Citrus crops face significant threats from foliar diseases such as citrus greening, which require a timely and accurate diagnosis to mitigate yield losses. This study addresses the problem of automated classification of citrus leaf diseases by proposing a transfer learning approach based on MobileNetV2 and MobileNetV3-Small, enhanced with Grad-CAM for visual explainability. Through three experimental configurations, Studio-only, mixed, and MobileNetV3 mixed acquisition and Grad-CAM-integrated, the study demonstrates robust classification performance, with MobileNetV3 achieving 89.3% precision and an F1 macro score of 0.895 under heterogeneous conditions. The integration of Grad-CAM confirms that model decisions are based on disease-specific visual cues, enhancing trust and interpretability in real-world deployments. Despite strong results, the work acknowledges the limited diversity of disease classes and environmental conditions in the dataset. Future research will address these limitations by expanding the dataset, incorporating additional imaging modalities such as hyperspectral data, and validating the models in real-time field environments. These findings underscore the potential of lightweight and explainable deep learning models for sustainable and scalable agricultural diagnostics.