Abstract:
Sugarcane is an important tropical agricultural crop, which is under the threat of several leaf diseases like mosaic disease, red rot, rust, yellow leaf and mawa (banded blight) that causes a huge loss in yield. Traditional manual inspection-based disease detection is timeconsuming and error-prone, it cannot be used for many varieties of large-scale monitoring. This paper demonstrates a new approach based on deep learning to automatically identify six sugarcane leaf diseases. We experimented with several deep learning-based model, such as EfficientNet-B0, DenseNet-121, MobileNetV2, ResNet50 and hybrid architecture which is a fusion of dense net and resnet and vision transformer (ViT). Results revealed that the accuracy with CNN-based models worked well in >90%. The performance result of the Hybrid super model between DenseNet and ViT is also significantly better than others with 96.7% validation accuracy, 96.5% test accuracy and F1 score weighted=0.97(Far from them). This shows the potentiality to integrate local convolutional features with global context features from transformers for enhancing classification performance. This type of a solution could offer an efficient scalability and reliability for precision agriculture and contribute to reducing reliance on manual disease detection and to sustainable crop growing. Moreover, the current method is suitable for detecting in real-time and practical for farmers to eliminate disease at early stage of crop plant filed, so it can greatly limit area-affected plants and agricultural production. Incorporating deep learning into disease management is a part of the global trend toward “smart” farming, which intends to use technology-enabled tools for more precise, information-based decision-making. 2 Significance statement This paper opens up the road towards revolutionizing crop management and ensure food security in tropical zones by leveraging AI-based solutions.
