Dr Angus Carnegie, NSWDPI
Urban and peri-urban trees in major cities provide a gateway for exotic pests and diseases (hereafter “pests”) to establish and spread into new countries. Consequently, they can be used as sentinels for early detection of exotic pests that could threaten commercial, environmental and amenity forests. Post-border biosecurity surveillance for the early detection of exotic forest pests relies on monitoring of host trees — or sentinel trees — around high-risk sites, such as airports and seaports. In the event of an exotic pest incursion, surveillance of all known hosts in a defined area is necessary to determine whether eradication is feasible, and if so, was it subsequently successful. However, in Australia, there are few publicly available spatial databases of urban street and park trees, so locating and mapping host trees is primarily conducted via ground surveys. This is time-consuming and resource-intensive, and generally does not provide complete coverage.
Advances in remote sensing technologies and machine learning provide an opportunity for semi-automation of tree species mapping to assist in biosecurity surveillance. In this study, we obtained high resolution (>12 cm), 10- band, multispectral imagery using the ArborCamTM system mounted to a fixed-wing aircraft over Bayside Council in Sydney, Australia, which encompasses Port Botany and Sydney International Airport — two major entry pathways for invasion of exotic pests. We mapped 630 Pinus trees and 439 Platanus trees on-foot, validating their exact location on the airborne imagery using an in-field mapping app. These genera were chosen as they are hosts for several high priority pests for Australia. Using a machine learning, convolutional neural network workflow, we were able to classify the two target genera with a high level of accuracy in a complex urban landscape. Overall accuracy was 92.1% for Pinus and 95.2% for Platanus, precision (user’s accuracy) ranged from 61.3% to 77.6%, sensitivity (producer’s accuracy) ranged from 92.7% to 95.2%, and F1-score ranged from 74.6% to 84.4%. Our study validates the potential for using ArborCam imagery and machine learning to assist in tree biosecurity surveillance. We encourage biosecurity agencies to consider greater use of this technology.