Localized Smart Interpretation - a data driven semi-automatic geological modelling method

Abstract

INTRODUCTIONWhen setting up geological models today, an enormous amount of information may be available, e.g. airborne electromagnetic data (AEM-data), borehole data, radiometric data, seismic data, geological information, etc. However, it is only possible to incorporate a limited amount of the accessible information in the geological modelling process due to the very time demanding task of manual interpretation work. We suggest a methodology to target this problem, which infers a linear model that describes the relation between geological interpretation and the information available to the geologist making the interpretation. Once such a model has been inferred, it can be used, in a semi-automatic way, to perform new (i.e., predict) geological interpretation based on and consistent with the interpretation made by the geological expert. METHODIdeally the method constructed will build a statistical model f (d,M), describing the relation between what the geologist interprets, d=(d 1, d 2, …, d M ) T , and the quantifiable information (i.e., attributes) available to the geologist, M = (m 1, m 2, …, m N ). Each vector m i contains the i'th attribute, which can be any available information that can be quantified, such as geophysical data, the result of geophysical inversion, elevation maps etc. The vector d contains actual interpretations, such as for example the depth to the base of a ground water reservoir. First, the statistical model f(d,M) should be inferred by examining sets of actual interpretations made by a geological expert d and the attributes used to perform the interpretation M. This makes it possible to formulate a probability distribution f(d,M) that describes the relation between attributes and the geological interpretation. As the geological expert proceeds interpreting, the number of interpreted data points from which the statistical model is inferred, increases and, consequently, the accuracy of the statistical model increases. When a model f(d,M) has been successfully inferred, it is possible to predict how the geological expert would perform an interpretation d pred given some external information M ext through f(d pred |M ext ).In this paper we assume a linear relation between d pred and M ext such that:and seek to find the d pred , which gives the highest probability value of the conditional probability distribution (d pred |M ext ). In equation (1) SUMMARYLocalised Smart interpretation (LSI) is a method that infers a statistical model, which describes a relation between the knowledge of a geologist (as quantified by geological interpretation) and the available information (such as geophysical data, well log data, etc.) that a geologist uses when he/she interprets. This model is then used to perform semi-automatic geological interpretation wherever the same kinds of attributes, as used for the initial interpretation, are available. The statistical model is inferred using a combination of a regularized least squares method and cross validation. In this study, we demonstrate the applicability of...

Semiautomatic mapping of permafrost in the Yukon Flats, Alaska

“…First, the location frozen soil is found by using unsupervised K-means clustering [see, e.g., Bishop, 2006;Everitt et al, 2001]. Then, the thickness of the permafrost is estimated by using supervised regression, also known as Smart Interpretation (SI) [Gulbrandsen et al, 2015].…”

“…Smart Interpretation is based on a linear regression technique and is a method developed for geological layer modeling [ Gulbrandsen et al , ]. Here we make use of this method extended from linear to second‐order polynomial regression.…”