You are on page 1of 14

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa,

September 1994.

PRACTICAL GEOSTATISTICS FOR ON-SITE ANALYSIS A COAL EXAMPLE M.A. Nopp Department of Geology, Rhodes University, Grahamstown, 6140, Republic of South Africa* *Current address: Snowden Mining Industry Consultants, P O Box 2207, Brisbane Qld 4001 mnoppe@snowden.com.au SYNOPSIS Geostatistical analysis provides a powerful tool for enhancing the prediction and decisionmaking capabilities of mine planners and geologists. Much of the literature from the coal industry has focused on results for very large mining blocks, generally equivalent to annual production areas. This work is typically conducted by geostatistical experts, and seldom by the geologist-on-site who tends to apply more traditional techniques. While traditional approaches provide reasonable results for long-term estimation purposes, they do not provide all the requirements for short-term predictions and reconciliations associated with a mine-production environment. The current economic climate demands the optimal utilisation of available reserves. This is achieved by applying effective sampling patterns and procedures and by ensuring that resulting production estimates closely reflect actual achievements. Geostatistical analysis provides the only method for quantifying differences between alternative sampling philosophies as well as providing the errors associated with various reserve tonnage and quality estimates. Providing the fundamentals of geostatistics are understood, a number of geostatistical software packages are available to allow geological data to be analysed by the geologist-on-site. The results and some benefits of such an exercise are presented for calorific value and seam thickness for a particular coal seam. Basic on-site geostatistical processing of data should be encouraged by professional geostatistitions since the geological controls responsible for the geostatistical attributes are often best appreciated by the geologist-on-site, thus providing verification of the results and ultimately increasing the awareness and confidence in such geostatistical methods. INTRODUCTION The majority of practising geologists are involved with mining and/or exploration ventures. As such a significant proportion of the geologists time is devoted to designing and implementing drilling and sampling programmes and using the resultant data to predict (and reconcile) reserve tonnages and qualities on a variety of scales, ranging from life-of-mine areas to daily production figures. This paper highlights the availability and benefits of some basic geostatistical techniques for improving the results emanating from routine processing of geological data. The intention is not to provide a detailed review of available geostatistical techniques and benefits, nor to compare available software packages. The aim is rather to summarise the
Nopp, 1994 Page 1 of 1

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

results and conclusions gained from a laymans exposure to geostatistics, in the hope that those concerned with routine manipulation of geological data, predominantly by applying traditional statistical methods, will be encouraged and empowered to take advantage of the benefits available from geostatistical analysis. Geostatistical data analysis and the resultant benefits for metallic and diamond deposits are widely appreciated and applied, while the opposite appears true for other non-metallic deposits specifically coal deposits. This is in spite of a number of papers published during the late 1970s and the 1980s and reviewed by Armstrong (1989). The latter paper provides a concise review of geostatistical applications in the coal industry. Geostatistical papers relating specifically to South African coal deposits include those by Wood (1976, 1979, 1985), Watson (1977), and Clark (1991). WHY GEOSTATISTICS? The experimental semivariogram provides the only measure of whether a deposit, or part of a deposit, is best analysed using geostatistical methods or whether classical statistics would suffice. This is possible since the experimental semivariogram defines the extent of correlation between data points, and clearly if no correlation or structure is evident from the available data set, then classical statistics will provide adequate answers to most questions. This may be the case where a pure nugget effect exists, where sample sites are relatively widely spaced, or where reserve estimates are for very large mining blocks. David (1977) comments there is no need to apply geostatistical estimation techniques if the range of influence is les than twice the dimensions of the block being estimated. Recent thinking is that if a structure exists, it is best to make use of it when making any estimates (M. Armstrong, pers. comm., 1992). Since ore deposit formation is almost always controlled by spatial factors such as distance from source, or depositional floor topography, results for adjacent points will generally show a degree of correlation. For coal deposits, quality and seam thickness zonation is clearly evident on a variety of scales, and can generally be traced for hundreds of metres if not a few kilometres (Wood, 1976, 1979; Armstrong, 1989; Noppe, 1992). It is also common for such zonation to be anisotropic, with a parameter showing more variation in one specific direction. When estimates are being made for mining blocks on a similar or smaller scale to these trends, that is for short or medium-term planning purposes, then this correlation between nearby points definitely needs to be taken into account. Geostatistics provides the best possible weighting for samples used in reserve estimation so as to produce the lowest possible error of estimation. Of importance to the person involved with routine tonnage and quality estimates is that geostatistics provides the only estimation technique that quantifies the error associated with making such estimates, and therefore allows confidence limits to be placed on these estimates. Conversely, this measure of estimation error allows sampling patterns and spacings to be selected to provide a required degree of confidence in some specific parameter. Coal-related parameters that appear suited to geostatistical analysis include seam thickness, proximate analyses, calorific value, and many of the trace element concentrations. In addition, wash yield, sulphur and some physical parameters such as ash fusion temperature and swell index may also be suitable.

Nopp, 1994 Page 2 of 2

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

The benefits mentioned above highlight why geostatistics, particularly experimental semivariogram construction, should be as much a part and routine of coal reserve estimation and evaluation as the construction of data histograms and geological maps and sections. Obtaining a well-defined experimental and model semivariogram is the basis on which more involved geostatistical techniques rest. The semivariogram parameters (nugget, sill, range, anisotropy orientation) themselves allow sampling and drilling procedures to be refined, while also allowing realistic reporting of estimates by providing errors of estimation, or degrees of confidence, in such estimates. The background to semivariogram construction is succinctly outlined by Clark (1979 a, b). A COAL EXAMPLE The following example illustrates how readily geostatistical techniques can be used to answer two basic queries, namely: What is the error associated with estimating mean values for different sized mining blocks for a given sampling configuration? How can various sampling configurations by objectively compared to select that which will cost-effectively provide estimations at a required confidence level? The second query simply involves iteration of the method applied to the first problem. Data Two parameters were selected, namely seam thickness and specific heat energy, locally referred to as calorific value. The resultant semivariograms achieve sills over the area of study (stationary condition) and thus allow ordinary kriging methods to be applied. Seam elevation data was reviewed but displayed a marked trend and has not been included for this illustration. The database used consisted of 172 actual sample results from a 2x4 kilometre area of the No. 4 Seam of the Highveld Coalfield, with samples distributed as shown in Fig. 1. Factors have been applied to protect the confidentiality of the original data, while this example is summarised from a more detailed account by Noppe (1992). The coal seam was modelled as a two-dimensional body. In addition calorific value and other variable data should be accumulated per sample site using variable x seam thickness x seam density so as to account for changes in thickness and/or density (M. Armstrong, 1989). Data Processing Initial contour plans and histogram construction can be prepared by a variety of means and these assist in defining areas or domains where the deposit behaves in a consistent manner. It is important to identify such domains and where possible work within them when making detailed estimates, since like is then being related to like. Such domains are generally many square kilometres in area for coal deposits. The bulk of the analyses for this example, were performed using public-domain GEO-EAS (Version 1.2.1) software (Englund and Sparks, 1988), while kriging estimations were conducted using PLAYKRIG software. GEO-EAS has some limitations, particularly in terms of the number of data-pairs it can compute, namely 16384, but study areas and lag intervals can be varied to ensure that all data within a given area are applied.

Nopp, 1994 Page 3 of 3

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 1 Sample positions for coal data Following this an iterative procedure allows the construction of the best possible experimental semivariogram in any number of orientations. Once a robust experimental semivariogram is achieved it is then relatively simple to fit a modelled semivariogram using GEO-EAS. The semivariogram model parameters may then be checked and refined using a cross-validation procedure, before the parameters are applied for kriging estimation of block values. Geostatistical estimation makes use of the kriging technique to determine the variance of the estimation error distribution. The kriging procedure also allows a set of sample weighting coefficients to be determined for a given block and sample configuration, so as to minimise the block estimation variance and produce the best linear unbiased estimator (Brooker, 1979). Geostatistical estimation theory relies on three main criteria, namely: Blocks to be estimated must lie within the same geological domain for which the semivariogram has been calculated Sufficient data pairs are used to ensure the experimental semivariogram reliably reflects the variability in the area Mineralisation is homogenous with no distinct trends in the data (stationary condition) so that special kriging techniques are not required.

Nopp, 1994 Page 4 of 4

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Without knowing the actual mean value of a block, ordinary kriging provides estimates of between-block variance, block mean variance and sample weighting coefficients for a given block and sample configuration. In addition the variance within a given block may also be calculated. Since the errors, or kriging variance, associated with estimating the mean value for a block are generally normally distributed, these may readily be interpreted in terms of confidence limits for the block mean estimate. Clearly the probability of a specific block mean exceeding a certain cut-off value may then also be determined. Traditional Analysis Classical statistical results for calorific value and seam thickness are presented in Fig. 2. Although slightly skewed, these omnidirectional distributions approximate the normal distribution with a mean of 24.73 MJ/kg for calorific value (variance 0.258 [MJ/kg]) and a mean of 1.76m for seam thickness (variance 0.0239m). This initial processing does not indicate the presence of more than one domain and obviously is not able to highlight the presence of anisotropic conditions. Contour plans of the variables indicated clear anisotropy in the case of seam thickness. Semivariogram Analysis The limited number of close-spaced data presented some problems in fixing any nugget effect, although close-spaced sampling experience from other coal deposits was applied to improve these estimates. Additional close-spaced sampling should actually be carried out to more accurately determine the nugget effect. The nugget effect is very important for determining estimation variances for point-estimates, but is less important when estimating block means. The latter is particularly valid if the nugget effect is small compared to the sill (David, 1977). Experimental and modelled semivariograms for calorific value and seam thickness are presented in Fig. 3 and Fig. 4. Directional semivariograms were initially calculated in four orientations (0, 45, 90 and 135) in addition to an omnidirectional semivariogram. By varying the calculation orientations on a smaller scale, the exact axes of any anisotropy could be established and quantified. Prior to cross-validation the semivariogram model parameters were established as follow: Calorific Value (exponential and isotropic model) Nugget (Co) = 0.01 (MJ/kg) Sill (C1 + Co) = 0.22 (MJ/kg) Range (a) = 500m Seam thickness (spherical and anisotropic model) Nugget (Co) = 0.005m Sill (C1 + Co) = 0.023m Maximum range (a1) = 750m (azimith = 25) 250m (azimith = 115) Minimum range (a2) = These models are illustrated schematically in Fig. 5 for calorific value and Fig. 6 for seam thickness.

Nopp, 1994 Page 5 of 5

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 2 Classical statistical results for calorific value and seam thickness

Nopp, 1994 Page 6 of 6

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 3 Experimental and model semivariograms for calorific values

Nopp, 1994 Page 7 of 7

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 4 Experimental and model semivariograms for seam thickness

Nopp, 1994 Page 8 of 8

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 5 Geostatistical exponential model for calorific value distribution

Fig. 6 Geostatistical spherical model for seam thickness distribution

Nopp, 1994 Page 9 of 9

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

The respective models should then undergo an iterative process of cross-validation and parameter refinement until the model provides the best results according to the following crossvalidation checks, namely: Good graphical fit to experimental data points Mean error of estimation of the variable close to 0 Minimum mean kriging standard deviation Mean ratio of estimated to kriging variance (Z-score) close to 0, with a standard deviation close to 1. The cross-validation results for the above models are given in Table 1. PARAMETER Mean error Mean kriging (standard deviation) Mean z-score (standard deviation) CALORIFIC VALUE MODEL -0.002 0.356 -0.001 (1.165) Table 1 Cross-validation results Application of Results The geostatistical structural model parameters of the nugget effect, ranges of influence, and anisotropy axes are of utmost importance when designing sampling projects or when estimating point or mining-block values. The model parameters also aid the selection of search ranges and sample weightings for computer-based contouring packages. The structural models may be interpreted at this early stage by applying some rules of thumb. For example some authors believe that an optimal sample spacing should not exceed 67% of the range of influence, while others quote a figure of 80% (Sarkar et al., 1988). The range of influence also provides a minimum search radius value for estimation purposes, while anisotropy orientations may indicate the need for specifically orientated sampling patterns. Since the use of the 80% of the range of influence rule for sample spacing would require some 30% fewer samples than spacing samples at 67% of the range of influence, it is necessary to quantify the benefits of different sampling densities and patterns. This is achieved by determining the estimation error associated with various sample configurations by applying kriging techniques. Estimation Accuracy and Sample Configurations A variety of two-dimensional block sizes were selected as being representative of mining blocks in modern, high-extraction collieries. The blocks represent monthly, weekly or even daily areas for which predictions may be required. A selection of possible sample configurations were then chosen to allow estimation of the effect of the sample density and sample pattern on block estimates. The various configurations used are illustrated in Fig. 7. THICKNESS MODEL 0.000 0.120 -0.001 (1.150)

Nopp, 1994 Page 10 of 10

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 7 Example of sample and block configurations used for kriging variance estimation
Nopp, 1994 Page 11 of 11

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

PLAYKRIG was applied to provide quantitative results using ordinary kriging. The volumevariance relationships for the various blocks at the given model parameters are listed in Table 2. BLOCK mxm CALORIFIC VALUE (MJ/kg)2 BETWEENBLOCK 200 x 200 200 x 20 100 x 100 10 x 10 0.173 0.185 0.190 0.208 WITHINBLOCK 0.047 0.035 0.030 0.012 SEAM THICKNESS m2 BETWEENBLOCK 0.015 0.016 0.017 0.018 WITHINBLOCK 0.008 0.007 0.006 0.005

Table 2 Volume-variance relationship: between-block and within-block variances Assuming a normal distribution for the kriging variance, the estimation error for the block means was calculated as follows (95% confidence level): Estimation error = +- (1.96)(kriging variance)0,5 This estimation error was then expressed as a percentage of the expected block mean to produce a percentage estimation error at a given confidence level. The results are summarised graphically in Figures 8 and 9, where the percentage estimation errors are plotted against the specific sample con-figurations for the different sized mining blocks, for calorific value and seam thickness. If the data and block estimates were not normally distributed, the confidence limits on the mean block estimate could also be expected to be skewed.

Fig. 8 Estimation errors of the block means for various block and sample configurations for calorific value
Nopp, 1994 Page 12 of 12

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

Fig. 9 Estimation errors of the block means for various block and sample configurations for seam thickness Although obvious, it is worth noting that at sample densities of around 10 samples per 100 hectares, each sample not required would save some 10% in sampling and analytical costs. The seam thickness results in Fig. 9 clearly indicate that sample density could be reduced by 58%, from some 12 samples per 100 hectares (400 x 200I configuration) to 5 samples per 100 hectares (750 x 250I configuration), with a resultant increase in the percentage estimation error of only 0.9% (from 7.0% to 7.9% for a 1.76m thick seam for a 100m x 100m mining block). The I in the sample configuration description indicates the grid is orientated at some angle from due north. The quantification of various scenarios as illustrated above allows for objective decisionmaking, not only for selection of the most suitable sample configurations for a specific purpose of parameter, but also by placing a degree of reliability or confidence on geological estimates. CONCLUSIONS The above example highlights only the preliminary benefits that can be gained by applying geostatistical techniques to routine data processing. Of significance is that the above processing and results can be achieved in a matter of days providing the geologists or planners-on-site are familiar with basic geostatistical principles and have the basic software at their disposal. Experimental and model semivariogram construction should be a standard part of any analysis of geological data, whether for a prospect or operating mine. Even a poorly-defined semivariogram provides some quantitative data regarding structure and correlation between sample sites, and it is probably advantageous at an early stage of prospect investigation to gather sufficient data to allow construction of representative semivariograms for the parameters of interest.
Nopp, 1994 Page 13 of 13

Paper in Proceedings of the Mining Geostatistics Conference, Geostatistical Association of South Africa, Kruger National Park, South Africa, September 1994.

The benefit of well-designed drilling programmes reaches beyond the potential cost savings of optimally-designed grids. The improved estimation and contouring ability following from the application of geostatistical techniques will aid geological predictions, mine planning, mine operation, marketing, and plant operation, while boosting client confidence in a mines ability to meet its commitments. In conclusion, it must be remembered that geostatistics is a tool available for improving predictions and assisting with objective decision-making. As such, the geo in geostatistics must always receive a high priority and geostatistics should not be applied blindly, without a thorough appreciation of the geological controls responsible for its attributes. ACKNOWLEDGEMENTS Thanks to Elbie Pretorius, Gertie Horburger and Annike Briedenhann for assisting with the manuscript preparation and presentation, and Richard Shelford and others for their constructive comments. REFERENCES Armstrong, M., 1989. Review of the applications of geostatistics in the coal industry. Geostatistics, Vol. 2. Armstrong M. (Ed.), Kluwer Academic Publishers, 995-1005. Brooker, P.I., 1979. Kriging. Eng. Mining Journal, New York, 180(9), 148-153. Clark, I., 1979a. The semivariogram Part 1. Eng. Mining Journal, New York, 180(7), 90-94. Clark, I., 1979b. The semivariogram Part 2. Eng. Mining Journal, New York, 180(8), 92-97. Clark, I., 1991. A geostatistical approach to estimation of particulate emission levels from coal fired power stations. In: Abstracts; Conference on South Africas coal reserves: their origin, characterisation and utilisation. Geol. Soc. S.Afr. (Eastern Transvaal Branch), Nov. 1991. David, M., 1977. Geostatistical ore reserve estimation. Elsevier, Amsterdam, 364pp. Englund, E. and Sparks, A., 1988. GEO-EAS (Geostatistical Environmental Assessment Software) users guide. U.S. Environmental Protection Agency, Las Vegas, Sept. 1988. Noppe, M.A., 1992. Geological controls for coal exploration and mining. MSc. (Exploration Geology) dissertation (unpubl.), Rhodes Univ., Grahamstown, 99pp. Sarkar, B.C. OLeary, J. and Mill, A.J.B. 1988. An integrated approach to geostatistical evaluation. Mining Magazine, Sept. 1988, 199-207. Watson, M.I., 1977. A study in variogram estimation in the No. 2 Seam, Delmas Colliery. S.Afr. Chamber of Mines, Report No. 55/77. Wood, I.D., 1976. Geostatistical investigation in the No. 2 Seam in the Witbank area. Report 2, S.Afr. Chamber of Mines, Report No. 99/77, 15pp. Wood, I.D., 1979. The geostatistical evaluation of low-ash coal reserves in the No. 2 Seam, Witbank area. S.Afr. Inst. Min. Metall, July 1979, 348-354. Wood, I.D., 1985. An application of geostatistics in the Eastern Transvaal coal fields. Trans.Geol.Soc.S.Afr., 88(1), 81-82.

Nopp, 1994 Page 14 of 14

You might also like