Abstract:
Wheat is a vital staple crop worldwide, yet its yield is frequently reduced by leaf diseases that threaten food security. Traditional diagnosis through expert inspection is costly, time-consuming, and often inaccessible to farmers, while existing deep learning models, though effective, are computationally intensive and less practical for deployment in real-world agricultural settings. To address this gap, this thesis proposes WheatCNet, a lightweight convolutional neural network tailored for efficient and accurate wheat disease detection. The model was trained and evaluated on a dataset covering five classes—Black Point, Fusarium Foot Rot, Healthy Leaf, Leaf Blight, and Wheat Blast—and benchmarked against several pretrained CNN architectures, including VGG16, VGG19, DenseNet121, Xception, and MobileNetV2. Although these models performed strongly, WheatCNet surpassed them, achieving an overall accuracy of 99.80% with near-perfect precision, recall, and F1-scores across all classes. Reliability was further confirmed through confusion matrices, accuracy– loss curves, and ROC–AUC analysis, with WheatCNet achieving AUC values of 1.000 for every class. Interpretability was enhanced by integrating Gradient-weighted Class Activation Mapping (Grad-CAM), which verified that predictions were based on disease-specific regions. Combining high accuracy, efficiency, and explainability, WheatCNet is suitable for deployment on resource-constrained devices, enabling real-time field diagnosis. The study demonstrates WheatCNet’s potential as a practical and sustainable solution for precision agriculture, capable of reducing pesticide misuse, supporting farmers’ decision-making, and contributing to food security.
