Analytical techniques for plant phenomics - procedures to evaluate observable plant characteristics - are a crucial part of any approach to achieving desirable agronomic and biological traits. Traditional phenotyping techniques have been done manually, and plant features have been measured through time consuming efforts that are prone to errors and inaccuracies. Advances in sensor technologies have paved the way for faster and more efficient phenotyping, and new methods have been adapted from other scientific disciplines such as high-resolution X-Ray three-dimensional (3D) computed tomography (CT).
A crucial step in the analysis of CT root phenotyping data is segmentation: the identification and classification of a scan’s data points (3D voxels) as “root” or “non-root”. Unlike roots in transparent mediums, roots in non-transparent mediums are particularly difficult to segment from their surrounding materials (water, air, stones, debris, etc.) as root and non-root voxels have overlapping CT values.
In this work we discuss a convolutional neural network (CNN) approach for the segmentation of 3D CT data that addresses the challenge of non-destructive root segmentation in high-throughput plant imaging systems. Our proposed solution uses an encoder-decoder neural network architecture, and is primarily implemented using the MapReduce paradigm, which segments CT plant scans by processing subvolumes: full volumes are recursively split so that we can segment many subvolumes rather than a single monolithic volume.
We present some initial results of the analyses of several different crop species such as cassava, potato, bean, and maize, as well as discuss the challenges of obtaining and developing accurate training data for our network.