<![CDATA[This study conducts a comparative evaluation of seven Convolutional Neural Network (CNN) architectures, namely MobileNet, NASNet Mobile, MobileNetV2, VGG16, ResNet50, InceptionV3, and InceptionResNetV2, in classifying early blight and late blight diseases on potato leaves using a transfer learning approach. The dataset used is the Potato Disease Leaf Dataset (PLD), with all models initialized using pre-trained ImageNet weights and trained on images sized 224 × 224 pixels. To increase data diversity and reduce overfitting, real-time image augmentation was applied through horizontal flip, vertical flip, and ±5° rotation. The training process utilized the Stochastic Gradient Descent (SGD) optimizer with a learning rate of 0.001, a batch size of 2, for 30 epochs, and the Binary Crossentropy loss function. Evaluation was performed using accuracy, precision, recall, and F1-score metrics on an independent test dataset. The experimental results indicate that ResNet50 achieved the best performance with an accuracy of 97.16%, loss of 0.1153, precision of 94.63%, recall of 100.00%, and an F1-score of 97.24%, outperforming VGG16 (96.45%) and MobileNet (95.04%). In contrast, InceptionV3, NASNet Mobile, and InceptionResNetV2 demonstrated lower training stability and generalization capability on this dataset. These findings confirm that the residual connection mechanism in ResNet50 plays a significant role in improving the discrimination of visual features in leaf disease detection, while MobileNet offers an effective compromise between accuracy and computational efficiency, making it potentially suitable for implementation in plant disease detection systems on resource-constrained devices.]]>
