Runners play a crucial role in vegetative propagation and provide a cost-effective method for expanding strawberry production, excessive growth presents significant challenges. Uncontrolled runner development diverts energy from the mother plant, leading to smaller, lower-quality fruit and ultimately reducing overall yield. Accurate and efficient identification of strawberry runners is essential for optimizing crop management and propagation. This dataset was prepared to enhance the automation of strawberry runner identification in field environments by developing a deep learning-based approach. A diverse dataset of strawberry images was collected, encompassing a wide range of varieties, different growth stages, and multiple growing seasons. Images were captured using three different platforms and cameras at varying altitudes to improve dataset diversity: 0.5 m above the ground level (AGL) for ground imaging (GI), 10 m AGL for aerial imaging (AI10), and 5 m AGL for aerial imaging (AI5). All raw images were preprocessed to segment individual plants. out and saved. Runners were annotated using polygonal masks to delineate their boundaries, with a single class designated for runner identification.

We collected data from strawberry breeding trials conducted at the University of Florida (UF) Gulf Coast Research and Education Center (GCREC) in Wimauma, Florida (27° 45' 40" N, 82° 13' 49" W). Videos and images were collected from two separate trials conducted over two consecutive growing seasons (Table 1). Each trial included multiple strawberry genotypes, with replicated plantings to ensure experimental consistency (Table 1).

Table 1. Number of strawberry genotypes and plants used for image collection in different trials across two consecutive field seasons.

Video Collection Using Ground Imaging (GI) Platform

4K videos at 60 fps were collected using an RGB camera (24MP, full-frame, EOS R8, Canon, Japan) mounted onto a ground vehicle (Amiga, Farm-Ng, CA, US). The camera was mounted on a bar at approximately 0.5 m AGL, positioned about 0.3 m above the plant canopy, and angled 45 degrees down from the horizon, facing forward in alignment with the vehicle’s movement direction. All the videos were collected during the 2023-24 strawberry season and processed using the Python script ground_video_preprocessing.py, which trims the footage to remove background to retain a single plant in view, extracts high-resolution frames from the video, and saves images for annotation.

Image Collection Using Aerial Imaging Platform at 10 m AGL (AI10)

A quad-copter unmanned aerial vehicle (UAV, EVO-II, Autel Robotics, WA, US) equipped with an RGB camera (20MP 1-inch CMOS, Sony, Japan) was used to capture strawberry plant images at 10 m AGL from the same clonal trials as the GI during the 2023-24 strawberry season, aiming to enhance data collection efficiency. The onboard camera was downward-facing, capturing top-view images of the strawberry plants. These images were then processed using a Python script UAV_image_preprocessing.py to segment out each individual strawberry plant for annotation.

Image Collection Using Aerial Imaging Platform at 5 m AGL (AI5)

To balance image resolution and data acquisition efficiency, we utilized another UAV (Mavic Pro 3, DJI, China) equipped with an imaging sensor array to capture images at 5 m AGL. A different UAV was selected for AI5 compared to AI10, primarily due to the superior performance of its main imaging sensor (20 MP, 4/3-inch CMOS, Hasselblad, Sweden). The onboard camera was downward-facing, capturing top-view images of the strawberry plants. Using this platform, we collected mid-resolution aerial images from density trials conducted during the 2024-25 strawberry season. Unlike images collected by GI and AI10, these mid-resolution aerial images were only collected during the early growth stage, when the number of runners typically reaches its peak. These images were processed the same way as the images from AI10.

Table 2. Summary of Data Collection Platforms and Data Collection Date and Time

Image Annotation and Augmentation

Runners were annotated using polygons on all the preprocessed video frames and images. All the training data were augmented by applying horizontal flipping and 90-degree rotations (none, clockwise, or counterclockwise) to increase dataset diversity.

Table 3. Training and Validation Data Sets Summary

For ground videos, the proportions of the left and right soil borders varied slightly, resulting in minor differences in the dimensions of the cropped videos. However, the selected video frames were approximately 1,464 × 1,152 pixels in size. The segmented aerial images captured from 10 m and 5 m had resolutions of 218 × 218 pixels and 392 × 392 pixels, respectively.