Mustafa Tasci
Rice cultivation, the foundation of global food security, is vulnerable to leaf diseases, which significantly reduce both yield and quality. Early and accurate detection is crucial for effective and sustainable disease management. This study presents an efficient and deployable real-time system for rice leaf disease detection, optimized for edge-computing environments. A novel Dual-Branch Lightweight Attention MobileNetV2 (DBLA-MobileNetV2) architecture is proposed and compared with a baseline MobileNetV2 model fine-tuned via transfer learning. DBLA-MobileNetV2 employs a dual-branch structure to enhance multi-scale feature extraction, where one branch focuses on fine-grained visual patterns and the other captures high-level semantic information. Integrated Squeeze-and-Excitation (SE) attention mechanisms further improve channel-wise feature representation and increase accuracy, while maintaining computational efficiency. To demonstrate real-world applicability, we developed a portable hardware prototype using an NVIDIA Jetson Nano Core Processing Unit. The device was designed as an arm-mounted unit, enabling direct field deployment without reliance on a cloud infrastructure. The proposed system was tested in six rice fields at various growth stages. The experimental results showthat DBLA-MobileNetV2 achieved 97.9% precision and 12.40 FPS under the quantization of FP16 using the TensorRT framework. The combination of lightweight architecture and mobile implementation provides a practical pathway for precision agriculture, allowing rapid in-field disease diagnosis and timely agronomic interventions to enhance productivity.