spatial statistics with image analysis · spatial statisticsspatialnon-gaussiancorrupted pixelsmore...

Spatial Statistics Spatial Non-Gaussian Corrupted Pixels More

Spatial Statistics with Image Analysis

Johan Lindstrom1

1Mathematical StatisticsCentre for Mathematical Sciences

Lund University

LundOctober 11, 2018

Johan Lindstrom - [email protected] Spatial Statistics 1/35


Outline

Spatial Statistics with Image AnalysisBayesian statisticsHierarchical modellingEstimation Procedures

Spatial StatisticsStochastic FieldsGaussian Markov Random FieldsImage ReconstructionEnvironmental Data

Non-Gaussian DataExamples

Corrupted PixelsNDVI

Learn more


Spatial Statistics Spatial Non-Gaussian Corrupted Pixels More Bayesian statistics Hierarchical Estimation

Spatial Statistics

Many things vary in space and obsevations maydepend on what happens at nearby locations. Tomodell and analys such data we need spatialdependence.

Spatial Interpolation

Given observations at some locations (pixels),y(ui), i = 1 . . .nwe want to make statements about the value atunobserved location(s), x(u0).

Stationary Stochastic Processes (FMSF10)in 2+ dimensions!



Bayesian modelling

We assume that there is some unknown truth, that wewould like to find out about. This “reality” can bemeasured, usually with measurement variation, and oftenonly partially.

Bayesian modelling

A Bayesian model consists ofI A prior, “a priori”, model for reality, x, given by the

probability density π(x).I A conditional model for data, y, given reality, with

density p(y|x).

The prior can be expanded into several layers creating aBayesian hierarchical model.



Bayes’ Formula

How should the prior and data model be combined to makestatements about the reality x, given observations of y?

Bayes’ Formula

p(x|y) = p(y|x)π(x)p(y)

=p(y|x)π(x)∫

x′∈Ω p(y|x′)π(x′) dx′

p(x|y) is called the posterior, or “a posteriori”, distribution.

Often, only the proportionality relation

p(x|y) ∝ p(x,y) = p(y|x)π(x)

is needed, when seen as a function of x.



Hierarchical Models

I We often have some prior knowledge of the reality.I Given knowledge of the true reality, what can we say

about images and other data?I Construct a model for observations given that we know

the truth.I Given data, what can we say about the unknown

reality?This is the inverse problem.



Bayesian hierarchical modelling (BHM)

A hierarchical model is constructed by systematicallyconsidering components/features of the data, and how/whythese features arise.

Bayesian hierarchical modelling

A Bayesian hierarchical model typically consists of (at least)

Data model, p(y|x): Describing how observations ariseassuming known latent variables x.

Latent model, p(x|θ): Describing how the latent variables(reality) behaves, assuming known parameters.

Parameters, π(θ): Describing our, sometimes vauge, priorknowledge of the parameters.



Estimation Procedures

Maximum A Posteriori (MAP): Maximise the posteriordistribution p(x|y) with respect to x.

I Standard optimisation methodsI Specialised procedures, using the model

structure

Simulation: Simulate samples from the posteriordistribution p(x|y). Estimate statisticalproperties from these samples. The samples canbe seen as representative “possible realities”,given the available data.

I Markov chain Monte Carlo (MCMC)I Gibbs sampling


Spatial Statistics Spatial Non-Gaussian Corrupted Pixels More Fields GMRF Reconstruction Environmental

Image Reconstruction

Spatial Interpolation

Given observations at some locations (pixels),y(ui), i = 1 . . .nwe want to make statements about the value atunobserved location(s), x(u0).

The typical model consists of a latent Gaussian field

x ∈ N (μ,Σ) ,

observed at locations ui, i = 1, . . . ,n, with additive Gaussiannoise (nugget or small scale variability)

yi = x(ui)+ εi εi ∈ N(

0,σ2ε

).



Stochastic Fields

To perform the reconstruction (interpolation) we need amodel for the spatial dependence between locations(pixels).

1. Assume a latent Gaussian field

x ∈ N (μ,Σ) .

2. Assume a regresion model for μ = Bβ.

3. Assume a parametric (stationary) model for thedependence (covariance)

Σi,j = C(x(ui), x(uj)) = r(ui,uj;θ) = r(∥∥ui − uj

∥∥ ;θ).

r(ui,uj;θ) is called the covariance function.



Matern covariances functions

I One of the most common families of covariancefunctions is named after Bertil Matern, who worked forStatens Skogsforskningsinstitut (Forest Research Institute of Sweden).

I Variance σ2 > 0, scale parameter κ > 0 and shapeparameter ν > 0

rM(h) =σ2

Γ(ν)2ν−1 (κ‖h‖)νKν(κ‖h‖), h ∈ Rd,

I A measure of the range is given by ρ =√

8ν/κ.



0 20 40 60

0

0.2

0.4

0.6

0.8

1

nu=0.5

0 20 40 60

0

0.2

0.4

0.6

0.8

1

nu=1.0

0 20 40 60

0

0.2

0.4

0.6

0.8

1

nu=2.0

0 20 40 60

0

0.2

0.4

0.6

0.8

1

nu=4.0



nu=0.5

5 10 15 20 25 30

5

10

15

20

25

30

nu=1.0

5 10 15 20 25 30

5

10

15

20

25

30

nu=2.0

5 10 15 20 25 30

5

10

15

20

25

30

nu=4.0

5 10 15 20 25 30

5

10

15

20

25

30



A local model

Instead of specifying the covariance function we couldconsider the local behaviour of pixels. A popular model isthe conditional autoregressive, CAR(1) model.

xij =1

4 + κ2

(xi−1,j + xi+1,j + xi,j−1 + xi,j+1

)+ ε,

ε ∈ N(

0,1τ2

).

This corresponds to a model for x where

x ∈ N(

0,Q−1),

where Q is called the precision matrix



Matern covariances

The Matern covariance family

The covariance between two points at distance ‖h‖ is

rM(h) =σ2

Γ(ν)2ν−1 (κ‖h‖)νKν(κ‖h‖)

Fields with Matern covariances are solutions to a StochasticPartial Differential Equation (SPDE) (Whittle, 1954),

(κ2 −Δ

)α/2x(u) =W(u).



Lattice on R2

Order α = 1 (ν = 0):

κ2

1

︸︷︷︸(C)

+

−1−1 4 −1

−1

︸︷︷︸≈−Δ (G)

Order α = 2 (ν = 1):

κ4

1

︸︷︷︸(C)

+2κ2

−1−1 4 −1

−1

︸︷︷︸≈−Δ (G)

+

12 −8 2

1 −8 20 −8 12 −8 2

1

︸︷︷︸≈Δ2 (G2=GC−1G)



Spatial models for data

GMRF representations of SPDEs can be constructed foroscillating, anisotropic, non-stationary, non-separablespatio-temporal, and multivariate fields on manifolds.

(κ2 −Δ)(τ x(u)) =W(u), u ∈ Rd





(κ2 −Δ)(τ x(u)) =W(u), u ∈ Ω





(κ2 eiπθ −Δ)(τ x(u)) =W(u), u ∈ Ω





(κ2u +∇ ·mu −∇ ·Mu∇)(τux(u)) =W(u), u ∈ Ω



Image Reconstruction II

Model with observations, y, and latent field, x,

y|x ∈ N(

Ax,σ2I)

x ∈ N(μ,Q−1

).

and Q = κ2C + G or Q = κ4C + 2κ2G + GC−1G.

Interpolation using a GMRF

E (x|y) = μ+1σ2 Q−1

x|y A> (y − Aμ)

V (x|y) = Q−1x|y =

(Q +

1σ2 A>A

)−1



Image Reconstruction



Global Temperature — Data

January 2003 July 2003



Global Temperature — Reconstruction

Global mean: 15◦C.



Satellite Data — Vegetation

January 1999

July 1999



Satellite Data — Trend in Vegetation David Bolin

K2 Estimate

−0.4

−0.2

0

0.2

0.4

Independent estimates

K2 Estimate

−0.3

−0.2

−0.1

0

0.1

Correlated estimates


Spatial Statistics Spatial Non-Gaussian Corrupted Pixels More Examples

Non-Gaussian Data

Bayesian hierarchical modelling

A Bayesian hierarchical model typically consists of (at least)

Data model, p(y|x): Describing how observations ariseassuming known latent variables x.

Latent model, p(x|θ): x ∈ N(μ,Q−1

).

Parameters, π(θ)

So far we have assumed Gaussian observations

y|x ∈ N(

Ax,σ2I)

However we could (almost) as easily handle

yi|x ∈ F (g(Ax); θ)



Larynx Cancer — Count data

Given counts of larynx cancer cases, yi, and population ineach region, Ei, we want to estimate the risk of cancer.

Counts of Larynx Cancer Population (truncated)

yi|xi ∈ Po (Ei exp(xi))



Insurance Claims — Count data Oscar Tufvesson

Given the number of insurance claims, yi, we want toestimate the risk of an accident.

0,7500 1,5000

exp(X*beta)0,7500 1,5000

exp(u)

yi|ηi ∈ Po (Ei exp(ηi))

η = Bβ+ x



Parana Rainfall — Positive data

January Precip.

5

10

15

20

January log(Precip.)

1

2

3

0 10 200

0.05

0.1

0.15

Gamma

Y5 10 15

5

10

15

20

1 2 30

0.5

1

Normal

log(Y)1 2 3

1

2

3

yi|xi ∈ Γ(

b,exi

b

)




Spatial Statistics Spatial Non-Gaussian Corrupted Pixels More NDVI

Image Reconstruction — Corrupted Pixels

I Typically we don’t know which pixels are bad.I A better model is then

I Assume an underlying image, x.I Assume an indicator image for bad pixels, z.I Given the indicator we either observe the correct pixel

value from x or noise.

I Use Bayes’ formula to compute the distribution for theunknown image (and indicator) given observations andparameters.



Image Reconstruction — Corrupted pixels

image reconstruction

bad pixels − estimatebad pixels



Normalized difference vegetation index (NDVI)

NDVI =RNIR − RRED

RNIR + RRED

I RRED is the amount of reflected red light (0.58−0.68 μm)I RNIR is the amount of reflected near-infrared light

(0.72− 1.00 μm)Johan Lindstrom - [email protected] Spatial Statistics 31/35


Smoothed version of the NDVI Data

Smooth the data to fill in missing values and remove noisedue to cloud cover, etc.

1986 1988 1990

140

160

180

200

Important ecological questions:I Plant phenology (start and end of season)I Plant productivity (integral)



Smoothing of Satellite Based Vegetation Measurements

NDVI May 1993 Smooth

1991 1992 1993 1994 1995 1996

140

160

180

200

220

240



Smoothing of Satellite Based Vegetation Measurements

1991 1992 1993 1994 1995 1996 1997

140

160

180

200

220

240

Obs

Gaussian

Gaussian 1D

Nig 1D



Learn more!

What?Spatial statistics with image analysis, FMSN20

When?HT2-2018, October–December

Where?Information and Matlab files will be available atwww.maths.lth.se/matstat/kurser/fmsn20masm25/

(currently the 2017 webpage, updated soon)

Who?Lecturer: Johan Lindstrom

MH:319


spatial statistics with image analysis · spatial statisticsspatialnon-gaussiancorrupted pixelsmore...

Documents