Vines-DB: An RGB image dataset for multi-species ornamental vine segmentation

Technical Summary: Vine-DB Dataset for Multi-Species Ornamental Vine Segmentation

Problem Statement
Canopy cover area is a critical trait in horticulture, directly influencing irrigation management, crop growth assessment, and plant performance evaluation. However, traditional methods for measuring canopy cover, such as manual tracing or semi-automated thresholding tools (e.g., ImageJ), are labor-intensive and difficult to scale. These conventional approaches struggle significantly with structurally complex species like ornamental vines, which exhibit twining growth habits, overlapping foliage, and variable canopy architectures. There is a need for automated, deep learning-based solutions to facilitate field phenotyping for these challenging plant structures.

Methodology
The authors developed Vine-DB, a high-resolution RGB image dataset specifically designed for multi-class instance segmentation of ornamental vines.

• Data Collection: The dataset comprises 1,218 original high-resolution images captured at the Utah Agricultural Experiment Station's Greenville Research Farm in Logan, Utah, during the 2023 and 2024 growing seasons (July–October).
• Subjects: The study focuses on seven ornamental vine species: Akebia quinata ('Shirobana'), Campsis radicans ('Monbal'), Campsis × tagliabuana ('Madame Galen'), Hydrangea anomala petiolaris, Lonicera × heckrottii ('Goldflame'), Parthenocissus quinquefolia, and Wisteria floribunda ('Texas Purple'). These were grown on 4 ft × 8 ft trellises.
• Imaging Protocol: Images were captured using an iPhone 16 Pro (48 MP) mounted on a tripod at a height of 0.6 m and a distance of 1 m from the canopy. To ensure consistency, images were taken between 10:00 AM and 12:00 PM under natural daylight against solid black or white Styrofoam backdrops to minimize background noise and improve contrast.
• Annotation: All original images were manually annotated using the Roboflow platform. Trained annotators created polygon-based instance segmentation masks for eight classes: the seven vine species and a background class. Inter-annotator agreement was assessed on a subset of 150 images, yielding a mean Intersection over Union (IoU) of 0.89 ± 0.06.
• Preprocessing and Augmentation: Original images were resized to 640 × 640 pixels using letterbox padding, and pixel values were normalized to the range [0, 1]. For model development, data augmentation techniques (random rotation, flipping, brightness/contrast adjustment, and cropping) were applied, expanding the working dataset to 2,307 images.
• Dataset Split: The augmented dataset was divided using stratified sampling into training (2,019 images), validation (192 images), and test (96 images) subsets. Images from the same plant were kept within a single subset to prevent data leakage.

Key Contributions

1. Specialized Dataset: Vine-DB provides the first high-resolution, field-based RGB dataset for seven distinct ornamental vine species, capturing temporal variations in canopy development across multiple months.
2. High-Quality Annotations: The dataset includes expert-annotated, polygon-based instance segmentation masks for eight classes, enabling pixel-level analysis and supporting the training of advanced deep learning models.
3. Reproducibility and Transparency: The repository clearly distinguishes between original, non-augmented images and augmented derivatives, ensuring transparency in dataset composition. It provides data in COCO instance segmentation format and optional YOLO-format labels.
4. Benchmarking Utility: By capturing realistic field conditions, including natural variability in illumination and seasonal growth, the dataset serves as a robust benchmark for evaluating computer vision models in precision horticulture and urban ecology.

Results and Data Characteristics
The paper presents the dataset specifications rather than training specific deep learning models to report performance metrics. The core dataset contains 1,218 original images derived from 168 unique plant specimens (24 plants per species). The final working dataset for model development consists of 2,307 images. The dataset successfully captures the morphological diversity and structural complexity of the target species, with annotations verified for high consistency (IoU ~0.89).

Significance and Claims
The authors position Vine-DB as a foundational resource to support the transition from manual image analysis to automated deep learning-based field phenotyping. The dataset is claimed to be valuable for:

• Developing and benchmarking computer vision models for complex plant architectures.
• Enabling applications such as automated canopy cover estimation, species identification, and scalable field phenotyping.
• Facilitating research into temporal changes in canopy development and plant responses under field conditions.
• Supporting transfer learning and model generalizability when integrated with other datasets.

The paper maintains a modest scope, noting limitations such as the single geographic location, the restriction to the active growing season (excluding dormancy), and potential minor inconsistencies in annotations due to morphological similarities among species. The primary contribution is the provision of a standardized, high-quality dataset to accelerate research in precision horticulture and urban ecology.