Title: Predicting the Impact of Global Climatic Change on Land Use Patterns in Europe
1Predicting the Impact of Global Climatic Change
on Land Use Patterns in Europe
- Andy Turner
- Centre for Computational Geography
- University of Leeds, Leeds, UK
- andyt_at_geog.leeds.ac.uk
2Thanks are due to
- Stan Openshaw
- Ian Turton
- Tim Perree
3Contents
- Why model agricultural land use change for 75
years hence? - Existing models
- A neurocomputing approach
- Assembling the data
- Running the models
- Results
- What next?
4Why?
- It is an important subject
- It potentially affects many millions
- It emphasizes what little we know
- It provides a first attempt that others will have
to beat later - Someone had to try and do it!
- It provides a good example of how GIS can be used
to model environmental systems showing both the
strengths and weaknesses
5Some Background
- This research was part of the EU Medalus III
project - Medalus Mediterranean Desertification and Land
Use - Wide range of environmental research topics
mainly concerned with modelling hill slope
erosion, hydrological systems, water management,
ecosystems, and climatic change in semi-arid
Mediterranean climate zones - Study Area The Mediterranean climate region of
the EU
6Main objective was to incorporate a
socio-economic systems modelling component into
physical environmental models of LAND DEGRADATION
7The research challenge!
- To identify ways of predicting the likely impacts
of climatic change on agricultural land use
patterns for around 25 to 50 years time - In order to
- raise awareness of land degradation problems
- inform political and public debate
- contribute to a pro active framework for action
8The hardness of this challenge should not be
under-estimated!
9Modelling Challenges
- model contemporary agricultural land use patterns
based on a range of climatic, physical and
socio-economic variables - obtain and forecast these variables in order to
predict future agricultural land use - translate the land use changes into a land
degradation risk indicator - combine various land degradation risk indicators
to produce a synoptic forecast of land degradation
10Previous Research
- Very little research on long-term land use
prediction - Most of what little exists is non-spatial or at a
very coarse level of geography - Some micro-studies exist at the level of
individual farms BUT these cannot yet be scaled
up to the EU level or used to make long term
forecasts easily
11The CLUE modelling framework
- The CLUE (Conversion of Land Use and its Effects)
model of Veldkamp and Fresco (1997) is probably
the best and most relevant of existing models - A multi-scale stepwise regression model
- Its relates land use change to socio-economic and
biophysical factors - Operates at 7.5 km2 for Costa Rica and 32 km2
scale for China
12CLUE Model
- linear
- It is run recursively
- It produces nice computer movies
- Its runs at too coarse a scale to be useful
- It will probably have dreadful error propagation
properties
13Modelling Design Checklist
- Highest possible level of spatial resolution
- Consistency in coverage and application
- Make forecasts for 75 years hence
- Link with other Medalus III Projects and models
- Incorporate the principal driving factors and
processes - Produce outputs that can be instantly understood
by Joe Public - Provide a framework that can be refined later
14Building a Synoptic Prediction System (SPS)
- Objective was to build a GIS based computer
modelling system able to link changes in the
climate with associated physical and
socio-economic changes in order to make synoptic
land degradation forecasts for the entire
Mediterranean climate region of the EU - It was to function in a manner similar to a
long-term weather forecast
15SPS Modelling
- A model was required to link
- climate (temperature and rainfall)
- soil characteristics (permeability, texture,
fertility, parent material) - biomass
- elevation
- population densities
- to predict current and future patterns of
agricultural land use
16Synoptic Prediction System
biomass
height
Physical
climate
soil
F U T U R E
N O W
IMPACT
Classification
17SPS is limited by the following
- available data from other Medalus teams and
elsewhere - almost complete absence of space-time data series
- lack of knowledge of all the principal mechanisms
thought to be at work - the need to incorporate a broad range of inputs
to ensure plausibility - the necessity of working at a fine level of
spatial detail
18Other Problems
- Relationship between environment and land use is
mediated by - technology
- market forces
- historical traditions
- inertia
- culture
- various economic factors
- behavioural aspects
19but
20but
21there is little that can be done about any of
this!
22The problem was HOW to OPERATIONALISE this
schematic model in the best possible way
23In essence it is a kind of non-linear regression
model
- The inputs can be converted into outputs via
either - mathematical equations
- statistical equations
- fuzzy rules
- neural networks
24SPS is based on a neurocomputing approach
25Do not PANIC!
- the basic idea is very simple
26Its just an artificial neural network!
- they are now quite common
- not much to them
- they are not black magic
- its just a black box that performs a function
similar to regression - they cannot bite!
27A representation of an Artificial Neuron
28A representation of a 6x4x1 simple network
29Neural Networks Offer several advantages
- they are universal approximators
- they are equation free
- they are highly non-linear
- they are robust and noise resistant
- probably offer the best levels of performance
- they can model hard problems
- widely applicable modellers
30Neural Headaches
- They are essentially black box models
- Training can be problematic
- over-training
- length runs
- Choice of Architecture is subjective with an
element of black art or luck or intuition - Often a presumption of prejudice against because
of the lack of process understanding
31Some Key Assumptions
- the training data were representative
- the predictor variables were appropriate
- the effects of missing variables were implicit in
the available variables - the neural net architectures were reasonable
- that there is a systematic relationship between
environmental variables and land use that is
modellable
32Building a SPS
- Step 1. Assemble the data for a common EU wide
geography for - present day
- Step 2. Obtain or make forecasts for these data
for - 75 years time
33Building a SPS (Part 2)
- Step 3. Construct Neural Nets to model the
relationships between climate-soil-biomass-elevati
on-population density in order to predict present
day land use - Step 4. Compute estimates for 75 years time using
neural nets trained for the present
34Building a SPS (Part 3)
- Step 5. Create maps of changes
- Step 6. Consider modifying the predictions and
forecasts to reflect knowledge using fuzzy logic - Step 7. Repeat everything to test different
change scenarios - Step 8. Make estimates of uncertainties using
Monte Carlo simulation
35Step1 Assemble data for a common EU wide
geography
- Not easy!
- A major reason for the lack of models linking
environmental and socio-economic variables is the
lack of a common data geography - Environmental data tends to be grid-square based
for small areas whilst socio-economic data tend
to be for far larger and irregular polygons
36Data required to predict agricultural land use
- Soil Type
- Soil Quality
- Biomass
- Temperature seasonal
- Precipitation seasonal
- digital elevation model
- population density
37Why these variables?
- They are clearly related in some way to
agricultural land use patterns - They reflect the research by other Medalus teams
- They were available in some form
- They were available or could be estimated
381 Decimal Minute EU database
- Decided to use grid-squares
- Best scale was about 1km2 or 1 decimal minute of
resolution - Most environmental data can be manipulated into a
1 DM cell format using GIS - BUT.. socio-economic data need to be interpolated
from a coarser to a finer geography
39EVERY data set caused problems and required its
own set of GIS operations in order to create the
data base
40Estimating and Interpolating population data
- First task was to develop a means of creating
population (and other socio-economic) data for 1
DM cells for the EU when the best available data
was at NUTS 3 level of geography - For example, in UK there are 64 NUTS-3 regions
and 150,000 1 DM cells - The task was to interpolate from 64 to 150,000
cells!
41The Interpolation Problem
Nuts3
1DM
42Methodology
- Use available GB census data as target data
- Nothing as good available for anywhere else in
the EU - Test out different methods of estimating these
data from EU wide predictor variables - Apply the best method to rest of EU
43Review of Existing methods
- There are SOME existing methods that can be used
- A very old simple method
- uniform area shares
- Various surface interpolation methods
- Toblers pycnophylactic surface
- RIVMs Goodchild et al (1993) method
44RIVM population density surface
45RIVM Smart Interpolation
- Weighting factors were used to create a
population potential surface - auxiliary data sources were used to modify the
weights Sea, Roads, Rivers - An estimate of population was made using the
weights - Best of the existing methods
- Errors are large
46Maybe it is possible to do better using a neural
net to perform the interpolation
- Extend the RIVM approach to use a broader set of
digital variables - Train a neural net on UK data
- Apply to rest of EU
- Modify to meet accounting constraints based on
NUTS-3 control totals
47What Spatial Data for the EU is available that
can help?
- Data available for all of EU are
- Bartholomews 11000 000 digital map data with
various layers ( DCW) - Other spatial data (DTM, slope, land cover)
- NUTS 3 coverages
- RegioMap and Eurostat Statistical Data forecasts
at NUTS3 level - Satellite data (eg. Night-time lights data)
48Population Predictor Variables
- distance to
- built up areas
- airport
- parks
- river and canals
- towns by size
- location of
- built up areas
- place names
- density of
- communication networks
- various roads
- railways
- height above sea level
- night-time lights
- RIVMs population
49Communications network density
50Motorway and dual carriageway density
51Main and minor road network density
52Railway network density
53Night-time lights frequency
54Distance from extra large towns
55Distance from large towns
56Distance from medium sized towns
57Distance from small towns
58Distance from internatonal airports
59NUTS3 population from Eurostat
60Although the errors are still large, the
Population interpolation maps look good!
6123x20x20x1 prediction
6223x20x20x1 prediction close up of Italy
63(continue..) Step1 Assemble data for a common
EU wide geography
- Climatic Data based on global climatic change
models - results for a network of 50 weather stations
- linear interpolation to 0.5 DM grid
- imported into ArcInfo and aggregated to 1 DM
cells - Spatial interpolation errors probably less than
forecast errors!!!!!!
64Seasonal Temperature Data
N O W
F U T U R E
Forecast
65Seasonal Precipitation Data
N O W
F U T U R E
66Climatic Biomass Potential
Height above Sea level
67Step 2. Obtain or make forecasts for these data
for 75 years time
- We used other Medalus III project partners
forecasts for climate and climatic biomass
potential - Population forecasts made by changing the
accounting constraints to reflect Eurostat
forecasts - Note the convoy effect in that the various data
only need be a similar degree if inaccuracy!
68Step 3. Construct Neural Nets to model the
relationships between climate-soil-biomass-elevati
on-population in order to predict present day
land use
- used a feed forward multi-layer perceptron
- training data based on 20,000 randomly selected
cells - trained using a hybrid approach
- genetic optimiser to start training
- fine tuning using a conjugate gradient method
69SPS Neural Net
- Aim was to model current land use
- Various architectures investigated
- Best had 18 inputs, a single hidden layer with 50
neurons, and 1 output neuron - Net trained on present and then given the same
inputs for the forecast years - Results appear promising!
70Results
71Dominant Arable Landuse
Observed
Predicted
Forecast
72Dominant Tree Landuse
Observed
Predicted
Forecast
73Dominant Waste Landuse
Observed
Predicted
Forecast
74Deficiencies!
- many social, economic, and political processes
are only implicitly present - neural net modelling would be better if the data
inputs were better - uncertainty levels remain unidentified
- mixture of data sources with very different error
and uncertainty levels
75(yet more grave) Deficiencies!
- a major assumption that global climatic change is
equivalent to a shift in the boundaries of
agricultural capability - there is an assumption that technology and
behavioral influences remain constant as implicit
in the training data - land use categorization is very crude
76Good Points?
- a first attempt at socio-environmental modelling
- a common methodology for the EU
- brave (maybe foolish) attempt at broad-brush
forecasts for 50 years ahead - framework can be used to yield improved results
77Good Points (more??)
- results can be readily updated as improved data
become available - sets a benchmark that subsequent models will have
to beat - difficult to see how else the research objectives
can be achieved - offers a context for discussion and debate
- results are understandable
78Step 5. Create maps of changesStep 6. Consider
modifying the predictions and forecasts to
reflect knowledge expressed as fuzzy rules
79Fuzzy Interpretation of Impacts
- Important to handle the uncertainty in the
predictions and data - Fuzzy logic is a good way to achieve this
- Also allows incorporation of intelligent rules of
thumb to add realism to computer model results - Can be further extended as required
80Schematic of Fuzzy Land Degradation Interpreter
8116 Fuzzy Rules
- If landuse_now is arable and landuse_future is
- arable then land degradation is possible
- trees then land degradation is unlikely
- waste then land degredation is serious
- other landuse then land degradation is probable
82- If landuse_now is trees and landuse_future is
- arable then land degradation is possible
- trees then land degradation is possible
- waste then land degredation is serious
- other landuse then land degradation is probable
83- If landuse_now is waste and landuse_future is
- arable then land degradation is possible
- trees then land degradation is possible
- waste then land degredation is extensive
- other landuse then land degradation is possible
84- If landuse_now is other and landuse_future is
- arable then land degradation is possible
- trees then land degradation is possible
- waste then land degredation is severe
- other landuse then land degradation is unlikely
85Fuzzification of landuse data
1.0
Non-arable fuzzy landuse class
degree of member- ship
Arable fuzzy landuse class
0.0
0.0
1.0
Arable Landuse
86Land degradation membership function
fuzzy classes
possible
probable
unlikely
extensive
severe
serious
degree of member- ship
land degradation scale
1.0
0.0
87A Map of Land Degradation
88Step 7. Repeat everything to test different
change scenariosStep 8. Make estimates of
uncertainties using Monte Carlo simulation
89Conclusions
- 50 YEARS is a long time BUT the topic is SO
IMPORTANT that it is important attempts are made
to make these types of predictions - Predicting affects of not yet visible global
climatic change on land use in 50 years time has
one outstanding advantage... - when the true results are known I will not be
around to see them!!
90The results presented today are
- Preliminary
- Subject to change
- They need to be improved upon
- They are almost certainly WRONG
- They are probably very WRONG
- Its even conceivable they could be COMPLETELY
WRONG
91andyt_at_geog.leeds.ac.uk
- http//www.geog.leeds.ac.uk/staff/a.turner
- http//www.medalus.leeds.ac.uk/SEM/home.htm