Deep learning enables edge deployment for citrus leaf disease recognition in natural orchard scenes.

Abstract

Background Accurate and efficient detection of citrus leaf diseases is important for orchard monitoring, early intervention, and intelligent disease management. However, practical application in natural orchard environments remains challenging because of complex backgrounds, large variation in symptom scale, strong interclass similarity, and limited edge computing resources. Methods In this study, a lightweight object detection model, SKM-YOLOv11, was developed based on YOLOv11n for citrus leaf disease recognition in natural scenes. A self-built dataset was constructed from public images and field images, containing 6414 images across six classes. StarNet-S050 was introduced as the backbone, C3k2-Star was designed to enhance feature fusion across scales, and a lightweight shared detection head, MNS-Head, was constructed to reduce prediction redundancy. Results Compared with YOLOv11n, SKM-YOLOv11 increased Precision, Recall, and mAP @0.5 by 2.92, 1.47, and 0.95 percentage points, respectively. Meanwhile, FLOPs, parameter count, and model size were reduced by 31.7%, 32.7%, and 34.55%, respectively. Edge deployment on Jetson Orin NX Super achieved an inference speed of 72.45 frames per second. Conclusion The proposed model has strong potential for real-time citrus disease screening on resource-constrained devices and provides a feasible solution for edge-based intelligent disease management in natural orchards.

Plant phenotyping relevance

柑橘葉の病害状態を画像から認識する軽量物体検出モデルを開発し、データセット、精度比較、エッジ実装性能まで評価しており、植物表現型（病害状態）の取得・推定手法が中心である。

Code and data availability

The supplied blocks describe a self-built citrus leaf disease dataset (partly from the public AI Studio platform) and the SKM-YOLOv11 model, but contain no data availability statement, no public repository URL, no deposited code or trained model, and no actionable link to the dataset or analysis code. The AI Studio-sim

No evidence-backed public reproduction asset is currently recorded.