Crop production in subtropical regions such as Fujian, China, is highly vulnerable to the rapid spread of pests and diseases due to warm and humid climatic conditions. In this study, we propose a lightweight, deep-learning-based detection system tailored for citrus disease diagnosis, focusing on three categories: Huanglongbing (HLB), citrus canker, and healthy leaves. A curated image dataset was constructed and used to train several object detection models, with You Only Look Once (YOLO) v3 and YOLOv4 variants showing exceptional performance. The best-performing model, YOLOv3 in its fine-tuned Phase 2 version, achieved average precision scores of 98.6% for HLB, 97.4% for citrus canker, and 98.7% for healthy leaves. These results validate the system’s ability to accurately distinguish disease states under field conditions. The proposed framework supports early-stage detection, significantly reduces labor burden, and is optimized for deployment on edge devices, enabling real-time monitoring in agricultural environments. This work demonstrates a scalable and efficient solution for intelligent citrus crop management.