The binary segmented volumes were labelled and an intensity radius
threshold of 15 to 100 applied followed by a small spot filter to remove
round segmented objects with a radius of <50 pixels (134 μm)
for visual clarity of the fault plane (Figure 3a). The local aperture at
each voxel of the segmented voids was computed from the diameter of the
largest ball containing the voxel and entirely inscribed within the
object (Hildebrand and Ruegsegger, 1997). Even with the segmentation
method described, there is still significant under-sampling of the void
population, particularly at the narrower end of the aperture range
(Figure 3b). Further work in this area is required and would benefit
from machine learning approaches (Andrew, 2018).
We present the data according to a co-ordinate system (\(x,y,z\)) where\(z\) is the vertical axis, which is parallel to the loading direction
and corresponds to the direction of axial stress (\(\sigma_{1}\)). The
other two (\(x\) and \(y\)) are the horizontal axes, which are
perpendicular to the loading direction and correspond to the confining
stress (\(\sigma_{2}=\sigma_{3}\)) with their directions arbitrarily
assigned but consistent between the two experiments. Void orientations
are given in terms of their dip \(\phi\) (deviation from horizontal) and
strike \(\theta\) (deviation from \(y\)).