Weeds - Phase 1 - Research Highlights

Wheel Cacti detection

 

What the report is about

This project examined the role of unmanned aerial vehicles in detecting, classifying and mapping infestations of wheel cactus, Opuntia robusta, over large areas of rangelands in outback Australia.

This report presents details of the flight trials performed; details of the development of cactus detection, classification and mapping algorithms; and the results of these algorithms.

Click here to access the full report.

Methods used

In the months leading up to September 2009 the focus was on preparing a fixed-wing unmanned aerial vehicle and an associated sensor payload-consisting of a terrain-facing monocular vision camera and onboard navigation sensors-for data collection over rangelands populated by cacti.

Trial flights near Oraparinna, in the Flinders Ranges of South Australia, were performed in September 2009 by staff and students from the Australian Centre for Field Robotics. Several flights were performed over an area of 1.2 by 2.0 kilometres, and data from the onboard sensor payload were logged and used to develop and test algorithms for the detection and mapping of cacti.

In conjunction with the flight trails, a ground-based survey of cacti in the flight area was performed by ACFR staff and students in order to provide ground-truth and comparison data for validating the detection and mapping algorithms.

Results/key findings

Results from the project demonstrate the applicability and capability of UAV systems using cheap sensors (such as a colour vision camera) for identifying and mapping cacti in remote locations. The success of the performance of the classification scheme was, however, only limited: the scheme demonstrated a good ability to detect cacti but only while providing a fairly large number of false detections.

Nevertheless, such a system in its current form could, for example, provide valuable information for a weed expert, acting as a first stage data-processing tool for identifying possible cactus locations but requiring the user to analyse and remove false detections.

Recommendations

Several avenues of future work for improving the system present themselves.

The first would be to examine vision-based feature descriptors that can better capture the visually distinguishing characteristics of the cacti. Apart from the colour and texture properties already exploited in this project, other possible features might be a type of shape detection-that is, a circular or elliptical Hough transform-in order to account for the ‘wheel-like' leaves of the cactus.

A second avenue might be to vary imaging characteristics such as the capture resolution of the imagery data or to make observations of cacti from a more horizontal perspective by, for example, using a hovering vehicle to hover lower and closer to the ground. Wheel cacti are more naturally distinguishable from the horizontal perspective and so might be more easily classified if a second, lower flying vehicle were to be used to provide extra imagery.