Abstract
INTRODUCTION: Detection of rice panicles and recognition of rice growth stages can significantly improve precision field management, which is crucial for maximizing grain yield. This study explores the use of deep learning on mobile phones as a platform for rice phenotype applications. METHODS: An improved YOLOv8 model, named YOLO_Efficient Computation Optimization (YOLO_ECO), was proposed to detect rice panicles at the booting, heading, and filling stages, and to recognize growth stages. YOLO_ECO introduced key improvements, including the C2f-FasterBlock-Effective Multi-scale Attention (C2f-Faster-EMA) replacing the original C2f module in the backbone, adoption of Slim Neck to reduce neck complexity, and the use of a Lightweight Shared Convolutional Detection (LSCD) head to enhance efficiency. An Android application, YOLO-RPD, was developed to facilitate rice phenotype detection in complex field environments. RESULTS AND DISCUSSION: The performance impact of YOLO-RPD using models with different backbone networks, quantitative models, and input image sizes was analyzed. Experimental results demonstrated that YOLO_ECO outperformed traditional deep learning models, achieving average precision values of 96.4%, 93.2%, and 81.5% at the booting, heading, and filling stages, respectively. Furthermore, YOLO_ECO exhibited advantages in detecting occlusion and small panicles, while significantly optimizing parameter count, computational demand, and model size. The YOLO_ECO FP32-1280 achieved a mean average precision (mAP) of 90.4%, with 1.8 million parameters and 4.1 billion floating-point operations (FLOPs). The YOLO-RPD application demonstrates the feasibility of deploying deep learning models on mobile devices for precision agriculture, providing rice growers with a practical, lightweight tool for real-time monitoring.