UAV-derived Amaranthus retroflexus plant locations and plot boundaries from two site-specific weed management trials, Oregon, 2025

Description

This dataset contains georeferenced locations of individual Amaranthus retroflexus L. plants mapped across two vegetable field trials at the North Willamette Research and Extension Center, Aurora, Oregon, in 2025, together with the plot boundaries and the trained instance-segmentation models used to generate them. The data were collected to address whether conventional quadrat sampling can adequately characterize weed density and spatial distribution for evaluating site-specific weed management. Because the benefit of site-specific control depends on how weeds are distributed within a field, mapping every detectable plant provides a reference against which sampling-based estimates can be assessed. Plant locations were derived from low-altitude unmanned aerial vehicle imagery (approximately 1.5 mm ground sampling distance) using a YOLO26 instance-segmentation model. Each point represents the centroid of a detected plant. The dataset comprises 6,279 detections across eight plots in a beet trial and 6,182 across eight plots in a turnip trial, with each plot covering 707 m². Plot-level densities range from 0.16 to 3.11 plants m⁻² in the beet trial and 0.21 to 2.45 plants m⁻² in the turnip trial. Both fields show strong spatial aggregation, with clustering detectable from below 0.1 m out to several meters after accounting for plot-level differences in density. The point patterns can be used to examine weed spatial structure, to simulate alternative sampling designs, or as training or validation reference for weed detection work. Files include point shapefiles of detected plant locations for both trials, polygon shapefiles of the plot layouts, and the trained model weights: a base model trained on beet-trial imagery and a fine-tuned model derived from it for the turnip trial. All spatial layers use WGS 1984 UTM Zone 10N (EPSG:32610), with coordinates in meters. The dataset of annotated image tiles is included to document the annotation format and allow assessment of label quality. Tiles are 640 × 640 px extracts from the orthomosaics with 20% overlap, and labels are polygon annotations in YOLO segmentation format. Tile counts, including background tiles containing no A. retroflexus instances, are given in the associated publication. These locations are model-derived rather than field-verified. Detection performance on independent test sets is reported in the associated publication and should be considered when reusing the data.

Institutions

• Oregon State University

  Oregon, Corvallis

Categories

Funders

• United States Department of Agriculture

  Government of the United States of America

  District of Columbia, Washington

  Grant ID: Grant ID: ODA24007GR

Licence