Detection and Targeted Pesticide Application for Tea Diseases Based on YOLOv12 and 3D Vision

Abstract

As a globally significant economic crop, tea yield and quality are strongly affected by disease control. Traditional manual inspection is inefficient and subjective, while non-targeted chemical application may increase pesticide use and environmental burden. To support lesion-level precision plant protection, this study proposes a tea disease detection and targeted pesticide application framework that combines an improved YOLOv12 detector with RGB-D three-dimensional (3D) perception and robotic execution. A self-built dataset containing 2121 source images from four tea disease categories was first partitioned at the original-image level into training, validation, and test subsets, and data augmentation was applied only to the training subset to prevent source-level leakage. The proposed YOLO-EMA-SoftNMS-DSConv (YOLO-EPS) model integrates Efficient Multi-Scale Attention (EMA), Distribution Shift Convolution (DSConv), and Soft Non-Maximum Suppression (Soft-NMS). Under the leakage-controlled protocol, YOLO-EPS achieved 94.28% mAP@0.5 and 87.16% mAP@0.5:0.95. Five independent runs with fixed source-level partitions yielded 94.24 ± 0.20% mAP@0.5, indicating stable training behavior. Detector-only inference time was 1.8 ms per image on an RTX 5090 desktop GPU using a Blackwell-compatible PyTorch/CUDA software stack. The system further maps two-dimensional lesion detections to RGB-D point clouds and robot coordinates and uses local surface geometry to support pose-aware targeted spraying. Controlled 3D localization and edge-platform latency experiments were conducted separately from the desktop detector benchmark. The proposed framework provides a basis for integrating disease recognition, spatial perception, and robotic targeted application in tea plantations.