Using artificial intelligence for skeleton weed detection from low level aerial imagery

Document Type

Conference Proceeding

Publication Date

8-2024

Conference Title

23rd Australasian Weeds Conference - Breaking the Cycle: Towards Sustainable Weed Mangement

Place of Publication

Brisbane

ISBN

978-0-646-70156-1

Disciplines

Artificial Intelligence and Robotics | Weed Science

Abstract

Chondrilla juncea (Skeleton weed) has been under an eradication program in Western Australia since 1974. The low-density, wide-spread nature of this slender forb has presented challenges for aerial mapping in the last couple of decades. Recent advances in highresolution imagery captured by drones and small object detection by deep learning neural networks have allowed effective surveillance of this weed. High quality ground truth data was required for detection accuracy levels that match on-ground surveys. Correct identification of individual plants directly on drone imagery was difficult and there is a risk of missing plants which corrupts training dataset.

This paper presents the methods developed to: (1) gather high quality ground truth data, (2) build an initial reference dataset, (3) provide a framework for efficiently 209 collecting fine-tuned data, and (4) producing a deep learning model. Ground level imagery of nearcentimetre geolocated quadrats was captured with a smartphone in a way that facilitated geospatial alignment with drone imagery. Individual plants were annotated using QGIS (an open-source software). These were used to project bounding-box annotations onto individual training tiles taken from drone imagery. We are adapting the workflow to integrate it into the department’s GIS infrastructure to allow labelling and validation of detected plants by field officers. This ground-truth data framework with in-house management of detection models and finetuning halves the costs of processing compared to onground surveillance.

Link to Full Text

Share

COinS