Hyperspectral imagery combined with machine learning to differentiate genetically modified (GM) and non-GM canola

Brianna Rose West

Canola, also known as rapeseed (Brassica napus L.), is an oilseed that produces a healthy food-grade oil, canola meal by-product, and biofuel. It is the fourth most grown grain in Australia. Genetically modified (GM) canola currently represents approximately twenty percent of national canola production; hence, with clashing public and industry perceptions of genetically modified organisms (GMOs), transparency and traceability must be enabled throughout the supply chain to protect markets and relationships with consumers. GM canola must not cross-contaminate non-GM canola as our largest export market, Europe, has extremely strict protocols on GMOs. GM and non-GM canola cannot be differentiated by the human eye, with polymerase chain reaction (PCR) methods currently the main alternative, which is expensive and time-consuming. This thesis evaluates the potential to differentiate GM from non-GM canola using the novel, rapid, and non-destructive technique of hyperspectral imaging combined with machine learning. Hyperspectral imagery captures and processes wavelengths beyond simply red, green, and blue. It has a pre-existing multitude of uses including the characterisation and variety identification of other grains. In this study 500 images each of non-GM and GM canola seeds were captured. Seeds were placed on a black background with two lights sources. Images were captured from the 400nm to 1000nm wavelengths, a total of 80 bands, at a 25-millisecond exposure time. These images were run through a convolutional neural network in Keras for analysis. The high dynamic range and raw files were combined into a NumPy file for the hyperspectral image generator. Contrary to expectations, however, the models using the bitmap image files performed similarly to the models receiving the hyperspectral images. Regardless, both produced high validation accuracies around 90%, indicating a detectable phenotypical difference between the two, and further studies could lead to the development of a new approach to GM canola detection.

Hyperspectral imagery combined with machine learning to differentiate genetically modified (GM) and non-GM canola

Files and links (1)

Abstract

Details

Metrics