Advantages of multi-region kriging over bi-region techniques for computed tomography-scan segmentation

Quantifying the structure of soil is essential for developing effective soil management for farming and environmental conservation efforts. One approach to quantify soil structure is to scan intact soil cores by X-ray computed tomography (CT), which allows using computer vision algorithms to identify internal components within the soil. One commonly used approach is the colour-based segmentation of CT-scan soil images into two regions - matter and void - for the purpose of determining the soil porosity. A key problem with this approach is that soil CT images tend to be rather complicated, and thus this type of bi-region segmentation is a non-trivial problem, with algorithms following this type of bi-region approach typically performing unreliability across a variety of image sets. In this work, a technique is proposed that identifies an optimal number of regions present in the soil, rather than just two. It is claimed that this more sophisticated representation of soil structure leads to a more accurate representation than traditional bi-region segmentation; however, it is reducible to a bi-region segmentation yielding the required estimation of porosity with more accuracy and robustness than traditional methods. It is also proposed that segmentation is performed using a multi-region kriging algorithm, which establishes relationships between distance and regions that allows the segmentation to overcome many of the artefacts and noise issues associated with CT scanning. Our experiments focused on layer-by-layer segmentation and results demonstrated that the proposed approach produced segmentations consistent across a variety of scanned cores and were visually more correct than current state-of-the-art bi-region techniques.