Detection of yellow rust (Puccinia striiformis f.sp. tritici) disease using deep learning algorithms

This study aims to develop a deep learning-based image classification method for fast, reliable and low-cost detection of yellow rust disease, which causes yield and quality losses in wheat production, in field conditions. The research material includes healthy and yellow rust-infected wheat leaf images obtained from different wheat lines grown at Sivas Science and Technology University Agricultural R&D Center in the 2023-2024 production season. The images were taken with an iPhone 12 model phone and a total of 1948 leaf images consist of 376 healthy and 1572 yellow rust disease images. The classification of the images was carried out in SBTÜ Plant Protection Laboratory. The study was carried out using the Python programming language in the Google Colab virtual environment. The data were scaled to 75x75 pixels, and data augmentation techniques including rotation, shifting and zooming were applied to the healthy data set to eliminate class imbalance. ESA, one of the deep learning methods, was used for training the data. In the study, 40 pre-trained models in the Keras library were tested for 10 epochs in order to be tested quickly. As a result of the preliminary evaluation, MobileNet, Xception and EfficientNetV2B2 models stood out compared to other models. In the second stage, these 3 successful models were re-trained with different activation functions (ReLU, GELU, SWISH, LeakyRELU). As a result of the training, the MobileNet model stood out with 97.69% accuracy rate with the LeakyReLU activation function combination. In the last stage, in order for the deep learning models to obtain more stable results and to improve the results, the number of epochs was increased to 25 and tested with different optimizers. As a result of this stage, the most successful result was achieved with the MobileNet + Relu + Nadam combination with 98.21% accuracy value.