An Android-smartphone application for rice panicles detection and rice growth stages recognition using lightweight YOLO network

Huiwen Zheng ¹ Changjiang Liu ¹ Lei Zhong ¹ Jie Wang ¹ Junming Huang ¹ Fang Lin ¹ Xu Ma ^2,3 Suiyan Tan ^1*

• ¹ College of Electronic Engineering, South China Agricultural University, Guangzhou, China
• ² College of Mechanical and Electrical Engineering, Xinjiang Agricultural University, Urumqi, Xinjiang Uyghur Region, China
• ³ College of Engineering, South China Agricultural University, Guangzhou, China

Detection of rice panicles and recognition of rice growth stages can timely facilitate precision field management and is of great significance in pursuing maximum grain yield. This study explores usage of deep learning on mobile phones as a platform for rice phenotype applications.In this study, an improved YOLOv8 model, namely YOLO-Efficient Computation Optimization (YOLO ECO), was proposed to detect rice panicles at booting, heading and filling stages, and to recognize growth stages. YOLO ECO introduced key improvements, with C2f-Faster Block-Effective Multi-scale Attention (C2f-Faster-EMA) replacing the original C2f module in the backbone, adoption of SlimNeck to reduce neck complexity, and using a Lightweight Shared Convolutional Detection (LSCD) head to enhance efficiency. Then, an android application, YOLO-RPD was developed to apply for rice phenotype detection under the complex field environment. The performance impact of YOLO-RPD using models with different backbone networks, quantitative models and input image size were comparatively analyzed. Experimental results showed that YOLO ECO outperformed the traditional deep learning models with average precision values of 96.4%, 93.2%, 81.5% at booting stage, heading stage and filling stage, respectively. Furthermore, YOLO ECO exhibited advantage in detecting the occlusion and small panicle, and demonstrated significant 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 Giga floating-point operations (FLOPs). The YOLO-RPD application demonstrates the feasibility of deploying deep learning models on mobile devices for precision agriculture. YOLO-RPD provides rice growers with a practical and lightweight tool for real-time monitoring.