Citrus Disease & Pest Detection Using Yolo Models with Self -Attention Yolov8

Many of the diseases of citrus are equally devastating and early detection is essential for timely orchard management. In this work, a citrus disease detection and guiding pipeline, CitraYOLO, is presented, which covers comparative object-detection studies, the introduction of CitraYOLO's unique lightweight self-attention YOLOv8 network, and a Flask web application. The experimental part showcases Faster RCNN and seven YOLO family variants: YOLOv3, YOLOv5s6, YOLOv5x6, YOLOv8n, YOLOv11, YOLOv26 and a special Light SA-YOLOv8 model. The proposed detector incorporates Fast_C2f, BRA and ASF straight into the YOLO pipeline. Structured citrus dataset, consisting of 4602 training images and 657 validation images across five classes: Melanose, black spot, canker, greening and healthy, were used for training. Models are compared to each other by precision, recall, mAP@0.5, F1- confidence curve, precision-recall and confusion matrix. According to the comparison results, all the metrics of the top three models (YOLOv5x6, YOLOv11, YOLOv3) are complementary.Based on the comparison results, YOLOv5x6 has the highest precision (0.774), YOLOv11 has the highest recall (0.895) and YOLOv3 has the highest mAP (0.829). A modified version of SA-YOLOv8, which is a lightweight efficient model with a balanced architecture, outperforms YOLOv8 with 0.756 precision, 0.782 recall and 0.800 mAP. The chosen model is incorporated with Flask for real world deployment and SQLite based user authentication, image upload, confidence display and local disease description, cause, remedy and prevention counsel