Search for a command to run...
Black rot, smut, wilt and grassy shoot are some of the leaf diseases that have a great influence on the productivity of sugarcane as a commercially important crop. The timely and proper detection of the disease is necessary to minimize the loss of yield and facilitate sustainable agriculture. Manual inspection techniques are slow, subjective, and inappropriate with large-scale cultivation, and the current convolutional neural network (CNN)-based systems tend to fail in practice in the real field due to differences in brightness, background distractions, and similar disease patterns. In order to alleviate such restrictions, this paper suggests a multi-disease classification model that uses a Vision Transformer (ViT) and Contrast-Limited Adaptive Histogram Equalization (CLAHE) to process the images. CLAHE enhances local contrast and suppresses noise, thereby improving the visibility of disease-specific features. The dataset is split into training, validation and testing subsets in 70: 15: 15 proportion. The accuracy, precision, recall, and F1-score are used to analyze the proposed model and compare it to the traditional CNN architectures ResNet50 and VGG16. Through experimental findings, it is proven that the ViT-based framework outperformns CNN models in achieving better classification performance of all the types of diseases. These results show the suitability of transformer-based architectures in efficient sugarcane disease detection in natural agricultural fields.