Development of a Hybrid Pretrained Deep Learning Model with Explainable AI for Tomato Disease Detection

Tomato cultivation plays a pivotal role in ensuring global food security and sustaining national economies. However, tomato crops are frequently threatened by a variety of diseases caused by bacterial, viral, and fungal pathogens, leading to significant yield losses. Conventional disease identification approaches, which primarily rely on manual inspection by agricultural experts, are time-consuming, error-prone, and inaccessible to smallholder farmers, particularly in resource-limited settings such as rural areas of Tigray situated in the Northern Part of Ethiopia. This thesis proposes a hybrid deep learning model augmented with Explainable Artificial Intelligence (XAI) techniques to enhance the accuracy, interpretability, and practical applicability of automated tomato disease detection systems. A benchmark dataset comprising 17,920 annotated images representing ten distinct tomato disease classes was utilized for training and evaluation purposes. Two pre-trained convolutional neural network (CNN) architectures, namely ResNet-50 and MobileNetV2, were selected for comparative analysis based on their performance and computational efficiency. Based on the experimental results of the proposed method, the ResNet-50 model achieved a training accuracy of 99.25% and a validation accuracy of 94%, whereas MobileNetV2 attained 92.6% training and 87.4% validation accuracy. To improve the generalization capabilities of the models and mitigate overfitting, several data augmentation strategies, including rotation, flipping, and scaling, were employed. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize the class-discriminative regions within input images to provide interpretability and revealing potential biases in model predictions. The results demonstrate that the integration of deep learning with XAI techniques yields an effective and transparent solution for tomato disease detection. The proposed approach offers valuable insights for deploying intelligent diagnostic tools in precision agriculture, particularly benefiting smallholder farmers by facilitating early disease detection and improved crop management.