Deep Learning-Based Identification and Classification of Sugarcane Diseases Using Advanced Image Processing Techniques
Deep Learning-Based Identification and Classification of Sugarcane Diseases Using Advanced Image Processing Techniques
Authors:
Yaawar Hussain*1, Rohit Rahat Masih*2
*¹Research Scholar, Department of Computer Science and Engineering, Institute of Technology and Management Aligarh University: AKTU Lucknow. Email: yawarhussain7011@gmail.com
*² Assistant Professor, Department of Computer Science and Engineering, Institute of Technology and Management Aligarh University: AKTU Lucknow. Email: rohitmasihcs@gmail.com
Abstract—Sugarcane (Saccharum officinarum) is a critical global crop supplying more than 80% of the world's sugar and a significant portion of bioethanol. Fungal, bacterial, and viral diseases cause annual yield losses estimated between 15% and 40%, yet field diagnosis remains largely manual and subjective. This paper proposes a comprehensive deep learning framework for the automated identification and classification of six economically important sugarcane diseases: Healthy, Red Rot, Bacterial Blight, Yellow Leaf Syndrome, Mosaic, and Rust. Five convolutional neural network (CNN) architectures—a custom four-block CNN, VGG16, ResNet50, MobileNetV2, and EfficientNetB3—are trained and rigorously benchmarked on the publicly available Sugarcane Leaf Disease Dataset comprising 5,700 annotated leaf images. Advanced image preprocessing combines Contrast-Limited Adaptive Histogram Equalisation (CLAHE), green-channel enhancement, and a ten-transform Albumentations augmentation pipeline. A two-phase transfer-learning strategy (feature-extraction followed by selective fine-tuning of the final 30 layers) consistently outperforms single-phase training. EfficientNetB3 achieves the highest individual accuracy of 96.5% and a weighted AUC of 0.998. A weighted soft-voting ensemble further raises accuracy to 97.1%. Gradient-weighted Class Activation Mapping (Grad-CAM) is employed to produce per-class saliency maps that align with domain-expert knowledge of disease symptoms. t-SNE visualisation confirms that the learned representations form well-separated, compact clusters, demonstrating strong discriminative capacity. The best model is exported to TensorFlow Lite (FP16) with a 2.3× size reduction, enabling deployment on low-power edge devices for in-field disease scouting.
Index Terms—Sugarcane disease classification, convolutional neural networks, transfer learning, EfficientNet, Grad-CAM, image augmentation, precision agriculture, plant pathology.