Disease Classsification in Sugarcane using Deep Learning Technologies

Abstract

This document presents a study on improving sugarcane disease detection in Sri Lanka, where productivity is significantly hindered by diseases such as White Leaf Disease, Red Rot, Rust, and Yellow Leaf Disease. Although these diseases severely impact yield, there is currently limited research in Sri Lanka addressing their automated detection. Traditional manual diagnosis methods are often time-consuming, error-prone, and require substantial agricultural expertise. To address these challenges and the lack of ensemble-based approaches in existing work, this study leverages deep learning techniques, employing four state-of-the-art Convolutional Neural Network (CNN) architectures VGG16, ResNet50, InceptionV3, and EfficientNetB0. A hard-voting weighted ensemble approach is applied to combine these models, enhancing both accuracy and reliability in automated disease classification. A custom dataset of 1,543 annotated sugar cane leaf images, collected from the Palwatta Sugar Company, was used with data augmentation techniques employed to improve model generalization. Individually, VGG16, ResNet50, and EfficientNetB0 achieved 95% accuracy, while InceptionV3 reached 94%. The ensemble method further boosted performance, achiev ing an overall F1 score of 99.57%, demonstrating the effectiveness of this approach for robust and reliable sugarcane disease detection in the Sri Lankan context.