[Paper Review] 4Weed Dataset: Annotated Imagery Weeds Dataset

Varun Aggarwal, Aanis Ahmad+2 more|arXiv (Cornell University)|Mar 29, 2022

Smart Agriculture and AI4 citations

TL;DR

This paper introduces the 4Weed Dataset, a large-scale, annotated RGB imagery collection of 618 images from four common weeds—Cocklebur, Foxtail, Redroot Pigweed, and Giant Ragweed—acquired under real field and greenhouse conditions. The dataset supports training deep learning models for both image classification and object detection, with bounding box annotations enabling accurate early-season weed identification in corn and soybean fields.

ABSTRACT

Weeds are a major threat to crops and are responsible for reducing crop yield worldwide. To mitigate their negative effect, it is advantageous to accurately identify them early in the season to prevent their spread throughout the field. Traditionally, farmers rely on manually scouting fields and applying herbicides for different weeds. However, it is easy to confuse between crops with weeds during the early growth stages. Recently, deep learning-based weed identification has become popular as deep learning relies on convolutional neural networks that are capable of learning important distinguishable features between weeds and crops. However, training robust deep learning models requires access to large imagery datasets. Therefore, an early-season weeds dataset was acquired under field conditions. The dataset consists of 159 Cocklebur images, 139 Foxtail images, 170 Redroot Pigweed images and 150 Giant Ragweed images corresponding to four common weed species found in corn and soybean production systems.. Bounding box annotations were created for each image to prepare the dataset for training both image classification and object detection deep learning networks capable of accurately locating and identifying weeds within corn and soybean fields. (https://osf.io/w9v3j/)

Motivation & Objective

• To address the lack of high-quality, field-validated weed imagery datasets for training deep learning models in precision agriculture.
• To enable accurate early-season identification of four major weed species that significantly reduce corn and soybean yields.
• To create a balanced, annotated dataset under complex real-world field conditions to improve model generalization and performance.
• To support the development of automated weed detection systems using object detection and image classification deep learning architectures.
• To provide a publicly accessible, well-documented dataset to accelerate research in computer vision for agricultural weed management.

Proposed method

• Acquired 618 RGB images using a UTV-mounted array of six Logitech 920 webcams and handheld Sony WX350 and mobile cameras (Pixel 3, iPhone 11 Pro) at Purdue University’s ACRE farm and greenhouse.
• Captured images at 14 inches above ground level with resolutions ranging from 1080×720 to 4032×3024 pixels under diverse field and controlled lighting conditions.
• Used the LabelImg tool (Tzutalin, 2015) to manually create precise bounding box annotations for each weed instance in every image.
• Balanced the initially imbalanced dataset by acquiring additional images of Cocklebur and Foxtail using mobile cameras to achieve near-equal class distribution.
• Organized the dataset into four classes: 159 Cocklebur, 139 Foxtail, 170 Redroot Pigweed, and 150 Giant Ragweed images.
• Designed the dataset for dual use in image classification and object detection deep learning models, enabling both species recognition and spatial localization.

Experimental results

Research questions

• RQ1 Can a large-scale, field-validated, and well-annotated weed imagery dataset improve the accuracy of deep learning models in early-season weed identification?
• RQ2 How does dataset balance and image diversity under real field conditions affect the performance of object detection and classification models?
• RQ3 To what extent can a publicly available, annotated dataset like 4Weed reduce data scarcity challenges in precision agriculture research?
• RQ4 What is the impact of varying image resolutions and acquisition methods (mounted vs. handheld) on model generalization in weed detection tasks?
• RQ5 Can the 4Weed Dataset serve as a reliable benchmark for training and evaluating deep learning models in real-world agricultural settings?

Key findings

• The 4Weed Dataset comprises 618 high-resolution RGB images of four common weeds—Cocklebur (159), Foxtail (139), Redroot Pigweed (170), and Giant Ragweed (150)—acquired under real field and greenhouse conditions.
• The dataset was balanced through targeted image acquisition using mobile cameras, resulting in a nearly equal distribution across all four weed species.
• Bounding box annotations were created using the LabelImg tool, enabling direct training of both object detection and image classification deep learning models.
• Images were captured at multiple resolutions (1080×720 to 4032×3024 pixels), enhancing model robustness to scale variations.
• The dataset was collected at 14 inches above ground level using a UTV-mounted camera array, simulating real-time field scouting conditions.
• The dataset is publicly available at https://osf.io/w9v3j/ and has already been used in prior research to train deep learning models for weed identification.