Deep Learning-Based Pore Segmentation of Thin Rock Sections for Aquifer Characterization Using Color Space Reduction

The conventional way of obtaining hydraulic parameters of aquifers is through aquifer tests; by performing the interpretation of data whose acquisition requires a fairly complex logistics in terms of equipment and personnel A completely different approach that employs the processing and analysis of digital images of thin rock sample micrographs, has proven to be a promising alternative approach for obtaining estimates for hydraulics parameters being simpler and cheaper than the traditional methodology. This approach involves sampling of rocks followed by thinning out and imaging of rock samples, image segmentation, three-dimensional reconstruction and flow simulation. This methodology has been applied to the analysis of several aquifers. Here, we propose a method for segmenting and classifying thin rock sections into pores and background (comprised of rocks and cement), by using a procedure that performs color space reduction on the CIELab space and pixel classification using convolutional networks. The high values of accuracy, specificity, sensitivity and Dice coefficient obtained in our experiments indicate that this methodology can be safely applied in our hydraulic parameters estimation system. The best results were obtained by the SegNet network with the color space redaction, the values of 0.9609 ± 0.0339 of accuracy, 0.9609 ± 0.0339 of specificity, 0.9707 ± 0.0339 of sensitivity and Dice coefficient equivalent to 0.8873 ± 0.1382.