Impact of different segmentation methods on digital rock analysis results, including trainable deep learning-based segmentation
Christian Hinz , Andreas Griesser , Jens-Oliver Schwarz , Christian Wagner
Digital rock analysis (DRA) provides crucial input data for reservoir models. DRA is less expensive and much faster than the conventional determination of these parameters in the laboratory, cutting experimental time from one year down to two weeks. Among other factors, the current economic situation is the driving force behind the digitalization of laboratory-based rock analysis. The DRA workflow consists of three basics steps which depend on each other and are influenced by all preceding steps: image acquisition, image processing, and property predictions. Among them, the image processing step has the largest impact on the final DRA results.
The DRA software GeoDict® is used to perform the image processing and property prediction steps. We compare different image processing and segmentation methods and their subsequent influence on the DRA results. For this, we use global thresholding with user-defined thresholds and automatic OTSU-based thresholds, a watershed-based multiphase segmentation, and a trainable deep learning-based segmentation.
These different segmentations result in 3D voxel-grid structure models of digital rocks. For each of these structures, we perform digital Routine Core Analysis (RCAL) with GeoDict®, including determination of porosity, pore throat size distribution, and permeability. Finally, we compare the DRA results for a digital reservoir rock sample [1] with reference results [2].
