melles
DESCRIPTION
Third International Workshop on "Geographical Analysis, Urban Modeling, Spatial Statistics"TRANSCRIPT
Sampling Optimization Trade-offs for Long-term Monitoring
of Gamma Dose Rates
S.J. Melles, G.B.M. Heuvelink, C.J.W. Twenhöfel, and U. Stöhlker
Presented at ICCSA, GEOG-AN-MODJuly 2, 2008, Perugia, Italy
Where do we go?
Motivation: why important, what can we gain, and current research needs?
Methods: Regression kriging context, mukv similar, can get a measure of uncertainty prior to sampling as long as you have complete info on predictors
Optimization: Simulated annealing (algorithm to min. Obj. Function developed in the field of condensed matter physics and based on the analogy of controlled cooling of a metal in order to achieve optimal overall strength, avoiding local weakness).
What have we learned? not related to sleeping
Motivation & motivating questions
1. Design of a sampling scheme important for the measure E.g.: Gamma radiation (minimize costs cant sample everywhere
Air pollution Environmental variables of interest
2. Need an objective assesment of the quality of a monitoring network
3 main things to consider: purpose, decision criteria, constraints
3. How to determine which approach to apply? Design-based DETERMINISTIC (probability), Model-based STOCHASTIC (simulation/interpolation in geostatistics) , Geometric (linear transects)
4. How to optimize? (Practical situations – have a decision problem:
Optimization problems are mathematical translations of decision problems, simulations)
Design-based vs. Model-baseddeterministic probability based, global vs. stochastic model-based, local
1. Target ‘design-unbiased’?2. Accuracy quantified objectively 3. Random sampling feasible?4. Reliable model available?5. Substantial spatial autocorrelations?
Part 2:
Methods to predict or describe spatial variability in
environmental variables
Spatial variability in geostats
z(s) = a realization of an underlying random function Z(s)
Spatial variability in Z(s) is related to natural, deterministic processes (e.g. gamma radiation is affected by soil type, altitude, etc.)
An exhaustive process description is not possible Stochastic methods are commonly used to describe
remaining spatial structure in the data (and map environmental variables for risk management)
Spatial variability in geostats – in order to do statistics and make inferences we assume
Second order stationarity E[Z(s)] = m Expected value of the random
variable C(h) = E[(Z(s + h) – m) (Z(s) - m] Covariance of two
random variables h distance apart Intrinsic stationarity allows us to make inferences in cases where the
mean γ(h) = ½ E[(Z(s + h) – Z(s))2]
Regression krigingoften w/ enviro data, we have known deterministic trends that influence our var of
interest, and for which the parameters are unknown (a spatial regression technique)
Trends in the mean Hybrid spatial modelling technique
Regression Interpolation of regression residuals
Example: variogram & point cloud
Experimental variogram
distance
semivariance
0.2
0.4
0.6
500 1000 1500
0 500 1000 1500
01
23
4
distance
semivariance
Sem
ivar
ianc
e γ(
h)
Distance (h)
1500 1000 500
Distance (h)
1500 1000 500
0.6
0.4
0.2
Part 2:
Example: γ-dose rate radiation
Example: mean long term γ-dose rates
Self-effect of probe Anthropogenic
Hospitals, nuclear power plants, research
Natural Cosmic (atmospheric
pressure & altitude) Airborne (precipitation) RN, Bi,
Lb attached to aerosols
Terrestrial (soil type and soil moisture)
Spatially correlated residualsnotice directionality in residuals
0
frequ
ency
60
0 40-40
catsoilelevsZ )(
Modelled with aniso variogram spherical and linear component in all directions, but spherical is dominant in swesterly
and linear in ortho
Regression kriging prediction
30
80
nSv/hr
Kriging prediction error variance
65
85
nSv/hr
Part 3:
Optimization
How to optimize?
Minimize a CRITERION (the objective function) Prediction errors due to
REGRESSION MODEL PARAMETER ERRORSINADEQUATE SPATIAL COVERAGE
What about other decision criteria? E.g., dynamic and most important case???POPULATION DENSITYDISTANCE TO NPPS
ncCqqqCqcCqq
cCcCCsmeanRKSE
TTTT
nT
/))()()(
)(()((
01
011
01
0
10
10100
2
residual component
trend component
Optimization by simulated annealing
1. Start with current design2. Moving one device at a time3. Compute objective function
(criterion)
4. Compare criterion previous design
5. Accept if current is lower, but not always…
6. New sampling design constructed; loop back to step 1
Minimizing the objective function (RKSE)
Current design Optimized design
Examining the trade-offs?
Costs
Part 4:
What have we learned? Where to from here?
What have we learned?
Fairly minor changes to current network could be made
(improve quality of mapped predictions of gamma dose rate or remove stations w/ no decrease in ‘’quality’’, particularly at borders)
Most room for improvement in border regions
Examination of trade-offs useful to the extent that actual costs are evaluated and monitoring purpose is adequately captured by
objective function
Doesn’t that look like even spatial coverage sampling?
Current thinned design( 50% stations)
Optimized design(50% stations)
Npp
Where to from here:
Shortcomings I: Spatial correlation structure is an integral over
the whole range of values Does not deal well with extreme values Tends towards spatial coverage sampling with
larger sample sizes Assumes known trends and variogram
Where to from here:
Shortcomings II: Sensitive to boundary conditions Time consuming Dynamic case is primary purpose of monitoring…
As purposes differ, so do criteria. What to do with constraints (e.g. NPPs,
population centers)Next stage, multi-criteria optimization with
weights
6,)min(1
pwi
ppi
pi
THANK YOU!
Appendix
P of accepting a worsening design
T
feP
Delta f is the change in MUKV
Coming up with the costs
Ask the experts? But, no one wants to say… Are you an expert? Is anyone an expert? Consider the main purpose of the network…
Coming up with weights?
Ask the experts? Are you the expert? Is anyone an expert? Consider the main purpose of the network…
Dynamic case