Milan ŽukoviÄ. Institute of Physics. P.J. Å afárik University in KoÅ¡ice. Dionisios Hristopulos. Geostatistics research unit. Technical University of Crete ...
Modified planar rotator model for efficient spatial prediction Milan Žukovič Institute of Physics P.J. Šafárik University in Košice
Dionisios Hristopulos Geostatistics research unit Technical University of Crete
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
1
Outline
Spatial prediction problem “Interaction” based approach
Modified planar rotator model
Results
Spatial correlations
Prediction of missing data
Summary
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
2
Spatial prediction problem -
-
Typical problem in geostatistics: Available sample data ZS={zi} at GS={(xi, yi)} Task: to predict values ZP at prediction points GP Classical geostatistical methods (e.g. kriging) Computationally intensive for large data Usually assume multivariate normal distribution Require human (subjective) input
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
3
“Interaction” based approach
Short-range interaction based approach [1]
-
-
Computationally efficient Applicable to gridded and scattered data Requires minimal user input Limited to Gaussian data
Spin model based on “energy matching” approach [2]
-
Heuristic assumption of the existence of relevant correlations Discrete predictions even for continuous data
-
-
[1] D.T. Hristopulos, Gibbs random field models for geostatistical applications, SIAM J. Scient. Comput., 24(6):2125-2162, (2003). [2] M. Žukovič and D T. Hristopulos, Phys. Rev. E 80 011116 (2009); Journal of Statistical Mechanics: Theory and Experiment P02023 (2009).
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
4
Planar rotator model
Hamiltonian H = – J ∑si.sj = – J ∑cos(i- j)
where J – exchange interaction constant si – spin at site i (2d unit vector) i – turn angle of si
Types of vortices
Vortex pair (dipole)
A. Kaser at al., 2D XY model (web project) Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
5
Mapping Zi ↔ i max
Zmax
Modified planar rotator
Spin angles
Gesostatistical data Zmin
2π
Energy H
π
HΔ=0 = HΔ=2π Degeneracy!
0
min
IC-MSQUARE, Mykonos, Jun 5-8, 2015
6
Modified planar rotator model
Hamiltonian
where q ≤ 1/2
H* = – J ∑cos[q(i- j)]
(a) q = 1
Modified planar rotator
(b) q = 1/2
IC-MSQUARE, Mykonos, Jun 5-8, 2015
7
Modeling spatial correlations Matérn correlation function where
Variogram function where
[3] B. Minasny, A.B. McBratney, Geoderma 128 192 (2005)
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
8
Results Spatial correlations Variation in time for temp. T = 0.01 and (a) τ1 = 10, (b) τ2 = 102,
(c) τ3 = 103, (d) τ4 = 104
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
9
Results Spatial correlations Variation in time for temp. T = 0.0001 and (a) τ1 = 10, (b) τ2 = 102,
(c) τ3 = 103, (d) τ4 = 104
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
10
Results Prediction 1. 2.
conditional MC simulations in equilibrium at T estimated from samples predictions obtained as average values from Nc eqilibrium realizations
(a) Lidar elevation data (83% missing) Modified planar rotator
(b) Prediction map (T=10-5, Nc=100)
IC-MSQUARE, Mykonos, Jun 5-8, 2015
11
Results Prediction performance Synthetic data with Matérn covariance (ξ = 5, ν = 2.5) and p% missing values p=30%
p=90%
L=1024
L=256
L=64
Original
Reconstructed p=60%
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
12
Results Prediction performance L=64 p = 30%
p = 60%
p = 90%
AAE
0.1850
0.3525
1.2711
ARE [%]
-0.02
-0.10
0.01
AARE [%]
0.39
0.76
2.68
RASE
0.2543
0.5015
1.7325
L=1024
p = 30%
p = 60%
p = 90%
AAE
0.1757
0.3140
1.0973
ARE [%]
-0.05
-0.09
-0.39
AARE [%]
0.37
0.67
2.34
RASE
0.2302
0.4232
1.4886
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
13
Results Computational performance Relaxation process
In Matlab® on desktop computer with 1.93 GB of RAM and Intel® Core ™2 CPU 6320 @ 1.86 GHz 1.86 GHz
(a) Number of MC sweeps
Modified planar rotator
(b) CPU time
IC-MSQUARE, Mykonos, Jun 5-8, 2015
14
Summary Modified planar rotator model
Spatial correlations
Short-range nature – typical for many geophysical and environmental applications Great variability in the parameter space of temperature and relaxation time
Spatial prediction by conditional MC simulations
Universal - no assumptions about data distribution Automatic – no user imput required Efficient – suitable for large (remote sensing) data Accuracy – ??? more testing needed
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
15
Thank you for your attention!
Modified planar rotator
IC-MSQUARE, Mykonos, Jun 5-8, 2015
16