A High-Accuracy Deep Learning Approach for Wheat Disease Detection

Objectives: The purpose of the present investigation is to develop a reliable and practically deployable solution for farmers to detect wheat diseases. The proposed research aims to detect four major diseases namely, leaf rust, wheat loose smut, crown, and root rot, in the wheat crops, which is of paramount importance for controlling the spread and ensuring quality wheat production. Methods: This research uses pre-trained deep-learning (DL) models for detecting and classifying wheat diseases. The proposed investigation explores the usefulness of EfficientNetB0, VGG16, and ResNet50 for the detection of the said wheat diseases. The developed methodology is tested using a large dataset containing 4906 images of wheat plants with varying diseases. Findings: The EfficientNetB0 model achieved a 99.37% accuracy rate, outperforming the other models. The data augmentation and Multistage fine-tuning techniques, explored in this research enhance the performance of the model. Novelty: The literature reports a relatively low accuracy, for the detection of wheat diseases, which needs to be further increased from the perspective of reliability and practical deploytion. Moreover, the reported methods are verified on a limited dataset and diseases. It is also necessary to consider varying levels of infections of all commonly occurring wheat diseases. The proposed research methodology is experimented on a huge varying dataset of four diseases and obtains a high accuracy, thus addressing the gaps in the existing literature. The proposed DL-based approach exploiting the data augmentation and fine-tuning techniques for enhanced and reliable performance marks a leap towards practical deployment, making it novel.