Sugarcane sentinel: a Transfer Learning framework for sugarcane leaf disease detection with real-time web and mobile deployment

Abstract

Sugarcane is the most significant cash crop in India and serves as a primary source for sugar, jaggery, ethanol, and other value-added by-products. Due to its high economic significance, yield loss caused by leaf diseases poses a serious challenge to sugarcane production. An early detection of such diseases is crucial for minimizing crop loss and improving productivity. Recent advances in deep learning have enabled automated plant disease detection. However, developing robust models for sugarcane leaf disease detection remains challenging due to limited labeled datasets. The proposed work addresses these challenges by integrating publicly available sugarcane leaf image datasets to construct a comprehensive dataset of 8,355 images across nine classes, including seven disease categories, healthy and dried leaf class. Transfer learning was employed by fine-tuning several pre-trained models from the EfficientNet, MobileNet families, and Vision Transformer. Model performance was further enhanced using class imbalance handling and hyperparameter optimization. Among the evaluated models, EfficientNet-B0 demonstrated the best performance, achieving over 99 percent accuracy on both training and test datasets, along with macro-averaged precision, recall, and F1-score of 99 percent. To enable real-world deployment, an Android mobile application was developed supporting offline on-device inference. The system achieved real-time performance with an average inference time of 80.83ms. Memory profiling indicated stable execution, with Java heap usage of 4-12 MB and native memory briefly peaking at 95 MB before stabilizing at 49-54 MB. Pilot field testing conducted using 264 samples under uncontrolled conditions achieved approximately 89.02 percent accuracy, demonstrating the robustness and practical applicability of the proposed system.