Deep learning models have achieved high accuracy in plant disease detection, but their deployment on resource-constrained devices such as smartphones or edge sensors remains challenging due to large memory and compute demands. In this work, we investigate model compression techniques for potato disease classification, focusing on dynamic dropout regularization (DDROP), structured pruning, and self-distillation, as well as hybrid approaches that combine these methods. Using a benchmark dataset of potato leaf images, we systematically evaluate the trade-offs between model size, inference speed, and classification accuracy. Our results show that DDROP combined with self-distillation effectively reduces model complexity while preserving high accuracy, and pruning further decreases parameter counts with minimal performance degradation. The proposed pipeline enables efficient deployment on low-power devices for real-time agricultural applications. This study highlights the potential of compressed and regularized models to make potato disease detection more accessible, scalable, and suitable for precision farming at the edge.