Sei sulla pagina 1di 7

11/8/2018

Spatial Interpolation & Geostatistics Kriging – Step 1

Describe spatial variation with Semivariogram


“Point cloud” Map
+ + + + Lag + + ++ 4
+ + + + + + +
3
+
(Zi – Zj)2 / 2

Mean 5
++ + + + + + + + + 2
++ + ++ + + + + + + 5
+ +
+ + ++ + + (Zi – Zj)2 / 2
+ + + + + + 6

++ + + + 7

Lag + + + + + 6
++ + + + 5

y 
+ + +
+ + 7
Distance between pairs of points +
x 
Distance between pairs of points

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 1 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 2

Kriging – Step 2 Kriging – Step 2


Summarize spatial variation with a function
Several choices possible; curve fitting defines different Key features of fitted variogram:
Range Nugget: semivariance at
types of Kriging (circular, spherical, exponential, gaussian,
d=0
Semivariance

etc.)
+ + Range: d at which
+ + + semivariance is constant
+ + + + Sill
(Zi – Zj)2 / 2

+ + + + + Sill: constant
+ + + + + semivariance beyond
+ + + the range
+ + + + + Nugget
+ +
+ + + Distance (d)
+
+ +
Distance between pairs of points
Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 3 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 4

1
11/8/2018

Kriging – Part II ESRI Geostatistical Analyst Products

Goal: predict values where no data have been collected  Map types:
Relies on first establishing:  Prediction – contours of interpolated values
DEPENDENCY – z is, in fact, correlated with distance  Prediction Standard Errors – show error distribution, as quantified by
STATIONARITY – z values are stochastic (except for spatial minimized RMS error (see below)
dependency they are randomly distributed) and have no other
dependence – use “detrending” or transformation tools if not  Probability – show probability of values exceeding a specified
Gaussian threshold
DISTRIBUTION – works best if data are Gaussian. If not they have to
first be made close to Gaussian.  Quantile – show where thresholds overestimate or underestimated
predictions

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 5 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 6

ESRI Geostatistical Analyst Products Some Kriging Products


 Prediction map – interpolated values
Maps: (e.g. max. ozone concentration, 1999)
 Probability map- showing where critical values
Prediction
exceeded
Prediction Standard Errors
Prediction Map Probability Map
Probability
Quantile

Prediction map of radioceasium soil contamination levels in of Belarus Dark orange and red indicate a probability greater than 62.5% that
after the Chernobyl nuclear power plant accident. radiocesium contamination exceeds the critical threshold in forest berries.

Figure from ESRI “Intro. to Modeling Spatial Processes Using Geostatistical Analyst” Figures from ESRI “Using Geostatistical Analyst”
Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 7 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 8

2
11/8/2018

Spatial Dependence:
Kriging – Part II Semivariogram and Semivariogram Surface

 Goal: predict values where no data have been collected Strong Spatial Dependence Weak Spatial Dependence
 Relies on first establishing:
DEPENDENCY – z is, in fact, correlated with distance
STATIONARITY – z values are stochastic (except for spatial dependency they
are randomly distributed) and have no other dependence – use “detrending”
or transformation tools if not Gaussian
DISTRIBUTION – works best if data are Gaussian. If not they have to first be
made close to Gaussian.

Figures from ESRI “Intro. to Modeling Spatial Processes Using Geostatistical Analyst”

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 9 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 10

Spatial Dependence:
Cross-Validation Diagnostic Kriging – Part II
Use a subset of the data to test measured vs. predicted values  Goal: predict values where no data have been collected
 Relies on first establishing:
Weak Spatial Dependence DEPENDENCY – z is, in fact, correlated with distance
Strong Spatial Dependence
STATIONARITY – z values are stochastic (except for spatial dependency they
are randomly distributed) and have no other dependence – use “detrending”
or transformation tools if not Gaussian
DISTRIBUTION – works best if data are Gaussian. If not they have to first be
made close to Gaussian.

Figures from ESRI “Intro. to Modeling Spatial Processes Using Geostatistical Analyst”

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 11 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 12

3
11/8/2018

1. SPATIAL DEPENDENCY 2. STATIONARITY - Randomness


Test with semivariogram & cross-validation plots  Data variance and mean is the same at all localities (or within a
neighborhood of nearest points); data variance is constant in the
neighborhood of investigation
 Correlation (covariance) depends only on the vector that separates
localities, not exact locations, number of measurement or direction
• Violates Stationarity
– detrend first

Semivariogram and temperature


distribution map of winter
temperature data for the USA.

Figures from ESRI “Intro. to Modeling Spatial Processes Using Geostatistical Analyst”
Figure from ESRI “Using Geostatistical Analyst”

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 13 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 14

California Ozone Demo. ArcGIS Kriging Processing Steps

Data in “Geostat_demo” folder 1. Add and display the data


2. Explore the data’s statistical properties
3. Select a model to create a surface – make a prediction map!
4. Assess the result
5. Compare to other models

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 16 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 17

4
11/8/2018

Data Exploration:
Data Exploration Examine the Distribution
 Normal (Gaussian) distribution? (central value, spread, symmetry;
1. Examine the distribution – normal (Gaussian)? mean and median the same?)

Transformation to normal required? Transformation to normal required?


 Histograms and QQPlots  Histogram tool, QQPlot tool (compare real and
standard normal distributions)
2. Identify trends, if any
 Trend Analysis unimodal,
near-symmetric
straight line

3. Understand spatial autocorrelation and


directional influences
Semivariogram analysis
= nearly Normally Distributed data

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 18 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 19

Data Exploration: Data Exploration:


Identify Trends, If Any Spatial Autocorrelation & Directional Influences
Underlying trends affect Kriging assumption  Variogram Analysis:
of randomness – remove and work with  Look for correlation Exhibits Spatial Autocorrelation

“residuals” with distance


Trend Analysis tool  Look for directional
trends among pairs
of points
Strong ~NE-SW “U-
shaped” Trend Semivariogram/
Covariance Cloud
tool
Weaker ~NW-SE linear
Directional Influence shown here – pairs in NNW-SSE direction
Trend show largest covariance over shortest distances
Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 20 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 21

5
11/8/2018

ArcGISKriging Processing Steps Mapping Ozone Concentration

1. Add and display the data 1. Incorporate results of Data Exploration into Model selection
2. Explore the data’s statistical properties  This example:
remove underlying trends discovered during data exploration that have a
3. Select a model to create a surface – make a prediction map! rational explanation. (Analysis is then performed on residuals and trend
4. Assess the result surface is added back into final surface) = “Detrending”
Remove directional trends between pairs of points – in certain directions
5. Compare to other models closer points are more alike than in other directions = “anisotropy removal”

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 22 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 23

Mapping Ozone Concentration –


Interpolation & Cross Validation ArcGIS Kriging Processing Steps

2. Define search neighborhood for interpolation 1. Add and display the data
(c.f. I.D.W.) 2. Explore the data’s statistical properties
 Use a search ellipse (or circle) to find nearest 3. Select a model to create a surface – make a prediction map!
neighbors; specify radii of ellipse, min. & max.
number of points per sectors 4. Assess the result – Cross Validation Plots
3. Examine Cross Validation plot 5. Compare to other models
Predicted vs. Measured for subset(s) of the data
“Mean error” should be close to zero
“RMS error” and “mean standardized error” should be small
“RMS standardized error” should be close to one.

Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 24 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 25

6
11/8/2018

Comparing Model Results Probality Mapping with Indicator Kriging


Cross validation comparisons:  Task: Make a map that show the probability
• “Mean error” should of exceeding a critical threshold, e.g. 0.12
be close to zero ppm ozone for an 8 hr. period
• “RMS error” and
“mean standardized
 Technique:
error” should be  Transform data to a series of 0s and 1s according
small to whether they are above or below the threshold
• “RMS standardized
error” should be
 Use a semivariogram on transformed data;
close to one. interpret indicator prediction values as
probabilities
Geo327G/386G: GIS & GPS Applications in Earth Sciences Geo327G/386G: GIS & GPS Applications in Earth Sciences
11/8/2018 Jackson School of Geosciences, University of Texas at Austin 26 11/8/2018 Jackson School of Geosciences, University of Texas at Austin 27

Potrebbero piacerti anche