Hyperspectral Imaging in Root Phenotyping

Hyperspectral imaging techniques have been widely applied in fields like remote sensing and the food industry, however, its application in the context of root phenotyping was only initiated recently. Due to the much higher information content of hyperspectral images compared to RGB images, a throughout assessment of its’ potential as a novel tool to advance root research and subsequently our understanding of root systems is key. In this study, we focused on maximizing the contrast between root and soil background (root segmentation) and to study differences between individual root segments using hyperspectral images (900-1700 nm). Principal component analysis (on 222 wave bands) was used for initial root and background labeling, deep convolution networks (DCNN) for pseudo labeling. High accuracy classification of root and soil background was based on the DCNN-labeled image; important wavelengths leading to the highest classification accuracy were selected by a combination of random forest and support vector machine approaches. Further clustering algorithms were applied for optimal cluster number determination and cluster separation to explore the composition pattern within roots. The potential of hyperspectral imaging for automatic root tracking and root phenotyping will be discussed.