Early detection of tomato leaf diseases is essential for reducing crop losses, improving yields, and promoting sustainable agriculture. This study evaluates the performance of three convolutional neural network architectures—ResNet50, VGG16, and VGG19—for automating disease detection using a dataset of annotated tomato leaf images, including healthy and diseased samples such as late blight, early blight, and leaf mold. The models were trained and tested using data augmentation and cross-validation, with performance assessed based on accuracy, precision, recall, and F1-score. ResNet50 outperformed VGG16 and VGG19, achieving the highest accuracy of 96.51, attributed to its residual connections, which enhance feature extraction and address vanishing gradient issues in deeper networks. Challenges included handling class imbalances and optimizing computational efficiency during training. The findings highlight ResNet50’s suitability for real-time disease detection, offering a scalable solution for precision agriculture. By integrating these models into IoT-based tools, farmers can implement timely interventions, reducing pesticide use and improving crop health. Future work should explore expanding datasets to include diverse environmental conditions and lightweight models for resource-limited environments. This study demonstrates the potential of deep learning to revolutionize crop management, contributing to food security and sustainable farming practices.