We used the dataset available at https://data.mendeley.com/datasets/9424skmnrk/1 for our project. These images were collected in Maharashtra, India manually. We got 518 images of sugarcane leaves infected with red rot and 522 images of healthy leaves.

To enhance the diversity of the dataset and improve model generalization, we employed a two step data augmentation process. Every image in the dataset was subjected to three rotations, thereby quadrupling the size of the dataset. The dataset was further expanded by applying additional augmentation techniques like RandomResizedCrop, sharpening, introducing gaussian noise, flipping etc. with probabilistic randomness, which resulted in doubling the dataset size.

We converted the image to grayscale and applied a binary threshold to create a binary image, which allowed us to identify and remove the dark areas around the leaf. Using this, we effectively cropped out the portion of the image containing the leaf.

We trained a Random Forest classifier on features extracted from a fine-tuned ResNet-18 model, specifically to detect red rot disease in sugarcane leaves.

The Vision Transformer (ViT) model with the base configuration (ViT-base-patch16-224) was fine-tuned on our dataset, leveraging pre-trained weights to enhance performance in detecting red rot disease in sugarcane leaves.

Finally, we fine-tuned VGG16, EfficientNet, and ResNet models on our sugarcane leaf dataset. This allowed us to leverage pre-trained knowledge and adapt the models for better performance in classifying the leaf images.