A Wheat Leaf Disease Detection Approach Using Deep Learning & Machine Learning Techniques

Wheat is a grass of the genus Triticum that is grown for its seeds. After rice & maize, it is the third most-produced cereal and second for human consumption, which makes it a primary food grain in many nations. Wheat production is significantly threatened by leaf diseases, which can cause yield losses of up to 40\% along with leaf rust covering 1–20% loss. Effective detection of wheat plant disease can mitigate the adverse impacts. However, it became extremely difficult to detect these diseases manually. While deep learning offers a promising solution, a systematic analysis of how crucial preprocessing steps impact the performance of different deep learning architectures remains largely unexplored. For classification, 5530 wheat leaf images with multi-class observations were collected from a valid source. To address this, our methodology systematically evaluates an integrated pipeline, comparing four deep feature extractors (DenseNet121, ResNet152, Xception, VGG19) when paired with both traditional classifiers (SVM, Random Forest, XGBoost) and a custom CNN classifier. Preprocessing methods such as gamma correction, unsharp masking, and random oversampling were used to improve the dataset quality. Performance was evaluated using overall accuracy, class-wise AUC-ROC, F1 score, recall, and precision. We achieved the highest overall accuracy of 99.43% using ResNet152 with a custom CNN classifier across five classes. A key contribution of this work is the comparative analysis showing that while traditional classifiers struggle with deep features, a fine-tuned, end-to-end deep learning approach significantly overcomes these limitations. The proposed approach not only advances automated disease detection in wheat farming but also provides a foundation for improving disease control methods in other crops. Furthermore, model interpretability using LIME has been employed to verify that the model focuses on relevant image features for its predictions.