Sugarcane Disease Classification: Fine-Tuning Deep Learning Models Through Transfer Learning

Sugarcane plantations are cultivated globally, with India being the 2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> leading producer. It is estimated that fungal diseases cause the loss of approximately 18–31% of these crops. Thus, identifying sugarcane leaf disease is vital in deep learning and agriculture. Pre-trained deep learning model using fine-tuning trained with sugarcane dataset of 2521 images representing five classes of leaf diseases, i.e., Yellow Leaf, Rust, Mosaic, Red Rust and Healthy. These images have been augmented to 22424 images for better results. Further, fine-tuning the model by varying learning rates, batch size, and epochs has resulted in high accuracy. In our work, different CNN models like MobileNetV2, InceptionV3, VGG16 and EfficientNetB0 models utilized transfer learning and fine-tuning approaches to recognize diseases from sugarcane leaf images. These models are selected due to their variable architectural differences, and their performances are later compared using different performance metrics. The results in this paper show considerable improvement, and the best accuracy of 99.63% was obtained using the EfficientNetB0 model for detecting disease in sugarcane plants. They can be helpful in the agriculture sector in the timely detection of a given disease.