This document discusses using remote sensing data and data assimilation techniques to calibrate cellular automata based land-use models. It aims to improve land-use simulations by lowering uncertainties compared to other automatic calibration methods. A simplified MOLAND land-use model for Dublin is used to test error propagation modeling and a particle filtering data assimilation approach. Preliminary results seem promising, but differences between spatial metrics from remote sensing and model outputs may hamper the analysis. The overall goal is to develop robust and reliable tools for land-use change modeling and calibration to inform policy contexts.

1. Uncertainty Analysis and Data
Assimilation of Remote Sensing Data for
the Calibration of Cellular Automata
Based Land-Use Models
Johannes van der Kwast UNESCO-IHE, the Netherlands
Lien Poelmans, Inge Uljee, Guy Engelen VITO, Belgium
Tim Van de Voorde, Casper Cockx, Frank Canters Vrije Universiteit Brussel, Belgium
Kor de Jong, Derek Karssenberg Utrecht University, the Netherlands

2. Introduction
» Land-use change models are becoming important instruments for the
assessment of policies aimed at
» improved spatial planning
» sustainable development
» scenario analysis
» Need for robust and more reliable tools
» Correct calibration and validation of land-use change models is of major
importance
13/07/2012 2

3. MOLAND land-use model for Dublin
13/07/2012 3

4. MOLAND land-use model for Dublin Land use
Land use
& Interaction
Stochastic
at time T+1
weights
perturbation
t
v 1 ln rand
Suitability &
0 0.5 1
& Transition Rule
Change cells to land use for
Time Loop
which they have the highest
transition potential until the
demands are met.
Transition
Accessibility
Zoning Potentials
& & =
13/07/2012 4

5. Historic calibration
» Land-use change models are typically calibrated using a historic
calibration
Model initialisation Hindcast Forecast
1990 2000 2030
not Ok Ok
Actual map 1990 Actual map 2000
parameters Courtesy of EC JRC
13/07/2012 5

6. Land-use data for calibration
» Dynamic land-use change models require for their calibration time series
of high quality and consistent land-use information.
» Remote sensing data can be used to
» Correct inconsistencies in land-use maps available for calibration
» Produce land-use information at more time steps
» Provide additional land-use information to improve calibration
13/07/2012 6

7. Remote sensing data for calibration
1994 1997 New image
RS New
RS
Data RS
Data
Data
Model initialisation Hindcast Forecast
1990 1994 1997 2000 2030
Model
simulation
Source: MAMUD project
Actual map 1990 Actual 13/07/2012
map 2000 7

8. Spatial Metrics
» Spatial metrics:
» Quantitative measures to describe structures and patterns in the
landscape
» Calculation at different levels of abstraction, e.g. patch, class, moving
window or landscape scale
» Examples of spatial metrics are: fractal dimension, contagion, edge
density, patch density, adjacency event
13/07/2012 8

9. Uncertainties in predicted land use
» A major shortcoming in the historic calibration of land-use change models
is that uncertainties are neglected. Uncertainties mostly exist in:
» Model parameters
» Reference data used for calibration of the model
» This leads to uncertainties in the prediction of land use
1994 1997 New image
RS New
RS
Data RS
Data
Data
1994 1997
13/07/2012 9

10. Spatial metric
Inferred land
use
Image
interpretation
Calibrated
model parameters
Model
initiation
Predicted land
use
Spatial metric
13/07/2012 10

11. Objectives
» Main objectives of the Belspo STEREO II ASIMUD project:
» Improve land-use simulations: lower uncertainties compared to
other automatic calibration methods
» Development of an automatic calibration method using remote
sensing data in an innovative data-assimilation approach
» Robust and reliable tools for land-use change modelling and
calibration for use in policy contexts will be facilitated and promoted
» The probability maps of simulated land use will be valuable additional
data for end users to assess planning policies
13/07/2012 11

12. Calibration with data-assimilation algorithm
» Data-assimilation algorithms
» integrate observations of the state of a system with the modelled state (the
hindcast) to produce the best estimate of the parameter values and state
variables.
» balance the uncertainty in the observation data and in the hindcast.
» provide calibrated parameters as probability distributions
» We apply the Particle Filter, a robust Monte Carlo based method, implemented in
a Python framework
» Data assimilation is often used in atmospheric chemistry models, weather
forecasting, hydrological modelling, GPS technology and astronomy
» Relatively new in the field of land-use change modelling
13/07/2012 12

13. Workflow
1. Model in error propagation mode
(Monte Carlo simulations)
» Uncertain model parameters
2. Model in data assimilation mode
(Particle Filter)
» Uncertain observations
Observations Observations
13/07/2012 13

14. Simplified MOLAND land-use model for Dublin
Simplified
land-use Original MOLAND land-use
model categories
Residential continuous dense
urban fabric, Residential
Population continuous medium dense
related urban fabric, Residential
classes discontinuous urban fabric,
Residential discontinuous sparse
urban fabric
Employment Industrial areas, Commercial
related areas, Public and private
classes services, Port areas,
Arable land, Pastures, Forests,
Semi-natural areas, Wetlands,
Non urban Artificial non-agricultural
vegetated areas, Construction
sites
Road and rail networks and
associated land, Abandonment,
Other Mineral extraction sites, Airport,
Water bodies, Restricted access
areas, Dump sites
13/07/2012 14

15. Simplified MOLAND land-use model for Dublin
» Neighbourhood influence rules: 5 parameters  2 parameters
(exponential function)
(1, a)
(b, c)
(d, 0)
(0, inertia)
13/07/2012 15

16. Quantification of uncertain input parameters
Sill (s) Range (r)
From To Mean Mean
min (SD) max min (SD) max
50.5 0.41
Population Population 1 (25) 100 0.12 (0.2) 0.7
-25.0 0.205
Population Employment -100 (25) 50 0.01 (0.2) 0.4
-50.0 0.355
Employment Population -100 (25) 0 0.01 (0.2) 0.7
50.5 0.455
Employment Employment 1 (25) 100 0.16 (0.2) 0.75
Range = 0.41
Sill
Sill
Range = 0.7
Range = 0.12
13/07/2012 16

17. 1. Error propagation - Probability maps
1990 1997 2001 2006 2010
Employment related urban Population related urban
13/07/2012 17

18. 1. Error propagation - Spatial metrics
PD – Patch Density
Number of urban patches (patches/100ha)
13/07/2012 18

19. 1. Error propagation – Spatial metrics
PLADJ – Percentage of Like
Adjacencies PLADJ = 0 PLADJ = 65
Degree of aggregation of the urban patches
PLADJ = 0: urbanised area is maximally disaggregated
PLADJ = 100: one large urban patch
Population class Employment class
13/07/2012 19

20. 2. Data assimilation – Particle Filter
Step A:
» Apply Bayes’ equation to
realizations of the model
» Results in a ‘weight’ assigned
to each realization
Step B:
» Clone each realization a Step B
number of times proportional to Step A
the weight of the realization
13/07/2012 20

21. Step 1: Apply Bayes’ equation to each realization (particle) i
Prior: PDF of
Prior: PDF of model
observations
realization i
Prior: PDF of observations
given the model realization i
Posterior: probability
distribution function
(PDF) of realization i
given the
observations
13/07/2012 21

22. Calculating weights
æ 1é (i) ùö
( ) = exp ç- ëyt - Ht ( x t )û Rt ëy t - Ht ( x t )û÷
(i) ù -1 é
T
p yt x (i)
t
è 2 ø
Weight of
particle
Measurement
operator = 1
Measurement error
Model
variance
Median value realization
and covariance of
of spatial
observations
metrics for
observations at
time step t
13/07/2012 22

23. 2. Data assimilation - Particle filter
Observations Observations
13/07/2012 23

24. 2. Data assimilation - Particle filter
Number of copies or clones
1997
2001
2006
Population Class
13/07/2012 24

25. 2. Probability maps with data assimilation
1990 1997 2001 2006 2010
Employment related urban Population related urban
13/07/2012 25

26. Conclusions
» Monte Carlo framework for error propagation modelling and particle
filtering was applied to a simplified version of the MOLAND model for
Dublin
» First results seem promising
» However, relative big gap between spatial metrics calculated from the RS-
based land-use maps and the MOLAND land-use map may hamper the
analysis
13/07/2012 26

27. http://www.asimud.be
13/07/2012 27