Wheat Leaf Dataset

Ethiopia has a huge resource for planting several crops, thus wheat is one of the major crops which
feed the population, but the crop has been infected by viruses, Bacteria, and Fungi. In this study,
the two major diseases namely Stripe Rust and Septoria wheat leaf image were used for
experimentation.
Currently, the wheat disease mentioned above is a major headache not only for
farmers but also for plant pathologists. Furthermore, the pathologist uses the naked eye observation
for the detection of wheat disease, sometimes it is challenging to detect without using the
laboratory material. Several types of research conducted disease identification and detection using
the traditional machine learning algorithms. Thus algorithms have drawbacks since the feature
extraction is expert-based and the amount of data processing required is high relative to the machine
learning. Because of this, the deep learning methodology was used to detect wheat disease.
The approach has three main phases. The first phase is to collect the dataset from the wheat farm, and the image has three categories i.e. ‘Healthy Wheat Leaf’, ‘Strip Rust’, and ‘Septoria Disease’. Then the dataset is partitioned using the 80%-10%-10% approach which is used for training, validation, and testing respectively. The second phase is to design a neural network by experimenting with the best hyperparameter. Finally, the best model was selected and tested with unseen image data.
The dataset contains 208 stripe rust affected leaf, 102 healthy leaf and 97 septoria affected leaf disease-infected image pictures. From this, 80% of the images were used for training, l0% were used for validation, and the remaining 10% were used for testing. During training, the data augmentation technique is used to generate more images to fit the proposed model. The experimental result demonstrates that the proposed model is effective for the detection of wheat leaf disease (Strip Rust and Septoria). The pretrained model used for experimentation are VGG19, InceptionV3, MobileNet, and EfficientNet. Among mentioned pretrained models MobileNet has achieved the best result and the model can successfully classify the given image with a testing accuracy of 90% with images captured in the real wheat farm with a heterogeneous environment.