Abstract
There is currently no publicly available tea chrysanthemum dataset to the authors’ knowledge. Consequently, we provide an image dataset for six varieties of tea chrysanthemums in three camera view angles obtained under complex outdoor scenes. This open-source image dataset aims to greatly promote the development of tea chrysanthemum detection methodology, particularly for automated harvesting applications.
Introduction
Tea chrysanthemums are high-value crops in China, utilized for both medicinal and edible purposes. A critical challenge in their production is the narrow harvesting window, typically lasting only 25 days. The increasing difficulty in securing skilled labor during this peak period necessitates the development of robotic harvesting solutions. However, computer vision models for such robots often struggle in real-world agricultural environments. Most existing datasets are captured in controlled laboratory settings, failing to account for "unstructured" outdoor conditions like shifting sunlight, shadows, wind, and obstructions from leaves or support structures. This paper introduces a robust, diverse dataset designed to bridge this gap.
Methodology
Data collection was conducted across multiple Chinese provinces using a consumer-grade smartphone (Xiaomi Mi 10) to ensure the dataset reflects hardware common in mobile agricultural robotics.
Multi-Angle Data Acquisition
Images were captured from three distinct orientations to help models recognize flowers regardless of the robot's approach:
- Parallel (~0°): Top-down view of the canopy.
- Oblique (~45°): Angled view.
- Lateral (~90°): Side view.
Environmental Context and Processing
To quantify lighting conditions, radiant illuminance was recorded using a solar power meter, averaging 47.1 W/m². High-definition videos were processed using FFmpeg, extracting one frame every 20 frames to maintain diversity while avoiding redundancy. For the labeled subset, 18,000 images were manually annotated using LabelImg, producing bounding boxes in XML format compatible with standard machine learning pipelines.
Dataset Characteristics
The dataset encompasses six major varieties: Bo-chrysanthemum, Chuju, Gongju, Hangbaiju, Jinsihuangju, and Wuyuanhuangju.
- Total Volume: 81,276 images at 1080 × 1920 resolution.
- Labeled Subset: 18,000 images (3,000 per variety).
- Raw Data: Includes 453 ultra-high-resolution images (5760 × 3240) and 1080P/60FPS video files.
The dataset specifically includes challenging conditions such as direct light, backlight, shadows, occlusion by plant organs, and overlapping flowers.
Discussion and Key Insights
A significant feature of this dataset is the intentional inclusion of "noise"—common farm fixtures like bamboo poles and ropes used for plant support. Training models to distinguish these from the crop is vital for reducing false positives in field operations.
Furthermore, the dataset reflects variety-specific growth densities. For instance, Jinsihuangju and Hangbaiju exhibit different sizes and overlap patterns. By limiting labeled images to 3,000 per variety, the authors follow the "3,000 Rule," which suggests an optimal threshold for training high-precision models like TC-YOLO without encountering diminishing returns from redundant data.
Conclusion
The release of this dataset provides a comprehensive benchmark for testing neural network architectures (e.g., YOLO, Faster R-CNN) in smart agriculture. Hosted on IEEE Dataport, it is accessible for global research. Future work should expand the dataset to include varying weather states, such as overcast or rainy conditions, to further improve model robustness across different spectral qualities.
Summary updated