A Comparative Study of Transfer Learning-Based Deep Learning Models for Tomato Leaf Disease Detection

Tomatoes are an important crop around the world, both for their nutritional value and their economic value. However, if you do not catch the different diseases that can affect tomato leaves early, they can cause lower yields and lower quality. Small- and medium-sized producers need to be especially careful about how they classify these diseases. This study comparatively assessed deep learning models employing transfer learning for the visual diagnosis of tomato leaf diseases. We used a number of pretrained models, such as DenseNet121, MobileNetV2, ResNet50, EfficientNetB0, VGG16, VGG19, and InceptionV3. One of four optimization algorithms was used to train each model: Adam, Stochastic Gradient Descent, Nadam, or AdamW. The experiments utilized the publicly available Tomato Leaf Disease Detection dataset, comprising 10 classes (9 diseases and 1 healthy class) and 10,000 images equally distributed across categories. An 80/20 training–testing split was applied, and an independent external test set was used to assess model generalization. Data augmentation, dropout, early stopping, and learning rate schedulers were all part of the training process. The performance was measured by the model’s accuracy, loss values, and confusion matrices. The AdamW optimization algorithm worked best with the EfficientNetB0 model, which got 99.72% accuracy on the validation set and 99.68% accuracy on the test set. That means that transfer learning techniques could be used to make quick, accurate, and reliable diagnostic systems in the field of agricultural technology.