Wenjunliang Zhang, Muhammad Amirul Aiman Asri, Norrima Mokhtar, Shunta Kimura, Ryosuke Harakawa, Masahiro Iwahashi, Heshalini Rajagopal, Rahmadwati, Takao Ito, Siti Sendari, Pringgo Widyo Laksono
To support fast field monitoring and practical deployment, we benchmark the classification heads of YOLOv5, YOLOv8, and YOLOv11 in nano and medium variants on New Paddy Doctor, a public rice-disease dataset. From its 10 annotated categories, we select an eight-class leaf subset with 6,627 images covering Bacterial leaf blight (BLB), Bacterial leaf streak (BLS), Rice blast, Brown spot, Downy mildew, Hispa damage, Tungro, and healthy leaves. Using a unified 224×224 training and evaluation protocol, we report Top-1 accuracy, Macro-F1, Weighted-F1, and confusion matrices, and we compare model complexity by parameters and FLOPs. On our test set, YOLOv8-m attains the highest accuracy at about 99.9%, YOLOv11 variants reach about 99.8%, while YOLOv5 achieves about 95%. We also examine the balance between accuracy and computational cost and provide deployment recommendations. The data splits and key configurations are released to facilitate reproducibility. © The 2026 International Conference on Artificial Life and Robotics (ICAROB2026).
Department of Electrical Engineering, Faculty of Engineering, Universiti Malaya, Malaysia; Faculty of Engineering, Nagaoka University of Technology, Japan; MILA University, Malaysia; Universitas Brawijaya, Indonesia; Hiroshima University, Japan; Universitas Negeri Malang, Indonesia; Industrial Engineering, Universitas Sebelas Maret, Indonesia
