A Comparative Study of VGG-like and ResNet-like CNN Architectures Trained from Scratch for Tomato Leaf Disease Classification on the PlantVillage Dataset

Abstract

Tomato (Solanum lycopersicum) is a high-value horticultural commodity whose productivity is highly susceptible to various leaf diseases with often similar visual symptoms. This study compares two Convolutional Neural Network (CNN) architectures trained from scratch, namely VGG-like and ResNet-like, for classifying ten categories of tomato leaf conditions consisting of nine disease classes and one healthy class. The dataset used is the tomato subset of the PlantVillage dataset published on Kaggle. Input images of 224 × 224 pixels were normalized to the range [0, 1] and augmented through random rotation of 40°, horizontal and vertical shifts of 0.2, shear of 0.2, zoom of 0.2, horizontal and vertical flipping, and brightness variation within the range [0.8, 1.2]. The VGG-like architecture consists of four 3 × 3 convolutional blocks with 64, 128, 256, and 512 filters, each followed by batch normalization, max-pooling, and dropout, and ends with two dense layers. The ResNet-like architecture consists of an initial 7 × 7 convolutional layer followed by ten residual blocks grouped into four stages with 64, 128, 256, and 512 filters, ending with global average pooling and two dense layers. Both models were trained using the Adam optimizer with a learning rate of 1 × 10⁻⁴, categorical cross-entropy loss function, a batch size of 32, and a maximum of 50 epochs, with EarlyStopping, ReduceLROnPlateau, and ModelCheckpoint callbacks. Experimental results show that the VGG-like model achieved approximately 91% validation accuracy with an average per-class accuracy of 94.8%, while the ResNet-like model achieved approximately 94% validation accuracy with an average per-class accuracy of 96.3%. The ResNet-like architecture demonstrated superior performance, particularly for visually challenging classes such as Tomato mosaic virus and Spider mites. These findings indicate that residual connections provide a significant advantage for tomato leaf disease classification when models are trained from scratch, although the VGG-like architecture remains a viable option when architectural simplicity is prioritized.