Transcript Slides - IIASA

Experiences in assessing
deposition model uncertainty and
the consequences for policy
application
Rognvald I Smith
Centre for Ecology and Hydrology,
Edinburgh
Concentration (measured)
>
MODEL
>
Deposition estimate
MODEL – field programme of flux measurements
Substantial degree of confidence – but not quantified
Ammonia concentrations provided by a combination
of model and measurement – local variability
FLUX measurements
Comparison of national
model output against
measurements would help
provide uncertainty
measure
DRY – 2 sites
WET – lack of co-located
rainfall amount and
precipitation
concentration collection
Sensitivity and Uncertainty Analyses on dry and wet
deposition
Dry – concentration always important component, but also
some model parameters were very influential
Wet (seeder-feeder model) used for more extensive study
Demonstrated usefulness of techniques but also raised
questions
What is the important output?
each 5km square (>10000 for UK)
some groups of squares to make regions
a smaller area like a hectare
Many sensitivity analyses assume there is one, or possibly
a few, summary statistics as important output – need
to look for 2D area-based approaches.
It appeared that biased output was probably the norm
from the deposition models.
- even simple models are non-linear
- current preferred parameter choices may not be
optimal.
Bias is not a problem if
- it can be estimated with reasonable accuracy, or
- the flux estimate is so far away from a ‘test level’
that it can be ignored.
but it is a bigger issue when it can be cumulated:
regional/national budgets
inside transport models
(bias may be applied at each time step)
It proved to be extremely difficult to get good
estimates of uncertainty for the inputs or the
parameters.
SA/UA identified
‘important’ sources of uncertainty
a number of important interactions within the model
– these should be used to identify where further work is
required on the inputs and parameters
Spatial interpolation
PROBLEM: models require values of parameters and input
variables everywhere.
NH3 concentration driven by
local sources:
background 1 g NH3 m-3
(blue and green)
near source 50 g NH3 m-3
(purple and black)
approx 3km image
With a wide distribution of farm animals, impossible to
interpolate ammonia only from measurements.
Smoothly varying fields, e.g. SO4 in rain
30 site networks:
kriged interpolation CV about 30% for most areas
[magnitude confirmed by other studies]
2 x standard error approximation
=> concentration in many areas is the mapped value 60%
Little mechanism to reduce this in the deposition models,
so the flux uncertainty will be greater in almost all cases.
Summary:
Scale effects on deposition
terrain: valley v hilltop (rain, wind, temperature …)
stochastic rainfall (even on flat areas)
local sources, especially with a cleaner atmosphere
Interpolation or modelled concentration uncertainty
Deposition/Flux model uncertainty

1)
Uncertainty in any statistic which focuses on
small ecosystem areas and is derived from a
national or European scale model will be
large.
2)
Any reasonable assessment of uncertainty in
the deposition estimates will take a
substantial effort.
A possible way forward considers these points:
 Focus on specific statistics for which an uncertainty
estimate is required.
 Modelling studies can give some insight into scale
uncertainties.
 Accept that predictions for small areas will be
extremely uncertain.
 Consider a result in probability terms over larger
areas and accept the sacrifice, at present, of small area
information.
 Look to simplifying the structure where possible, for
example by smoothing.
 Massive simulations are now possible, but are still
expensive.
 There is no off-the-shelf satisfactory solution.