Abstract for presentation at The 13th Australasian Remote Sensing and Photogrammetry Conference

Non-native plants and tree species are posing a significant threat to biodiversity, ecosystem functioning and productivity on both local and global scales. A key requirement for effective management of invasive species is the ability to reliably identify their location and distribution across landscapes. Remote sensing, and in particular the species level data that can be extracted from it, offers great potential as a source of this information. This study tested the use of four-band (red, green, blue and near infrared) digital multispectral imagery (DMSI) acquired at a resolution of 1m for mapping the non-native and highly invasive Prosopis species (mesquite) in the northwest Pilbara region of Western Australia. Various per-crown statistics were calculated on the canopies of all woody assemblages for all bands. The statistics offering the greatest degree of separation between species were selected using transformed divergence analysis. Classification of shrub and tree species was performed by integrating the selected per-crown statistics into a neural network architecture. Overall and per category accuracy was calculated using the KHAT value and conditional kappa, respectively. KHAT was found to have a value of 0.71, and the conditional kappa value for mesquite was found to be 0.99. These encouraging results highlight the ability of neural networks for weed mapping and can be tailored to a wide range of high-resolution remotely sensed datasets and alternative ecological applications.