- Hi you all,

Recently I published an article by Dr. Stephen Henley in my newsletter (Earth

Science Computer Applications) called: "Geostatistics - Cracks in the

foundations?". I am looking for comments on the statements he makes about

limitations of geostatistical methods for building ore body models - and

whether these same limitations apply for other types of data. I would like to

publish comments in a future issue. Feel free to make comments - but let me

know if you do *not* want your comments published in the newsletter.

A copy of the article can be found on the InfoMine/Softwaremine site at:

<http://www.infomine.com/softwaremine/articles/>www.infomine.com/softwaremi

ne/articles/

I will be happy to send sample issues of the newsletter if you send me your

mailing address.

Thanks for any help you can give.

Betty

**************************************************

Betty Gibbs Earth Science Software Information

Gibbs Associates

P.O. Box 706, Boulder, CO 80306

Ph & Fax: 303-444-6032

e-mail: gibbsb@... http://www.earthsciswinfo.com

--

* To post a message to the list, send it to ai-geostats@...

* As a general service to the users, please remember to post a summary of any useful responses to your questions.

* To unsubscribe, send an email to majordomo@... with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list

* Support to the list is provided at http://www.ai-geostats.org - Dear Betty, et. al.,

Betty Gibbs wrote:>

I am ignorant of geostatistical methods on ore bodies. However, I have

> Recently I published an article by Dr. Stephen Henley in my newsletter (Earth

> Science Computer Applications) called: "Geostatistics - Cracks in the

> foundations?". I am looking for comments on the statements he makes about

> limitations of geostatistical methods for building ore body models - and

> whether these same limitations apply for other types of data. I would like to

been using variations of geostatistical techniques on geographic data,

especially digital elevation models. Recently, a paper of mine

"Representing multiple spatial statistics in generalized elevation

uncertainty models: moving beyond the variogram" was recently accepted

to be published in the International Journal for Geographic Information

Science.

The paper discusses at length the limitations of traditional

geostatistical techniques for data with measurable surfaces. Of course,

I propose a solution that deals with those limitations. (Otherwise, it

wouldn't be "moving beyond the variogram".)

I don't know how these techniques might be useful to ore bodies, but you

did ask about "limitations for other types of data".

The following contains the abstract.

sincerely, chuck

ABSTRACT: Spatial data uncertainty models (SDUM) are necessary tools

that quantify the reliability of results from geographic information

system (GIS) applications. One technique used by SDUM is Monte Carlo

simulation, a technique that quantifies spatial data and application

uncertainty by determining the possible range of application results. A

complete Monte Carlo SDUM for generalized continuous surfaces typically

has three components: an error magnitude model, a spatial statistical

model defining error shapes, and a heuristic that creates multiple

realizations of error fields added to the generalized elevation map.

This paper introduces a spatial statistical model that represents

multiple statistics simultaneously and weighted against each other. This

papers case study builds a SDUM for a digital elevation model (DEM).

The case study accounts for relevant shape patterns in elevation errors

by reintroducing specific topological shapes, such as ridges and

valleys, in appropriate localized positions. The spatial statistical

model also minimizes topological artifacts, such as cells without

outward drainage and inappropriate gradient dis tributions, created by

the realization heuristic. Multiple weighted spatial statistics enable

two conflicting SDUM philosophies to co-exist. The two philosophies are

`errors are only measured from higher quality data and `SDUM need to

model reality. This article uses an automatic parameter fitting random

field model to initialize Monte Carlo input realizations followed by an

inter-map cell swapping heuristic to adjust the realizations to fit

multiple spatial statistics. The inter-map cell swapping heuristic

allows spatial data uncertainty modelers to choose the appropriate

probability model and weighted multiple spatial statistics which best

represent errors caused by map generalization. This article also

presents a lag based measure to better represent gradient within a SDUM.

This article will cover the inter-map cell swapping heuristic as well as

both probability and spatial statistical models in detail.

--

Chuck Ehlschlaeger 212-772-5321

Dep. of Geography fax: 212-772-5268

Hunter College chuckre@...

695 Park Ave. secure: chuckre@...

New York, NY 10021 http://www.geo.hunter.cuny.edu/~chuck/

"I'm not satisfied with any of the explanations yet, including my own."

-- Dr. Robert Wayne, evolutionary biologist studying the red wolf

--

* To post a message to the list, send it to ai-geostats@...

* As a general service to the users, please remember to post a summary of any useful responses to your questions.

* To unsubscribe, send an email to majordomo@... with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list

* Support to the list is provided at http://www.ai-geostats.org - Dear AI-GEOSTATS members,

Back in April, Betty Gibbs introduced us to an article by Dr. Stephen

Henley entitled "Geostatistics - Cracks in the foundations?". I

responded to AI-GEOSTATS with a message describing an article I wrote

for the International Journal of Geographic Information Science (IJGIS).

http://www.geo.hunter.cuny.edu/~chuck/RGSIBG/beyondVariogramSmallMaps.pdf

is a digital copy of the final draft of that paper.

Other people, including Steve Krajewski, Isobel Clark, Brian Gray, Dan

W. McCarn, and Digby Millikan responded to the Infomine newsletter at

http://www.infomine.com/softwaremine/articles/ with some well thought

out responses to Dr. Henley's paper. Dr. Henley then provided good

feedback to these comments. I learned a lot from "Geostatistics - Cracks

in the foundations?" and about different perspectives from these views.

I noticed that there were two main goals not explicitly stated within

these discussions. Many geostatistical analyses have the goal of

estimating the sum of the magnitude of a variable over a specific region

where the slope of the variable isn't very important. Other analyses,

for example, Dan McCarn's description of uranium deposits, need a good

estimate of a variable's slope in order to provide useful application

results. The following is a snippet from Dan's comment:

"For roll-front uranium deposits, a geostatistical approach is avoided

because of the deposit characteristics. The leading edge of the roll has

generally the highest concentration of uranium and is bounded by a very

distinct discontinuity on the reduced side. The trailing edge displays a

more gradual tapering of grade on the oxidized side."

Many geostatisticians are trying to design a better "probability model"

than the BLUE model that traditional geostatistics is based on. They

correctly realize that non-linear magnitude models are necessary to

better represent spatial structure.

However, many applications are more sensitive to the slopes and shapes

of the goal surface than the magnitude of goal surface values. Digital

elevation models, for example, are often used for viewshed analysis,

watershed delineation, and countless other applications where the value

of the surface is irrelevant except for determining slope and shape. I

described the process of estimating the appropriate slopes and shapes to

be "fitting the spatial statistical model". Roll-front uranium deposits

seem to need a combination of a well-fitting non-linear probability

model along with a well-fitting spatial statistical model. Comments

about the variogram and madogram statistics in InfoMine suggest that

geostatisticians are focusing their efforts on trying to create a

probability model that will generate appropriate spatial structures for

a specific application.

Variograms, and its many variants, are really only useful for

representing the spatial structure of the magnitude of a variable, but

not the slopes and shapes of that variable. My IJGIS paper provides a

specific example of a semivariogram's inability to represent specific

shapes. Other spatial statistics, with no specific relationship to the

probability model, are necessary to recreate these shapes and slopes

during stochastic simulation.

In order to model these surfaces, I designed a modular system where the

probability model can be designed independently from the various spatial

statistical models. I use an "inter-map cell swapping" algorithm to fit

the spatial statistical model while preserving the probability model.

This modularization provides several benefits: The most important is

that it is relatively easy to design probability models without worrying

whether they generate the appropriate shapes. The same is true for

spatial statistical models.

For the applications I am trying to represent uncertainty for, this

technique seems to be meeting my objectives. Several people have

requested digital copies of the IJGIS paper. I have finally gotten

around to putting it on the net.

sincerely, chuck

Chuck Ehlschlaeger wrote:>

--

> Dear Betty, et. al.,

>

> Betty Gibbs wrote:

> >

> > Recently I published an article by Dr. Stephen Henley in my newsletter (Earth

> > Science Computer Applications) called: "Geostatistics - Cracks in the

> > foundations?". I am looking for comments on the statements he makes about

> > limitations of geostatistical methods for building ore body models - and

> > whether these same limitations apply for other types of data. I would like to

>

> I am ignorant of geostatistical methods on ore bodies. However, I have

> been using variations of geostatistical techniques on geographic data,

> especially digital elevation models. Recently, a paper of mine

> "Representing multiple spatial statistics in generalized elevation

> uncertainty models: moving beyond the variogram" was recently accepted

> to be published in the International Journal for Geographic Information

> Science.

>

> The paper discusses at length the limitations of traditional

> geostatistical techniques for data with measurable surfaces. Of course,

> I propose a solution that deals with those limitations. (Otherwise, it

> wouldn't be "moving beyond the variogram".)

>

> I don't know how these techniques might be useful to ore bodies, but you

> did ask about "limitations for other types of data".

>

> The following contains the abstract.

>

> sincerely, chuck

>

> ABSTRACT: Spatial data uncertainty models (SDUM) are necessary tools

> that quantify the reliability of results from geographic information

> system (GIS) applications. One technique used by SDUM is Monte Carlo

> simulation, a technique that quantifies spatial data and application

> uncertainty by determining the possible range of application results. A

> complete Monte Carlo SDUM for generalized continuous surfaces typically

> has three components: an error magnitude model, a spatial statistical

> model defining error shapes, and a heuristic that creates multiple

> realizations of error fields added to the generalized elevation map.

> This paper introduces a spatial statistical model that represents

> multiple statistics simultaneously and weighted against each other. This

> papers case study builds a SDUM for a digital elevation model (DEM).

> The case study accounts for relevant shape patterns in elevation errors

> by reintroducing specific topological shapes, such as ridges and

> valleys, in appropriate localized positions. The spatial statistical

> model also minimizes topological artifacts, such as cells without

> outward drainage and inappropriate gradient distributions, created by

> the realization heuristic. Multiple weighted spatial statistics enable

> two conflicting SDUM philosophies to co-exist. The two philosophies are

> `errors are only measured from higher quality data and `SDUM need to

> model reality. This article uses an automatic parameter fitting random

> field model to initialize Monte Carlo input realizations followed by an

> inter-map cell swapping heuristic to adjust the realizations to fit

> multiple spatial statistics. The inter-map cell swapping heuristic

> allows spatial data uncertainty modelers to choose the appropriate

> probability model and weighted multiple spatial statistics which best

> represent errors caused by map generalization. This article also

> presents a lag based measure to better represent gradient within a SDUM.

> This article will cover the inter-map cell swapping heuristic as well as

> both probability and spatial statistical models in detail.

Chuck Ehlschlaeger N 40 46' 07.7", W 73 57' 54.4"

Dep. of Geography 212-772-5321, fax: 212-772-5268

Hunter College chuckre@...

695 Park Ave. secure: chuckre@...

New York, NY 10021 http://www.geo.hunter.cuny.edu/~chuck/

"We don't have time to have meetings about how to fix

problems, just fix them." -- George Tenet, CIA Director

--

* To post a message to the list, send it to ai-geostats@...

* As a general service to the users, please remember to post a summary of any useful responses to your questions.

* To unsubscribe, send an email to majordomo@... with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list

* Support to the list is provided at http://www.ai-geostats.org - Dear Chuck:

On reflection, I guess that I should have been less "absolute" in my

statement (You never know how your own words may come back to haunt you!). A

number of companies (including COGEMA) have very successfully applied

geostatistical techniques to the evaluation of ore zones in roll-front

environments. One of the approaches is to use a curvilinear block which

matches the sinuosity of the ore body geometry (a significant handicap if you

can't model the shape). Although it doesn't directly overcome the problem of

leading-edge discontinuity of the ore body, with practice, well justified

estimates can be obtained. I am intrigued, however, by your suggestion of

using a well-fitting non-linear probability model. But for two companies

that I've worked with, other means to obtain ore reserve estimates have been

applied. In other sandstone U environments besides roll-fronts, the ore

tends to be more or less continuous, and spatial statistical models are quite

appropriate. One of these ore bodies, however, had a disappointingly very

"flat" variogram which had very little spatial structure over a large range

and direction. This more or less justified their use of an "average" grade

adjusted by windsorising the high outliers when specific criteria are met.

R = Thickness * Grade * Area * density * probability of encountering ore

zones in the area

I think that the above is "out of the book" for Russian & Kazakh

practitioners.

Regards,

Dan ii

In a message dated 10/27/2001 5:18:11 PM Mountain Daylight Time,

chuckre@... writes:

<< "For roll-front uranium deposits, a geostatistical approach is avoided

because of the deposit characteristics. The leading edge of the roll has

generally the highest concentration of uranium and is bounded by a very

distinct discontinuity on the reduced side. The trailing edge displays a

more gradual tapering of grade on the oxidized side." >>

Dan W. McCarn, AIPG CPG #10245, Wyoming PG#3031

10228A Admiral Halsey NE

Albuquerque, NM 87111

+1 (505) 822-1323

+1 (505) 710-3600

http://resumes.yahoo.com/dwmccarn/geologist

11 Sept 2001

--

* To post a message to the list, send it to ai-geostats@...

* As a general service to the users, please remember to post a summary of any useful responses to your questions.

* To unsubscribe, send an email to majordomo@... with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list

* Support to the list is provided at http://www.ai-geostats.org - Re: the article by S. Henley ("Geostatistics-cracks in the foundations?")

There was a debate of sorts in Mathematical Geology in

1987 but I don't think it contributed much to understanding geostatistics

including its strengths and weaknesses. Here are a few more thoughts.

GENERAL COMMENTS

All of the concerns raised in the article are well-known in the geostatistical

literature, in particular in Mathematical Geology and Water Resources

Research as well as in the publications emanating from the various

APCOMS. Unfortunately the problems, strengths and weaknesses of

geostatistics are not well described in the article, in some cases there are

clear errors or mis-representations.

I. Although geostatistics is clearly and deservedly associated with G. Matheron

(and some of his early students such as A. Journel, J.-P. Chiles,

P. Delfiner as well as others from the Fontainebleau center), his

work is similar to or duplicates the work of Matern and Ghandin. Matheron

acknowledges this in some of his writings. In addition as almost always

happens, his work builds on work done earlier by many people in a

number of fields. Finally, geostatistics is not a stagnant field, it

has developed and evolved due to the work of many people, not always

in directions directly related to Matheron's original work (this is

not to belittle or denigrate Matheron's contribution but to recognize

that geostatistics as it exists today is not the same as 30 years ago).

Henley's article seems to imply that little has changed in the interim.

II. Although there are aspects of the applications in hydrology that

incorporate state equations derived from basic principles, geostatistics

has most often been used for the analysis of spatial data when no

state equations are available. This means that the problem is really

ill-posed, i.e., the solution is not unique and to obtain a unique solution

one must impose some form of model. The stochastic model implicit in

Matheron's work serves this purpose. However, one should not be mis-led

in thinking that this is totally artificial. There are clear connections

with Bayesian statistics (see the work of Wahba as early as 1970),

Thin Plate Splines and the more general interpolation methodology

known in the numerical analysis literature as "Radial Basis Functions"

which is a deterministic approach to the problem. Again Henley seems to

ignore this background.

MORE SPECIFIC COMMENTS

III. Henley makes a great deal out of the point that kriging is a "linear

method". This is true and not true. The kriging estimator (Simple,

Ordinary, Universal) for the value at a non-data location or the

spatial average over an area or volume (e.g., average grade in a mining

block) is a linear combination of the data. Written as an interpolating

function it is NOT linear, i.e., not a linear function of the position

coordinates. Moreover in the case of multivariate normality, the

Simple Kriging estimator is the conditional expectation (which is THE minimum

variance estimator, linear or otherwise). Of course multivariate

normality is a strong condition.

IV. Henley claims that it is necessary that the error distribution be normal,

this is absolutely wrong.

V. While there are individuals who would identify themselves as "geostatisticians"

it is more likely that individuals using geostatistics as well as

contributing to new developments in the field would call themselves:

mining engineers, petroleum engineers, geologists, soil scientists,

hydrologists, statisticians, mathematicians, ecologists, geographers,

etc. Hence the constant reference to (and laying blame on) "geostatisticians"

is mis-leading.

VI. Henley fails to distinguish between "estimation variance" and

"kriging variance" (the latter being the minimized value of the former,

i.e., the weights in the kriging estimator are obtained by this

minimization).

VII. Henley fails to distinguish between the sample/experimental

variogram and the (theoretical) model. The sample variogram is AN

estimator of the model (and certainly a number of authors would

say that it is not the only choice). Ultimately however it is the

variogram (theoretical) that is of interest and which is use in

the kriging equations (to obtain the coefficients in the kriging

estimator).

VIII. It is true that there are similarities between the Inverse

Distance Weighting (IWD) estimator and the kriging estimator. (1) As

noted by Henley, both are weighted averages of the data (but in the

case of IWD the coefficients are always non-negative, this is not

true for the kriging estimator), (2) in the usual formulation, IWD

is isotropic (only distance is used, not direction), (3) IWD is

not a "perfect/exact" interpolator as is the kriging estimator since

one can estimate at a data location using the data value at that location

in IWD (this would involve a zero distance), (4) while IWD in some

sense incorporates the spatial correlation between the value at the location

to be estimated and a data location (all pairings), IWD does NOT

incorporate the spatial correlation between the values at pairs of

data locations hence some information is ignored in IWD. Finally

one might attempt to optimize the choice of the exponent in IWD, one

size does NOT fit all (see a paper by Kane et al, Computers & Geosciences

1982).

IX. It is certainly true that kriging will "smooth" the data, howeve

this is true of all interpolation methods/algorithms. This is one

reason why some advocate the use of conditional simulation as an

alternative.

X. All statistical techniques are subject to problems resulting from

a lack of data, many of the problems Henley identifies or asserts are

related to insufficient data. Unfortunately, data costs money and one

will almost never have enough (of either). The related problem is how the

data is collected. In more typical applications of statistics, one

"designs" the experiment to ensure that the underlying statistical

assumptions are satisifed. This simply will not work in geostatistics.

XI. Henley correctly identifies "stationarity" or rather the lack of it

as a critical problem. However he does not quite describe it correctly. The

constant mean condition (which is only part of the definition of

second order stationarity) pertains to the implied random function, not

to the data. Since one has data from at most one realization of the

random function one can not statistically "test" this assumption. This

is one of the places where the lack of data is a serious problem,e.g,

to decide whether to partition the region of interest into separate

regions to obtain "stationarity" on each separately. Note that

the condition on the variogram/covariance (being only a function of

the separation vector) is critical and not implied by the constant

mean condition.

XII. The statement "The problem with this method was that the semivariogram

itself was sensitive to the form of the deterministic surface.

Therefore, it required a number of iterations of kriging, variogram

computation, and model fitting, to converge towards a consistent

solution." contains a germ of truth but it is also confused and mis-leading.

What Henley is presumably referring to is that if one fits a trend surface to

the data and then computes a sample variogram using the residuals, the

resulting sample variogram is biased (see a paper by N. Cressie in

Mathematical Geology for example and a much earlier paper by another

author in the proceedings of the NATO conference of 1975). Then there

was a dissertation in the Dept Hydrology (J. Samper) about 1987) on

a maximum likelihood method (an iterative application) for fitting the

sample variogram to residuals. This problem is not completely resolved

because as is frequently the case in geostatistics or rather in the

application of geostatistics there is a discrepancy between theory and

application. It is well known (see Matheron's 1971 Fountainebleau Summer

School Lecture Notes) that the optimal estimator of the drift, i.e., the

non-constant mean of the random function, is obtained by kriging. However

to apply this means that one must first have the variogram model, yet

one can not model the variogram using the original data. The problem is

circular. In practice the problem in fact often addressed by fitting a

trend surface to the data and fitting the variogram to the residuals.

The real problem is the fact that the sample variogram only estimates

the variogram if the mean of the random function is constant. When the

mean is not constant and in particular when the fitted trend surface is

not just a constant, the sample variogram will exhibit a very rapid growth

rate (quadratic or greater). There are no valid variogram models with

this property. Hence if the sample variogram exhibits this growth condition

this property is taken as evidence of a non-stationarity. Note that this

property or characteristic is not absolute, it may only appear for large lag

distances and hence one may be able to fit a variogram model to the

sample variogram using only the information for short lags.

XIII. The statements "These methods have found little practical application

because of their complexity, and the inherent instability of their

solutions. Furthermore, the resulting kriging system is no longer

linear and thus loses its ideal "BLUE" properties." are partially correct and

partially incorrect.

It is well known that the form of the kriging estimator and the kriging

equations when using generalized covariances is exactly the same as for

Universal Kriging. It is also well known that the kriging estimator

(written in the dual form), when using the right choice of a generalized

covariance, is the same as the thin plate spline. One must use the so-called

"spline covariance", this was implemented in the BLUEPACK software back

in the 80's and is also in the current ISATIS. It is NOT correct to

say that the resulting kriging system is not linear, the form of the kriging

estimator is unchanged and the kriging equations ARE still linear. The

kriging estimator is still a BLUE contrary to Henley's claim.

There are several reasons why generalized covariances are not widely used

(for a particularly good presentation see the recent book by Delfiner

and Chiles). One is that to write software is much more complicated, i.e.,

to determine the order of the non-stationarity. A second is that as

contrasted with the family of known variograms there are only a few

known generalized covariances (although every variogram corresponds to

a generalized covariance) and these are all isotropic. Hence in practice

one is likely to revert to using a variogram. Third, it is not so difficult

to use geostatistics/kriging because of the availability of both free

and moderately inexpensive software, the theoretical questions are largely

taken care of in writing the software. There is essentially no free

software that incorporates the use of generalized covariances (ISATIS is

rather expensive because it is intended primarily by petroleum and mining

companies)

XIV. The first half of the statement "Although regionalised variable theory

does not require normal (gaussian) distribution of the data, it does

assume normal distribution of error terms." is almost true but the second half

is completely false. Henley consistently fails to distinguish between

properties of a particular data set and the properties of the random function.

The kriging equations are derived without making any distributional

assumptions. It is true that kriging (and any other interpolation method)

will tend to smooth the data. This characteristic is exagerated when the

distribution of the data is not approximately symmetric (but not necessarily

normal) and if the distribution of the data has "fat tails". Both lognormal

kriging and indicator kriging are ways to deal with this problem.

In claiming that the kriging estimator is no longer "BLUE" when using the

a log transformation, Henley is presumably referring to the fact that

if one simply exponentiates the result then there is a bias. Under an

assumption of multivariate lognormality one can compute the bias and hence

compensate for it. This has received considerable attention in the geostatistics

literature. The problem of course is that is not possible to absolutely

know whether the underlying random function (for a particular data set) is

multivariate lognormal (univariate lognormal is a much weaker condition).

All one can do is to determine whether that assumption is reasonable.

Finally, it is quite plausible to ask whether there might be a better way

to approach the problem. There is nothing really wrong with Matheron's

work, the problem is knowing whether the assumptions implicit in its

use are valid. Since those assumptions pertain to the random function

, RATHER THAN THE DATA, it is not possible to completely answer the

question. The real question, particularly for the practitioner is whether

geostatistics produces useful results. "useful results" does not have

a very precise meaning in general but it will have a very real meaning

to the practitioner. Geostatistics is not a cure-all nor is it useful

for all problems. It is not a black box scheme and must be used with

some care. In some cases, i.e., for some data sets and some objectives,

a user need only have access to a software package such as GEOEAS,

G-STAT or even the add-ons available in ARCVIEW, S-PLUS, etc. In other

cases it may be necessary to seek the advice and assistance of someone

more experienced and with a stronger understanding of the mathematics/

statistics in order to adequately apply the geostatistics to the data

set.

For a slightly different take on the same general question see a recent book by M.L. Stein

Donald E. Myers

http://www.u.arizona.edu/~donaldm

--

* To post a message to the list, send it to ai-geostats@...

* As a general service to the users, please remember to post a summary of any useful responses to your questions.

* To unsubscribe, send an email to majordomo@... with no subject and "unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list

* Support to the list is provided at http://www.ai-geostats.org