An Improved Lightweight YOLO Model for Real-Time Blueberry Growth Stage Detection to Facilitate Smart Irrigation

Intelligent irrigation, guided by real-time crop water status, is paramount for sustainable, high-yield agriculture. For high-value crops like blueberries, accurately detecting phenological growth stages is a key prerequisite for precise water management. However, vision-based monitoring in field environments is challenging due to dynamic lighting, severe occlusion from foliage, and subtle inter-stage variations. To address these issues, this paper proposes a lightweight and efficient object detection model, YOLO-Berry, for real-time blueberry growth stage identification. We first established a comprehensive field-image dataset of blueberries, encompassing multiple growth stages under diverse environmental conditions. The proposed model enhances the YOLO11n baseline by incorporating a Bidirectional Feature Pyramid Network (BiFPN) for superior multi-scale feature fusion, a Convolutional Block Attention Module (CBAM) to focus on critical regions, and lightweight Ghost and C3Ghost structures to minimize computational costs. Experimental results show that YOLO-Berry achieves a mean Average Precision (mAP) of 96.7%, a 6.5 percentage point increase over the baseline. Moreover, it demonstrates a 2.7% gain in Precision, a 5.2% rise in Recall, and a 17.9% reduction in parameters, while maintaining a rapid inference speed of 15 ms per frame. The detection outputs are used to generate growth stage statistics, which inform an automated irrigation decision-making system and facilitate early yield prediction. This work provides a robust technological foundation for vision-guided precision irrigation in modern smart agriculture.