Ground-truthed and high-resolution drone images of the leafy spurge weed plant (Euphorbia esula)

TL;DR

This paper presents a high-resolution drone dataset of leafy spurge plants in Montana to help develop early warning systems for weed invasion.

Contribution

The novel contribution is a publicly available, ground-truthed drone image dataset for leafy spurge detection with Hugging Face integration.

Findings

A 1.3 cm resolution drone image dataset was created with 1000 ground-truthed points for leafy spurge presence.

Two vision models achieved 0.85 test accuracy, showing classification is feasible but challenging.

The dataset is released as a Hugging Face Dataset for easy access and model development.

Abstract

This dataset comprises 1.3 cm resolution aerial images of grasslands in western Montana, USA, captured by a commercial drone. Many scenes contain leafy spurge (Euphorbia esula), introduced to North America, now widespread in rangeland ecosystems, which is highly invasive and damaging to crop production and biodiversity. Technicians surveyed 1000 points in the study area, noting spurge presence or absence, and recorded each point’s position with precision global navigation satellite systems. We cropped tiles from an orthomosaic image at these locations. We publicly release these images and metadata as a Hugging Face Dataset, accessible in one line of code. Our aim is to invite the research community to develop classifiers as early warning systems for spurge invasion. We tested classification performance for two contemporary vision models and achieved 0.85 test accuracy. This demonstrates the feasibility yet difficulty of this classification task.

Figures