeavor propels precision horticulture through AI-robotics synergy, mitigating pollinator exigencies with prospective multi-modal extensions for volumetric localization.