CE 397 Statistics in Water Resources

Exercise 9
Geostatistical Analysis
by:
John Shaw, Yao You, Ruth Haberman and David Maidment
University of Texas at Austin
Ari! "##$
Contents
%ntrodu&tion'''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' (
)asi&s of *ri+in+''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' "
,ase Study''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' -
.eostatisti&a! Ana!ysis'''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' /
Sim!e *ri+in+ and %nversed Distan&e 0ei+ht Methods''''''''''''''''''''''''''''''''''''''''''''"1
To be turned in'''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''''' 2#
Introduction
%n this exer&ise we wi!! ex!ore Satia! %ntero!ation Methods
(
3i+ure ( : The intero!ated va!ue at the unmeasured ye!!ow oint is a fun&tion of
the nei+hborin+ red oints 43rom Ar&.%S He! Menu5'
A very basi& rob!em in satia! ana!ysis is intero!atin+ a satia!!y &ontinuous
variab!e from oint sam!es' 3or exam!e, many satia!!y ex!i&it
hydro!o+i&6watershed mode!s re7uire &ontinuous surfa&es of temerature' Three
&ommon!y used intero!ation methods to mode! satia!!y distribution from oint
data are %nverse Distan&e 0ei+htin+ 4%D05, s!ine and ordinary 8ri+in+'
The %D0 is sim!e and intuitive deterministi& intero!ation method based on the
rin&i!e that sam!e va!ues &!oser to the redi&tion !o&ation have more in9uen&e
on redi&tion va!ue than sam!e va!ues farther aart' Usin+ hi+her ower assi+ns
more wei+ht to &!oser oints resu!tin+ in !ess smoother surfa&e' :n the other hand,
!ower ower assi+ns !ow wei+ht to &!oser oints resu!tin+ in smoother surfa&e' 0e
otimi;e the ower arameter usin+ Ar&.%S' The ma<or disadvanta+e of %D0 is the
=bu!!>s eye? e@e&t 4hi+her va!ues near observed !o&ation5 and ed+y surfa&e' S!ine is
deterministi& intero!ation method whi&h Ats a mathemati&a! fun&tion throu+h inut
data to &reate smooth surfa&e' A s!ine &an +enerate suB&ient!y a&&urate surfa&es
from on!y a few sam!ed oints and retain sma!! features 4Anderson, "##15' S!ine
wor8s best for +ent!y varyin+ surfa&es !i8e temerature' %n Ar&.%S S!ine is Radia!
)asis 3un&tion'
Un!i8e %D0 and s!ine, 8ri+in+ is method based on satia! auto&orre!ation' %t uses a
semivario+ram'
Basics of Kriging
*ri+in+ was deve!oed in the ($/#s by the 3ren&h mathemati&ian .eor+es
Matheron' The motivatin+ a!i&ation was to estimate +o!d deosited in a ro&8 from
a few random &ore sam!es' *ri+in+ has sin&e found its way into the earth s&ien&es
and other dis&i!ines' %t is an imrovement over inverse distan&e wei+htin+
be&ause redi&tion estimates tend to be !ess bias and be&ause redi&tions are
a&&omanied by redi&tion standard errors 47uantiA&ation of the un&ertainty in the
redi&ted va!ue5'
The basi& too! of +eostatisti&s and 8ri+in+ is the semivario+ram' The semivario+ram
&atures the satia! deenden&e between sam!es by !ottin+ semivarian&e a+ainst
searation distan&e 4semivarian&e wi!! be ex!ained in the next ara+rah5' The
remise of any satia! intero!ation is that &!ose sam!es tend to be more simi!ar
than distant sam!es 4this is a!so &a!!ed satia! auto&orre!ation5' This roerty of
satia! data is im!i&it!y used in %D0' %n 8ri+in+, one must mode! the satia!
auto&orre!ation usin+ a semivario+ram instead of assumin+ a dire&t, !inear
re!ationshi with searation distan&e'
"
Semivarian&e e7ua!s oneCha!f the s7uared di@eren&e between oints searated by a
distan&e dDEd 4assumin+ no dire&tion referen&e5' As the distan&e between
sam!es in&rease, we exe&t the semivarian&e to a!so in&rease 4a+ain, be&ause near
sam!es are more simi!ar than distant sam!es5' This is true, however, on!y u to
some +iven searation distan&e' 3or this distan&e and u, oints are unre!ated'
Stated another way, if F#m is this &riti&a! searation distan&e, two oints searated
by F#m are !i8e!y to be <ust as simi!ar 4or deendent on one another5 as sam!es
searated by (##, "##, 2##, or any distan&e +reater than F#m'
Suose we have the semivario+ram shown in 3i+ure "' 0hat information does the
!ot rovideG 0e!!, the semivarian&e between sam!es searated by no distan&e is
about ('FHC-' This is &a!!ed the nu++et' 0hat it says is that if you measure the
variab!e at !o&ations very, very &!ose to one another, the va!ues measured mi+ht be
7uite di@erent' 0hy wou!d this haenG Suose you had a +o!d nu++et in the
midd!e of an otherwise +o!dCfree ro&8' %f you sam!e <ust on the ed+e of the nu++et
you +et a hi+h +o!d estimate' %f you sam!e <ust outside of the ed+e, you +et no
+o!d in your estimate' The resen&e of a nu++et in the semivario+ram therefore
te!!s you that, assumin+ no measurement error, the variab!e is not satia!!y
&ontinuous'
The semivario+ram a!so te!!s us that oints searated by /#,### m are !i8e!y to
have the same avera+e di@eren&e as oints searated by (##,###, (F#,###,
"##,### m or any distan&e above /#,###m' /#,### m is the ran+e of the
semivario+ram and su++ests the area of in9uen&e for any +iven oint' An
unmeasured !o&ation &an be redi&ted based on its nei+hborin+ sam!es &!oser than
/#,###m' A sam!e &o!!e&ted /(,### m away from the sam!e wi!! !i8e!y have no
in9uen&e on the a&tua! va!ue at the unmeasured !o&ation'
0hen you !oo8 at the mode! of a semivario+ram, you>!! noti&e that at a &ertain
distan&e, the mode! !eve!s out' The distan&e where the mode! Arst 9attens out is
8nown as the ran+e Sam!e !o&ations searated by distan&es &!oser than the ran+e
are satia!!y auto&orre!ated, whereas !o&ations farther aart than the ran+e are not'
The va!ue that the semivario+ram mode! attains at the ran+e 4the va!ue on the yC
axis5 is &a!!ed the si!!' The artia! si!! is the si!! minus the nu++et
2
3i+ure " : The semivario+ram is used to mode! the satia! re!ationshis between
sam!es searated by some distan&e, d
3or 8ri+in+ estimation, the semivaro+ram mode! 4the ye!!ow !ine in A+ure "5 is used
to obtain estimates for the wei+htin+ arameters of H7uation (' This ro&ess is
done automati&a!!y by the +eostatisti&a! ana!yst on&e the user is satisAed with the
semivario+ram' %f you are interested in the derivation of the wei+htin+ arameters
4or any of the other toi&s dis&ussed here5, A!ied .eostatisti&s by Hdward H'
%saa8s and R' Mohan Srivastava is an ex&e!!ent resour&e' :r for the more
mathemati&a! fo!8s, try Statisti&s for Satia! Data by Ioe! A',' ,ressie'
Goals of tis Exercise
Iow we 8now the basi&s of satia! intero!ation' JetKs use our 8now!ed+e to deve!o
a di+ita! e!evation mode! of the 0ax Ja8e :ut!et Area usin+ the ($/2 bathymetri&
data'
Co!"uter Re#uire!ents
This exer&ise is to be erformed in HSR% Ar&MAL $'x with the .eostatisti&a! Ana!yst
extension' The data for this exer&ise is at:
htt:66www'&e'utexas'edu6rof6maidment6Stat0R"##$6Hx$6Hx$';i 4((M)5'
The data used in this exer&ise were ta8en dire&t!y from a ($/2 bathymetry &hart of
the 0ax Ja8e :ut!et Area whi&h is !o&ated in the de!ta of the Mississii River in
Jouisiana, as shown in 3i+ure 2'
Case Study
-
3i+ure 2 : from 0e!!ner et a!' "##F
The Wax Lake Outlet is a man-made channel that connects the Atchafalaya River to Atchafalaya
Bay in southern Louisiana. The Atchafalaya River is connected to the Mississippi river in northern
Louisiana y the Old River !ontrol "tructure. At the Old River !ontrol "tructure# approximately $%& of
the Mississippi's (ater and suspended sediment is routed into the Atchafalaya asin. As sho(n in )i*ure
+# this sediment is uildin* su-areal deltas at the mouths of the Atchafalaya River and the Wax Lake
Outlet at rates of rou*hly ,%km
-
su-areal land per year. .nderstandin* the mechanisms of this delta
*ro(th is seen as a key to etter remediation of (etland loss in coastal Louisiana.

3i+ure - : Wellner et al$ %&''()
We have uncovered a athymetry chart of Atchafalaya Bay from ,/0$. We can make an estimate
of ho( much sediment has accumulated in the delta since then y sutractin* a current di*ital elevation
F
($M
-
"##
"
model from an elevation model created from this athymetric chart. This exercise (ill turn this chart into
a di*ital elevation model.
3i+ure - : from USA,H
Geostatistical Analysis
Goals of tis Exercise
Iow we 8now the basi&s of satia! intero!ation' JetKs use our 8now!ed+e to
deve!o a di+ita! e!evation mode! 4DHM5 of the 0ax Ja8e :ut!et Area usin+ the ($/2
data'
Co!"uter Re#uire!ents
/
This exer&ise is to be erformed in HSR% Ar&MAL $'2 with the Geostatistical Ana!yst
extension' This is avai!ab!e in the ,HCJearnin+ Resour&e ,enter in Room 2'2#(' The
data for this exer&ise is at:
htt:66www'&e'utexas'edu6rof6maidment6Stat0R"##$6Hx$6Hx$';i
The data used in this exer&ise were ta8en dire&t!y from the ($/2 bathymetry &hart
of the 0ax Ja8e :ut!et Area' Additiona! data &ame from IHDL!us Re+ion 1
htt:66www'hori;onCsystems'&om6nhd!us6HS,CwthMS'h
Getting Started
%n this exer&ise we need to use satia! ana!yst and +eosatia! ana!yst too!bar' :en
Ar&.%S' 0e need to enab!e this extension in the Arst !a&e' :en *ools
Extension$
,he&8 the +S"atial analyst, and +Geos"atial analyst, box' Iow we &an use a!!
the fun&tions in both too!bars'
The di+iti;ed water deths from the bathymetri& &hart are &ontained in the
sreadsheet -e"t$xls and &onsist of !atitude and !on+itude oints asso&iated with
a deth of water at that oint, measured in feet from the water surfa&e to the bed'
Sin&e these data were maed in ($/2, the !atitudes and !on+itudes wou!d have
M
been referen&ed to the Iorth Ameri&an Datum of ($"M 4IAD"M5, the standard for
earth datums at that time'
These data were extra&ted from the Hx&e! A!e and &onverted to Ar&.%S as oint
events, to be&ome the feature &!ass DethHvents' This feature &!ass was &ou!ed
with another for ./-Water0ody reresentation to rovide satia! &ontext for the
oint set' The IHD dataset is resented in +eo+rahi& &oordinates with the Iorth
Ameri&an Datum of ($12 4IAD125' )e&ause satia! intero!ation has to be done in
a ,artesian 4x,y5 &oordinate system rather than a 4!atitude, !on+itude5 &oordinate
system, a!! the data have been ro<e&ted into the State L!ane ,oordinate system for
South Jouisiana, usin+ the IAD12 datum' The DethHvents were Arst &onverted
from IAD"M to IAD12 datum before this ro<e&tion to State L!ane ,oordinates was
done' The resu!tin+ dataset is stored in the 0axJa8e:ut!et feature dataset in the
Hx$ +eoddatabase, as shown be!ow:
:en Ar&.%S, and add the -e"tE1ents and ./-Water0ody !ayers to the ma
dis!ay' The fo!!owin+ view wi!! aear on your s&reen after some maniu!ation of
the !e+ends and feature &!ass !abe!s:
1
Ioti&e the oint !abe!s, whi&h identify the deth in feet' Most of the oints are
between " and M, with an area at the very bottom of the dataset with mu&h
more extreme deths 4/" feet, "$ feet, (- feet5' These are not an anoma!y,
but re9e&t a dee tren&h o@Cshore'
Ex"loratory S"atial -ata Analysis
Hx!oratory Satia! Data Ana!ysis 4HSDA5 is a ro&ess of understandin+ the
roerties of a satia! dataset in order to best mode! the data usin+ +eostatisti&s'
The word =Hx!ore? shou!d te!! you that HSDA is more of an adventure than a stri&t N
you must fo!!ow the ath at a!! times N ro&edure' %n this exer&ise we show how the
.eostatisti&a! Ana!yst too!s &an be used to understand the ou!ation distribution of
$
the attribute of interest and how to understand the !ar+eCs&a!e atterns in the
dataset throu+h 2D visua!i;ation' This is not a &omrehensive !ist of HSDA
ro&edures, but a +ood start' Throu+h this ro&ess, 8ee in mind that the better
one understands the satia! &hara&teristi&s of the data, the better 8ri+in+ mode! one
&an bui!d to intero!ate the data, and, &onse7uent!y, the better estimates one wi!!
rodu&e'
/istogra!
:en Geos"atial analyst Ex"lore -ata /istogra!$
Se!e&t the -# bars and &he&8 the statisti&s otion to view the statisti&s of the
surfa&e e!evations' *ee the transfor!ation as .one' Se!e&t layer as
-e"tE1ents and Attri0ute as -e"t' A simi!ar window to be!ow wi!! aear on
s&reen if you in&rease the number of bars to -#'
(#
0e wi!! now ana!y;e the histo+ram for the surfa&e e!evations at the 0ax Ja8e
:ut!et' The uer ri+ht &orner shows the statisti&s of the surfa&e e!evations' The
histo+ram shows that our data is not erfe&t!y norma!!y distributed but the
s8ewness is not !ar+e' 0e shou!d remember that we did not fu!!y fo!!ow this
assumtion when ana!y;in+ these data' :ne of the &ross&he&8s of norma!
distribution of data is that mean shou!d be &!ose to the median' %n our &ase mean is
F'M2#" feet and median is / feet, so we arenKt too far o@' 0e &an &onsider our data
as norma!!y distributed' User &an transform the data into log or 0ox2cox
distribution if it is not norma!!y distributed'
33 %3uantile23uantile) "lot
Another way to understand the dataKs distribution is by usin+ the Iorma! OO L!ot
too!' %n a OO 4Ouanti!eCOuanti!e5 !ot we test whether data are norma!!y distributed
by !ottin+ it a+ainst a dataset with a 8nown norma! distribution' %f the !ot is !inear
a!on+ the !ine YPQ, then the data fo!!ow a norma! distribution'
:en Geos"atial analyst Ex"lore data .or!al 33 "lot$
The OO !ot shows the !inear re!ationshi between !o+4DethsRHvents5 and the
standard norma! distribution' The data is not exa&t!y !inear on the uer be&ause of
oints from the !ar+e tren&h'
((
So, !ets !o+Ctransform the data to a!!ow for that:
:8, this !oo8s better'
("
*rend analysis
The trend ana!ysis too! rovides a 2D !ot of the sam!es and a re+ression on the
attribute in the QS and YS !anes' The urose of the too! is to visua!i;e the data
and to observe any !ar+eCs&a!e trends that the mode!er mi+ht want to remove rior
to estimation' %t is best to 8ee the 8ri+in+ mode! as sim!e as ossib!e and to on!y
remove a trend if it si+niA&ant!y imroves redi&tion errors'
:en Geos"atial analyst Ex"lore data *rend analysis$
The s&reen wi!! !oo8 simi!ar to the A+ure be!ow' 0e &an &hoose the di@erent +rah
otions and rotate the !o&ation to see the trend in the data' %t !oo8s !i8e these data
have some si+niA&ant trends in them 4whi&h we wou!d exe&t be&ause water 9ows
downhi!!T5'
(2
The trend ana!ysis too! rovides a 2D !ot of the sam!es and a re+ression on the
attribute in the QS and YS !anes' The urose of the too! is to visua!i;e the data
and to observe any !ar+eCs&a!e trends that the mode!er mi+ht want to remove rior
to estimation' %t is best to 8ee the 8ri+in+ mode! as sim!e as ossib!e and to on!y
remove a trend if it si+niA&ant!y imroves redi&tion errors'
4erfor! Kriging inter"olation
:en Geos"atial analyst Geos"atial Wi5ard
Se!e&t the Kriging as !etod and -e"tE1ents as in"ut data' Se!e&t attri0ute
as -e"t$
(-
,!i&8 .ext to ro&eed'
Jets use a Jo+ transformation of the data, and a 3irst :rder trend remova! to a!!ow
for the trends that our ear!ier ex!oratory data ana!ysis su++ested'
(F
0e wi!! &hoose the most wide!y used ordinary 6riging method and "rediction
!a"' As dis&ussed in
,!i&8 .ext to ro&eed'
YouK!! see a ma that shows the Atted !inear trend in the data
(/
Hit .ext, and !ets &hoose the defau!t S"erical mode! for the intero!ation
(M
:ther &ommon!y used mode!s are exonentia! and .aussian' Ja+ si;e and number
of !a+s are the imortant arameters to be se!e&ted' These two arameters are
used to +rou the number of airs of data' 3or our ana!ysis we wi!! use the defau!t
va!ues' 3or +iven !a+ si;e and number of !a+s, Ar&.%S automati&a!!y &a!&u!ates the
nu++et, ran+e and si!!' Semivario+ram for arti&u!ar dire&tions &an be investi+ated
usin+ so7 searc direction too!'
,!i&8 .ext to ro&eed'
(1
,han+e .eig0ours to include to 8& and in&!ude at !east 8' stations' Sin&e we
have the tota! 38( data oints, (" and minimum (# stations is an aroriate
number for intero!ation' ,!i&8 on the third &ir&!e for sector ty"e'
,!i&8 .ext to ro&eed'
($
These are the error statisti&s for the intero!ation' This is based on oneC!eaveCout at
a time &ross va!idation method' 0e &an save the &ross va!idation tab!e to erform
further statisti&a! ana!ysis'
,!i&8 9inis$
You +et a reort of the ana!ysis'
"#
This is the summary of intero!ation arameters &hosen' ,!i&8 :6$
Iow you wi!! see the fo!!owin+ di+ita! e!evation ma 4DHM5 of the 0ax Ja8e :ut!et'
"(
0e wi!! save this as a raster dataset' Rigt clic6 on the :rdinary Kriging !ayer'
Se!e&t -ata Ex"ort to raster$
""
Se!e&t outut raster as fo!der and A!e name as -e"ts;E1ents
Add this !ayer to the ma'
You &an twea8 the symbo!o+y to show sha!!ow water as !i+ht and dee water as a
dar8er b!ue by ri+ht &!i&8in+ on the raster !ayer, and +o to 4ro"erties
Sy!0ology< and &hoosin+ an aroriate &o!or ram'
"2

This is what you wi!! see'
"-
You &an a&tua!!y see the dee area at the bottom of the DHM, and the sha!!ow areas
that have been &aused by sediment deosition'
Jets reeat the ro&ess and +et the 4rediction Standard Error =a"
"F
"/
"M
You &an &!i&8 on the ma at any oint in the intero!ation and use the %dentify too!
to +et the redi&ted deth and its standard error of estimate: in this &ase /'1 ft U6C
2'# ft'
Si!"le Kriging and In1ersed -istance Weigt =etods
Iow +o ba&8 to Geos"atial analyst Geos"atial Wi5ard$ Sti!! &hoose Kriging
and &!i&8 next' This time &hoose Si!"le Kriging as the intero!ation method' The
rest is the same as :rdinary Kriging ro&edures' You wi!! +et a resu!t simi!ar to the
fo!!owin+ 4after over!ay the ori+ina! data oints on to5:
"1
Ioti&e that this ma a+rees erfe&t!y with the data oints where the :rdinary
*ri+in+ method doesnKt' This is be&ause in Sim!e *ri+in+ the va!ues at those
ori+ina! data oints are 8et the same whi!e in :rdinary *ri+in+ every oint on is
ma is reassi+ned with an intero!ated va!ues' Dis&uss in the homewor8 whi&h one
is more reasonab!e here'
0e +o ba&8 to Geos"atial analyst Geos"atial Wi5ard a+ain, and this time
!etKs try %nverse Distan&e 0ei+ht method' This method does not assume any
distribution of the data' Try with di@erent arameters and &reate one you !i8e the
most 4re&ord the arameters you used for your Ana! resu!t5' The fo!!owin+ is one
exam!e of the outut'
"$
:bvious!y this method does not 8ee the ori+ina! va!ues at the sam!e sites' You
&an !earn more about this method by readin+ the do&ument rovided by the
software' Dis&uss in your homewor8 whether this method is better than other two
methods for this exer&ise'
*o 0e turned in
1. A histogram, normal QQ plot, trend analysis, for the data.
2. A semivariogram, method summary, a DEM, and a Prediction Standard Error
Map developed using rdinary !riging, Simple !riging, and "nverse Distance
Method. Discuss the di#erence $et%een these techni&ues. 'ote that Simple
!riging assumes %e have 1(() con*dence in data, %hich %e do not have for
the 1+,- $athymetry. rdinary !riging does not assume that %e trust the
data 1((), and does a $etter .o$ of /smoothing out0 the output.
-. 1ry the same process, $ut see %hat happens if you don2t log transform the
data or remove its trends. Do you thin3 the result is signi*cantly di#erent
from %hat you got earlier4
2#
2(