Plant diseases represent a persistent challenge to global food security, with annual crop yield losses reaching as high as 40%. Timely and reliable identification of these diseases is therefore essential for reducing economic losses and ensuring sustainable farming practices. Although deep learning has achieved remarkable success in image-based disease classification, much of the existing research emphasizes complex convolutional neural networks (CNNs) that are unsuitable for resource-constrained environments. In this work, we present a systematic evaluation of two widely adopted transfer learning models ResNet50 and MobileNetV2 applied to the New Plant Disease dataset (~87,000 images spanning 38 categories). Our experiments examine the role of augmentation, learning rates, batch sizes, and fine-tuning strategies in model performance. Results demonstrate that MobileNetV2 not only achieved superior accuracy (98.33%) with strong precision, recall, and AUC values, but also required seven times fewer parameters and significantly reduced training time compared with ResNet50. Error analysis further revealed MobileNetV2’s ability to differentiate between diseases with overlapping symptoms. Importantly, its lightweight architecture supports real-time implementation on mobile devices, drones, and IoT systems, offering clear advantages for field deployment. Unlike prior studies that emphasize raw accuracy, this research highlights efficiency, robustness, and deployability. Overall, our findings establish MobileNetV2 as a practical and scalable solution for next-generation precision agriculture and low-cost disease monitoring systems (pp.11-23).