HR: 1340h
AN: MR33A-0143    [Abstracts]
TI: Image Analysis And Pattern Recognition For Porosity Estimation From Thin Sections
AU: * Richa, R
EM: richa7@stanford.edu
AF: Stanford Rock Physics Laboratory, Department of Geophysics, Stanford, CA 94305
AU: Mukerji, T
EM: mukerji@pangea.stanford.edu
AF: Stanford Rock Physics Laboratory, Department of Geophysics, Stanford, CA 94305
AU: Keehm, Y
EM: ykeehm@gmail.com
AF: Stanford Rock Physics Laboratory, Department of Geophysics, Stanford, CA 94305
AU: Mavko, G
EM: mavko@stanford.edu
AF: Stanford Rock Physics Laboratory, Department of Geophysics, Stanford, CA 94305
AB: Estimating porosity from thin sections is one of the key steps in many different rock physics and petrologic analyses. The estimated porosity is a critical input for computing transport properties of rocks from thin sections. The porosity estimate and its uncertainty depend, amongst other things, on the image analyses techniques used. In this poster, we present the results of exploring different image analysis algorithms for estimating porosity from thin section. The general methodology for calculating porosity from thin section involves conversion of a colored image to a binary image. The average of the binary image gives us the porosity. As most thin sections use blue epoxy impregnation, the conversion to binary image requires computationally identifying pixels that are blue. One of the challenges is to capture the variability of the color value, all of which are nominally blue. We compared two different color spaces, RGB and HSV color space, which can be used to specify the blue color. Two different approaches were tried for converting the colored image in different color spaces to a binary image. The first approach involved using thresholds for conversion. A single dimension threshold based on the intensity histogram as well as a multiple dimension threshold based on (RGB) and (HSV) pixel values were explored. In general, multiple dimension thresholding in HSV space gave better results but the choice of threshold is subjective. The second approach involved statistical pattern recognition and classifying of grains and pores. We tested both discriminant analysis and neural network classification. A training data was defined using different groups of pixels from selected pore and grain regions of the thin section. The trivariate training data consists of the range of HSV values for each group (grain or pore). A misclassification error was calculated for the different classification algorithms as the fraction of the observations in the training data that are misclassified. The quadratic discriminant method seems to give the best results and least error. The misclassification error was about 12.6%. The neural network classification depends upon the residual error to be achieved. Different instantiations of the neural network give slightly different porosities for a specified residual. The variance of the estimated porosities decreases with decrease in residual error, but the trade-off is an increased bias in the estimate. This is the expected bias-variance trade-off behavior. In general, the HSV color space gave better results in specifying the blue color than the RGB color space. The multiple dimensional threshold works better than the single threshold. It also proved to be a simpler, though subjective, method than statistical discriminant analysis. Though discriminant analysis gave good results, it involved preparation of training data from the thin section, which adds to processing time. Nevertheless, it can be useful for identifying different types of grains and hence may be useful for computational methods that require not only classification of pore space but also different grain types.
DE: 0515 Cellular automata
DE: 0540 Image processing
DE: 0555 Neural networks, fuzzy logic, machine learning
DE: 5100 PHYSICAL PROPERTIES OF ROCKS
DE: 5114 Permeability and porosity
SC: Mineral and Rock Physics [MR]
MN: Fall Meeting 2005