Muhammad Amirul Aiman Asri, Wenjunliang Zhang, Norrima Mokhtar, Afdhal Haziq Noramly, Raza Ali, Takao Ito, M. Aziz Muslim, Siti Sendari, Pringgo Widyo Laksono, Tsutomu Ito
Pixel accurate segmentation of paddy leaf lesions is vital for field diagnostics, yet many prior studies examine only one disease and use inconsistent pipelines, which limits comparability. We present an experimental study and a unified benchmark on the public Kaggle New Paddy Doctor dataset with 2,444 RGB images at 480×640 covering Bacterial Leaf Blight (804 images), Brown Spot (936 images), and Hispa (704 images). We manually annotated pixel level masks for all images. Under one protocol, we evaluate U-Net, U-Net++, and DeepLabV3 with a ResNet-50 backbone in both per disease training and pooled training. DeepLabV3 with ResNet-50 attains the best average per disease Dice of about 0.70, with stronger scores on Bacterial Leaf Blight at about 0.775 and Hispa at about 0.730, while Brown Spot remains more challenging at about 0.599. Training a single pooled model gives slightly lower performance than training per disease, with a similar pattern for IoU. We also examine how stable the model rankings are across diseases and provide qualitative examples of typical failure modes such as background leakage and boundary errors. These results offer clear baselines and practical guidance on when pooling is helpful, and which architecture provides the best balance between accuracy and efficiency for deployment. © The 2026 International Conference on Artificial Life and Robotics (ICAROB2026).
Department of Electrical Engineering, Faculty of Engineering, Universiti Malaya, Lembah Pantai, Kuala Lumpur, Malaysia; Balochistan University of Information Technology, Engineering and Management Sciences (BUITEMS), Quetta, Pakistan; Graduate School of Advanced Science and Engineering, Hiroshima University, Japan; Department of Electrical Engineering, Faculty of Engineering, Universitas Brawijaya, East Java Indonesia, Malang, Indonesia; Faculty of Engineering, Universitas Negeri Malang, Indonesia; Industrial Engineering, Universitas Sebelas Maret, Indonesia; Ube National College of Technology, Japan
