Detecting myrtle rust (Austropuccinia psidii) on lemon myrtle trees using spectral signatures and machine learning

Hundreds of species in one of Australia’s dominant plant families, the Myrtaceae, are at risk from the invasive pathogenic fungus Austropuccinia psidii. Since its arrival in Australia in 2010, native plant communities have been severely affected, with highly susceptible species likely to become extinct from recurring infections. While severe impact on Australian native and plantation forestry has been predicted, the lemon myrtle industry is already under threat. Commercial cultivars of lemon myrtle (Backhousia citriodora) are highly susceptible to A. psidii. Detecting and monitoring disease outbreaks is currently only possible by eye, which is costly and subject to human bias. This study aims at developing a proof-of-concept for automated, non-biased classification of healthy (naive), fungicide-treated and diseased lemon myrtle trees by means of their spectral reflectance signatures. From a lemon myrtle plantation, spectral signatures of fungicide-treated and untreated leaves were collected using a portable field spectrometer. A third class of spectra, from naive lemon myrtle leaves that had not been exposed to A. psidii, was collected from a botanical garden. Reflectance spectra in their primary form and their first-order derivatives were used to train a random forest classifier resulting in an overall accuracy of 78% (kappa = 0.68) for primary spectra and 95% (kappa = 0.92) for first-order derivative-transformed spectra. Thus, an optical sensor-based discrimination, using spectral reflectance signatures of this as yet uninvestigated pathosystem, seems technically feasible. This study provides a foundation for the development of automated, sensor-based detection and monitoring systems for myrtle rust. Download paper here

Recommended citation: Heim RHJ, Wright IJ, Chang H-C, Carnegie AJ, Pegg GS, Lancaster EK, Falster DS, Oldeland J. 2018. Detecting myrtle rust ( Austropuccinia psidii ) on lemon myrtle trees using spectral signatures and machine learning. Plant Pathology 67: 1114–1121.