Geologically constrained fi fl - oucreate.commatthewpranter.oucreate.com/article_pdf/Allen_Pranter_AAPG_2016.pdf · Geologically constrained electrofacies classification of fluvial

Geologically constrainedelectrofacies classification offluvial deposits: An examplefrom the Cretaceous MesaverdeGroup, Uinta and PiceanceBasinsDaniel B. Allen and Matthew J. Pranter

ABSTRACT

Statistical classification methods consisting of the k-nearestneighbor algorithm (k-NN), a probabilistic clustering procedure(PCP), and a novel method that incorporates outcrop-basedthickness criteria through the use of well log indicator flags areevaluated for their ability to distinguish fluvial architectural ele-ments of the upper Mesaverde Group of the Piceance and UintaBasins as distinct electrofacies classes. Data used in training andtesting of the classification methods come from paired cores andwell logs consisting of 1626 wireline log curve samples eachassociated with a known architectural element classification asdetermined from detailed sedimentologic analysis of cores (N = 9).Thickness criteria are derived from outcrop-based architecturalelement measurements of the upper Mesaverde Group. Throughan approach that integrates select classifier results with thick-ness criteria, an overall accuracy (number of correctly predictedsamples/total testing samples) of 83.6% was achieved for a four-class fluvial architectural element realization. Architectural el-ements were predicted with user’s accuracies (accuracy of anindividual class) of 0.891, 0.376, 0.735, and 0.985 for thefloodplain, crevasse splay, single-story channel body, and mul-tistory channel body classes, respectively. Without the addi-tional refinement by incorporation of thickness criteria, thek-NN and PCP classifiers produced similar results. In both thek-NN and PCP techniques, the combination of gamma ray andbulk density wireline log curves proved to be the most usefulassemblage tested.

AUTHORS

Daniel B. Allen ~ Department ofGeological Sciences, University of Colorado,2200 Colorado Avenue, Boulder, Colorado80309; present address: Apache Corporation,2000 Post Oak Blvd #100, Houston, Texas77056; [email protected]

Daniel B. Allen is a geologist at ApacheCorporation working in south Louisiana.He received his B.S. degree in geologyfrom the University of Texas at Austin andhis M.S. degree in geology from theUniversity of Colorado at Boulder.

Matthew J. Pranter ~ ConocoPhillipsSchool of Geology and Geophysics,University of Oklahoma, 100 E. Boyd Street,Sarkeys Energy Center, Norman, Oklahoma73019; [email protected]

Matthew J. Pranter holds the Lew and MyraWard Endowed Chair in ReservoirCharacterization and is a professor of geologyand geophysics at the University ofOklahoma. He received his B.S. degrees ingeology and geological engineering fromthe Oklahoma State University and ColoradoSchool of Mines, respectively; his M.S.degree in geology from Baylor University;and his Ph.D. in geology from the ColoradoSchool of Mines. He was previously a geologyprofessor at the University of Colorado atBoulder and a senior research/reservoirgeologist with ExxonMobil UpstreamResearch Company andConoco. His researchinterests are in reservoir characterization andmodeling, sedimentary geology, andpetroleum geology.

ACKNOWLEDGMENTS

Funding was provided by the industrysponsors of the Reservoir Characterizationand Modeling Laboratory and the WilliamsFork Consortium: Anadarko, Chevron,ExxonMobil, Occidental (Oxy), QEPResources, Suncor Energy, and WPX Energy.Additional funding was provided through theAAPG Foundation Grants-in-Aid (Norman H.Foster Memorial Grant) and the SEPMEndowment Grant. Software was providedby IHS (Petra); Fugro-Jason (Powerlog3.3 and FaciesID); and eGAMLS Inc.

Copyright ©2016. The American Association of Petroleum Geologists. All rights reserved.

Manuscript received December 8, 2014; provisional acceptance November 18, 2015; revised manuscriptreceived April 15, 2016; final acceptance May 13, 2016.DOI:10.1306/05131614229

AAPG Bulletin, v. 100, no. 12 (December 2016), pp. 1775–1801 1775

mailto:[email protected]

mailto:[email protected]

http://dx.doi.org/10.1306/05131614229

INTRODUCTION

The classification of architectural elements is a vital task in theinvestigation of fluvial depositional systems and reservoirs. Thistask is commonly conducted through the study of core samplesor through observations made from outcrop. Numerous suchreservoir-scale studies (Ellison, 2004; Cole and Cumella, 2005;German, 2006; Panjaitan, 2006; Pranter et al., 2007, 2009; Cole,2008; White et al., 2008; Pranter and Sommer, 2011) of the fluvi-al deposits of the upper Mesaverde Group of the Piceance andUinta Basin province of western Colorado and eastern Utah haverecorded detailed sedimentologic analysis, dimensional charac-teristics (e.g., width, thickness, andwidth-to-thickness ratio), andpaleocurrent data to describe and classify the different fluvialarchitectural element types common to these deposits. Theseefforts to classify the fluvial architectural elements and recordtheir distributions have resulted in a better understanding ofthe fluvial depositional system of the upper Mesaverde Group.Other studies have, at least in part, if not fully, relied on themanual interpretation of fluvial architectural elements in welllogs through commonly recognized well log curve motifs (Rider,2002) to assist in the investigation of the sequence stratigraphyof the upper Mesaverde Group (Shaak, 2010) and the staticconnectivity of the reservoir-quality sandstones that composethis interval (Hewlett, 2010; Sloan, 2012; Pranter et al., 2014).

Though extensive study has been conducted through theanalysis of core and outcrop exposures as well as the manualinterpretation of fluvial deposits from well logs, these methodshave their constraints. Most wells have well logs but are notcored, whereas outcrops suitable for study can be geographicallylimited. Though comparatively more economical than coringwells to aid in interpretation of fluvial deposits, manual inter-pretation of these deposits using well logs can be a tedious, time-consuming, and subjective task for even the most experienced ofwell log analysts and geologists.

As a response to these constraints, this study explores an al-ternative approach to the classification of fluvial architectural el-ements through the application of two statistical classificationmethods: (1) the k-nearest neighbor algorithm (k-NN) and (2)a probabilistic clustering procedure (PCP). In addition, a thirdapproach is used to refine the results of the classifiers, whichincorporates outcrop-based, architectural element thickness cri-teria. These methods are evaluated for their ability to extrapolatethe fluvial architectural elements as distinct electrofacies classesfrom training wells (wells with core and log data used to train theclassification method) to testing wells through the comparison ofwireline log curve measurements. Data used in the training andtesting of the classifiers come from (N = 1626) wireline log curvesamples (values) associated with known architectural element

(previously Eric Geoscience, Inc.) (GeologicAnalysis via Maximum Likelihood System[GAMLS]). We especially thank Mike Ulandand iReservoir for the very helpful advice andsupport. We thank Eric Eslinger for hisassistance with GAMLS. The authorsappreciate the constructive reviews of AAPGEditor Michael L. Sweet, Robert M. Cluff, andtwo anonymous reviewers.

1776 Geologically Constrained Classification of Fluvial Deposits

classifications as determined from detailed sedimento-logic analysis of cored wells (N = 9) located throughoutthe study area (Figure 1). The results of this study pro-vide a methodology for making interpretations of thefluvial architectural elements using well logs that iscarried out in a cost-effective, timely, objective, andreproducible manner.

The discrimination of depositional facies fromwell logs dates to the mid-1950s, when workersat Shell used spontaneous potential well log curveshapes to distinguish depositional facies of themodern Mississippi Delta (Serra, 1989). Studies in the1970s (Serra and Sulpice, 1975; Rider and Laurier,1979; Serra and Abbott, 1982) attributed character-istic shapes of additional well log curves such as gam-ma ray (GR), bulk density (RHOB), neutron porosity

(NPHI), and dipmeter logs to depositional facies(Sullivan et al., 2003). With the ever increasingneed for timely and cost-effective methods of faciesclassification, the decades since the 1980s have seena growth in the application of various statistical ap-proaches to automated electrofacies prediction. Theterm “electrofacies” was used by Serra and Abbott(1982, p. 117) and is defined as the set of log responsesthat characterizes a bed and permits it to be dis-tinguished from others, a task typically carried outusing simplistic graphical techniques (Doveton andPrensky, 1992). Wolff and Pelissier-Combescure(1982) were early practitioners of the multivari-ate statistical approach to automated facies whenthey used principle component and cluster analysisto estimate the occurrence of lithofacies. The use of

Figure 1. Map of the studyarea in the Piceance and UintaBasins of northwestern Colorado(CO) and eastern Utah (UT), re-spectively. Cored wells used intraining and testing classifiermodels are shown as well asnoncored wells, which were usedfor demonstrating the predictionand mapping of electrofaciesclasses across the region. Lat =latitude; Long = longitude.

ALLEN AND PRANTER 1777

multivariate statistics continued with the implemen-tation of discriminant function analysis by Busch et al.(1987) and Delfiner et al. (1987) to estimate the oc-currence of lithofacies (Dubois et al., 2007). Perhapsthe most popular method of electrofacies classificationin recent years (e.g., Kapur et al., 1998; Grotsch andMercadier, 1999; Saggaf and Nebrija, 2000; Russellet al., 2002) has been the nonmultivariate statisticalapproach of the artificial neural network, whichDubois et al. (2007) showed to have an advantage inits ability to correctly predict electrofacies classes whencompared with other commonly used classificationmethods. A summary of commonly performed stat-istical approaches to electrofacies classification can befound in Doveton (1994). As in this study, many ofthe previously mentioned studies use core-defineddepositional facies to provide ground truth to thepredictive capabilities of the classification methodsto facilitate acceptance of the results among othergeoscientists.

This study examines the statistical approachesof the k-NN algorithm and a PCP. The k-NN algo-rithm (Cover and Hart, 1967) is attractively noted asbeing one of the simplest and most intuitive classi-fication methods (Bremner et al., 2005; Hall et al.,2008). Unlike complex artificial neural networkclassifiers, this classification method is a simple look-alike contest where unknown objects are matchedaccording to similar objects with known classes(Dubois et al., 2007). Though simplistic, in a studycomparing the ability of four commonly used classi-fiers’ abilities to predict selected lithofacies of thePermian Council Grove Group in the Hugoton andPanoma fields in southwest Kansas and northwestOklahoma, Dubois et al. (2007) found the k-NNclassifier to perform comparatively well. The k-NNclassifier has also been applied to distinguishing be-tween ground cover classes as captured by satelliteimagery and is used as a tool in internet searchengine functions (McRoberts et al., 2002; Haapanenet al., 2004; Beaudoin et al., 2005; Yeung et al.,2008). The PCP is a maximum likelihood model-based neural system (but not an artificial neuralnetwork) (Eslinger and Boyle, 2011). It is the mainclustering engine within the Geologic Analysis viaMaximumLikelihood System� software. The PCP hasbeen previously employed in past studies in the dis-crimination of lithofacies of the Mississippian BarnettShale at Newark East field, Fort Worth Basin, Texas

(Vallejo, 2010), and has been applied to the estimationof petrophysical properties (Eslinger and Boyle, 2011).

TECTONIC AND STRATIGRAPHIC SETTING

The Campanian-age Uinta and Piceance Basins arelocated in northeastern Utah and northwestern Col-orado, respectively (Figure 1). The Uinta Basin isasymmetrical with a west–northwest-trending axis.It is roughly 120mi (193 km) in length and nearly 100mi (161 km) wide. The basin is bounded by the Uintauplift to the north, the Wasatch Plateau to the west,the Rafael and Uncompahgre uplifts to the south, andthe Douglas Creek arch to the west (Spencer, 1995;Stancel et al., 2008). The Piceance Basin is asymmet-ric with a northwest–southeast-trending axis. It is 100mi (161 km) in length and 40–50 mi (64–80 km)wide (Spencer, 1995). The basin is bounded to thenorthwest by the Uinta uplift, to the north by the Ax-ial arch, to the east by the White River uplift, to thesoutheast by the Elk Mountains and Sawatch uplift, tothe south by the Gunnison uplift and San Juan vol-canic field, and to the southwest by the Uncompahgreuplift and is separated from the Uinta Basin by theDouglas Creek arch to the west (Johnson, 1989). Dur-ing the Cretaceous, both the Piceance and Uinta Ba-sins were part of the much larger Cretaceous RockyMountain foreland basin that formed as a result ofthrust loading along the Sevier orogenic belt, whichlies on the western margin of the Uinta Basin. FromLate Cretaceous through the Eocene, the forelandbasin was separated into several smaller structuraland sedimentary basins by rising Laramide uplifts.Subsidence of the Piceance Basin began in the LateCretaceous (late Campanian) and ended during themiddle Eocene (Johnson and Finn, 1986; Johnson,1990; Johnson and Roberts, 2003). In contrast,subsidence of the Uinta Basin did not begin untilthe Paleocene and continued into the late Eoceneand possibly early Oligocene (Johnson and Finn, 1986).

During the Late Cretaceous, the area that is nowthe Piceance and Uinta Basin province was locatednear the western shoreline of the Western Interior Cre-taceous seaway (Hettinger and Kirschbaum, 2002).Sediments eroded from the Sevier orogenic beltin eastern Nevada and western Utah formed a broadpiedmont of coalesced alluvial fans that graded east-ward into alluvial plain, coastal plain, deltaic, and


marine settings that compose the strata of theMesaverde Group (Cole, 2008; Stancel et al., 2008).

This study follows the stratigraphic terminologyof Hettinger and Kirschbaum (2002, 2003) for theMesaverde Group in the southern part of the Piceanceand Uinta Basins (Figure 2). This study focuses onthe upper part of the Mesaverde Group in both thePiceance and Uinta Basins, which includes the sim-ilarly deposited Williams Fork Formation and Farrerand Tuscher Formations, respectively (Lawton, 1986;Hettinger and Kirschbaum, 2002). The WilliamsFork Formation is 3600–5155 ft (1097–1571 m)thick along the Grand Hogback and thins westwardto approximately 1200 ft (366 m) at the Colorado–Utah state line (Hettinger and Kirschbaum, 2002).In the southwestern Piceance Basin, the WilliamsFork Formation is informally subdivided into lower(sandstone-poor), andmiddle–upper (sandstone-rich)intervals based on lithofacies, architectural elements,and net-to-gross ratio (Cole and Cumella, 2005;Pranter and Sommer, 2011; Keeton, 2012). As ob-served in Coal Canyon, the lower Williams ForkFormation, which also contains the Cameo–Wheelercoal zone at its base, is a relatively low net-to-grossratio system (15% average net-to-gross ratio) that large-ly consists of mudrock and isolated channel form sand-stone bodies (channel bars). This lower interval isinterpreted to have been deposited by anastomosing tomeandering streams in a mostly coastal plain settingto marginal marine setting (Lorenz and Rutledge,1987; Johnson, 1989; Hemborg, 2000; Pattersonet al., 2003; Cole and Cumella, 2005; Pranter et al.,2007). The middle and upper Williams Fork For-mation are distinguished by a relatively high net-to-gross ratio system (50%–80% average net-to-grossratio) containing abundant amalgamated sheet-likechannel-form sandstone bodies and associatedmudrocks that are interpreted to have been depos-ited in a low-sinuosity braided alluvial plain envi-ronment (Patterson et al., 2003; Cole and Cumella,2005; German, 2006).

In the Uinta Basin, the Farrer and Tuscher For-mations successively overlie the Neslen Formation,a prograding delta complex that includes tidal andcoastal plain deposits, and are overlain unconform-ably by the Paleogene Wasatch Formation (Stancelet al., 2008). The Farrer Formation extendswest fromthe Utah–Colorado border to Soldier Canyon whereit grades into the laterally equivalent Price River

Formation. It has been measured at 950 ft (290 m)thick at Tusher Canyon in the southeast part of thebasin and thins westward to 131 ft (40 m) at itswestern limit. The Farrer Formation consists of fining-upward single-story sandstone bodies, multistorysandstone bodies, and thick siltstone sequences thatare interpreted to have been deposited by a moderate-sinuosity meandering fluvial system in an uppercoastal plain environment (Lawton, 1986; Hettingerand Kirschbaum, 2002). Net-to-gross ratio begins toincrease in the upper part of the Farrer Formation, andthe gradational contact with the overlying TuscherFormation was placed by Lawton (1983, 1986) tobe where the sandstone content exceeded 50%. TheTuscher Formation, which extends westward fromthe Utah–Colorado border to near Green River, Utah,ranges from 919 ft (280 m) in thickness at TuscherCanyon to 358 ft (109 m) at its western limit(Lawton, 1986; White et al., 2008). The sandstonedominant Tuscher Formation is characterized bythick amalgamated sheetlike sand bodies with thinsiltstone intervals, as well as interspersed thick lat-erally discontinuous sandstone bodies that featurelateral-accretion surfaces. The Tuscher Formation isinterpreted to have been deposited by northeast-flowing meandering and braided fluvial systems(Lawton, 1986).

METHODOLOGY AND DATA SET

Study Area

Data used in the training and testing of electrofaciesmodels come from 1626 samples that are associatedwith known architectural element classifications asdetermined from the detailed sedimentologic analysisof cores (N = 9, total footage 1692 ft [516 m]). Coreswere selected for the study on the basis of (1) acces-sibility, (2) geographic distribution, (3) stratigraphic dis-tribution, and (4) quality (length of core, continuity ofcored intervals, and few rubblized zones) (Figures 1,3). Each sample is also associated with as many asfour available measured properties, which consist ofthe wireline log curves (1) GR, (2) RHOB, (3) deepresistivity (ILD), and (4) NPHI. These measuredproperties were selected based on their mutual pres-ence for all of the study cores. To evaluate the effectiveprediction of electrofacies classes in noncored wells,


Figure 2. Schematic cross section showing the stratigraphic relationships and nomenclature of the Piceance and Uinta Basins. Thestudy interval focuses on the upper part of the Mesaverde Group consisting of the Farrer and Tuscher Formations (Fm.) in the UintaBasin and the Williams Fork Fm. in the Piceance Basin (modified from Hettinger and Kirschbaum, 2002, courtesy of US GeologicalSurvey). Ls. = limestone; Maastricht. = Maastrichtian; Mbr. = member; Paleo. = Paleocene; Sh. = shale; Ss. = sandstone.


the study cores were divided into two subsets: atraining set (5 cores, 440 samples) and a testing set (4cores, 1186 samples).

To demonstrate the applicability of the meth-ods investigated in this study, additional noncoredwells (N = 216) were selected throughout the studyarea in which the most successfully trained classifierwould be employed to exhibit the batch prediction

of electrofacies classes (Figure 1). These wellswere visually inspected and selected on the basis of(1) geographic coverage, (2) stratigraphic cover-age, (3) robustness of wireline log curve assem-blage, (4) data quality (few obvious data spikes,few borehole breakouts, and modern wireline logs),and (5) nondeviated well paths through the intervalof interest.

Figure 3. Locations of cored intervals (black bars) in cored wells. Well locations are shown on Figure 1. Wells have equal spacing; nohorizontal scale is used. Thickness values are in feet. Fm = formation; GR = gamma ray log (API units); ILD = deep induction log (ohm m);RHOB = bulk density log (g/cm3); Ss. = sandstone.


Data Editing and Normalization

Prior to the creation and testing of the trainingmodels, several preprocessing subtasks were per-formed on the wireline data to help ensure reliableelectrofacies class assignments: (1) core-to-log depthshift corrections were made to all well log curves;(2) well log curves were visually inspected to removeobvious data errors or spikes; (3) where necessary,logs were resampled to a common increment of 0.5 ft(0.15 m); (4) and GR curves were normalized usingtwo-point histogram shifting.

Classifying Architectural Elements

Detailed sedimentologic descriptions (lithology, grainsize, texture, sedimentary structures, and contacts)of the study cores were used to determine architec-tural elements to investigate their potential to begrouped into distinct electrofacies classes. To cap-ture the detail necessary for architectural elementdescription, a relatively fine–scale description of litho-facies is required. Eleven lithofacies were describedin the study cores on a 0.39-in. (1-cm) basis and in-clude (1) highly fissile mudstone; (2) laminated tomottled mudstone; (3) carbonaceous mudstone; (4)convoluted sandy siltstone; (5) wavy laminated sandysiltstone; (6) argillaceous siltstone; (7) convolutedsilty sandstone; (8) ripple laminated sandstone; (9)low- to high-angle cross-bedded sandstone; (10)convoluted sandstone; and (11) structureless, cryp-tically bioturbated sandstone. These lithofacies aresimilar to those described by past workers of thefluvial deposits of the Mesaverde Group (Cole andCumella, 2005; Harper, 2011; Keeton, 2012;Sloan, 2012) and are similar to lithofacies that areuniversally recognized in fluvial deposits because ofthe common physical processes that control deposi-tion of clastic fluvial lithofacies (Miall, 1978, 2010).Architectural elements were classified through rec-ognition of distinctive assemblages of these lithofaciesin addition to the nature of lower and upper bound-ing surfaces, internal geometry, and scale (thickness)(Miall, 1985). Based on these characteristics as ob-served in the study cores and by comparing themto observations made in outcrop in past studies ofthe fluvial deposits of the Mesaverde Group, thefollowing architectural elements were interpreted:

(1) floodplain, (2) crevasse splays, (3) single-storychannel bodies, (4) multistory channel bodies, and(5) amalgamated channel bodies (Cole and Cumella,2005; Pranter et al., 2009; Pranter andSommer, 2011)(Figure 4, Table 1).

Floodplain deposits represent the dominantlyfine-grained lithologies surrounding ancient fluvialchannels andmeander belts and are dominated by thehighly fissile mudstone, mottled mudstone, and car-bonaceous mudstone lithofacies (Bridge, 2006).Abundant carbonaceous root traces sometimes asso-ciated with iron-oxide mottling were commonly ob-served in floodplain deposits throughout the studycores suggesting humid climatic conditions withseasonal flash-flooding and a fluctuating water table(Retallack, 2001; Flaig et al., 2011). Pedogenic fea-tures such as small calcite nodules do occurwithin thefloodplain facies assemblage but are very rare, sug-gesting little time for soil development in a rapidlyaggrading floodplain setting (Smith and Rogers,1999). Floodplain deposits commonly exhibit rela-tively highGRvalues, loweffective porosity, and bulkdensity values that are less than sandstone-rich ar-chitectural elements.

Crevasse splays develop during high-runoffevents when the channel bank is breached and sedi-ment spreads out to be deposited onto the floodplainarea (Bridge and Tye, 2000; Miall, 2010). Crevassesplays tend to feature a coarsening-upward grain sizetrend that typically grades from basal mud-richfloodplain lithofacies into wavy to convolute sandysiltstones and up into very fine– to fine-grained ripplelaminated sandstones. These deposits are commonlycapped by a sharp contact with overlying floodplainlithofacies and commonly feature a high amount ofbioturbation in the form of rooting and probableinsect burrowing. Crevasse splays of the MesaverdeGroup have been observed in outcrop at CoalCanyon, Colorado, to have an average thickness of3 ft (0.85 m) (Cole and Cumella, 2005). Crevassesplays commonly exhibit relatively higher GRvalues as compared with point bar deposits. TheGR values can decrease upward and reflect theprogradational nature of the deposit. They com-monly have lower effective porosity and bulk densityvalues that are less than sandstone-rich architecturalelements.

Single-story channel bodies typically represent iso-lated point bar deposits, which are common in high-


Figure 4. Classification of sandstone body types of the Williams Fork Formation. (A) Crevasse splay, (B) single-story channel body,(C) multistory channel body, and (D) amalgamated channel complex examples are shown with a schematic map view of the depositionalsetting and cross sectional view of the body. Classification is modified from Cole and Cumella (2005). Figure is modified from Pranter et al.(2009), with permission of AAPG.


sinuosity rivers (Ellison, 2004; Cole and Cumella,2005; Pranter et al., 2007, 2009; Miall, 2010; Pranterand Sommer 2011). These deposits commonly fineupward and consist of fine- to medium-grainedcross-bedded to ripple-laminated sandstone litho-facies with mudchip lags at the base sourced bynearby cutbank erosion. As measured in MesaverdeGroup outcrops at Coal Canyon, Colorado, single-story channel bodies average 9 ft (3 m) in thickness(Cole and Cumella, 2005). Point bar deposits com-monly exhibit relatively lower gamma-ray values ascompared with other architectural element. The GRvalues can decrease upward and reflect the upward-fining character of the deposit unless several de-posits are amalgamated. Single-story channel bodiescommonly have relatively higher effective porosityand bulk density values that are slightly greater thancrevasse splays and floodplain deposits.

Multistory channel bodies consist of fine- tomedium-grained cross-bedded to ripple-laminatedsandstone lithofacies, both of which have the po-tential to be convoluted or bioturbated. These depositsare characterized by stacked individual channels thatare made recognizable by the presence of typicallyapproximately two to five internal scours, whichmay have a mudchip lag present at their base. Thiscomplex internal architecture suggests that thesechannel bodies, which are observed in outcrop at CoalCanyon,Colorado, to average 14 ft (4m) in thickness,were deposited by dynamic fluvial channels in a ro-bust sinuous channel system with significant me-ander belts (Cole and Cumella, 2005). Amalgamatedchannel bodies are composed of stacked assemblagesof single-story and multistory channel bodies. InMesaverde Group outcrops at Plateau Creek Canyon,Colorado, they have been observed to be thicker

Table 1. Key Fluvial Architectural Elements of the Mesaverde Group Based on Cores from the Piceance Basin, Colorado and Uinta Basin,Utah

Architectural Element ClassPrinciple FaciesAssemblage Description Interpretation

Floodplain MF, ML, MC, STAM Mud-dominated facies assemblage thattypically features high degrees ofbioturbation in the form of carbonaceousroot traces, which are commonly iron-stained.

Floodplain

Crevasse splay STSC, STWR, SSTC, SR Typically, a coarsening-upward faciessequence with a gradational basal contactand a sharp overlying contact.Bioturbation in the form of rooting andinsect burrowing is common.

Crevasse splay

Single-story channel body SL, SC, SS, SR Singular, fine- to medium-grained sand-dominated fining-upward sequence withsharp basal contact commonly featuringmud clasts.

Isolated point bar deposit inhigh-sinuosity fluvial system

Multistory channel body SL, SC, SS, SR Consists of multiple stacked individualchannel bodies made recognizable by thepresence of between two and five internalscour surfaces.

Stacked individual channels ina robust sinuous channelfluvial system

Amalgamated channel body SL, SC, SS, SR A complex consisting of stacked multistoryand single-story channel bodies typicallymuch larger in scale than a singlemultistory channel body.

Fluvial channel deposits in alow- to medium-sinuositybraided fluvial system

Abbreviations: MC = carbonaceous mudstone; MF = highly fissile mudstone; ML = laminated to mottled mudstone; SC = convoluted sandstone; SL = low- to high-angle cross-bedded sandstone; SR = ripple laminated sandstone; SS = structureless, cryptically bioturbated sandstone; SSTC = convoluted silty sandstone; STAM = argillaceoussiltstone; STSC = convoluted sandy siltstone; STWR = wavy laminated sandy siltstone.


(average thickness of 26 ft [8 m]) and more laterallyextensive (average width of 870 ft [265 m]) thanmultistory channel bodies and are interpreted to havebeen deposited by a low- to medium-sinuosity braidedriver system in an alluvial plain setting (Lorenz andNandon, 2002; Patterson et al., 2003; German, 2006).

Measuring Success of Classification Methods

A means of judging the trained classifiers’ ability tocorrectly predict electrofacies classes in test wells isneeded to evaluate and compare classifier perform-ances. A commonly used visualization tool, typicallyused in evaluating supervised learning procedures inthe field of artificial intelligence, known as a confusionmatrix was selected for this purpose (Ting, 2011). Aconfusion matrix displays information about the actualand predicted classifications present in a classificationsystem, where each column represents instances ofthe predicted class and each row represents instances ofthe actual class (Kohavi andProvost, 1998;Ting, 2011).Data organized in this manner not only display cor-rectly predicted classes but also allow for character-ization of erroneous interclass predictions (Foody, 2002).

To quantitatively compare performances be-tween variably trained classifier outcomes, fourmetrics of success were derived from the content ofthe confusion matrix: overall accuracy, user’s ac-curacy, predicted volume, and average deviation.Overall accuracy is the simplest and one of the mostcommon accuracy measurements used in confusionmatrix analysis (Congalton, 1991). It is calculated bydividing the number of correctly predicted classes(the sum of the major diagonal in the confusion ma-trix) by the total number of predicted classes. To mea-sure the accuracy of individual class predictions, asimilar calculation to overall accuracy is derivedfrom the confusion matrix whereby the number ofcorrectly predicted instances of a particular class isdivided by the total number of actual samples thatexist for that class. This success metric is known asuser’s accuracy and is commonly employed in con-junction with overall accuracy when an emphasis isplaced on the accuracy of individual class predictionsas is the case in this study (Janssen and van der Wel,1994; Foody, 2002). In addition to measures of ac-curacy, it is also desirable that the number of pre-dicted samples of a particular class is similar to the

number of actual samples for that class so that themodel accurately represents the actual volumetricdistribution of facies (Dubois et al., 2007). To evaluatethe volumetric distributions of individual classes, thetotal number of predicted samples for a particular classis divided by the number of actual samples for thatclass. This ratio is multiplied by 100 to give a measurein terms of percentage of the correctness of thevolumetric distribution for individual classes termedin this study as “predicted volume.” To evaluate cor-rectness of the volumetric distribution of facies for themodel as a whole, an average is taken of the absolutedifferences of the predicted volumes of each individualclass from the ideal percentage (100%). This measureof central tendency is known as the average deviation.

The k-Nearest Neighbor Approach

In the k-NN algorithm, the training phase consistssimply of assigning and storing class labels to trainingsamples, which are vectors in n-dimensional space. Inthe classification phase, an unclassified sample (a testsample or query point) is plotted among the trainingdata in n-dimensional space and is compared with auser-defined constant number (k) of the most similartraining samples (nearest neighbors). The test sampleis then classified according to the most commonlyoccurring class out of this k number of nearest train-ing samples (majority rules) (Cover and Hart, 1967;Dubois et al., 2007) (Figure 5). This algorithm issimilar to, but not to be confused with, the populark-means clustering algorithm.

Training samples (N = 440) with known archi-tectural element classifications were selected at ran-dom throughout the four training cores where dataquality was deemed satisfactory (high quality of as-sociated wireline data, reliable core-to-log depth shiftcorrection, and confidently chosen architectural ele-ment classifications). The n-dimensional space that thetraining samples are plotted in corresponds to thenumber of different well log curves that are used, witheach sampling point being associated with the well logvalues at its respective depth. To investigate the ef-fectiveness of the well log curves chosen for this study(GR, RHOB, ILD, and NPHI) in distinguishing be-tween architectural element classes, the performanceof the classifier as trained by seven different well logcurve assemblageswas evaluated. These sevenwell log


curve assemblages consist of (1) GR and RHOB; (2)GR and ILD; (3) GR and NPHI; (4) GR, RHOB, andILD; (5) GR, RHOB, and NPHI; (6) GR, NPHI, andILD; and (7)GR, RHOB, ILD, andNPHI. TheGR logcurve was left static in these assemblages because it ispresent in all of the selected noncored study wells.

Determining the number of nearest neighbors (k)to examine when classifying the test sample can be adelicate choice. If the value of k is too small, there isthe potential for outlying sampling points to have agreater influenceon test sample classification. If the valueof k is too large, there is the tendency for classes thatare associated with larger sampling populations tohave an overwhelming influence on test sample classi-fication (Drummond et al., 2010). Because there isnot a widely accepted formula for determining theoptimum k value and attempts at such are compli-cated, the example of Drummond et al. (2010) wasfollowed where a series of 5 different k values (5,10, 15, 20, and 25) were tested (Hall et al., 2008).These k values were used in conjunction with 7 dif-ferent log curve assemblages to create 35 uniquelytrained classifier models. After initial evaluation ofthe trained classifiers’ ability to predict the five originalarchitectural element classes as described in core, a

simplified four-class architectural element realizationwas created and tested in the manner described above.

Probabilistic Clustering Procedure

In the PCP, an initial model is created wherein core-defined architectural element classes and their asso-ciatedwell log curve values are stored as sampling points.For consistency, the same core depth sampling pointsas well as the same well log curve assemblages usedin training the k-NN classifier were used to train thePCP classifier. In the imposed model, the frequencydistribution of each selected well log curve is segregat-ed into a user-defined number of pseudo-Gaussiandistributions that represent each desired electrofaciesclass (N = 4 and 5 in this study). These pseudo-Gaussian distributions are plotted as clusters or modesin n-dimensional space, with n being dependent onthe number of well log curves used (Vallejo, 2010;Eslinger and Everett, 2012). During the clusteringprocess a probability density function is employed tocalculate the likelihood that each sample belongs toa particular electrofacies class. The sampling pointsobtained from core description that are used inthe calibration of the model are initially assigned aprobability of 1.0. Data from the testing wells werewithheld from the calibration process so that theywould not influence the initialmodel. To incorporatethe data from the test cores so that electrofacies classpredictions can be made, at least one adjusting cal-culation (one iteration) must be made in which anew probability density function is computed. Thisiteration not only assigns electrofacies class proba-bilities in the test cores but may also change theclass assignment probabilities of the training samples.Generally, when greater numbers of iterations arecomputed, there exists a higher chance that the trainingsamples will be reassigned to a new electrofacies class(Vallejo, 2010; Eslinger and Boyle, 2011). For thisreason, all of the models tested using the PCP in thisstudy were only allowed the minimum requirementof one iteration. The final classification given to samplesin the testing wells is determined by the electrofaciesclass that is calculated to have the highest probability.

Thickness Criteria Approach

Past studies of the Mesaverde Group in the Piceance Ba-sin have established relationships between the fluvial

Figure 5. A k-nearest neighbor algorithm example. Test sam-ples (stars) are plotted among 20 randomly selected trainingsamples in a bulk density (RHOB) versus gamma ray cross plot. Ifthe first nearest neighbor (k = 1) is chosen to determine the classof the test sample, as demonstrated by the smaller circles, thenboth samples are classified as single-story channel bodies. However,if k is increased to 3, the test sample on the left would remainclassified as a single-story channel body, whereas the test sampleon the right would be classified as a crevasse splay. This dem-onstrates the potential for confusion at class boundaries.


channel body and crevasse splay architectural elementsand their respective thicknesses (e.g., Cole andCumella,2005; Pranter et al., 2009).

These relationships were incorporated in the re-finement of classification results by applying a meth-od involving an indicator flag, which pairs theresults of the electrofacies classifiers with thicknesscriteria. For the thickness criteria approach, thethickness of a user-defined well log value is count-ed from top to bottom, whereas unspecified valuesor gaps can be overlooked if they fall below a pre-defined thickness limit. If the well log value beingcounted passes a gross thickness requirement(thickness including gaps that fall below thepredefined thickness limit) and a net thicknessrequirement (thickness of user-defined value ex-cluding gaps), then an indicator flag is created de-noting a newly assigned class (Figure 6). This methodwas applied to the k-NN classification results forthe four-class architectural element realization tohelp distinguish between the crevasse splay, single-story channel body, and multistory channel bodyclasses. Thickness criteria are based on average thick-ness values of architectural elements as observed inoutcrop at Coal Canyon, Colorado, by Cole andCumella (2005). The workflow for the thickness cri-teria approach is as follows. (1) Test wells in whichelectrofacies have been predicted by the k-NN clas-sifier are used (as log curves in .las file format) with eachpredicted electrofacies corresponding to a represen-tative value (1 = floodplain, 2 = crevasse splay, 3 =single-story channel body, and 4 = multistory channelbody). (2) A multistory channel flag is created using agross thickness requirement of 12 ft (4m), which is setat roughly 2 ft (0.61 m) less than the average thicknessof a multistory channel body (14 ft [4 m]) such thatthe thickness requirement is slightly more inclusivethan the average thickness value (Figure 6). A netthickness requirement of 9 ft (3 m) was deemed ap-propriate based on visual inspection of classifier resultsin well log form, and the maximum acceptable gap wasset at 4 ft (1 m) based on previous use of this method(Figure 6). (3) After the multistory channel body flaghas been established, a three-class electrofacies logcurve is created, which consists of (i) the floodplainelectrofacies; (ii) a new class that combines the crevassesplay electrofacies, the single-story channel body elec-trofacies, and any interval that was previously clas-sified as a multistory channel body but did not pass

the thickness requirements necessary to be consideredpart of the multistory flag; and (iii) the multistorychannel body flag just created. (4) A single-storychannel body indicator flag is then applied to thisnew simplified electrofacies log curve using a grossthickness requirement of 7 ft (2 m), which, like wasdone for the multistory channel body indicator flag,is set at approximately 2 ft (~0.6 m) less than theaverage thickness of a single-story channel body (9 ft[3 m]) such that the thickness requirement is slightlymore inclusive than the average thickness value. A netthickness value of 4 ft (1m)was chosenbased on visualinspection of classifier results, and the maximum ac-ceptable thickness gapwas kept to 0.5 ft (0.15m) so asto eliminate any small erroneous classification spikeswhile not allowing the grouping of closely stackedsimilar electrofacies. Within this three-class elec-trofacies log curve, if an interval of values belongingto the second class (ii) does not meet the minimumthickness requirements to be flagged as a single-storychannel body, it is classified as a crevasse splay. (5) Afinal electrofacies log curve is created that merges thetwo channel flags with the floodplain and crevassesplay classes. A coal flag was also created with thecriteria GR < 75 API and RHOB < 2.1 g/cm3 andmerged along with these classes into the final elec-trofacies log curve.

Figure 6. Example illustrating the process by which a well logindicator flag for a multistory channel body is created.


RESULTS

Through an approach that combined selected clas-sifier results with the thickness criteria approach, anoverall accuracy of 83.6% was achieved. The in-dividual architectural elements of the simplified four-class architectural element realization were predictedwith user’s accuracies of 0.891, 0.376, 0.735, and0.985 for the floodplain, crevasse splay, single-storychannel body, and multistory channel body classes,respectively. Such a methodology could provide analternative approach to the classification of fluvial

architectural elements in commonly available welllogs (Figure 7).

Prediction Performance of the k-NearestNeighbor Method

Evaluation of the variably trained (varying log curveassemblages and k values) k-NN classifiers’ ability tocorrectly predict the occurrence of the five architec-tural elements as they exist in the testing wells ren-dered a best case overall accuracy of 63.8% withthe floodplain, crevasse splay, single-story channelbody, multistory channel body, and amalgamatedchannel body yielding individual user’s accuracies of0.897, 0.624, 0.026, 0.228, and 0.904, respectively(Figure 8). Although the k-NN classifier fared quitewell in its ability to predict the floodplain, crevassesplay, and amalgamated channel body architecturalelements, the single-story channel body and multi-story channel body architectural elements were rare-ly predicted correctly. An examination of interclassconfusion in all of the k-NN tests revealed a highdegree of confusion between the single-story channelbody and crevasse splay architectural elements with132 out of the 189 total single-story channel bodysamples being incorrectly classified as the crevassesplay class. A high degree of confusion between themultistory channel body and amalgamated channelbody architectural element was also noted with 92out of 158 multistory channel body samples beingincorrectly classified as the amalgamated channelbody class. This confusion is also represented by thehigh predicted volumes of the crevasse splay andamalgamated channel body classes and the low pre-dicted volumes of the single-story channel body andmultistory channel body classes, which contributed toan average deviation value of 62.6 (Figure 8). A newsimplified four-class architectural element realizationwas developed in which the highly confused and geo-logically similar multistory channel body and amalgam-ated channel body classes were combined intothe same multistory channel body class. Evaluation ofthe k-NN classifier’s ability to predict the simplifiedfour-class realization resulted in an improved best caseoverall accuracy of 74.5% with the floodplain, crevassesplay, single-story channel body, and multistory chan-nel body yielding individual user’s accuracies of 0.926,0.539, 0.069, and 0.932, respectively (Figure 9).Simplification of the architectural element classes also

Figure 7. The well log curve of the far right column features theresulting architectural elements as predicted in a section of testwell, Kerr–McGee 21 Natural Buttes, by the coupling of the welllog indicator flag approach with the results of the k-nearestneighbor classifier as trained by the gamma ray (GR) and bulkdensity (RHOB) well log curves. This is compared with a well logcurve to its left in which the actual core-described architecturalelements are featured.


led to an improvement in interclass confusionmarkedby a more desirable predicted volume for the newlyaggregated multistory class and a consequently loweraverage deviation (45.9) compared with the pre-viously tested five-class architectural element realiza-tion. However, the confusion between the single-storychannel body and crevasse splay architectural ele-ment classes still remained highwith 111 out 189 totalsingle-story channel body samples being incorrectly clas-sified as the crevasse splay class (Figure 9). Observationsof the success of the varyingwell log curve assemblageand k values in distinguishing the original five ar-chitectural element classes were ignored in favor offocusing on the significance of these variables in dis-tinguishing between the classes of the more geologicallyrealistic four-class architectural element realization.

The range in overall accuracy values produced bythe seven well log curve assemblages tested was notlarge, andnoonewell log curve assemblagepredicted anindividual architectural element class markedly betterthan the others. However, the well log curve assem-blages of GR and RHOB and GR, RHOB, and ILDare associated with the identical highest overall accu-racies achieved. The addition of the ILD well log curveseemed to have had little to no effect on the outcomeof the predictions, indicating that the GR and RHOBcurves are primarily responsible for the success of thepredictions (Figure 10). Testing of the five different kvalues (5, 10, 15, 20, and 25) with the various well logcurve assemblages revealed a general increase in overallaccuracy up to the k value of 20 followed by a declinein overall accuracy at the k value of 25 in the majority

Figure 8. (A) User’s accuraciesfor the five-class architectural-element realization produced bythe k-nearest neighbor (k-NN)classifier trained by the log-curve assemblage gamma ray(GR), bulk density (RHOB),neutron porosity (NPHI), and kvalue of 20—one of two variablytrained k-NN classifiers to pro-duce the highest overall accuracyachieved (63.8%) for the five-classarchitectural element realization.(B) The confusion matrix asso-ciated with the k-NN classifier thatwas trained using the well logcurves GR, RHOB, true resistivity(RT), and NPHI and a k value of 20,which produced an overall accu-racy of 63.8% achieved when at-tempting to predict the fivearchitectural element classes.Highlighted cells represent thenumber of correctly predictedsamples out of the actual classsum, whereas the other cellswithin the row represent themisclassified architectural ele-ments. AM = amalgamatedchannel body; CS = crevassesplay; FP = floodplain; MS =multistory channel body; SS =single-story channel body.


of cases (Figure 11). This trend in overall accuracy ismirrored by the user’s accuracies of the individualarchitectural elements with the exception of the single-story channel body, which commonly decreases inuser’s accuracy with successively larger k values. Theclassifiers trained by the combination of these optimalvariables of k = 20 and the well log curve assemblageof GR and RHOB are responsible for producing thebest case results for the four-class architectural ele-ment realizations that were previously mentioned.

Prediction Performance of the ProbabilisticClustering Procedure Classifier

Like the k-NNmethod, when evaluated for its abilityto distinguish the original five architectural elements,

the PCP proved capable of predicting the floodplain,crevasse splay, and amalgamated channel body ar-chitectural elements with moderate to high user’saccuracies, which contributed to a best case overallaccuracy of 62%. However, the single-story channelbody and multistory channel body were rarely pre-dicted correctly by the PCP (Figure 12). Analysis ofinterclass confusion revealed that the single-storychannel body was commonly incorrectly predictedas either the crevasse splay or amalgamated channelbody architectural elements. A high degree of con-fusion between the multistory channel body andamalgamated channel body architectural elementswas also observed with 93 out of 158 multistorychannel body samples being incorrectly classified asthe amalgamated channel body class. This confusion

Figure 9. (A) User’s accuraciesfor the four-class architecturalelement realization produced bythe k-nearest neighbor (k-NN)classifier trained by the log curveassemblage gamma ray (GR),bulk density (RHOB), and k valueof 20—one of two variably trainedk-NN classifiers to produce thehighest overall accuracy achieved(74.5%) for the four-classarchitectural-element realization.These user accuracies are similarto those achieved by the otherwell log-curve assemblages forthe k-NN classifier. (B) The con-fusion matrix associated with thek-NN classifier that was trainedusing the well log curve assem-blage GR, RHOB, and a k value of20, which produced an overallaccuracy of 74.5% achievedwhen attempting to predict thefour architectural elementclasses. Highlighted cells repre-sent the number of correctlypredicted samples out of theactual class sum, whereas theother cells within the row rep-resent the misclassified architec-tural elements. CS = crevassesplay; FP = floodplain; MS =multistory channel body; SS =single-story channel body.


is also represented by the high predicted volumes ofthe crevasse splay and amalgamated channel bodyclasses and the low predicted volumes of the single-story and multistory channel body classes, whichcontributed to an average deviation value of 52.6(Figure 12). As was done during the testing of thek-NN method, the PCP was tested for its abilityto predict classes in a simplified four-class archi-tectural element realization, which combines thehighly confused and geologically similar multistorychannel body and amalgamated channel body classes.Once again, the simplified four-class architecturalelement realization lent itself to the alleviation ofsome degree of interclass confusion denoted by alower average deviation value (42.7) attributed tomore ideal predicted volumes of the newly aggregatedmultistory class (Figure 13). Though success metricsfor the multistory class improved, the single-storychannel body is still almost never predicted cor-rectly (Figure 13). Of the seven well log curve as-semblages tested, none produced dramatically betteroverall accuracies than the others; however, althoughthe difference is small, the well log curve assemblagesGR and RHOB and GR, RHOB, and ILD are asso-ciated with the highest overall accuracies achieved bythe PCP classifier of 72.8% (Figure 10). Again, the ILDwell log curve seemed to have little to no effect on theaccuracy of the predictions suggesting the GR and

RHOB well log curves are primarily responsible forthe success of the predictions. With the PCP clas-sifier, none of the seven well log curve assemblagesshowed an advantage over the others in their abilityto predict the individual architectural element classeswith the exception that those assemblages that in-cluded the RHOB curve were more effective at pre-dicting the crevasse splay architectural elementclass than those that did not include it.

Refinement of Classifier Results with theThickness Criteria Approach

Though both classifiers were shown to be capable ofpredicting the floodplain, crevasse splay, and multi-story channel body classes with satisfying accuracy inthe simplified four-class architectural element real-ization, attempts by both classifiers failed to accurate-ly predict the single-story channel body class, whichwas always highly confused for the crevasse splayand/or multistory channel body classes. After apply-ing the thickness criteria approach to the classifierresult having the best overall accuracy (k-NN clas-sifier trained by well log curve assemblage GR andRHOB and k = 20), the previously troublesome single-story channel body class was now predicted with a0.735 user’s accuracy, bringing the overall accuracy to

Figure 10. Overall accuracyfor each well log curve assem-blage for the four-class architec-tural element realization usingthe k-nearest neighbor (k-NN)and probabilistic clustering pro-cedure (PCP) methods. GR =gamma ray; ILD = deep re-sistivity; NPHI = neutron porosity;RHOB = bulk density.


a new high of 83.6%. The improvement in accuracymetrics is owed to a reduction in interclass confusionmade possible by the incorporation of thicknesscriteria. The single-story channel body class was nowmuch less commonly misclassified as the crevassesplay class, a change that is reflected in more idealpredicted volumes and average deviation value(Table 2). The user’s accuracies of the other archi-tectural element classes remained relatively unchangedcompared with their success in the k-NN method,with the exception of the crevasse splay, which de-creased slightly. A depiction of the well log curvesgenerated by the thickness criteria approach ascompared with well log curves representing archi-tectural elements as described in the test cores isshown in Figure 14.

DISCUSSION

Combining Alike Architectural Elements

The resulting architectural element predictions ofthe two classification methods displayed many sim-ilarities in terms of the accuracies that were achieved

as well as the interclass confusion that was revealedupon examination of the confusion matrices. Whentested for their ability to distinguish the five archi-tectural elements that were originally described incore, both the k-NNandPCP classifiers showed a highdegree of confusion between the multistory channelbody and amalgamated channel body classes. In thisstudy, the use of the term amalgamated channel bodywas influenced by the observations of this class byCole and Cumella (2005) at Main Canyon and Pla-teau Creek Canyon, Colorado, and further detailedstudy byGerman (2006)where itwas described as thestacked multistory channel bodies that were thicker(average thickness of 26 ft [8 m]) and more laterallyextensive (average width of 870 ft [1265 m]) than atypical multistory channel body. Beyond scale,which can be difficult to accurately assess in core,the geologic similarity of these two architecturalelement classes made their distinction in core asubjective process. The subjectivity of this distinctionis likely reflected in the high degree of confusionobserved between the two classes. Combining thesimilar multistory channel body and amalgamatedchannel body architectural elements into a single classto create a simplified four-class architectural elementrealization seemed a geologically realistic choice giventhat both of these architectural elements can poten-tially be deposited by low- tomedium-sinuosity fluvialsystems (White et al., 2008). The single-story channelbody class was found to be highly confused with thecrevasse splay class by both classifiers, and combiningthe two classes would have likely also led to improvedmeasures of interclass confusion. However, thesearchitectural elements were kept as discrete classesbecause of their unique depositional and reservoircharacteristics, which would carry significance if themethods used in this study were to be applied infuture mapping and modeling endeavors.

Comparing the Classifiers as Trained by theDifferent Well Log Curve Assemblages

Evaluation of the effectiveness of the two classifiers astrained by the different well log curve assemblageswas based upon their ability to predict the four-classarchitectural element realization. It was reasoned thatmeasures of accuracy would be more meaningful forthismore geologically realistic four-class realization as

Figure 11. Average overall accuracies achieved by each k valuefor the various well log curve assemblages. An average of theoverall accuracies for each k value over the different well log curveassemblages was taken to summarize the trend seen among themajority of the well log curve assemblages when the successivelylarger k values were tested.


compared with the five-class architectural elementrealization. In both the k-NN and PCP classifiers,the seven different well log curve assemblages test-ed resulted in a small range of overall accuraciesachieved by both classifiers with the log curve combi-nation of GR and RHOB being primarily responsi-ble for modestly higher overall accuracies achieved(Figure 10). To avoid the pitfall of making inferencesof the effectiveness of the well log curve assemblagesbased only on the overall accuracy metric, which canpotentially be misleading if a large testing group isvery well predicted (e.g., the floodplain or multi-story channel body testing groups), user’s accuracy ofthe individual architectural elements must also be

inspected. Any differences in the user’s accuracies ofthe individual architectural elements produced by thedifferent well log curve assemblages in the k-NNclassifier are negligible; they are just slightly moreaccurate when trained by the well log curve assem-blage GR and RHOB as suggested by the overallaccuracy metric. The agreement between the overallaccuracy and user’s accuracy metrics produced bythe different well log curve assemblages suggests thatthe log curve assemblage does not make an immensedifference but is slightly more accurate with the com-bination of GR and RHOB. This is a beneficial reve-lation for possible future application of this methodin the fluvial deposits of the Mesaverde Group

Figure 12. (A) User’s accu-racies for the five-class architec-tural element realizationproduced by probabilistic clus-tering procedure (PCP) classifiertrained by the log curve assem-blage gamma ray (GR), bulkdensity (RHOB), and neutronporosity (NPHI)—one of two logcurve assemblages to producethe highest overall accuracyachieved (62%) by the PCPclassifier for the five-class archi-tectural element realization.These user accuracies are similarto those achieved by the otherwell log curve assemblages forthe PCP. (B) The confusion matrixassociated with the PCP classifierthat was trained using the well logcurves GR, RHOB, and NPHI,which produced an overall ac-curacy of 62% achieved whenattempting to predict the fivearchitectural element classes.Highlighted cells represent thenumber of correctly predictedsamples out of the actual classsum, whereas the other cellswithin the row represent themisclassified architectural ele-ments. AM = amalgamatedchannel body; CS = crevassesplay; FP = floodplain; MS =multistory channel body; SS =single-story channel body.


where well log curve suites can vary substantially.When user’s accuracy was evaluated in the outcomesof the PCP classifier, it was found that the well logcurve assemblages that included the RHOB well logcurve were roughly 20% more effective at predict-ing the crevasse splay class than those that did notcontain RHOB. The approximately 20% swing ob-served in the user’s accuracy of the crevasse splayclass while the overall accuracy concurrently re-sponds to a much lesser degree highlights the po-tential for disarity between the two metrics and theneed for the simultaneous evaluation of both. Astudy in the nearby Mamm Creek field, PiceanceBasin, Colorado, also emphasized GR and RHOBas the most important well log curve assemblage in amethod devised to predict diagenetic facies in a single

well experiment in the upper part of the MesaverdeGroup (Ozkan et al., 2011). The authors speculatedthat aside from diagenetic facies, the method may alsobe useful in predicting the crevasse splay architecturalelement through its recognition of fine-grained andsometimes clay matrix–rich siltstones and sandstones(Ozkan et al., 2011).

As a whole, the top overall accuracy for eachwell log curve assemblage for the k-NN classifierwas, in most cases, slightly higher than the overallaccuracy produced by the same well log curve as-semblage in the PCP classifier (Figure 10); however,the results are similar. The k-NN classifier alsodid not show the same RHOB well log curve–dependent contrast in its ability to predict thecrevasse splay class as was observed in the PCP

Figure 13. (A) User’s accu-racies for the four-class archi-tectural element realizationproduced by probabilistic clus-tering procedure (PCP) classifiertrained by the log curve assem-blage gamma ray (GR) and bulkdensity (RHOB)—one of two logcurve assemblages to producethe highest overall accuracyachieved (72.8%) by the PCPclassifier for the four-class ar-chitectural element realization.(B) The confusion matrix asso-ciated with the PCP classifier thatwas trained using the well logcurve assemblage GR and RHOB,which produced an overall ac-curacy of 72.8% achieved whenattempting to predict the fourarchitectural element classes.Highlighted cells represent thenumber of correctly predictedsamples out of the actual classsum, whereas the other cellswithin the row represent themisclassified architectural ele-ments. CS = crevasse splay; FP =floodplain; MS = multistorychannel body; SS = single-storychannel body.


classifier. Similar to the findings of this study, thecomparative strength of the k-NN classifier to othercommonly used classifiers was also reported byDubois et al. (2007).

The Troublesome Single-Story Channel BodyArchitectural Element Class

The main source of interclass confusion in thefour-class architectural element realization seen inboth classifiers was caused by the consistent mis-classification of the single-story channel body archi-tectural element class. In the k-NN classifier it wasobserved that the troublesome class was better(though still poorly) predicted at low k values, andits accuracy became progressively poorer with eachsuccessive increase in k. A commonly recognized faultin the k-NN algorithm is that if the k value used istoo large, the classes with larger training populationshave the potential to overwhelm classes with smallertraining populations (Drummond et al., 2010). Toinspect if the relatively smaller training populationof the single-story channel body class was having anegative effect on the success of its prediction, a testwas conducted wherein the training populations ofthe different classes were equalized (N = 141 for eachclass), thus taking away the potential for class sizeinfluence. This resulted in a very slight increase inuser’s accuracy values (<6% improvement over theprevious top user’s accuracy) for the single-story chan-nel body. This minimal improvement indicates that ifthere were a training population sample size influence,it was likely insignificant. Equalization of the train-ing populations was also found to be unbeneficial tothe success of lithofacies prediction using the k-NNclassifier in a study by Dubois et al. (2007). Exami-nation of 50 random training samples each of thecrevasse splay and single-story channel body classescross-plotted against the useful well log curves of GRand RHOB illustrates the degree of similarity be-tween the two classes and helps explain why the po-tential existed for such confusion. Cross-plotting oftraining samples of the single-story channel bodyclass and the multistory channel body class also il-lustrates the potential for confusion among theseclasses.

Because diagenetic andmineral composition varyboth geographically and stratgraphically in the fluvial

deposits of the Mesaverde Group (Keighin and Fouch,1981; Johnson and Roberts, 2003; Stroker et al.,2013), without conducting detailed petrographic workit is only possible to broadly speculate on geologicdrivers for the single-story channel body interclassconfusion. It was observed in core that single-storychannel bodies could range from upper very fine tomedium sand grain size. This range in grain sizeoverlaps to some degree with both the crevasse splaygrain sizes (silt to fine sand) and the multistorychannel body grain sizes (upper fine to mediumupper sand). All three of these architectural elementsalso contain the ripple-laminated sandstone litho-facies to varying extents. These common charac-teristics are also recorded in outcrop descriptionsof Mesaverde Group fluvial architectural elementsby Cole and Cumella (2005), White et al. (2008),Harper (2011), Hlava (2011), and Keeton (2012).Though well logs do not measure grain size directlyand are most certainly blind to details that geologistscommonly describe in cores like ripple laminatedsandstones, it is possible that these overlapping at-tributes of grain size and lithofacies are also associatedwith changes in porosity and clay content. Although

Table 2. Confusion Matrix of the Combined k-Nearest Neighborand Thickness Criteria Approaches

Actual ClassPredicted Class Actual

Class SumFP CS SS MS

FP 408 29 7 14 458CS 34 53 0 36 141SS 9 35 139 6 189MS 0 6 0 392 158Column total 451 123 166 446 240Predicted volume 98% 87% 88% 112% 1186

The confusion matrix resulting from the application of the well log indicator flagapproach in refining the results of the k-nearest neighbor (k-NN) classifier thatwas trained using the well log curves gamma ray and bulk density and a k valueof 20. Gray highlighted cells represent the number of correctly predictedsamples out of the actual class sum, whereas the other cells within the rowrepresent the mispredicted architectural elements. The incorporation of thicknesscriteria greatly alleviated interclass confusion involving the misclassification ofthe SS architectural element class. This improvement is reflected in the moreideal predicted volumes achieved and the lowered average deviation value of9.6 as compared with the average deviation value of the prerefined k-NNclassifier of 45.9. This improvement in interclass confusion led to an overallaccuracy of 83.6%.

Abbreviations: CS = crevasse splay; FP = floodplain; MS =multistory channel body;SS = single-story channel body.


Figure 14. Actual and predicted archi-tectural elements in testing wells. Testingwells with architectural elements as clas-sified from core descriptions are in the lefttrack for each of the four wells. These arecompared with the architectural elementsas predicted through the combined ap-proach of integrating select classifier results(k-nearest neighbor) with outcrop-basedarchitectural element thickness criteria(right track for each of the four wells).


porosity and clay content may be difficult for thegeologist to adequately describe in hand sample, theydo affect the response of the GR and RHOB curvesused in this study. Perhaps shared similarities in po-rosity and clay content with parts of the crevasse splayand multistory channel body classes could have con-tributed to the frequentmisclassification of the single-story channel body.

Incorporation of Thickness Criteria andGeologically Constrained ElectrofaciesPrediction

The resulting improvement in single-story channelbody prediction via the thickness criteria approachmakes the methods described in this study a muchmore viable means of predicting the fluvial architec-tural elements of the Mesaverde Group in noncoredwells. An unfortunate byproduct of the thicknesscriteria approach was themoderate (16%) decrease inuser’s accuracy experienced by the crevasse splayclass. This is a consequence of the merging of thecrevasse splay and single-story channel body classesprior to running the single-story channel body thick-ness criteria code described in the methods, whichresulted in crevasse splays that directly overlie or un-derlie a single-story channel body being grouped withthe contiguous single-story channel body. This side ef-fect should be kept in mind during future applicationof this process.

No limit exists to how the thickness criteria of thisprogram can be manipulated to tailor the thicknesscriteria approach to a particular study.As an example,it may be desirable to attempt the prediction andmapping of the large amalgamated channel bodies,deemed “super stories” by some, which are proposedto represent deposition in incised valleys (Whiteet al., 2008). However, it should be cautioned thatthere is always the potential for the well in whichthe architectural elements are being predicted tohave penetrated a large sandstone body at its compar-atively thinner margin, which may result in its misclas-sification as an isolated single-story channel body.

To demonstrate the application of the methodsdescribed in this paper, the results of the k-NN clas-sifier trained by theGR andRHOBwell log curves anda k value of 20, which produced the highest overallaccuracy of all uniquely trained classifiers (74.5%),

were paired with the thickness criteria approach topredict the fluvial architectural elements of the four-class realization in theMesaverdeGroup fluvial intervalof noncored wells (N = 216) throughout the studyarea. Mapping of the cumulative thickness ofthe single-story channel body architectural elementrevealed an area of relatively high cumulativethickness in the northwestern part of the study area(Figure 15). This thickening corresponds to the areain and around Natural Buttes field, Utah, in whichhighly discontinuous, lenticular sandstone bodieshave been recognized as the primary MesaverdeGroup reservoirs (Schmoker et al., 1996; Stancelet al., 2008). When cumulative thickness values ofthe multistory channel body class were mapped, aroughly east-to-west fairway was noted in the southernpart of the Piceance Basin, Colorado (Figure 15).

The extent of well control prevents investigationin this southernmost part; however, it is noted inoutcrop and well logs that a fairway of large amal-gamated channel bodies is present from PlateauCreek Canyon, Colorado, to Parachute field, Col-orado (Foster, 2010) (Figure 15). Perhaps the relativeincrease in cumulative thickness for the multistorychannel body class observed in the generated maps isrelated to this nearby occurrence of large amalgam-ated channel bodies. The cumulative thicknesses ofboth the single-story and multistory channel bodiesare observed to follow general Mesaverde Groupthickness trends where they decrease at the DouglasCreek arch, Colorado, and thicken toward the cen-ter of the Piceance Basin, Colorado (Hettinger andKirschbaum, 2002).

CONCLUSIONS

Although previous studies of the fluvial deposits ofthe upper Mesaverde Group have historically reliedon the analysis of core samples, outcrops, or the man-ual interpretation of well logs to develop architec-tural element classifications, this study explores analternative approach through the use of statisticalclassification methods. Through the combinationof selected classifier results with outcrop-based, ar-chitectural element thickness criteria, an overallaccuracy of 83.6% was achieved for the simplifiedfour-class architectural element realization. The in-dividual architectural elements were predicted with


user’s accuracies of 0.891, 0.376, 0.735, and 0.985for the floodplain, crevasse splay, single-story channelbody, and multistory channel body classes, re-spectively. Combining the thickness criteria ap-proach with the selected classifier results signified avast improvement in the prediction of the single-story

channel body architectural element class, which boththe k-NN and PCP methods on their own wereunable to distinguish as a distinct electrofacies class.

Significant differences in prediction performancewere not elicited by the classifiers as trained by theseven different well log curve assemblages with theexception that the PCP classifier predicted the cre-vasse splay class more competently when the RHOBwell log curve was used. Though the disparity wasmodest, the use of well log curves GR and RHOBwas associated with the highest overall accuraciesachieved by both the k-NN and PCP classifiers. Thecommon occurrence of these well log curves in thestudy area bodes well for future use of the studiedclassification techniques.Unique to the k-NNclassifier,the potential for improved prediction performancebased on choice of k value was also demonstrated,highlighting the benefits of experimenting with thisvariable.

When coupled with the outcrop-based thicknesscriteria approach, themethods described in this studypresent a feasible approach to the classification of thefluvial architectural elements from well logs for theMesaverde Group and formations in other basins.This has implications for futurework that could allowfor the creation of maps and three-dimensional res-ervoir models to be conducted in a timely fashionwith results that are objective and easily reproducible.

REFERENCES CITED

Beaudoin, A., L. Guindon, A. Leboeuf, J. E. Luther,C. H. Ung, S. Cote, and M. C. Lambert, 2005, Biomassmapping of Canadian northern boreal forests using ak-NN approach and sample plots from high resolutionQuickBird images, in D. Werle, ed.: 26th CanadianSymposium on Remote Sensing, Wolfville, Nova Scotia,Canada, June 14–16, 2005, paper 74, 5 p.

Bremner,D., E.Demaine, J. Erickson, J. Iacono, S. Langerman,P. Morin, and G. Toussaint, 2005, Output sensitive al-gorithms for computing nearest neighbor decisionboundaries: Discrete & Computational Geometry, v. 33,p. 593–604, doi:10.1007/s00454-004-1152-0.

Bridge, J. S., 2006, Fluvial facies models: Recent develop-ments, in H. W. Posamentier and R. G. Walker, eds.,Facies models revisited: Tulsa, Oklahoma, SEPM SpecialPublication 84, p. 85–170, doi:10.2110/pec.06.84.0085.

Bridge, J. S., and R. S. Tye, 2000, Interpreting the dimensionsof ancient fluvial channel bars, channels, and channelbelts from wireline-logs and cores: AAPG Bulletin, v. 84,no. 8, p. 1205–1228.

Figure 15. Cumulative thickness maps of channel body ar-chitectural elements for the study interval. The location of PlateauCreek Canyon, Colorado, the western extent of the amalgamatedchannel bodies that continue east to Parachute field, Colorado, isshown in Figure 1.


http://dx.doi.org/10.1007/s00454-004-1152-0

http://dx.doi.org/10.2110/pec.06.84.0085

Busch, J. M., W. G. Fortney, and L. N. Berry, 1987, Deter-mination of lithology from well logs by statisticalanalysis: SPE Formation Evaluation, v. 2, p. 412–418,doi:10.2118/14301-PA.

Cole, R. D., 2008, Characterization of fluvial sand bodies inthe Neslen and lower Farrer Formations (Upper Creta-ceous), Lower Sego Canyon, Utah, in M. W. Longmanand C. D. Morgan, eds., Hydrocarbon systems and pro-duction in the Uinta Basin, Utah: Rocky Mountain As-sociation of Geologists and Utah Geological AssociationPublication 37, p. 81–100.

Cole, R. D., and S. P. Cumella, 2005, Sand-body architecturein the lower Williams Fork Formation (Upper Creta-ceous), Coal Canyon, Colorado, with comparison to thePiceance Basin subsurface: Mountain Geologist, v. 42,no. 3, p. 85–108.

Congalton, R. G., 1991, A review of assessing the accuracyof classifications of remotely sensed data: RemoteSensing of Environment, v. 37, p. 35–46, doi:10.1016/0034-4257(91)90048-B.

Cover, T. M., and P. E. Hart, 1967, Nearest neighbor patternclassification: IEEE Transactions on Information Theory,v. 13, p. 21–27, doi:10.1109/TIT.1967.1053964.

Delfiner, P., O. Peyret, and O. Serra, 1987, Automatic de-termination of lithology from well logs: SPE FormationEvaluation, v. 2, p. 303–310, doi:10.2118/13290-PA.

Doveton, J. H., 1994, Geologic log analysis using computermethods: AAPG Computer Applications in Geology 2,169 p.

Doveton, J. H., and S. E. Prensky, 1992, Geological appli-cations of wireline logs: A synopsis of development andtrends: Log Analyst, v. 33, no. 3, p. 286–303.

Drummond, R. D., A. C. Vidal, J. F. Bueno, and E. P. Leite,2010, Maximum autocorrelation factors applied toelectrofacies classification: 2010 Society of ExplorationGeophysicists Annual Meeting, Denver, Colorado,October 17–22, 2010, p. 1560–1565, doi:10.1190/1.3513138.

Dubois, M. K., G. C. Bohling, and S. Chakrabarti, 2007,Comparison of four approaches to a rock facies classi-fication problem: Computers & Geosciences, v. 33,p. 599–617, doi:10.1016/j.cageo.2006.08.011.

Ellison, A. I., 2004, Numerical modeling of heterogeneitywithin a fluvial point-bar deposit using outcrop andLIDAR data: Williams Fork Formation, Piceance Basin,Colorado, master’s thesis, University of Colorado,Boulder, Colorado, 225 p.

Eslinger, E., and F. Boyle, 2011,Mineral andwell log balance forcomputing porosity in argillaceous rocks—Applications ingas shales: International Unconventional Oil-Gas Con-ference, Qinqdao, China, July 4–5, 2011, p. 1–31.

Eslinger, E., and R. V. Everett, 2012, Petrophysics in gas shales,in J.A. Breyer, ed., Shale reservoirs:Giant resources for the21st century: AAPG Memoir 97, p. 419–451.

Flaig, P. P., P. J. McCarthy, and A. R. Fiorillo, 2011, Atidally-influenced, high-latitude coastal plain: The LateCretaceous (Maastrichtian) Prince Creek Formation,North Slope, Alaska, in S. Davidson, S. Leleu, andC. North, eds., From river to rock record: The preservation

of fluvial sediments and their subsequent interpreta-tion: Tulsa, Oklahoma, SEPM Special Publication 97,p. 233–264, doi:10.2110/sepmsp.097.233.

Foody, G. M., 2002, Status of land cover classification ac-curacy assessment: Remote Sensing of Environment,v. 80, p. 185–201, doi:10.1016/S0034-4257(01)00295-4.

Foster, R., 2010, Sequence stratigraphy of the Upper Creta-ceous Middle Williams Fork Formation, Piceance Basin,northwestern Colorado: Implications for reservoir sand-stones, master’s thesis, University of Colorado, Boulder,Colorado, 78 p.

German, Q. A., 2006, Analysis of fluvial sandstone-bodycharacteristics and architecture in a high net-to-grosssystem: Upper Williams Fork Formation, Plateau CreekCanyon, Piceance Basin, Colorado, master’s thesis,University of Colorado, Boulder, Colorado, p. 1–42.

Grotsch, J., and C. Mercadier, 1999, Integrated 3-D reservoirmodeling based on 3-D seismic: The TertiaryMalampayaand Camago buildups, offshore Palawan, Philippines:AAPG Bulletin, v. 83, no. 11, p. 1703–1728.

Haapanen, R., A. R. Ek, M. E. Bauer, and A. O. Finley, 2004,Delineation of forest/nonforest land use classes usingnearest neighbor methods: Remote Sensing of Environ-ment, v. 89, p. 265–271, doi:10.1016/j.rse.2003.10.002.

Hall, P., B. U. Park, and R. J. Samworth, 2008, Choice ofneighbor order in nearest-neighbor classification: Annals ofStatistics, v. 36, p. 2135–2152, doi:10.1214/07-AOS537.

Harper, E. S., 2011, Fluvial architecture of the lowerWilliamsFork Formation (middle Mesaverde Group), DouglasCreek Arch, Colorado, master’s thesis, University ofColorado, Boulder, Colorado, p. 30–46.

Hemborg, H. T., 2000, Gas production characteristics of theRulison, Grand Valley, Mamm Creek, and Parachutefields, Garfield County, Colorado, turning marginallyeconomic basin-centered tight-gas sands into profitablereservoirs in the southern Piceance Basin: ColoradoGeological Survey Resource Series 39, p. 1–30.

Hettinger, R.D., andM.A.Kirschbaum, 2002, Stratigraphy ofthe upper Cretaceous Mancos Shale (upper part) andMesaverde Group in the southern part of the Uinta andPiceance basins: Utah and Colorado: US GeologicalSurvey Investigation Series I2764, p. 15–17.

Hettinger, R. D., and M. A. Kirschbaum, 2003, Stratigraphyof the Upper Cretaceous Mancos Shale (upper part)and Mesaverde Group in the southern part of the Uintaand Piceance basins, Utah and Colorado: US GeologicalSurvey Digital Data Series DDS-0069-B, 21 p.

Hewlett, A. C., 2010, Fluvial architecture and static con-nectivity of theWilliams Fork Formation, central MammCreek Field, Piceance Basin, Colorado, master’s thesis,University of Colorado, Boulder, Colorado, 211 p.

Hlava, K. S., 2011, Sequence-stratigraphic controls onreservoir-scale architecture of the middle MesaverdeGroup, Douglas Creek Arch, Colorado, master’s thesis,University of Colorado, Boulder, Colorado, 314 p.

Janssen, L. F., and F. M. van der Wel, 1994, Accuracy as-sessment of satellite derived land-cover data: A review:Photogrammetric Engineering and Remote Sensing,v. 60, p. 419–426.


http://dx.doi.org/10.2118/14301-PA

http://dx.doi.org/10.1016/0034-4257(91)90048-B

http://dx.doi.org/10.1016/0034-4257(91)90048-B

http://dx.doi.org/10.1109/TIT.1967.1053964


http://dx.doi.org/10.1190/1.3513138

http://dx.doi.org/10.1190/1.3513138

http://dx.doi.org/10.1016/j.cageo.2006.08.011

http://dx.doi.org/10.2110/sepmsp.097.233

http://dx.doi.org/10.1016/S0034-4257(01)00295-4

http://dx.doi.org/10.1016/j.rse.2003.10.002

http://dx.doi.org/10.1214/07-AOS537

Johnson, R. C., 1989, Geologic history and hydrocarbonpotential of Late Cretaceous age, low- permeabilityreservoirs, Piceance Basin western Colorado: US Geo-logical Survey Bulletin 1787-E, 51 p.

Johnson, R. C., 1990, Development of Laramide-style sub-sidence trends during Late Cretaceous foreland basindeposition, northwest Colorado (abs.): AAPG An-nual Convention, San Francisco, California, June3–6, 1990, accessed September 20, 2012, http://www.searchanddiscovery.com/abstracts/html/1990/annual/abstracts/0687b.htm.

Johnson, R. C., and T. M. Finn, 1986, Cretaceous throughHolocene history of theDouglasCreekArch, inD. S. Stone,ed., New interpretations of northwest Colorado geology:Colorado and Utah: Rocky Mountain Association of Ge-ologists Guidebook, p. 77–95.

Johnson, R. C., and S. B. Roberts, 2003, The Mesaverde totalpetroleum system, Uinta-Piceance, Utah and Colorado:US Geological Survey Digital Data Series DDS-69-B,version 1.0, Chapter 7 [CD-ROM].

Kapur, L., L. Lake, D. Sepehrnoori, D. Herrick, andC. Kalkomey, 1998, Facies prediction from core and logdata using artificial neural network technology, inTransactions of the 39th Society of ProfessionalWell LogAnalysts Annual Logging Symposium, Keystone, Colo-rado, May 26–28, 1998, paper PP, 11 p.

Keeton, G. I., 2012, Sedimentological and stratigraphiccharacteristics of fluvial sandstones based on outcropspectral-gamma-ray data and borehole images, WilliamsFork Formation, Piceance Basin, Colorado, master’sthesis, University of Colorado, Boulder, Colorado, 142 p.

Keighin, C. W., and T. D. Fouch, 1981, Depositional envi-ronments and diagenesis of some nonmarine UpperCretaceous rocks, Uinta Basin, Utah, in F. G. Ethridge andR. M. Flores, eds., Recent and ancient non-marine depo-sitional environments; models for exploration: Tulsa,Oklahoma, SEPMSpecial Publication 31, p. 109–125, doi:10.2110/pec.81.31.0109.

Kohavi, R., and F. Provost, 1998, Special issue on applicationsof machine learning and the knowledge discovery process:Boston, Kluwer Academic, 271 p.

Lawton, T. F., 1983, Tectonic and sedimentologic evolution ofthe Utah foreland basin, Ph.D. thesis, University ofArizona, Tucson, Arizona, 217 p.

Lawton, T. F., 1986, Fluvial systems of the Upper CretaceousMesaverde Group and North Horn Formation, centralUtah - A record of transition from thin-skinned tothick-skinned deformation in the foreland region, inJ. A. Peterson, ed., Paleotectonics and sedimentation inthe Rocky Mountain region, United States: AAPGMemoir 41, p. 423–442.

Lorenz, J. C., and G. C. Nandon, 2002, Braided river depositsin a muddy depositional setting: The Molina Member ofthe Wasatch Formation (Paleogene), west-central, Col-orado, U.S.A.: Journal of Sedimentary Research, v. 72,p. 376–385, doi:10.1306/100801720376.

Lorenz, J. C., and A. K. Rutledge, 1987, Late CretaceousMesaverde Group outcrops at Rifle Gap, PiceanceCreek Basin, northwestern Colorado, in S. S. Beus, ed.,

RockyMountain Section Geological Society of AmericaCentennial Field Guide 2, p. 307–310, doi:10.1130/0-8137-5402-X.307.

McRoberts, R. E., M. D. Nelson, and D. Wendt, 2002,Stratified estimates of forest area using satellite imagery,inventory data, and the k-nearest neighbor technique:Remote Sensing of Environment, v. 82, p. 457–468, doi:10.1016/S0034-4257(02)00064-0.

Miall, A. D., 1978, Lithofacies types and vertical profilemodels in braided river deposits: A summary, inA.D.Miall, ed., Fluvial sedimentology: Canadian Societyof Petroleum Geology Memoir 5, p. 596–604.

Miall, A. D., 1985, Architectural-element analysis: A newmethod of facies analysis applied to fluvial deposits:Earth-Science Reviews, v. 22, p. 261–308, doi:10.1016/0012-8252(85)90001-7.

Miall, A. D., 2010, Alluvial deposits, in N. P. James andR. W. Dalrymple, eds., Facies models 4, 4th ed., St.John’s, Newfoundland, Canada, Geological Associationof Canada, p. 105–137.

Ozkan, A., S. P. Cumella, K. L. Milliken, and S. E. Laubach,2011, Prediction of lithofacies and reservoir quality usingwell logs, Late Cretaceous Williams Fork Formation,Mamm Creek field, Piceance Basin, Colorado: AAPGBulletin, v. 95, no. 10, p. 1699–1723, doi:10.1306/01191109143.

Panjaitan, H., 2006, Sand-body dimensions in outcrop andsubsurface, lower Williams Fork Formation, PiceanceBasin, Colorado, master’s thesis, Colorado School ofMines, Golden, Colorado, 170 p.

Patterson, P. E., K. Kronmueller, and T. D. Davies, 2003,Sequence stratigraphy of theMesaverdeGroup andOhioCreek conglomerate, northern Piceance Basin, Colorado,inK.M. Peterson, T. M. Olson, andD. S. Anderson, eds.,Piceance Basin 2003 guidebook: Denver, Colorado,Rocky Mountain Association of Geologists, p. 115–129.

Pranter, M. J., R. D. Cole, H. Panjaitan, and N. K. Sommer,2009, Sandstone-body dimensions in a lower coastal-plain depositional setting: Lower Williams Fork For-mation, Coal Canyon, Piceance Basin, Colorado: AAPGBulletin, v. 93, no. 10, p. 1379–1401, doi:10.1306/06240908173.

Pranter, M. J., A. I. Ellison, R. D. Cole, and P. E. Patterson,2007, Analysis and modeling of intermediate-scale res-ervoir heterogeneity based on a fluvial point-bar outcropanalog, Williams Fork Formation, Piceance Basin, Colo-rado: AAPG Bulletin, v. 91, no. 7, p. 1025–1051, doi:10.1306/02010706102.

Pranter, M. J., A. C. Hewlett, R. D. Cole, H. Wang, andJ. R. Gilman, 2014, Fluvial architecture and connectivityof theWilliams Fork Formation:Use of outcrop analoguesfor stratigraphic characterisation and reservoir model-ling, in T. Good, J. Howell, and A. W. Martinius, eds.,Sediment-body geometry and heterogeneity: Analoguestudies for modelling the subsurface: Geological Society,London, Special Publications 2014, v. 387, p. 57–83, doi:10.1144/SP387.1.

Pranter,M. J., andN. K. Sommer, 2011, Static connectivity offluvial sandstones in a lower coastal-plain setting: An


http://www.searchanddiscovery.com/abstracts/html/1990/annual/abstracts/0687b.htm



http://dx.doi.org/10.2110/pec.81.31.0109

http://dx.doi.org/10.1306/100801720376

http://dx.doi.org/10.1130/0-8137-5402-X.307

http://dx.doi.org/10.1130/0-8137-5402-X.307

http://dx.doi.org/10.1016/S0034-4257(02)00064-0

http://dx.doi.org/10.1016/0012-8252(85)90001-7

http://dx.doi.org/10.1016/0012-8252(85)90001-7

http://dx.doi.org/10.1306/01191109143

http://dx.doi.org/10.1306/01191109143

http://dx.doi.org/10.1306/06240908173

http://dx.doi.org/10.1306/06240908173

http://dx.doi.org/10.1306/02010706102

http://dx.doi.org/10.1144/SP387.1

example from theUpperCretaceous lowerWilliams ForkFormation, Piceance Basin, Colorado: AAPG Bulletin,v. 95, no. 6, p. 899–923, doi:10.1306/12091010008.

Retallack, G. J., 2001, Soils of the past: An introduction topaleopedology, 2nd ed., Oxford, United Kingdom,Blackwell, doi:10.1002/9780470698716.

Rider, M., 2002, Facies, sequences and depositional envi-ronments from logs, in M. H. Rider, ed., The geologicalinterpretation of well logs, 2nd ed. Sutherland, Scotland,Rider-French Consulting, p. 226–238.

Rider, M., and D. Laurier, 1979, Sedimentology using a com-puter treatment of well logs: Transactions of the Society ofPetrophysicists and Well Log Analysts, 6th EuropeanSymposium, London, March 1979, Paper J, p. 1–12.

Russell, S. D.,M.Abkar, B. Vissapragada, andG.M.Walkden,2002, Rock types and permeability prediction fromdipmeter and image logs: Shuaiba reservoir (Aptian),AbuDhabi: AAPGBulletin, v. 86, no. 10, p. 1709–1732.

Saggaf, M. M., and E. L. Nebrija, 2000, Estimation of litholo-gies and depositional facies from wire-line logs: AAPGBulletin, v. 84, no. 10, p. 1633–1646.

Schmoker, J. W., T. D. Fouch, and R. R. Carpentier, 1996,Gas in the Uintah Basin, Utah—Resources in continuousaccumulations: Mountain Geologist, v. 33, p. 95–104.

Serra,O., 1989, Sedimentary environments fromwireline logs:2nd ed.: Sugarland, Texas, Schlumberger EducationalServices, p. 119–233.

Serra,O., andH. T. Abbott, 1982, The contribution of loggingdata to sedimentary sedimentology and stratigraphy:Society of Petroleum Engineers Journal, v. 22,p. 117–131, doi:10.2118/9270-PA.

Serra, O., and L. Sulpice, 1975, Sedimentological analysis ofshale-sand series from well logs: Transactions of theSociety of Petrophysicists and Well Log Analysts 16thAnnual Logging Symposium, New Orleans, Louisiana,June 4–7, 1975, paper W, p. 1–23.

Shaak, R. V., 2010, Stratigraphic architecture of shallow-marine to coastal-plain parasequences: Lower WilliamsFork Formation, southeastern Piceance Basin, Colorado,master’s thesis, University of Colorado, Boulder, Colo-rado, 163 p.

Sloan, J. A., 2012, Stratigraphic architecture and connectivityof a low net:gross fluvial system: Combining outcropanalogs andmultiple-point geostatistical modeling, lowerWilliams Fork Formation, Piceance Basin, Colorado,master’s thesis, University of Colorado, Boulder, Colo-rado, 266 p.

Smith, N. D., and J. Rogers, 1999, Fluvial sedimentology VI:Oxford, United Kingdom, Blackwell Science, Interna-tional Association of Sedimentologists Special Pub-lication 28: 488 p., doi:10.1002/9781444304213.

Spencer,C.W., 1995,Uinta-PiceanceBasin Province (020), inD. L. Gautier, G. L. Dolton, K. I. Takahashi, and K. L.Varnes, eds., 1995 national assessment of United Statesoil and gas resources—Results, methodology, and sup-porting data: U.S. Geological Survey Digital Data SeriesDDS-30, Release 2 [CD-ROM]

Stancel, S., J. J. Cuzella, L. MacMillan, and A. Ragas, 2008,Controls on gas production along the southern flank ofGreater Natural Buttes Field: A case study of the LoveArea, in M. W. Longman and C. D. Morgan, eds.,Hydrocarbon systems and production in the UintaBasin, Utah: Rocky Mountain Association of Geologistsand Utah Geological Association Publication 37, p.237–265.

Stroker, T. M., N. B. Harris, C. W. Elliot, and J. M.Wampler,2013, Diagenesis of a tight gas reservoir: UpperMesaverdeGroup, Piceance Basin, Colorado: Marine and Petro-leum Geology, v. 40, p. 48–68, doi:10.1016/j.marpetgeo.2012.08.003.

Sullivan, K. B., J. S. Campbell, and K. E. Dahlberg, 2003,Petrofacies: Enhancing the deepwater reservoir charac-terization effort in West Africa: Petrophysics, v. 44,p. 177–189.

Ting, K. M., 2011, Confusion matrix, in C. Sammut andG. I. Webb, eds., Encyclopedia of machine learning, 1sted.: New York, Springer, 209 p.

White, H., R. Cole, S. Stancel, C. Lee, and L. MacMillan,2008, Window outcrop analogs for Greater NaturalButtes field, inM.W. Longman and C. D. Morgan, eds.,Hydrocarbon systems and production in the UintaBasin, Utah: Rocky Mountain Association of Geologistsand Utah Geological Association Publication 37, p.209–235.

Wolff, M., and J. Pelissier-Combescure, 1982, FACIOLOG-Automatic electrofacies determination: 23rd AnnualLogging SymposiumTransactions: Society of ProfessionalWell Log Analysts, Corpus Christi, Texas, July 6–9,1982, paper FF, 22 p.

Vallejo, J. S., 2010, Prediction of lithofacies in thinly beddedBarnett Shale, using probabilistic methods and clusteringanalysis through GAMLS� (Geologic Analysis viaMaximum Likelihood System), master’s thesis, Uni-versity of Oklahoma, Norman, Oklahoma, 63 p.

Yeung, C.A.,N.Gibbins, andN. Shadbolt, 2008, A k-nearest-neighbor method for classifying web search resultswith data folksonomies: 2008 Institute of Electrical andElectronics Engineers/Web Intelligence Consortium/Association for Computing Machinery InternationalConference on Web Intelligence and Intelligent AgentTechnology, Sydney, New SouthWales, Australia, p. 70–76,doi:10.1109/WIIAT.2008.269.


http://dx.doi.org/10.1306/12091010008

http://dx.doi.org/10.1002/9780470698716


http://dx.doi.org/10.1002/9781444304213

http://dx.doi.org/10.1016/j.marpetgeo.2012.08.003

http://dx.doi.org/10.1016/j.marpetgeo.2012.08.003

http://dx.doi.org/10.1109/WIIAT.2008.269

Documents

Geologically constrained fi fl - oucreate.commatthewpranter.oucreate.com/article_pdf/Allen_Pranter_AAPG_2016.pdf · Geologically constrained electrofacies classification of fluvial