Citrus diseases significantly threaten global agricultural productivity, causing substantial economic losses. Early and accurate detection is crucial for effective disease management. This study presents an enhanced MobileNetV3-Small model optimized for classifying five common citrus leaf conditions: black spot, canker, greening, melanose, and healthy, tailored for resource-constrained edge devices. A comprehensive dataset was collected by merging publicly available images from Kaggle, Mendeley, and field-collected samples, which ensured diverse representations of disease severity and environmental variability. Key architectural improvements are the integration of a Flatten layer, Dropout, and L2 regularization, along with an optimal Dense layer configuration. The enhanced model achieved an accuracy of 98.85%, a compact size of 3.04 MB, and a rapid inference speed of 226.11 ms per image, outperforming ResNet101 and the baseline MobileNetV3-Small. These results demonstrate the model’s effectiveness for real-time, on-site citrus disease detection on mobile and embedded platforms, supporting cost-effective farming practices.