Automated Segmentation of Soils Using X-ray Tomography

Abstract:

X-ray tomography (CT) has long been a useful tool for three-dimensional imaging of compositionally heterogeneous objects. In the environmental sciences, CT is an efficient tool for the nondestructive inspection of sediment and soil cores. However, in order to extract parameters describing such properties as pore space and solid-phase distribution, the imaged volume must be segmented according to relevant categories. When done manually by image inspection, segmentation produces results that are often inconsistent, and applying the method over multiple images may be impractical. Modern machine learning techniques have been shown to be more accurate than humans at some vision tasks in fields of histology and remote sensing, and those techniques may be useful for environmental samples. We present a technique using deep learning to categorize a tomographic volume into solid and pore regions, while also identifying morphologically similar solid-phase regions within the imaged object. Finally, we show how the composition of these characteristic solid constituents may be estimated by propagating two dimensional XRF data through the segmented volume. This research was funded by the Chemical Imaging Initiative under the Laboratory Directed Research and Development Program at PNNL.