Evaluating the potential of hyperspectral imaging for remote detection of plum marbling viroid infested nursery and young orchard trees

The objectives of the study were:

1. Collect suitable in situ data (leaves) representing infected and healthy trees;
2. Extract spectral properties from leaves;
3. Employ feature selection and machine learning to build models for detecting infected leaves;
4. Assess the models and identify the bands that are most important; and
5. Report and evaluate the findings.

1. Collect a large (n=60) sample of leaves representing infected (n=30) and healthy (n=30) trees;
2. Split the sample into two sets, one (n=40) for model building/training and one (n=20) for model assessment/testing;
3. Image (scan) each leaf at the Central Analytical Facility (CAF);
4. Perform image segmentation and classification to extract sample objects (regions) on each image;
5. Experiment with different feature (band) selection methods to reduce the dimensionality of the dataset;
6. Employ the training objects to build machine learning (e.g. random forest, support vector machines, neural networks, boosting) models;
7. Assess the accuracy of the models;

The visible and near-infrared region (VNIR) of the electromagnetic spectrum was not sensitive to the viroid. However, using the short-wave infrared (SWIR) images, a detection accuracy of >90% was achieved. Using only two carefully selected bands (B1966 and B2511) a classification accuracy of 94% was achieved.

It is not clear whether the SWIR is sensitive to the viroid itself or the symptomatic effect that the viroid has on the harvested leaves. It is also not clear whether the model will work in operational orchards at different times of the season, as the relatively small sample size (30 leaves per class) collected from an experimental orchard on one day in December is likely not representative.

We did not think it would work, but it did! We strongly encourage the models to be tested on new data collected in operational orchards at different times of the season.