EGU23-2995
https://doi.org/10.5194/egusphere-egu23-2995
EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Machine learning-enabled quantification of segmentation uncertainty of X-Ray CT image-based analysis for vegetated soils

Zhenliang Jiang, Anthony Leung, and Jianbin Liu
Zhenliang Jiang et al.
  • The Hong Kong University of Science and Technology, Civil and Environmental Engineering, Hong Kong (zhenliang.jiang@connect.ust.hk)

Segmentation of X-ray computed tomography (CT) images of four-phase unsaturated rooted soils is challenging yet a crucial step for conducting the subsequent image-based analysis (IBA) for various properties, for instance, volume fraction of each phase, morphologies of pore and roots, pore fluid distributions and some engineering properties such as hydraulic conductivity. The accuracy and efficiency of phase segmentation have been widely investigated, but the segmentation uncertainty (SU), which is a measure of reproductivity or reliability, on IBA and how the uncertainty propagates at different stages of IBA have rarely been studied for rooted soils. In this study, we developed a machine learning (ML)-based technique, called the percentile-based segmentation method, to perform phase segmentation of CT images and quantify the uncertainty and propagation of phase segmentation at different stages of IBA. Two indicators were used: relative value (RE), which has been used in the literature and SU magnification factor (SU-MF), which is newly proposed in this study. X-ray CT images of soil samples with different particle sizes and cultivated with different plant species were taken by a micro-X-ray CT scanner. The images were then segmented using the proposed ML method. In the presentation, a detailed case study and sensitive analysis (e.g., different number of phases, plant species, sampling resolution, and simulation methods) will be presented. We will show that root volume is susceptible to SU yet has a marginal influence on CT-IBA as its fraction is relatively small compared to other phases. However, the volume of soil grains is less SU-sensitive, which could lead to a significant change in the IBA. Root architectures could substantially influence the SU. Increasing the segmentation percentile improves the reliability, but the accuracy reduces at the same time. Moreover, we will show that the newly proposed indicator, SU-MF, can reasonably reflect the SU propagation behaviour. Therefore, SU could significantly impact the CT-IBA of rooted soils, and SU propagation is phase- and parameter-dependent. The explored quantification and propagation of SU provide novel and practical perspectives for increasing the measurement reliability of the X-ray CT-IBA of rooted soils.

How to cite: Jiang, Z., Leung, A., and Liu, J.: Machine learning-enabled quantification of segmentation uncertainty of X-Ray CT image-based analysis for vegetated soils, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-2995, https://doi.org/10.5194/egusphere-egu23-2995, 2023.