Can machines learn to detect plum marbling?

A pilot project suggests that machine learning could enable devices to identify infected trees in the orchard. By Anna Mouton.

Plum marbling, caused by plum viroid 1, is named for the irregular skin discoloration associated with infection of Japanese plums. Fruit may also be small and hard with an irregular surface or have corky flesh.

Detecting plum marbling is difficult because infected trees may not develop abnormal fruit every year and appear otherwise normal. In addition, young trees without fruit don’t show signs. Laboratory tests can pick up viroids, but these assays are expensive, and results are not immediately available.

However, a Hortgro-funded project led by Prof. Adriaan van Niekerk, director of the Centre for Geographical Analysis in the Department of Geography and Environmental Studies at Stellenbosch University, demonstrated the potential for cost-effective screening in the orchard.

The research

Van Niekerk sampled leaves from Japanese plum trees previously inoculated with plum viroid 1 for research conducted by Dr Rachelle Bester and Prof. Hano Maree of Citrus Research International, seconded to the Department of Genetics at Stellenbosch University. Their work was featured in a previous Fresh Quarterly.

Leaves from infected and uninfected trees were imaged using different electromagnetic frequency bands. The electromagnetic spectrum includes electromagnetic radiation of various frequencies or wavelengths, from radio waves at low frequencies to X-rays and gamma-rays at high frequencies.

The plum leaves were imaged using the visible, near-infrared, and short-wave infrared bands, and a subset of the images was used to train a machine-learning model.

Visible and near-infrared images did not help detect infection, but the model could discriminate leaves from infected and uninfected trees with 94% accuracy based on two bands of the short-wave infrared spectrum.

The limitations

This project was a pilot study to test whether machine learning could detect plum marbling based on leaf imaging. Van Niekerk is quick to point out that the model may be spotting stressed trees as opposed to homing in on plum marbling specifically. However, plum marbling is not associated with reduced vegetative growth.

In addition, the pilot study was restricted to a small sample of leaves from a small number of experimental trees. The results must be replicated in commercial orchards under different conditions and at different times of year.

The potential

Even if the machine-learning model isn’t specifically detecting trees infected with plum viroid 1, it could be the basis of a rapid, low-cost technology to screen large tree numbers. Those trees flagged as positive by the model could then be tested with laboratory assays.

A screening method requiring only a few short-wave infrared bands lends itself to a relatively inexpensive handheld device. The device wouldn’t need an imaging sensor as it would take point measurements.

As the model only relies on two bands, the device collects far less data than, for example, a camera, speeding up measurement and processing. It should be possible to obtain results within 15 seconds or less.

Although Van Niekerk is cautious about reading too much into the findings of this small exploratory project, he hopes that further research will build on these results. Work by other scientists suggests it would be worthwhile pursuing.