LITHOFACIES CLASSIFICATION SYSTEM AND METHOD

FIELD

The present invention relates generally to characterization of geologic data and more particularly to automated interpretation of borehole log data.

BACKGROUND

[0002] Borehole data are collected via a number of techniques including, for example, resistivity/conductivity measurements, ultrasound, NMR, and radiation scattering. Conventionally, borehole data is analyzed by human interpreters in order to characterize a subsurface geological formation to allow decisions to be made regarding potential of the well or to determine information about the nature of the surrounding geologic area. Borehole data of this type may be used to replace or supplement the collection of cores for direct inspection.

in an aspect of the invention, there is provided a method of analyzing characteristics of a subterranean formation including applying a wavelet transform to measured data or to derived data obtained from the measured data, or to a portion thereof, the measured data obtained by measuring at least one physical property at each of a plurality of positions in a borehole that penetrates a region of the subterranean formation; selecting one or more scales for analyzing wavelet coefficients at each scale; performing a segmentation of the measured or derived data; calculating a distribution of wavelet coefficients for each scale, for one or more spatial directions of wavelet transform application, for the wavelet transform of the measured or derived data and for each segment that belongs to the portion on which the wavelet transform has been calculated; and for each segment, analyzing a distribution of the measured or derived data and the distribution of the wavelet coefficients for each scale in terms of their overlap with corresponding distributions of training samples to produce a geological interpretation of the subterranean formation.

[0004] In another aspect of the invention, there is provided a computer product having machine executable instructions, the instructions being executable by a machine to perform a method for

analyzing characteristics of a subterranean formation, the method including calculating a distribution of wavelet coefficients for each scale, for one or more spatial directions of wavelet transform application, for the wavelet transform of measured or derived segmented data and for each segment, the measured data obtained by measuring at least one physical property at each of a plurality of positions in a borehole that penetrates a region of the subterranean formation and the derived data obtained from the measured data; and for each segment, analyzing the distribution of the measured or derived data and a distribution of the wavelet coefficients for each scale in terms of their overlap with corresponding training sample distributions. |0005| fliese and other objects, features, and characteristics of the present invention, as well as the methods of operation and functions of the related elements of structure and the combination of parts and economies of manufacture, will become more apparent upon consideration of the following description and the appended claims with reference to the accompanying drawings, all of which form a part of this specification, wherein like reference numerals designate corresponding parts in the various Figures. It is to be expressly understood, however, that the drawings arc for the purpose of illustration and description only and are not intended as a definition of the limits of the invention. As used in the specification and in the claims, the singular form of "a", ^s 'an", and ''the ^* ' include plural referents unless the context clearly dictates otherwise.

BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 generally shows a method of analyzing characteristics of subterranean formation in accordance with an embodiment of the invention;

[0007] Figure 2 illustrates an example of a conventional borehole probe:

[0008 j Figures 3 a and 3b are side and front views respectively of a sensor portion of a probe as shown in Figure 2;

[0009] Figure 4 is a schematic illustration of a probe in communication with a computer in accordance with an embodiment of the invention;

[0010] Figure 5a shows gamma ray log (column 1), manually interpreted geological fades

(column 2), normalized resistivity traces (column 4), boundaries from a segmentation routine segwavc used for a test of the classification method (column 5), resistivity distribution (column

6), wavelet coefficient distributions for 1/16 foot scale (columns 7-9) and for 1/8 foot scale

(columns 10-12), and intemiediate classification results (columns 13-23) as a function of depth in the borehole (column 3), in accordance with an embodiment of the invention;

I] Figure 5b shows gamma ray log (column 1), manually interpreted geological facies (column 2), normalized resistivity traces (column 4), boundaries from a segmentation routine segwave used for a test of the classification method (column 5), resistivity distribution (column 6). wavelet coefficient distribution for 1/16 fool scale (column 7} and for 1/8 foot scale (column 8), and intermediate classification results (columns 9-19) as a function of depth in the borehole (column 3), in accordance with an embodiment of the invention; l\ Figure 6 shows a method for performing feature extraction and intermediate classification for each segment in accordance with an embodiment of the invention;

Figures 7a~b show classes for feature fusion and final classification in a deepwater environment with seven facies m ^" . accordance with an embodiment of the invention;

Figure 8 shows a method for performing feature fusion and final classification in accordance with an embodiment of the invention;

Figure 9 shows a method of feature fusion and final classification in accordance with an embodiment of the invention; and

Figure 10 shows a rule forming method in accordance with an embodiment of the

DETAILED DESCRIPTION

[0017] Borehole data are analyzed to determine a geological profile of a subterranean formation. Borehole data are provided by measuring and recording on well-logs physical quantities down the borehole. These physical quantities provide information about the different textures of the materials (facies) that constitute the geological profile of the subterranean formation. Analysis of borehole data generally requires a lot of human activity and may be very time consuming. Therefore, it is desirable to at least semi-automate this analysis. Borehole image analysis has been performed by means of Multi-Resolution Graph Based Clustering {see U.S. Pat. No. 6,295,504 to Ye and Rabiller). Besides the difficulties of adequate matching of automatic versus human interpretation that clustering introduces, there is also a major challenge in texture analysis in general to find a limited number of parameters to model textures that would allow

adequate discrimination and synthesis of visually similar samples. Current theories of texture discrimination incorporate the notion that two textures arc often difficult to discriminate (i.e. are similar) when they produce a similar distribution of responses in a bank of (orientation and spatial-frequency selective) linear filters (Hceger and Bergen, "'Pyramid Based Texture Analysis/Synthesis' ^" , Proc. ACM SIGGRAPH. August 1995; Liu and Wang, "Texture Classification Using Spectral Histograms ^5' , Image Processing, IKEE Transactions. Vol. 12. Issue 6, June 2003). In an aspect of the invention presented hereinafter, facies classification from well- logs makes use of a wavelet transform of well-log signals and is based on the similarity of distributions of wavelet transform coefficients of segmented data to the corresponding distributions of the facies training samples, whereby the training samples can be provided by an expert and unaltered throughout the application, In this way, classification is done strictly in terms of expert recognized facies and the difficulties of matching the results of an automatic interpretation to those of a human interpreter are significantly diminished.

[0018] The wavelet transform is a linear transform which convolves the original signal with a family of functions or wavelets of zero mean that satisfy a known admissibility condition (see Stephane MaI iat, ''A wavelet Four of Signal Processing", p. 82, Academic Press, 2 ^!ϊd Edition, 1999). The functions belonging to a wavelet family arc of a mutually similar shape but translated and/or dilated in the time (or depth) variable. The result of a wavelet transform is termed by wavelet coefficients which are indexed by position (at which the wavelet is centered) and scale (that characterizes the width of the wavelet), λ family of wavelets is obtained from the mother wavelet Vft) by shifting it by x and scaling it by a, as shown in equation (1):

[0019] Once a family of wavelets is chosen, then the continuous wavelet transform of a function βl) is defined as (see equation (2)): where W/(x,a) are the wavelet coefficients.

[0020] Most applications of wavelet families exploit their ability to efficiently approximate particular classes of functions with few non-zero wavelet coefficients (see "A wavelet Tour of Signal Processing" at p. 241), with coefficients of a small magnitude at finest scales generally

corresponding to smooth parts of the signal. Among other applications, wavelet transform is useful in characterizing abrupt signal changes (through coefficients of a large locally maximal magnitude around locations of such changes). In an embodiment of the invention, the wavelet transform applied to the original signal is discrete. However, it will be appreciated that a continuous wavelet transform may also be applied to the measured signals, A discrete wavelet transform differs from a continuous wavelet transform in that the convolution is not performed as an integral but as a sum over discrete points x, separated by the given scale a.

[0021] In an embodiment of the invention described herein, data segments are classified in terms of a defined similarity to strictly the fades training samples provided by an expert. In this way the difficulties of matching the results of an automatic interpretation to that of a human interpreter are significantly diminished. The faciεs that is studied may be as small as single foot lithofacies segments (as long as they have relatively well-defined measurable rock properties that set them apart from each other in an expert's view) or larger, depofacies segments of lithofacies aggregates.

[0022] Figure 1 generally shows a method of analyzing characteristics of subterranean formation in accordance with an embodiment of the invention. The method starts at procedure 10 where one or more physical quantities are measured at each of a plurality of positions in a borehole that penetrates a region of the subterranean formation. The physical quantities may be measured with the use of a probe such as that shown in Figure 2. f0023] The probe may be a multi-trace or multi-pad measurement probe. For example, Figure 2 illustrates a probe 100 for use in borehole characterization that includes a generally elongated shaft 120 having at one end a number of outwardly extending members 140, The outwardly extending members 140 may each include a sensor 160 (shown in more detail in Figures 3a and 3b) for interrogating a region 180 of a borehole. The illustrated sensor 160 includes a plurality of pairs of sensors 200 for monitoring a change in voltage that is developed across a portion 220 of region 180 when a current is injected through the current injector electrodes 240,

[0024] Though the sensor 160 shown in Figures 3a and 3b is constructed and arranged for measuring electrical characteristics of the borehole material, this is for illustrative purposes only and a wide variety of sensors may be employed in various embodiments of the present invention, In particular, it is envisioned that measurements of resistivity, ultrasound or other sonic waves, complex electrical impedance, neutron reflectivity or transmissivity, nuclear magnetic resonance, video imaging and/or spectrometry may be employed. Furthermore,

natural gamma ray emission may be measured. Consistent with this, the sensor 160 may be selected as appropriate for the measurement to be made, and may include, by way of non- limiting example, electrical sources and detectors, radiation sources and detectors, and acoustic transducers. As will be appreciated, it may be useful to include multiple types of sensors on a single probe 100 and various combinations may be usefully employed in this manner.

[0025] In use, the probe 100 is generally lowered into the borehole to be characterized. Upon reaching an appropriate depth, which may be the bottom of the hole, or a selected intermediate depth, the probe is retrieved and measurements are taken as the probe rises through the material, In many cases, the probe 100 will have four sensors 160 so that the hole may be characterized in each of four directions, which may be denominated 0, 90, 180 and 270, for example.

[0026] The operation of the probe 100 will be described using the example of the electrical sensors illustrated in Figures 3a and 3b. While resistivity measurements are described here in detail, the particulars will vary in accordance with the particular sensor employed in a given implementation. The probe 100 takes measurements by transmitting, from the electrodes 240, an electrical signal, which may be a DC or AC or otherwise varying current I, at least partially into a region 180 of material of the borehole. As will be appreciated, the penetration of the signal will depend in part on the nature of the measurement. A reflectivity measurement involves nearly no depth of penetration Into the material under interrogation, while a resistivity measurement requires the signal to propagate some distance through the material However, each of the foregoing degrees of penetration should be understood to be within the scope of the present invention,

[0027] In the special case of resistivity measurement in oil-based mud, the signal propagates through the material and is received by the sensors 200. The sensors 200 measure a change in voltage δV across a portion 220 of the region 180 that is indicative of a characteristic (in this case, resistivity) of the material that Is present in the subterranean formation being characterized. In an embodiment, the resistivity R of the material in the portion 220 may be calculated based on the formula: R = k(δV/Ij, where 1 is the current generated by the electrodes 240 and k is a geometrical factor that may be predetermined based on a geometry of the electrodes 240 and the sensors 200. As will be appreciated, this formula depends on the material being relatively uniform within the distance between the electrodes 240. For this reason, the scaling factor k may be implemented as a dynamic scaling, which may provide additional detail.

100281 As shown in Figure 4, borehole data collected by the probe are generally stored in a local memory device as in memorized logging-while-drilling tools or relayed via a wire, though the connection may be made wireless, to a computer 250 that may be, for example, located at a drilling facility where the data may be received via a bus 260 of the computer 250, which may bε of any suitable type, and stored, for example, on a computer readable storage device 270 such as a hard disk, optical disk, flash memory, temporary RAM storage or other media for processing with a processor 280 of the computer 250.

[0029| As shown in Figures 5a-h, column 4 under the heading "ALL TRACES", the borehole data may be converted to a graphical form, as image data 520 for display and/or further processing. The graphical data may be stored as full-color data (for example, 24-bit JPEG data) or may be single color or gray-scale data. Image data 520 of Figures 5a-b represent resistivity traces obtained at each of a plurality of positions in a borehole with the OBMI™ tool from Sehlumberger, Column 3 - heading "DEPTH FEET" - of Figures 5a-h shows the corresponding depth portion 515 (from x450 to x520 feet) of the borehole for which image data 520 are obtained, These resistivity traces have been normalized. A color is assigned to a resistivity value measured by the sensor to get a picture in resistivity. Image data 520 include four stripes 521a-d running down column 4 in Figures 5a-b. These four stripes represent data from the four sensor pads obtained with the probe 100, with each sensor pad providing five data traces. The darker areas of lhe image data 520 represent relatively conductive regions and the lighter areas represent relatively resistive regions. In clastic rocks, the darker, more conductive regions may be interpreted as shales, while the lighter, more resistive regions may be sandy regions. In carbonate rocks and in mixed lithoiogy, the texture of resistivity images and physical quantities other than an average resistivity become more important for distinguishing liihofacies as well as depofaeies.

Distributions of the physical properties measured by the probe 100 may be provided as a function of depth. For example, from the resistivity measurements (e.g. image data 520 in column 4 of Figures 5a-b), a resistivity distribution or histogram may be obtained as a function of bore-hole depth. The resistivity distribution 530 calculated for 1 foot depth intervals is shown in column 6 of Figures 5a-b. This resistivity histogram or distribution 530 captures the brightness of the images for the four traces 521a-d. Similar distributions can be obtained for other measured or calculated quantities for depth intervals of appropriate length.

[003Ij Referring back to Figure 1, once one or more physical properties are measured at each of a plurality of positions in a borehole thai penetrates a region of the subterranean formation (and optionally other derived quantities are calculated from the measured data), the method proceeds to procedure 20 where a wavelet transform is applied to the measured or derived data or to a portion thereof, In the case of resistivity images, the wavelet transform may be applied in the direction along the borehole axis, i.e. on each trace of each pad 521a-d independently, or on a data sequence formed from the interpolated data that lies on a curve (or a portion thereof corresponding to a tool pad image, if the tool has pads) constructed on a borehole view of the image (i.e. with each trace arranged in the order of spatial appearance around the borehole) as the cross-section of the borehole wall and a bedding plane or another plane of interest, In the ease when the borehole segment is a perfect cylinder, this curve is generally a sinusoid. The calculation of a wavelet transform along two directions is locally similar to a two-dimensional wavelet transform, but the direction of application that is not along the borehole axis can vary slowly to stay parallel to bedding.

[0032] In an embodiment of the invention, the wavelet transform is a discrete wavelet transform, though it is contemplated that a continuous transform could also be applied in another embodiment of the invention. The wavelet transform may be orthogonal. For example, a Daubechies wavelet of order 2 as a generating wavelet function may be used. It will be appreciated that other wavelet functions may be used in other embodiments of the invention.

[0033] At procedure 30 of Figure 1, a sequence of scales for the analysis of wavelet coefficients is determined in order to capture sharp changes in the signal at various scales. For example, wavelet coefficients may be obtained for three different scales: a first scale, a second scale twice as large as the first scale and a third scale twice as large as the second scale.

|0034] It is desirable that the selected scales of procedure 30 be comparable to the length scales of characteristic intrafacies variations of the measured data since the wavelet coefficients at such scales would be most closely related to the original signal changes occurring at that scale and thus related to a facies signature insofar as this signature is present in the measured data, In an embodiment, the length of the scales may be visually selected by looking at the fine-scale variations of the image data 520 that can be geologically interpreted (i.e. without considering the noise-related variations). That is, a few scales are chosen as representative of the intrinsic variabilities of facies at hand, which are also conditional on the resolution of the well-logs analyzed. For example, referring to Figure 5a, roughly ten thin beds can be seen on the first pad

data between x480 and x482 feet. For this reason in an embodiment, wavelet coefficients on scales 1/16 foot and 1/8 foot may be used to analyze resistivity image logs 520. Additional and/or different scales (e.g. 1/2 foot, ¹ A foot and 1/32 foot) may be selected for different subterranean environments. Selection of the scales may also be automated with a software or program that automatically defines the most relevant scale lengths. Such software or program may analyze the mean and dispersion of the wavelet energy at each scale from some previously chosen set of scales, over fixed depth intervals around various locations down the borehole. More representative scales to select then are those for which dispersion of wavelet energy per segment down the borehole is not small compared to the mean energy per segment, where the segment length, as will be discussed later, is appropriate for the data resolution and the minimal length of lithofacies units to be recognized by classification. f0035] After selecting the various scales to analyse the distribution of wavelet coefficients, the method of Figure 1 proceeds to procedure 40 where a segmentation of the measured or derived borehole data is performed. The borehole data (e.g. image data 520) are segmented based on, for example, a homogeneity of local areas. The well-logs, such as the resistivity data 520. are segmented according to any of the available segmentation methods found to be sufficiently sensitive for the data at hand, or by simple uniform splitting. The segmentation may include a step of depthwise readjustment of logs (including traces of a log from a multiple-pad tool) so that geological bed boundaries from different logs appear all at same depth. This step can be performed on a coarse scale in pre-processing, or concurrently with segmentation whereby the readjustment may vary locally. The latter process is suitable for local segmentation methods, but can be utilized in repetitions of a global segmentation method as well. The information on where segment boundaries are down the borehole is stored as 0 (no change) or 1 (segment boundary) in a one-dimensional array with an adequately sensitive sampling of the borehole (e.g., ¹ A ft. for deepwater turbiditεs). In an embodiment, a segmentation scale parameter may be selected such that segments are of a size appropriate for the analysis to be performed, i.e. proportional to the measurement resolution and preferably no larger than the minimal size of the lithofacies units desired to be recognized. This means that the segment size should be large enough to contain a large number of data points to form a smooth distribution, but not too large to comprise multiple lithofacies units. For example, a segment size between 1 and 3 feet is suitable for the analysis of an OBM1™ resistivity image histogram with 40 bins and recognition of lithofacies units from such a histogram. For single-trace and coarser-resolution measurements, such as a natural

gamma-ray detection (column 1 of Figures 5a-b), the segment size should be at least 10 times larger than the measurement resolution, thus excluding detection and recognition of thin-bedded facies from gamma-ray only, in an embodiment blocks larger than 3 feet can further be broken up into 3 foot long segments (with a remainder no larger than 3 feet). In a further embodiment, the length of each segment is at least about 30 times larger than the expected regular spacing of measurement points down the borehole.

|0036] In another embodiment, segmentation is done by specifying center locations of intervals of equal length which overlap by a predetermined fixed length (related to the difference between the measurement resolution multiplied by 10 and the minimal lithofacies unit length). This embodiment might be preferable for coarse-resolution measurements, because interval locations can be significantly closer than the interval length needed to have a smooth distribution of measured data. These intervals are taken into the classification part of the method in the same sense as if they were non-overlapping segments. For those regions of the borehole where the overlapping intervals are found to belong to different facies, the segmentation and classification procedures are iterated upon in order to refine the boundary locations. For a given pair of intervals, each new triplet of iterations consists of: 1} classifying these intervals at their current center positions, then 2) classifying after shifting the center locations of the overlapping intervals together, whereby performing the shift slightly first towards one borehole end and then 3) classifying after shifting towards the other borehole end by a small fixed iteration step length d. If a shift in one direction causes the classification of one interval to change while the other one is stable and the shift in the opposite direction does not result in different results from those for central position of the intervals, then the next iteration triplet will start from the central position of intervals shifted from the previous iteration triplet central location by the iteration step length d (or its fraction if fine precision is desired) in the direction of that interval which changed its facies membership. If both shifts (i.e. in both directions) result in the same number of intervals changing membership and if that number is either two or zero (either both, intervals change facies membership, or none), the iterative process is stopped and the boundary location for this pair of segments is chosen as the centerpoint between the center locations of the intervals in their central position for the last iteration triplet. In cases other than here described for the results of iteration triplets, it is concluded that either the intervals are too large so more than one boundary is affecting the classification (so several boundaries are selected instead of just one, with classification results corresponding to intervals centered halfway between these

boundaries), or that the iteration step length is too large, and the iterative process can be restarted with a smaller step length. After finding stable boundaries in the sense of this procedure, the overlapping segments that have the same classification results are merged.

[§037] After segmenting the well-logs, a distribution of wavelet coefficients for each wavelet scale selected at procedure 30, for one or more directions of wavelet transform application, and for each segment that belongs to the portion on which the wavelet transform has been calculated is determined at procedure 50. Referring back to Figure 5a, columns 7, 8 and 9 show distributions of wavelet coefficients for a 1/16 foot scale and for different segment sizes: 3 feet of data on a log scale (column 7) and 1 foot of data (columns 8 and 9). In column 8, the wavelet coefficient distributions are plotted on a log scale, while in column 9 they are plotted on a linear scale. Columns 10, 11 and 12 are analogous to columns 7, 8 aid 9, respectively, except that they refer to a 1/8 foot scale. Very different lilhofacies in terms of the variability of measured quantity used for classification (e.g. of a very different texture of resistivity image) have distributions of wavelet coefficients that are of substantially different shape. For example, as seen in Figure 5a, the shape of the 1/16 foot wavelet coefficient distribution in column 7 is substantially different around x454.5 feet and x493.5 feet. Likewise, the shape of the 1/8 foot wavelet coefficient distribution in column 10 is substantially different around x454,5 feet and X493.5 feet.

Then, after calculating a distribution of wavelet coefficients for each wavelet scale and segment, the method proceeds to procedure 55 where the distributions of wavelet coefficients for each scale are calculated for training samples of various selected known facies. Then, the method proceeds to procedure 60 where the distributions of wavelet coefficients and distribution(s) of borehole data (e.g. resistivity distribution 530 in column 6 of Figure 5a) for each segment are analyzed in terras of their overlap with the corresponding training sample distributions (calculated at procedure 55) in order to produce a geological interpretation of the subterranean formation. This analysis produces a geological interpretation of the subterranean formation in terms of units which are distinct both in their log signature and geologically.

[0039] In an embodiment of the invention, procedure 60 involves two main aspects. The first aspect consists of a feature extraction step (from matching each segment distributions of measured data and their wavelet coefficients to the corresponding distributions of the training sample data) and an intermediate classification for each segment of the borehole. The second aspect consists of a feature fusion step and final classification.

)J In an embodiment, the training samples, which are considered at procedure 55, are in the same set of measured data and are provided by the human interpreter through a reference to their beginning and end depth, The training samples may also be provided through a previously stored library, In this latter case, the distribution of wavelet coefficients for the training samples may already be available. At least one training sample is needed for each facies. In another embodiment of the invention, the training samples are determined by (1) identifying locations in the measured data where there is a local 1 or 2 foot scale (or comparable scale) wavelet coefficient maximum or minimum, the absolute value of which being among the several largest ones calculated on this data; (2) calculating the distribution of measured data or of small-scale (e.g. 1/8 foot or comparable scale) wavelet coefficients for the two 2 foot (or comparable scale) segments around each of such locations (one above and one below); (3) calculating the overlaps of the distributions on these segments pairwise; and (4) selecting a couple or a few of these segments with the smallest mutual overlaps of distributions to be the training samples. [0041] In order to perform the feature extraction siep of classification, distributions (e.g. histograms) of measured data and wavelet coefficients pertaining to the training samples are calculated first (procedure 55). Distributions (e.g. histograms) of wavelet coefficients, that are located within windows delineating training samples, are calculated for each of the few chosen wavelet scales (e.g. 1/16 and 1/8 foot scale) selected at procedure 30 and for each of a plurality of selected known facies (as well as for each direction if the more refined method is used). In an embodiment, seven training samples representative of known geological facies are considered. For image data, wavelet transform coefficients for separate traces of data are also included in the same distribution or histogram provided their positions are in the same - or appropriately depth adjusted - window. One training sample of about 3 feet in length of resistivity image data may suffice for each facies. It will he appreciated that if more training samples per facies are provided, one distribution is formed per scale per facies by adding the occurrences of coefficients in a certain magnitude interval in all training samples for that facies,

[0042] In an embodiment, these calculations obtained with the training samples representative of known geological facies provide three separate libraries or collections of distributions: (I) a first library including wavelet coefficient distributions for each of the training samples (e.g. seven training samples) at 1/16 foot scale, (2) a second library including wavelet coefficient distributions for each of the training samples (e.g. seven training samples) at 1/8 foot scale and (3) a third library including distributions of borehole data (e.g. amplitude histogram of a signal

or pixel histogram for an image) for each of the training samples (e.g. seven training samples), These libraries may then be used to provide an intermediate classification of each of the segments of the borehole.

[0043] Referring now to Figure 6, this figure shows a method of performing feature extraction and intermediate classification for each segment of the borehole in accordance with an embodiment of the invention. The method begins at procedure 610 where a distribution of wavelet coefficients for a property under consideration for each of the previously chosen representative scales (e.g. 1/16 and 1/8 foot scale) is formed for each segment down the borehole. This procedure corresponds to procedure 50 of Figiire 1 , Then, the method proceeds to procedure 620 where a similarity measure between the wavelet coefficient distribution of a given sample at a given scale and the distributions of each fades training sample wavelet coefficients at the same scale is calculated. This process is repeated for each scale and each segment, In an embodiment, the distributions of wavelet coefficients of the measured or derived data are compared to the first and second libraries (e.g. 1/16 and 1/8 foot scale) in order to identify the fades in the first and second libraries that have a distribution of wavelet coefficients that most closely corresponds to the distribution of wavelet coefficients obtained with the segmented data (see procedure 625).

[0044] A chi -square statistic may be used as a similarity measure to identify the matching training samples or facies, A chi-square statistic of two binned distributions is defined as (4):

where R - ^ R, and 5' ~ ]TS , and R ₁ and S ₁ are the respective numbers of the first and

second distribution events in the i-th bin (i.e data points of value belonging to the i-th bin), where i takes integer values from 1 to v in order to index those and only those pairs of bins where at least one of the two distributions has a non-zero number of events, and v is called the number of degrees of freedom. The bin size is chosen such that either v is much larger than 1 or the number of events in each bin is large. Under these conditions, the chi-square probability function Q(χ',v), which is an incomplete gamma-function expressing the probability that the sum of the squares of v random normal variables of unit variance and zero mean will be greater than χ ² . is a good approximation of the distribution of χ ² in equation 4 for the case of a null

hypothesis (Le. hypothesis that (normalized) Ri's and Si's are drawn from the same population) and its use is standard (see W. H. Press et a!., "Numerical Recipes in C", p. 621 , Cambridge University Press, 1999). A small value of Q(X ² A') indicates that the null hypothesis is rather unlikely (i.e. the given sample data and the training sample data have different distributions), In Figure 5a, columns 15 and 16 display Q(X ² ₅ v) as sticks for the segments shown in column 5 (seven sticks of different color - one for each fades - per segment) for scales 1/16 foot and 1/8 foot, respectively. Columns 19 and 20 of Figure 5a are analogous to columns 15 and 16, respectively, with the only difference being that a training sample for facies 8 present in the region around x500 ft was different. In Figure Sb ₅ columns 1 1, 12, 15, 16 and 19 display

|0θ45] The known geological facies with a wavelet coefficient distribution for the considered property at a given scale that matches most closely (as defined by the largest chi-square probability Q(χ ² ,v) at a given number of degrees of freedom v) the same scale wavelet coefficient distribution of the given segment is selected to be the result of intermediate classification of the segment based on the considered property's wavelet coefficient distribution at this scale (see procedure 625), and this facies is referred to as the facies indicated for this segment by the considered property's wavelet coefficient distribution at this scale, or as the comparison result of this segment with the considered property's wavelet coefficient distribution at this scale. Column 14 of Figure 5a displays results of such an intermediate classification for scales 1/8 foot, 1/16 foot and 1/32 foot, wherein the facies code (indexed 0-12) is the same as the one in column 2 from the manual interpretation, except for being reseated by an unessential factor 10. Column 18 of Figure 5a is analogous to column 14 with the sole difference being the training sample for the faeies 8. Columns 21-23 of Figure 5a are analogous to columns 17-19 except that segments larger than 3 feet were not partitioned into 3 foot blocks (with a remainder).

Then, after identifying the closest matches for the wavelet coefficient distributions at various scales (e.g. using the first and second libraries of distributions), the method proceeds to procedure 630 where a similarity measure is calculated between the data distributions of a measured borehole property or derived quantity for the given sample (e.g. gamma-ray amplitude histogram or resistivity image pixel histogram) and the training samples for each segment down the borehole. For example, in an embodiment, the resistivity distribution 530 for each segment shown in column 6 of Figure 5a is compared to the distributions of resistivity of the known faeies in the third library. The known geological facies with a pixel training distribution or

amplitude In the third library that most closely matches the pixel distribution or amplitude distribution of the given segment is determined to be the result of intermediate classification of this segment based on distribution of this data property (see procedure 635), or the facies indicated for this segment by the data distribution of this property.

[0047] Although the similarity measures have been performed with a chi-square statistic, it is contemplated that other statistics could be used in other embodiments of the invention,

[0048] In the event of uniform segmentation with non-overlapping intervals, additional cheeks may be performed to refine the segment boundaries. These checks may include finding the best matches for single foot segments within larger segments of uniform length and shifting boundaries if a single foot segment next to a boundary matches the same facies as the neighboring larger segment.

[0049] In an embodiment, the feature extraction from distributions of wavelet coefficients performed at an angle to the borehole axis proceeds in an analogous fashion,

[0050] Once the facies indicated by all considered distributions are determined for each segment, this information is used to make a final determination of the facies of each segment. This process, which may be referred to hereinafter as the feature fusion step, is applied for each segment down the borehole.

[θθ5I] The feature fusion process includes rules that specify which facies should be identified as the final result for a segment. The segment is characterized by certain feature flags and by facies kl, k2,., ,, kn indicated by distribution of features (measured properties or calculated/derived quantities) dl, d2,,, ,, dn respectively, where dl , d2,, . ,, dn are e.g. the measured or derived borehole data (e.g. in the third library) or wavelet transform coefficients of various measured properties or of derived quantities at various scales (i.e. those making the first and second libraries) from the wavelet transform performed along the borehole axis or along a different direction. The rules are based on discerning classes of all facies which satisfy some class property defined on the indicated facies kl ", k2' ₃ , .. ₅ tor of the facies' representative samples with distributions of features dl, d2,..., dn, respectively and on their feature flags.

10052] A feature flag of a segment is defined for a specific property (feature) pi and signifies if that property changes significantly when measured around the borehole axis in the same segment. In the case when pi is the resistivity data from a multiple pad probe, or the wavelet transform of such data, the pi feature flag signifies whether the distribution of pi varies

significantly from one pad to another, as evidenced from whether the fades matched to different pad samples in a segment in the intermediate classification based on feature pi are substantially different according to some rule. The rule that specifies substantially different fades for this purpose can be built-in for a particular oppositional environment. The presence of such different pad images in the same segment rather indicates another, discontinuous type of a faeies.

[0053] For example, in a shallow marine depositional environment, according to an embodiment of the invention, one of the feature tlags is calculated in the following way from the pad-to-pad variability of the 1/8 fool resistivity wavelet coefficient distribution calculated for separate pads. First, the comparison is made on the 1/8 foot wavelet coefficient distribution for each pad and the 1/8 foot wavelet coefficient distribution of the faeies training samples. The faeies indicated for a given pad by (or a comparison result with) the 1/8 ft resistivity wavelet coefficient distribution is the faeies with the best similarity measure. Further, a comparison result is deemed strong if the corresponding chi-squarε probability is higher than a certain threshold value, e.g. 50%. Second, if there are two pads that have strong comparison results which are faeies incompatible to coexist in a single continuous faeies across the borehole according to the rule of substantially different faeies, then the faeies flag indicates the discontinuous faeies. Substantially different faeies for a shallow marine environment are any of the homogeneous faeies when occurring on one pad and thick-bedded faeies on another pad. This procedure can be expanded to require that there does not exist a strong comparison result of a second-highest similarity measure (a strong second-best match) other than bedded or chaotic faeies on the pad matched with bedded faeies.

A class property defined on a faeies with representative training samples that have comparison results kl, k2,. , . _s kn with distributions of features dl, d2,... _> dn, respectively, and on feature pi flag, is a property that this class contains all faeies which may have the ordered (n+l)-tuplet (kl,k2,...,kn,pl flag) belong to a direct product (SIxS2x...xSnxSpl) of subsets S 1,...,Sn ₅ SpI, so that kl may belong to a subset of all faeies Sl, while k2 belongs to a faeies subset S2, and so on until kn belongs to a faeies subset Sn, and that the feature pi flag indicates a faeies subset, i.e. belongs to a faeies subset SpI. and that all faeies in at least one subset do occur in the corresponding comparison results of the representative samples for each of the faeies in this class. The latter specification is meant to avoid redundancy in the class definition, so that 1 ) only those faeies can be in the same class whose representative samples have similar appearance in terms of the faeies indicated for them by the distribution of a certain feature, and

that 2) the set of fades that belong to a class Is not a union of two or more subsets of facies which satisfy not just the class property but a more restrictive property of the facies indicated by this same feature distribution which was used to define the original class. Representative samples of a facies denote those samples of a facies that can reasonably be described by an expert to represent this facies. This ensures that outliers (as well as '"split decision ^" ' samples) are disregarded, arid the spread in comparison results that can be obtained on the representative samples of a facies is small, A class whose facies only occur with the comparison results described in their class property supports the method of elimination reasoning, which is described below in the feature fusion step (c) and shown in the example of a deepwater depositional environment class CJayered. Such a class may be referred to as an enclosure class. A class property and the knowledge of all facies which belong to that class is useful to identify a subset of facies to which the given segment's facies belongs if the facies indicated for this segment are consistent with this class property. For a single-fades class, its member facies m is identifiable by the corresponding class properly. If the corresponding class property is based on the facies indicated by just one feature d distribution, that facies m is identifiable by the feature d distribution.

In a deepwater depositional environment with, for example, seven facies, the rules in accordance with an embodiment of the invention are based on discerning various classes shown in Figures 7a-b.

] The rules are based on discerning the following classes:

57] (i) single-element class Cm consisting of all facies which satisfy the property that the comparison result kl of their representative samples with distribution of resistivity data may indicate this specific facies m, i.e. kl ^::: rn; these classes are labeled CL C5, C6, CS and Cl 1 after the number of the corresponding member facies, and the classes Cl and C5 are enclosure classes as described above;

(ii) two-element class Owe of all facies which satisfy the property that the comparison results kl, k2 and k3 of their representative samples with distributions of resistivity data, and the 1/8 foot scale and 1/16 foot scale resistivity wavelet transform coefficients distributions of resistivity data, respectively, may indicate facies 7 or 10: this is the class consisting of facies 7 and 10 and is an enclosure class as described above (i.e. the representative samples of these two facies have only these comparison results) ;

[G0S9J (iii) single-element class Cm 2 of all facies whose representative samples may have comparison results kl=7 and k2=m, with ni=6 _s 7, 8 or 10, respectively to Cm_2. with the resistivity data distribution and the 1/8 foot scale wavelet transform coefficient distribution, respectively; these classes are C6_2, C7_2, C8_2 and C10__2, named after the number of a corresponding member facies, and the class C7_2 is an enclosure class;

[0060] (iv) single-element class Cl 1 prime consisting of all facies whose representative samples may have comparison results kl as 7 and k2 belonging to the subset O ₅ S ₅ H } with the resistivity data distribution and the 1/8 foot scale resistivity wavelet transform coefficient distribution, respectively; this class consists of the facies 11 ;

[0061] (v) three-element class C_3 consisting of all facies whose representative samples may have facies kl=10 indicated by the resistivity data distribution; this class consists of the facies

10, 8 and 6;

|Oθ62| (vi) a single-element class Clast consisting of all facies whose representative samples may have comparison results kl=10, k2 belonging to the facies 1,5,6 or 7, and k3 belonging to the facies 1 ₅ 5 or 6, wherein kl, k2 and k3 are indicated by distributions of resistivity data and its 1/8 foot scale and 1/16 foot scale wavelet transform coefficients, respectively; this class consists of the facies 6;

[0063] In an embodiment, the rales specify the final facies identification based on matching the facies indicated for the given segment to the properties that define the different classes and narrowing down the number of facies whose representative samples may have the same indicated facies as the given segment. First, a primary property or feature dl distribution such as resistivity data distribution is selected if its comparison results are specified in the properties of all or nearly all classes and a large number of facies is identifiable by this distribution, The next best candidate for such a property may be termed a secondary property. For each given segment, the rules specify the final facies identification in the following way:

[0064] (a) it is determined which subsets SI the facies kl indicated for the given segment by dl distribution belongs to among all subsets appearing in the properties of facies indicated by dl distribution (for representative samples) used to define classes; for each such subset Sl, the class Cl is identified such that the representative samples of its facies may have comparison results with the feature dl distribution to belong to Sl and thai all facies in Sl are indicated by representative samples of each of the facies in Cl ;

[0065] (b) if there are single-element classes among all classes Cl found in a), then: (i) if there is a single-element class defined by the sole property that the comparison result of the fades representative samples with dl distribution may belong to the subset containing kl, then the element of this class is the facies determined as the final classification result for this segment; (ii) if the single element class with the smallest number of distributions needed to define the class property is also defined with other features d2....,dn, and possibly also feature flags, then: 1 ) if the other comparison results on this segment k2... ,, kn with d2,. ,., dn, respectively (and feature flags) are consistent with the property of this class, then the element of this class is the facies that corresponds to this segment: 2) if the other comparison results of the given segment are inconsistent with this class, the single-element class with the next-smallest number of property-defming distributions is considered in the same fashion; 3) if there are more than one single-element classes with the same number of property-defming features, these classes are considered in some chosen order of importance;

[§066] (e) if there are no single-element classes among classes Cl of faeies which satisfy that their representative samples may have a comparison result with feature dl distribution consistent with that of the given segment, the next-smallest classes are considered in the order analogous to that given in (b), and one uses a method of elimination of those facies in the class which belongs to other classes (and first considered are the enclosure classes) such that the representative samples of faeies in these other classes have comparison results that are inconsistent with those of the given segment;

[0067] (d) w ^r hen all indicated facies and feature flags are consistent with properties of representative samples of more than one facies, then a fall-back rule is used. In an embodiment, the fall-back rule in the criteria include the selection of the faeies which is given by the comparison result of the primary feature distribution if thai feature distinguishes these facies, or the next feature in the order of importance chosen in step (b)(ii)(3).

|0068] (e) if the query in (a) results in no classes Cl found, the facies k2 indicated by the secondary feature d2 is considered in the same way as kl was for the primary distribution, and steps (a)-(d) are carried out; the iteration (e) can be carried out until a facies is found with comparison results consistent with those of the given segment, or until all features used for classification are exhausted in searching for classes Cl for the chosen feature, in which case: (f) a fall-back rule is used that assigns to the segment a facies indicated by the primary feature distribution.

If the fades indicated by all considered distributions (e.g. resistivity and wavelet coefficient distributions at various scales obtained with the first, second and third libraries of training samples) were always the same fades, there would be nothing new to infer from the wavelet coefficient distributions and the intermediate classification based on the measured data would be the final classification (unless more than one of the measured properties and calculated quantities (other than the wavelet transform) are used). However, in the case of inhomogeneous fades with distinct textures shown in resistivity images or other data, wavelet coefficients of these data offer information thai is often not present in the measured data distributions.

[007OJ Referring now to Figure 8, this figure shows an application of the method of feature fusion and final classification (a-f) in accordance with an embodiment of the invention. In this application, the primary property distribution (in step (a)) is the resistivity image data distribution, As illustrated in Figure S. the method for performing feature fusion and final classification begins at procedure 815, where it is first determined whether the distribution of borehole data (e.g. image pixels or signal amplitudes) of the closest (known geological fades) match indicates one of the fades uniquely distinguishable by the resistivity distribution (or according to the step (b)(i), fades which are members of a sirsgle-facies class with the class property based only on the resistivity distribution). If the result of this inquiry 815 is true (yes), the segment is classified according to the distribution of borehole data (e.g. pixel distribution) only (procedure 820). If the result of this inquiry is false (no), the distribution of borehole data (e.g. pixel distribution) indicates a fades 7 or 10 (necessitating procedures 835 or 850), and the subsequent inquiries also use wavelet coefficient distributions to determine the correct class. Thus, steps (b)(ii) and (c)-(f) may be needed in this case.

[0071] The method proceeds to procedure 835 where an inquiry is made as to whether the distribution of borehole data of the closest match (e.g. pixel distribution) indicates facies 7. If the result of the inquiry 835 is true (yes), an inquiry is made as to whether the 1/8 foot wavelet distribution coefficients indicate facies 7 (procedure 840). If the result of the inquiry 840 is true (yes), the method proceeds to procedure 830 where facies 7 is selected for the given segment. This choice illustrates the application of step (b)(ii)(l) and the use of the class property for the class C7 2 in Figure?. If the result of the inquiry 840 is false (no), another inquiry is made as to whether the 1/8 foot scale wavelet coefficient distribution indicates a homogeneous facies 1,5 or 11 (procedure 841), ϊf the result of this inquiry (procedure 841 } is true (yes), the segment is classified as facies 11 (procedure 842). This choice illustrates the application of step (b)(ii)(l)

and the use of the class property for the class Cl 1 prime in Figure 7. If the result of this inquiry is false (no), the segment is classified according to the facies identified by the 1/8 foot wavelet coefficient distribution (procedure 845). The final classification in this case results from step (b)(ϊi)(l) applied to the class Cm 2 where m is equal to the facies indicated by the 1/8 foot scale wavelet coefficient k2~m.

If the result of the inquiry at procedure 835 is false (no), the method proceeds to procedure 850 where an inquiry is made as to whether the distribution of borehole data (e.g. pixel distribution) indicates the facies 10. This inquiry is not needed if only two facies of those indistinguishable solely by resistivity are desired to be identified, as by this point in the procedure only the facies 10 is left as a possible intermediate classification result from the resistivity data distribution. If the result of the inquiry 850 is false (no), the procedure could go on to query all classes of facies based on the facies indicated by wavelet coefficient distributions to find if any class has a property consistent with the comparison results on this segment, starting with a different primary distribution as outlined in procedures (a)-(e) (e.g. starting with the 1/8 foot scale wavelet coefficient distribution). However, with the only given classes being those in Figures 7a and 7b, no such class will be found and the fall-back rale of step (f) is used (procedure 851 - fall-back rule). Thus, if the result of the inquiry 850 is false (no), the facies indicated by the measured data distribution (procedure 820) is selected to classify the given segment (by the fall-back rale 851). If the result of the inquiry 850 is true (yes), the method proceeds to procedure 855 where an inquiry is made as to whether both the 1/8 and 1/16 foot wavelet distribution coefficients indicate facies 7 or 10. If the result of the inquiry 855 is true (yes), the facies 10 is selected as the final result (procedure 860). This is a result of the method of elimination in step (c) applied to the class Cwc to exclude the facies 7 of the class C7_2. If the result of the inquiry 855 is false (no), another inquiry (procedure 856) is made, to determine if both 1/8 foot and 1/16 foot scale wavelet coefficient distributions indicate one of the facies 1,5,6 or 7 and the 1/16 foot scale wavelet coefficient distribution does not indicate facies 7. If the result of this inquiry (procedure 856) is true (yes), facies 6 Is selected (procedure 857). This is a result of step (b)(ii)(l) applied to the class Clast. If the result is false (no), facies 8 is selected (procedure 858). The latter is a result of the method of elimination in step (c) applied to the class C 3 to exclude the facies 6 of the class Clast and the facies 10 of the class Cwc.

When a more detailed distinction of facies with intricate structure is desired, it may be desirable to include the coefficient distributions from the wavelet transform performed at an

angle to the borehole axis. The feature fusion may be more complicated in this case but conceptually transparent; for final classification in the case of a pixel histogram with an intermediate mean, one may rely on the wavelet coefficient histograms and consult the level of certainty (as expressed by similarity measures) from transforms at several angles. If one obtains a good match only at one angle (e.g. ±5"), this is likely to be a facies with directional order. If a few angles give good matches, the result will be linked to a poorly-ordered facies. [0074] Referring now to Figures 9 and 10, these figures more generally illustrate the method of feature fusion and final classification (a)-(f) and the rule forming.

[0075] In Figure 9, the first procedure 905 in forming rales consists of finding one measured or calculated property, which may be referred to as a primary feature, whose distribution alone identifies each facies of a large set or subset of facies used for training. If there are several such distributions, the one which appears in most class properties is chosen according to the step (a) of feature fusion (procedure 910), The next two procedures in forming rales are procedures 915 and 920 of Figure 9. In procedure 915, the characteristics Iengthscale cl, c2....,cn of spatial variability of data from the training samples of those facies that are not identifiable by the distribution of d 1 alone (here n is the number of such facies) is determined. In procedure 920, the scale wj of the scales used in the wavelet transform calculation that is the closest to a characteristic Iengthscale ej for each j-1 ,2,, , .. n of lengthscales which are found in procedure 915 is determined (procedure 925),

[0076] After an inspection similar io that of procedure 905 in Figure 9 of ordered pairs of facies kl and k2j indicated for representative samples by the dl distribution and the corresponding wavelet coefficient distribution on scale wj. respectively, (procedure 925), the method proceeds to procedure 930 where identifiable facies are found, For any (kl,k2j) for which two or three facies had representative samples with those indicated facies, this finding is noted in procedure 935 so that two-facies and three-facies classes can be formed with the property that they contain all facies with representative samples which may have the corresponding indicated facies (kl ,k2j ). These classes may be useful for the method of elimination in case no facies is identifiable by an ordered pair, triplet, and so on up to an n-tuplet of indicated facies for a given segment with various feature distributions.

[0077] If no facies is identifiable by an ordered pair of the facies (kl,k2) indicated for the given segment by any two distributions, the method proceeds to procedure 940 in Figure 9, where a facies is identified from triplets, quadruplets, and so on to n-tupiεts of indicated facies, relying

on rales that are formed by procedure 940 in Figure 9 analogously to the procedure 930. This is step (b)(ii)(2) of the feature fusion.

[0θ78] The facies that is most commonly associated (and thus identified) with having a given ordered n-tuplet of facies indicated by distributions of given measured or calculated (including wavelet transformed) quantities or indicated by feature flags is found from its higher occurrence than that of other facies within results of manual classification on some test set of representative samples which have this given ordered n-tupϊet of facies indicated by distributions of these given measured or calculated quantities or feature flags. Furthermore, a method of elimination from classes found in the procedure 935 of Figure 9 can be deduced in procedure 945 and applied in 1055 of Figure 10, and finally a fall-back rule is defined in procedure 950 and applied in procedure 1065 of Figure 10.

[0079] In the method of Figure 10, procedure 1005 consists of specifying a facies for a given segment based on the facies indicated with this segment's primary property distribution if the indicated facies belongs to the large set or subset of facies identifiable by this distribution alone. This corresponds to step (b)(i) of feature fusion in the case when identified and indicated facies are the same. If the result of the inquiry is true (yes), the facies identified by distribution dl is selected as the final result (procedure 1010). If the result of the inquiry is false (no), the method proceeds to procedure 1020,

[0080] Procedure 1020 of Figure 10 corresponds to the application of finding 930 of Figure 9: the rules specify a facies k2 for a given segment based on the facies kl indicated by this segment's primary feature distribution and facies k2j (for each j=l,...,n) indicated by the corresponding wavelet coefficient distribution on the scale wj if the facies k2 is identifiable by the ordered pair of indicated facies (kl,k2j) by these distributions and the facies indicated by the primary feature distribution does not belong to the large set or subset of facies identifiable by this distribution alone. Procedure 1020 of Figure 10 is an implementation of step (b)(ii)(l) of feature fusion with two distributions. If the result of the inquiry is true (yes), the facies identified k2 is selected as the final result (procedure 1025). If the result of the inquiry is false (no), the method proceeds to procedure 1040.

[§081] In procedure 1040 it is determined whether there is a single facies k identifiable by the ordered m-tuplet (k] ,k2....,km) of facies indicated on the given segment by distribution of dl and some other feature distributions, where kl is not identifiable by dl alone and there is no facies identifiable by the ordered pair of facies indicated by dl and any other distribution. If the

result of the inquiry is true (yes), the facies k is selected as the final result (procedure 1045). If the result of the inquiry is false (no), the method proceeds to procedure 1055 where it is determined whether there are comparison results of the given segment consistent with class properties of some two- or three-facies enclosure classes. If the result of the inquiry is true (yes), a method of elimination is used to identify the final result (procedure 1060). If the result of the inquiry is negative, a fall-hack rule is used (procedure 1065).

|θO82j The feature extraction and feature fusions according to the embodiment of the invention shown in FIGS. 1, 6 and 7 were tested on part of the actual well OBMI resistivity data (520, shown at column 4 of Figures 5a-b). Very good results were obtained in this case. The results of preliminary testing indicate an approximately 80% success rale on 50 samples roughly uniformly distributed over seven facies except that there were only a few samples of slurry and sandy mass flow, As both the correctly and the incorrectly identified facies span all of these seven, it is extremely unlikely that this success rate could have been due to either random picking or picking heavily biased to one or two faeies.

[0083] As noted above, column 1 of Figure 5a displays the gamma ray log, 505; column 3 shows the measured depth down the borehole and column 4 shows the normalized resistivity traces; column 5 marks segment boundaries from a segmentation algorithm; column. 6 of Figure 5a displays resistivity (i.e. pixel histogram) for each 1 foot of data and columns 7-9 display- wavelet coefficient distribution for 1/16 foot scale for each 3 feet of data on a log scale (column 7), 1 foot of data on a log scale (column 8) and 1 foot of data on a linear scale (column 9).

[0084] Column 2 of Figure 5a shows the geological facies index as identified by manual interpretation. In column 2, the numbers used to label facies are the same as those used in the description of the example with classes shown in Figures 7a~b .

[0085] Referring now to columns 13-16 of Figure 5a, these columns illustrate intermediate classification results that are obtained in accordance with the method shown in Figure 6, These results are stated as facies code divided by 10 and the corresponding similarity measures - with facies code (divided by 10) in column 13 for resistivity, in column 14 for wavelet coefficients on scales 1/32, 1/16 and 1/8 foot. Columns 15 and 16 display probabilities (Q(χ ² ,v)) for matching 1/16 foot scale and 1/8 foot scale, respectively, wavelet coefficient distributions to those of facies training samples.

As noted above, columns 17-20 of Figure 5a are analogous to columns 13-16 but differ in that the training sample for the facies 8 is different, As shown in Figure 5a, the displayed probabilities are the same in columns 19 and 20 as in columns 15 and 16 respectively with the exception of the probabilities of matching to facies 8 training sample (for example in the region shown below x500 feet in this figure),

[0087] Columns 21 to 23 of Figure 5a are analogous to columns 17-19, respectively, (with the same training samples), except that large segments were not partitioned into 3 foot blocks. There is no change in column 21 right below x519 feet, contrary to columns 13 and 17,

[0088] The representative test segments and final facies assigned on this portion of the image log based on the intermediate classification results in columns 13-14 of Figure 5 are shown in Table 1, Among all the segments obtained by a routine segmentation, the representative test segments are only those which are entirely within the geological facies segments (column 2 of Figure 5) and do not cross sinusoidally-shaped thick-bed boundaries on the OBMI image of column 4. In this fashion, the accuracy of classification is assessed. Table 1 also shows the comparison between the classification results obtained with the method according to an embodiment of the invention and the classification obtained by manual image classification and shown at column 2 of Figures 5a-b,

Table 1

189] Table 1 illustrates well the agreement of about 80% between the fades found by the automated method and the manually Interpreted fades for the sample segments of interest. This agreement was also found in testing the method on fifty samples. The distributions of wavelet coefficients for both 1/16 and 1/8 foot scales displayed in column 14 of Figure 5a indicate the correct class for segment [x505.75 ■■■■ x508]. Also, it is noted that a classification with larger segments such as that shown in column 21 of Figure 5a identifies the correct class for segment [x513.75 - x516.75j.

[0090] Although the invention has been described in detail for the purpose of illustration based on what is currently considered to be the most practical and preferred embodiments, it is to be understood that such detail is solely for that purpose and that the invention is not limited to the disclosed embodiments, but, on the contrary, is intended to cover modifications and equivalent arrangements that are within the spirit and scope of the appended claims. For example, though reference is made herein to a computer, this may include a general purpose computer, a purpose- built computer, an ASIC including machine executable instructions and programmed to execute the methods, a computer array or network, or other appropriate computing device. As a further example, it is to be understood that the present invention contemplates that, to the extent possible, one or more features of any embodiment can be combined with one or more features of any other embodiment.