A Simulation Study of Segmentation Methods On the Soil Aggregate Microtomographic Images.

Advances in X-ray microtomography open up a new way for examining the internal structures of soil aggregates in 3D space with a resolution of only several microns. However, processing of X-ray soil aggregate images in order to obtain reliable representations of pore geometries within aggregate pore remain to be established. Multiple approaches to the segmentation algorithms used to best separate grayscale images into pores and solid material. Additionally, ground-truth information with known pore geometries are needed to provide specific information that identifies the most accurate segmentation of microtomographic images. The study objectives are (i) to compare performance of several commonly used segmentation methods, namely, entropy based method, two clustering methods and indicator kriging method for soil images with various porosities as scenarios of the ground-truth standards; and (ii) to evaluate segmentation performance criterion, i.e., misclassification error (ME) and region non-uniformity measure (NU), the criterion that does/does not depend on the ground-truth image, in choosing the segmentation method optimal in terms of representation of soil pore characteristics. Simulations of the soil aggregate images were conducted on pore and solid spaces respectively, which are served as served as ground-truth images. Majority filtering was applied to smooth the resulting images. The results showed that through there is no any single method found to preserve pore characteristics in all the cases, indicator kriging method yielded most similar segmented images to the ground-truth images in most of the porosity cases. When the ground-truth information is available, misclassification error is not always better than region non-uniformity measure in assessing different segmentation results. Pore morphological characteristics, such as number of connected pores and pore boundary pixels, are very sensitive to small local variations (i.e., single isolated pore pixels). However, NU could be used as a criterion to select best segmented images with tolerant variation among pore characteristics comparing to the ground-truth images.