YOLO-TDC: A lightweight network for tea bud detection and counting of the large-leaf black tea in UAV imagery

Abstract

Accurate detection and counting of tea buds in unmanned aerial vehicle (UAV) imagery are essential for rapid tea plantation monitoring and yield estimation. However, nadir-view images of large-leaf tea canopies are often affected by uneven illumination, dense bud–leaf occlusion, complex backgrounds, and substantial scale variation, while existing detectors frequently struggle to balance detection accuracy and model compactness. To address these challenges, we propose YOLO-TDC, a lightweight network for tea bud detection and counting in UAV imagery. A UAV image dataset containing 2300 images and 43,885 annotated tea buds of the large-leaf black tea cultivar Yinghong No 9 was constructed. Based on YOLOv8n, YOLO-TDC incorporates a MobileViT backbone for lightweight global feature extraction, CARAFE for content-aware feature reconstruction, and a Lightweight Occlusion-aware Context Aggregation module for multi-scale contextual representation. A loss strategy combining SlideLoss and Normalized Wasserstein Distance was further employed to improve hard-sample learning and small-scale object localization. YOLO-TDC achieved 76.2% precision, 78.8% recall, and 84.6% [email protected], exceeding YOLOv8n by 6.1 percentage points in [email protected]. The model contained 2.42 M parameters, had a size of 5.3 MB, and achieved an inference speed of 51.0 FPS under the experimental hardware configuration. For counting, it achieved the highest $R²$ of 0.86 among the compared models and a mean relative counting accuracy ($A_{\text{count}}$) of 84.53% on field UAV images. These results demonstrate that YOLO-TDC provides an accurate and compact approach for UAV-based tea bud detection and counting and can supply bud-count information for subsequent yield estimation.