A soybean and weed image dataset collected using FarmBot in a controlled outdoor field environment: raw and curated image data for precision agriculture research

This dataset consists of images of soybean plants and weeds captured during a controlled outdoor field experiment conducted at Weihenstephan-Triesdorf University of Applied Sciences (HSWT) in Freising, Germany. The images were collected using a FarmBot precision agriculture robot equipped with an ELP USB4KHDR01-MFV camera featuring a Sony IMX317 sensor and a 5–50 mm optical zoom lens. All images were recorded in 4K Ultra HD resolution (3840 × 2160 pixels).

Soybean seeds were planted on August 14, 2025, following standard agronomic spacing guidelines, with 5 cm between seeds and 37.5 cm between rows across three parallel rows. Image acquisition began on August 24, 2025, which was eight days after the first signs of germination were observed. Data collection continued until September 20, 2025, covering a total period of 25 consecutive days. During each day, 60 images were captured—30 in the morning and 30 in the afternoon—using automated scripts developed with a Python-based API together with the FarmBot operating system interface.

From the initial collection of 1,037 images, a total of 641 high-quality images from the first 20 days of observation were retained after a quality assessment process. Images were selected based on criteria such as focus, consistent exposure, and clear separation between plant classes. Images from days 21 to 25 were excluded because the high density of weeds made it difficult to clearly distinguish between classes.

All selected images were annotated using the Roboflow platform with bounding box annotations created in collaboration with a domain expert. The annotations label instances of soybean plants and weeds within each image. The resulting dataset is intended to support research in computer vision applications for agriculture, including weed detection, crop growth monitoring, and automated precision farming systems. It can be used for training and evaluating models for tasks such as object detection, image segmentation, and image classification in agricultural environments.