Implementation of ResNet50 Based on Transfer Learning for Sugarcane Leaf Disease Detection

Abstract

Sugarcane (Saccharum officinarum) is a vital commodity in Indonesia’s sugar industry and is highly susceptible to leaf diseases such as Mosaic, RedRot, Rust, and Yellow, which significantly reduce yield quality and quantity. This study proposes an automatic disease classification system using the ResNet50 architecture with a transfer learning approach, offering a more systematic evaluation compared to previous studies that typically tested only a single configuration or focused on other crops. The dataset consists of 3,250 RGB images across five classes after preprocessing and augmentation to address class imbalance. Eight model configurations were evaluated by combining epoch values (20, 40) and learning rates (0.0001, 0.001, 0.01, 0.1). The best performance was achieved by the configuration with 20 epochs and a learning rate of 0.0001, producing an accuracy and F1-score of 97%. The model was further deployed into a Flask-based web application to demonstrate practical usability. However, this study is limited by the use of a single controlled dataset, so model performance may vary under real-field conditions such as different lighting, camera angles, and leaf damage severity. Future research should include field data evaluation to strengthen model generalization.