|
Tehran- June 28, 1999
Iranian Green Wave Front and Iranian Department of
Environment cosponsored a one-day workshop on soil protection in
Tehran. The lecturer was Dr. Karim Abbaspour from Swiss Federal
Institute of Technology. This workshop was attended by 30
researchers from Soil Research Institute, Watershed Management
Institute, Soil Branch of DOE and Iranian Fisheries
Organization. The main topic of workshop was allocated to the
role of modeling and risk analysis in saturated soil pollution.
A couple of computer models on contaminant movement in soil were
introduced. A summary of the workshop is presented below.
PROCEDURES FOR UNCERTAINTY AND RISK
ANALYSES AND APPLICATION TO ENVIRONMENTAL DATA
In the past twenty years, scientific research has moved much
closer to engineering practices than ever before in the areas of
characterization, control, and remediation of environmental
pollution. In particular, we can mention developments in the
following areas: (1) modeling of water and contaminant movement
in soil and groundwater, (2) employment of stochastic
simulations, (3) advances in geostatistical estimation and
simulation techniques, (4) employment of conditional simulation
procedures, (5) use of different data types (qualities) in
simulations, (6) adaptation of inverse simulation for parameter
identification, (7) use of Bayesian statistical framework to
accommodate engineers' prior knowledge, and (8) introduction of
data worth models. Since environmental protection has in recent
years become a major social, research, and political issue, a
more systematic and rational framework can better address all of
its complexities. Advances made in the above areas are linked
together in BUDA, making it a rational tool for decision
analysis. In the following sections the advances made in the
above areas are briefly discuss, but to set the stage, a general
description of information system in environmental engineering
areas is presented first.
Characteristic of Information System in
Environmental Studies
There are a few distinct features that separate the
information system in an environmental project from other
fields. These are: (1) environmental data exhibit spatial
correlation, (2) environmental data exhibit heterogeneity, (3)
most data have non-negligible errors, (4) statistics of the
parameters are uncertain, (5) collection of data is not the end
product, and (6) there are usually different types of data
available, but each of limited quantity. Therefore, any
algorithm designed to address an environmental engineering
project should be compatible with such an information system.
Flow and Transport Modeling
Until recently, flow and transport models for general and
specific applications were scarce and model users had to develop
their own models. This had severely limited the use of modeling
for practical application, but in recent years there has been an
explosion in user friendly general- and specific-purpose models
and the use of pre- developed models has become more and more
routine in academic and engineering institutions. Short courses
teaching the use of these models is prevalent in academic
centers and it does not take a long time to set up and run a
simulation program for a practical problem. Modeling can be used
for prediction, design, analysis of alternatives, and simulation
of worst case scenarios. Stochastic Simulation
Stochastic simulation is in recognition of two important
engineering facts. (1) Nature is heterogeneous, and (2) input
data is always limited and hence uncertain. Propagating the
input uncertainty gives rise to stochastic simulation, i.e.,
instead of producing one likely scenario based on unlikely
average input values we produce many likely scenarios based on
input values that look like reality.
Geostatistical Estimation
Geostatistical techniques are powerful estimation tools but
their use in geological, hydrogeological, and geotechnical
engineering have met with limited success due to large
variability and small size of data points in these fields. In
recent years, however, there has been a surge in the development
of techniques that make geostatistics more relevant to practical
environmental applications. Some of these include the
development of co- kriging procedures that allow linear
regression using data defined on different attributes, indicator
kriging which allows reproduction of spatial patterns of
categorical variables, indicator principle component kriging
which accounts for several categorical variables, and Markov-Bayes
model of soft kriging which utilizes soft or fuzzy data. The
author in a more engineering-friendly program, referred to as
Co_Est, where an alternative technique to co-kriging is
presented gives the latest development in this area.
Geostatistical Simulations
In addition to the above estimation methods, a number of
simulation techniques have also been developed. Stochastic
simulation, in this context, is defined as the process of
building alternative, equally probable, high resolution models
of spatial distribution for a random variable. Simulation
techniques, as opposed to estimation, are better suited to
engineering practices and risk assessment studies. These
techniques are intended to produce a better picture of reality
and to eliminate the unrealistic smoothing that is
characteristic of spatial averaging methods. Some of these
techniques include sequential Gaussian simulation which
generates realizations of a multivariate Gaussian field,
sequential indicator simulation which uses indicator techniques
to generate random fields, Markov-Bayes simulation which
accounts for soft indicator data, and simulated annealing
simulations where simulated annealing numerical technique is
used to generate alternate conditional stochastic images for
either continuous or categorical fields. The sequential
indicator technique has been used in many applications such as
reservoir mapping. In the above techniques, although the input
data can be expressed in different forms and with varying
qualities (hard, soft, fuzzy, etc.), the geostatistical
parameters depicting the spatial distribution of data (i.e.,
mean, variance, range or correlation length, and nugget) are
always assumed to be known with certainty. Since the true values
of these parameters are usually not known, parameter uncertainty
has become the focus of some works. The latest work in this area
is presented by the author in a program referred to as Bayesian
uncertainty simulation, or BUSIM, where uncertainty in
geostatistical parameters are explicitly accounted for in
stochastic simulations.
Conditional Simulations Conditional simulation is a powerful
tool that honors the measurement points at their measured
locations while producing multiple simulated random fields.
These techniques bring geostatistical applications much closer
to depicting the reality than unconditional methods.
Use of different data types
It is typical for an engineering situation to have data of
different types, but each of small sample size. An important
achievement in simulation programs is the ability to use data of
different types and qualities. This allows utilization of all
types of information that will effectively increase the sample
size.
Inverse Modeling
Inverse modeling is the process of using measured primary
outputs of a model to estimate input parameters of that model.
It is in recognition of the fact that in most practical
applications there is not enough input data to establish a
credible modeling result. The procedure generally involves
minimization of a square difference function of some measured
and simulated variable. The author developed a program, SUFI,
which performs the inverse modeling in BUDA.
Bayesian Framework
The main difference between the classical and Bayesian
statistical approach is the use of prior estimate of the form of
the probability density function and its statistics. This prior
estimate is by large subjective and could be based on limited
early data from the site or similar sites, or the engineer's
expert opinion. When additional data becomes available, they are
used to update the prior estimates of the statistics to
posterior estimates using Bayes theorem. Use of subjective
information based on experience is prevalent in engineering
practices and adaptation of Bayesian approach should, therefore,
come naturally to a practicing engineer. The Bayesian formalism
can help the quantification of subjective data and its
employment in a more rational basis. The implications in
adopting a Bayesian framework for an engineering project is that
the project iterates among analysis, field work, and decision
making, and collection of data is commonly based on the results
of a data worth model in pre-posterior analysis. Pre-posterior
analysis is the exercise of sampling from the prior
distributions of input data rather than a field sampling. By
using a Bayesian framework typical questions which are commonly
sought by engineers can be addressed, such as: what is the
quantity of risk based on the present information, how many more
samples are needed to decrease the risk to an acceptable level,
and what kind of data should be collected.
Program BUDA
BUDA (Bayesian Uncertainty Development Algorithm) is an
algorithm, which uses the tools mentioned above, for risk
analysis in environmental projects. The main components of BUDA
are: (1) problem definition, (2) uncertainty analysis, and (3)
risk analysis. The complexity of each component depends on the
nature of the problem at hand, but in general, problem
definition consists of application of a modeling protocol,
definition of a goal or objective function, definition of
domains of the problem, and finally probabilistic depiction of
parameters of the problem. Uncertainty analysis consists of
quantification of uncertainty, propagation of uncertainty, and
reduction of uncertainty. Risk calculation is problem dependent
and in general terms it can be defined as the product of
probability of failure and the cost of failure.
|
