Comparative Study of Lightweight Deep Learning Architectures for Potato Plant Disease Detection

Abstract

Potato leaf diseases pose a significant threat to crop productivity and global food security, necessitating accurate and reliable diagnostic systems for early detection. Although deep learning based image classification has shown promising results in plant disease recognition, many existing studies rely on simple train test splits, insufficient handling of class imbalance, and limited statistical analysis. This study presents a comprehensive evaluation of multiple pretrained convolutional neural network architectures for multi class potato leaf condition classification, including disease categories and a healthy class. DenseNet121, EfficientNetV2 S, InceptionV3, MobileNetV3 Small, ResNet50, and Xception were evaluated using a stratified K fold cross validation framework. Class imbalance was addressed through class weighted loss functions, and model performance was assessed using accuracy, macro averaged F1 score, and weighted F1 score reported as mean values with 95% confidence intervals. The experimental results indicate that ResNet50 achieved the best overall performance with a mean accuracy of 99.07% ± 0.38% and a macro F1 score of 98.24% ± 0.80%, demonstrating strong and consistent classification across all classes. Lightweight architectures such as MobileNetV3 Small also delivered competitive results with an accuracy of 97.77% ± 0.59%, highlighting their suitability for deployment in resource constrained agricultural environments. These findings emphasize the importance of statistically robust evaluation and imbalance aware training strategies for developing reliable deep learning based systems in precision agriculture.