A Comparative Study of Deep Learning Approaches for Early Detection of Sugarcane Diseases

Sugarcane, being one of the essential crops for the Indian agricultural economy, that has a significant contribution to the overall agricultural ecosystem of the country. However, the cultivation faces threats from variety of diseases and if, not treated in time bound fashion then it might result in poor yield and low quality, causing financial loss to the farmers. This study explores various approaches using deep learning to detect diseases that are affecting sugarcane cultivation. A dataset of over 6670 images across five classes (healthy, red rot, mosaic, rust, and yellow leaf) is taken. This study compares the performance of several deep learning models, including VGG-19, MobileNet-V2, EfficientNet, and Vision Transformer, with image processing techniques. Transfer learning is used with pre-trained models trained on the ImageNet dataset, and the Vision Transformer is fine-tuned during training. In this study, we found that Vision Transformer outperformed other deep learning models by achieving an accuracy of 98.0% and among the CNN models EfficientNet achieved an accuracy of 96.7%. The comparative analysis highlights the potential of various deep learning approaches for accurate disease detection in sugarcane cultivation. The research contributes to sustainable agricultural practices by providing a framework for early disease identification and prevention of crop losses through deep learning techniques.