Multi-format open-source weed image dataset for real-time weed identification in precision agriculture

Abstract

Weeds are considered obnoxious and a hindrance to crop yield. Due to their uneven spatial distribution pattern, a ground or aerial robot are deployed to spot spray herbicides. This herbicidal application depends entirely on the computer vision algorithms that assist with in-field weed identification prior to spot spraying. Therefore, to develop advanced computer vision algorithms, big data pertaining to agricultural weed dataset are required. In the past, public domain weed dataset have been released but mostly acquired using ground-based technologies. The dataset discussed in this paper is unique in that it incorporates data captured both from handheld camera and unmanned aerial system (UAS), thus catering to both ground-based and aerial-based weeding robots. This dataset comprises of 3,975 images featuring five different weed species commonly found in North Dakota: kochia (Bassia scoparia), common ragweed (Ambrosia artemisiifolia), horseweed (Erigeron canadensis), redroot pigweed (Amaranthus retroflexus), and waterhemp (Amaranthus tuberculatus). These images have been meticulously annotated in various formats to facilitate the development and advancements of computer vision algorithms. Furthermore, various augmentation techniques have been applied to ensure that the dataset closely represents the real-world field conditions. Additionally, this dataset is open-source to assist precision weeding technologies for real-time in-field weed identification followed by herbicidal spot spraying application, ultimately contributing to more efficient and sustainable agricultural practices.

Keywords

• Weed detection
• Deep learning
• Artificial intelligence
• Weed species
• Drone images
• Greenhouse images