Using GIS, remote sensing and a land cover change model to assess

foulchilianAI and Robotics

Oct 20, 2013 (3 years and 9 months ago)

175 views








Using GIS, remote sensing and
a land cover change model

to assess
urban sprawl dynamics and forecast future change in Austin, Texas





David Hunt



MSc Thesis submitted in
partial fulfilment of the requirements of the degree of Master of
Science in Geographic Information Science at Birkbeck, University of London


Date of submission: 12
th

September 2012



Total word count: 9982

2

A
bstract

The ability to map and monitor the exte
nt of urban sprawl over time has important
environmental, economic and societal relevance. In this study I developed a methodology
to map and describe urban growth using Landsat 5 data in the city of Austin for a three
epoch time
-
series dating from 1988 to

2010. The NDVI differencing change detection
method was employed to establish urban growth and filtered to avoid confusion with
agricultural variability. The results highlight significant urban growth throughout the Austin
area, with the city expanding at

approximately 14km
2

per year over the 22 year period.
Trends typical of urban sprawl were witnessed, such as greater urban growth away from
the city centre over time and a strong correlation between urban growth and roads. The
change maps were then input
into the Idrisi Land Change Modeler in order to make a
prediction for the amount and location of future urban sprawl. The distance from previous
sprawl and distance from roads were input in order to create a transition potential for the
model. A map for 20
10 was created and compared to the actual 2010 land cover map. The
model was able to predict urban growth and was very close to the actual land cover map in
terms of amount of sprawl. Finally a simulation was run for the year 2015. The study
highlighted th
e potential of using multi
-
temporal Landsat data and remote sensing
methods to provide an accurate and economical means to map and analyse urban growth
over time. The maps created, along with the prediction for the future could provide useful
tools in land

management planning and policy decisions to tackle urban sprawl.




3

Acknowledgements



4

Contents


Abstract

................................
................................
................................
................................

2

Acknowledgements

................................
................................
................................
..............

3

List of Figu
res

................................
................................
................................
.......................

6

List of Tables

................................
................................
................................
........................

7

Declaration

................................
................................
................................
...........................

8

1.

Introduction

................................
................................
................................
.....................

9

2.

Aims and
Objectives

................................
................................
................................
.......

11

2.1

Research Hypothesis

................................
................................
..............................

11

3. Literature Review

................................
................................
................................
...............

12

3.1 Urban Sprawl

................................
................................
................................
................

12

3.2 Remote Sensing

................................
................................
................................
...........

17

3.3 Change

Detection

................................
................................
................................
.........

19

3.4 Land Use/Cover Change Models

................................
................................
..................

21

3.5 IDRISI Land Change Modeler

................................
................................
........................

22

3.6 Multi
-
Layer Perceptron Neural Network

................................
................................
.....

23

3.7 Predicting Change


Markov Chain

................................
................................
..............

25

4. Study Area

................................
................................
................................
..........................

26

5.

Methodolog
y

................................
................................
................................
..................

30

5.1 Data Sources

................................
................................
................................
................

31

5.2 Image Pre
-
processing

................................
................................
................................
...

33

5.3 Change Detection

................................
................................
................................
.........

39

5.4 Validation and Accuracy Assessment

................................
................................
...........

46

5.5 Spat
ial Analysis

................................
................................
................................
.............

49

5.6 Land Cover Change Modelling

................................
................................
.....................

51

5.7 Transition Potentials

................................
................................
................................
....

53

6.

Results and Discussion

................................
................................
................................
...

61

6.
1 Map validation and accuracy assessment

................................
................................
...

61

6.2 General trends and dynamics of urban growth

................................
...........................

65

6.3 Austin population change

................................
................................
............................

74

6.4 Cr
eation of a change map for 1988
-
2010

................................
................................
....

75

6.5 Validation of change map

................................
................................
............................

76

6.6 Predicted urban growth map for 2015

................................
................................
........

79

6.7 Summary of Results

................................
................................
................................
.....

80

5

6.8 Limitations

................................
................................
................................
....................

83

7.

Conclusi
on

................................
................................
................................
......................

85

7.1 Future Research

................................
................................
................................
...........

86

Referenc
es

................................
................................
................................
.............................

88

Appendix

................................
................................
................................
................................

93

Appendix A


NDVI threshold selection

................................
................................
.............

93

Appendix B


Austin Census Tracts

................................
................................
....................

95




6

List of Figures


Figure 1:

Urban sprawl in Melbourne, Australia……………………………………………………

13

Figure 2:

Neural Network Diagram……………………………………………………………………….

24

Figure 3:

Location of Austin in the USA and Texas………………………………………………..

26

Figure 4:

Austin, Texas, with primary/secondary roads present
…………………
……..…

27

Figure 5:

Austin’s urban area expansion from 1970


2004…………………………………..

28

Figure 6:

Generalised work flow for creation of change detection maps……………..

30

Figure 7:

Austin urban area, as delineated by the US Census bureau, 2010………….

34

Figure 8:

Austin false colour composite for 1988…………………………………………………..

35

Figure 9:

Austin false col
our composite for 1995………………………………………………….
.

36

Figure 10:

Austin false colour composite for 2003…………………………………………………..

37

Figure 11:

Austin false colour composite for 2010…………………………………………………..

38

Figure 12:

VEGINDEX module in Idrisi


NDVI
selected…………………………………………..

39

Figure 13:

Austin NDVI map for 1988………………………………………………………………………

40

Figure 14:

Austin NDVI map for 1995………………………………………………………………………

40

Figure 15:

Austin NDVI map for 2003………………………………………………………………………

40

Figure 16:

Austin NDVI
map for 2010………………………………………………………………………

40

Figure 17:

Austin agricultural areas, identified from the Austin land use data set…
.

41

Figure 18:

Change detection map for epoch 1, between 1988 and 1995……………….
.

43

Figure 19:

Change detection map for epoch 2,

between 1995 and 2003……………….
.

44

Figure 20:

Change detection map for epoch 3, between 2003 and 2010……………….
.

45

Figure 21:

Accuracy assessment for errors of commission……………………………………
.

46

Figure 22:

Raster calculator to assess errors of
commission………………………………


47

Figure 23:

Accuracy assessment for increase in NDVI…………………………………………..

48

Figure 24:

Calculating the distance of urban growth away from the CBD…………….
.

49

Figure 25:

Generalised workflow for the future urban growth prediction
map…


51

Figure 26:

Change in urban growth between epoch 1 and epoch 2……………………


52

Figure 27:

Distance from roads calculated in Idrisi……………………………………………….
.

54

Figure 28:

Distance from previous sprawl calculated in Idrisi………………………………
.

55

Figure 29:

MLP N
eural Network sub model run…………………………………………………


57

Figure 30:

T
ransition potential map


from Austin to a decrease in NDVI…………


58

Figure 31:

Composite change in NDVI, from 1988


2010……………………………………
.

61

Figure 32:

Change in NDVI graphs for the three
change epochs…………………………
.

62

Figure 33:

Austin urban growth 1988


2010………………………………………………………
..

64

Figure 34:

Distance of urban growth from the Austin CBD…………………………………
..

65

Figure 35:

Distance of urban growth from CBD


epoch 1…………………………………


66

Figure
36:

Distance of urban growth from CBD


epoch 2…………………………………


67

Figure 37:

Distance of urban growth from CBD


epoch 3…………………………………


67

Figure 38:

Distance of urban growth from roads


epoch 1………………………………


68

Figure 39:

Distance of urban growth from
roads


epoch 2………………………………


69

Figure 40:

Distance of urban growth from roads


epoch 3………………………………


69

Figure 41:

Spatial trend of change over the three epochs…………………………………
..

71

Figure 42:

Urban growth in census tracts…………………………………………………………
….

72

Figu
re 43:

Austin population figures from 1990


2010……………………………………
….

73

Figure 44:

Change demand modelling dropdown in Change Prediction tab, LCM..

74

7

Figure 45:

Comparison of actual decrease in NDVI map for three epochs and
prediction created in LCM……………………………………………………………………



䙩gu牥r46:

Validation results…………………………………………………………………………………



䙩gu牥r47:

Validation map showing false alarms, misses and hits……………………
……



䙩gu牥r48:

P牥摩捴敤⁕牢慮⁧ o睴U慰⁦ 爠1988


2015……………………………………..



䙩gu牥r49:

Time series of urban growth in Austin…………………………………………………





List of Tables


Table 1:

Urban sprawl studies utilising a time
-
series of Landsat data
……………
….

18I19

呡扬攠2:

L慮T獡琠Im慧敲y⁵獥搠楮 瑨攠獴畤y
……………………………………………………


31I32

呡扬攠3:

US census bureau shapefiles……………………………………………………………
….



呡扬攠4:

䍩瑹 o映䅵獴楮V䝉G⁄a瑡Ⱐu獥T⁡猠牥晥r敮捥⁤慴愠 o爠rU攠獴畤y
……………..



呡扬攠5:

L慮T獡琠b慮T猠獵mm慲礠楮fo牭
慴楯n
………………………………………………….

33I34

呡扬攠6:

䍨慮g攠T整e捴楯n慰⁣慬au污瑩lnV
……………………………………………………..



呡扬攠7:

Cramer’s V results for the two selected explanatory variables
……………



呡扬攠8:

䅭AunWf⁎䑖䤠捨慮ge⁡ 牯獳V瑨攠W瑵Wy⁡牥愠景爠rUe 瑨牥攠捨cnge

数e捨c
…………………………………………………………………………………………………



呡扬攠9:

䅣捵牡捹ra獳敳Vm敮琠eo爠r散牥慳攠楮⁎䑖


amoun琠o映獰f捫汥
數p敲楥湣敤
………………………………………………………………………………………..



呡扬攠10:

啲U慮⁧牯w瑨⁥Wp敲楥湣敤 楮⁁ 獴楮
…………………………………………………..



呡扬攠11:

M敡渠
慮T m慸amum⁤楳瑡 捥co映畲f慮⁧牯睴U⁦ om⁴U攠䍂M
…………….



呡扬攠12:

M敡渠摩獴慮捥co映畲f慮⁧ o睴U⁡
睡y⁦ om⁡ p物m慲a o爠r散enT慲y
r
o慤
……………………………………………………………………………………………………..



呡扬攠13:

Pearson’s correlation and student’s t
-
瑥獴⁲ 獵V瑳
-

T楳i慮捥cof

u牢慮
g牯睴U⁦ om⁲o慤V
……………………………………………………………………………….



呡扬攠14:

却STy⁌im楴慴ion猠獵mm慲a
………………………………………………………………….

82I83





8

Declaration










9

1.

Introduction

During the last century there has been a huge increase in the amount of people living in
urban areas and an associated growth in the size of these. This process, termed
urbanisation, has greatly transformed landscapes
throughout

the world and continues to
be one of the most significant forms of land cover change in cities across the world. As
urban areas continue to grow and expand beyond their original boundaries the process
becomes k
nown as urban sprawl. A

general consensus regarding the definition and im
pact
of urban sprawl has not been achieved as the nature of this can vary in different parts of
the world (Huang et al, 2007)
. However,
it is acknowledged that it is generally characterised
by
an increase in developed land,

often

at a cost of natural land,

which
can cause a number
of environmental issues, such as an increased risk of flooding (Kaźmierczak and Cavan,
2011) and elevated air and noise pollution (Gairola and Noresla, 2010).

However,

a

loss of
open space is

but one of the

many

issues surrounding

urban sprawl. Urban sprawl has also
been criticised for being a financial and social drain on a city (Geller, 2003). Outlying
suburbs often require

more costly infrastructure

and
sprawl can create economic
disparities and social fragmentation across a cit
y (Brueckner
,

2000). The traditional city
centres

also pay a price whe
n residents leave for the outlying areas with development
outside of the CBD leading to the deterioration of m
any inner city areas (Brueckner,

2000).

Given the issues it has been known
to cause, it is critically important to monitor urban
sprawl. T
here are various methods in capturing change, such as the recording of
building
/planning

permits

and simple

visual inspections, though it can be difficult to
accurately capture and display temp
oral changes using these methods. Remote sensing has
been increasingly used to detect and
examine

urban land cover changes over time. Whilst it
is acknowledged that the complexity in urban environments present many challenges in
accurately
detecting land c
over change, remote sensing is being effectively used to monitor
changes and proves an effective tool in policy making for urban planning. The Landsat
series, which began in 1972, was designed to operate with the objective of tracking
10

changes in land cover

(Williams et al, 2006).

This is well suited to provide a synoptic view of
urban development and has been widely used to measure urban form and changes from
natural to impervious surfaces. Using multi
-
date satellite imagery such as the Landsat
series, it i
s possible to observe land
-
cover change over a period of time. Rather than just
providing a gross change over a long period,
satellite time
-
series can record the variability
of urban

development i
n space and time, thus allowing

a
comparison with

different
factors,
such as

economic
and demographic data

(Masek et al
,

2000)
.

Change detection, which
involves the use of multi
-
temporal dat
a to discriminate areas of land
cover change between
dates of imaging (Lillesand and Kiefer, 2000) can be performed to measure these changes
and act as
an efficient means of obtaining information on

temporal trends and

the

spatial
distribution of urb
an areas which are required
for unders
tanding, model
ling and projecting
land

cover
change (
Elvidge et al
, 2004).
M
odelling

of potential future land cover change can
help to provide an

insight into the land use change
dynamics

and can
allow us to
quantitatively predict where future change might

occur based on previous changes. Urban
planners require reliable information to assess the

possible

consequences of urban sprawl
and manage the spatial trajectory of an urban area over time (Griffiths et al, 2010) and
models can prove an effective tool in

planning for future growth scenarios.



11

2.

Aims and Objectives

The primary aim of this paper is to utilise the Landsat archive to investigate urban growth
in Austin, Texas over a 22 year period and

use a land change model to

provide an estimate
for future ur
ban growth. This will involve:



Developing a methodology to map and monitor urban growth

over time

through
change detection
,

using the NDVI differencing method
.



Assess the accuracy of the classified maps and change detection technique in
capturing
urban gro
wth



Measure the location, amount and direction of urban growth



Relate any

changes to the
population

record
to
assess the extent to which

land
cover changes observed by satellite reflect the underlying
population trends



Use the classified change detection
maps as an input into the Idrisi Land Change
Modeler multi
-
layer perceptron neural network in order to predict future land
cover change for the year 2015

2.1

Research Hypothesis

The research hypotheses are that the NDVI differencing change detection techni
que will
prove an effective method in measuring urban growth in Austin and it will provide reliable
information regarding the amount and location of change. The second main hypothesis is
that using the classified maps as an input into the Idrisi Land Chang
e Modeler will yield a
realistic and useful prediction of future urban growth in the city.




12

3. Literature Review

This chapter will start with an in depth look into urban sprawl. A review of previous
research into urban sprawl will be carried out with an

emphasis on the use of remote
sensing methods to effectively capture this. Methods of change detection will then be
assessed with a focus on the use of the Normalised Difference Vegetation Index in
assessing changes in land cover. The final part of this c
hapter will involve reviewing the use
of Land Use/Cover Change models with a focus on the Idrisi Land Change Modeler which
has been utilised in this study.

3
.1 Urban Sprawl


While the global population has grown dramatically during the last century, we

hav
e also
witnessed a ‘population implosion’: the

unprecedented concentration of people

into urban
areas around the globe

(Masek et al, 2000)
.

Although this trend started within cities in
Europe and North America it is now very much a global phenomenon, with
cities in
developing nation
s growing at rapid rates (Kraas
,

2007). Although a consensus on a
definition for urban sprawl has not been reached it is generally characterised by scattered
and uncontrolled development on the urban periphery. Urban growth was
long considered
a sign of regional economic vitality
(Yuan et al, 2005)
;

however it ha
s now become a major
concern in many areas around the world (Xian and Crane, 2005). Although cities must grow
spatially to accommodate an expanding population
,

the claim
is that often too much spatial
growth occurs. This excessive growth can cause a number of
issues

and urban sprawl has
become a particular problem throughout the United States (Xian and Crane, 2005). Downs
(1998) notes that

“suburban sprawl has been the dom
inant form of metropolitan area
growth in the US since the 1950’s”.


13

Figure 1
:

Urban sprawl in Melbourne, Australia. Source Wikimedia Commons


There are a number of causes of urban sprawl and it can be considered to be the outcome
of a complex array of
factors including technology, local geography, economics and politics
that interact in a number of different ways (Nechyba and Walsh, 2004).

Bruekner (2000)
argues that sprawl is stimulat
ed

by three main forces


a growing population, rising
incomes and fa
lling commuting costs.
Squires (2002) suggest
s that the development of
modern

private

transport has

undoubtedly

facilitated urban sprawl but other political,
economic and socia
l factors need to be considered, including t
he globalisation of the
economy and
subsequent decentralisation of jobs from traditional Central Business District
(CBD)

areas
. Rusk
(
1999
) notes that longer
-
term mortgages, coupled with subsidised
mortgage insurance has made home ownership open to more households than ever before

and

Squire
s

(2002)

notes that families

will

choose to live in communities that offer the
most attractive package of services that they can afford.

Schools, crime rates, access to
14

shops, transport links and the environment will all be considered by families deciding
upon
a location and these are frequently perceived to be better in a suburban environment.
Nechyba and Walsh (2004) believe there is

a
pull

of

people out of central cities because of

attract
ive features of suburbs and also a

push

of

people out of central c
ities
because of
inner city problems.

Geller (2003) notes that for many, the great American
Dream is
associated with large,

si
ngle
-
family homes, lots of land
and a feeling of independence.

Urban sprawl does provide this low density lifestyle and also a sep
aration from some of the
social problems that may be experienced in the inner city. For many, living in suburbanised
areas will have marked a vastly improved quality of life. The increased popularity of larger
houses which are more distant from the urban c
entre has contributed to sprawling cities
across the US and indeed also the wider world.


The perceived benefits of urban sprawl
are increasingly balanced

against ecosystem
impacts, i
ncluding loss of green spaces, degradation of air and
water quality and
a

number
of socioeconomic effects including

economic dispar
ities, social fragmentation and impacts
upon health (Squires, 2002). One of the most concrete costs associated is the loss of green
spaces.
Within and around urban areas, green spaces provide a huge

number of benefits
and play a crucial role in enhancing the

livability


of cities. Some of the many advantages of
urban green spaces include; air and water purification and mitigation of the impact of
environmental pollution, regulation of microclimate,

a

habitat for urban wildlife,
recreational, spiritual and therapeutic value as well as social integration (Gairola and
Noresla, 2010). Green space improves the environmental quality of life, promotion of
public health and can help to foster local sustainab
le development (Schopfer et al, 2005).
Therefore the preservation and maintenance of green spaces is well acknowledged and has
become an important policy issue in many cities (Rafie
e et al, 2009). The

loss and
fragmentation of green spaces

due to outward u
rban growth

greatly

diminishes these
benefits noted above
. Urban sprawl has also contributed to creating a car dependant
15

society and Frumkin (2002) notes that a number of negative effects of sprawl relate directly
to a heavy reliance on cars


including in
creased air pollution and car crashes causing
injuries and fatalities as well as a less active lifestyle. Cars are a leading source of air
pollution (
Kennedy and Bates
,

1989
). Nechyba (2004) states that

“In the United States in
2001, on
-
road v
ehicles accou
nted for 37%

of total nitrogen oxides, which play a major role
in the formation of ground
-
level ozone, particulate matter, haze and
acid rain; 27%
of
volatile organic compounds, which react with nitrogen oxides to form g
round
-
level ozone;
and 62%
of total
emissions of carbon monoxide, which is a particular threat for individuals
who suffer from cardiovascular disease”
.

The link is generally that urban sprawl is
associated with high levels of driving, driving contributes to air pollution and air pollution
c
auses increased morbidity and mortality amongst a population. Other negative health
impacts of sprawl include an increasingly sedentary lifestyle, with often less exercise being
taken due to a reliance on a car. Ewing et al (2008) conducted a study into ur
ban sprawl
and health and reports that

residents
of sprawling counties were likely to walk less during
leisure time, weigh more, and have greater prevalence of hypertension”
.
There is a body of
evidence assessing

the

health impacts of

urban

sprawl, in particular links with obesity
(Frumkin 2002, Zhao and Kaestner 2010 and Bernell et al 2003) and findings do point
towards a sprawling city potentially being a factor in increased obesity rates.

The effects of sprawl extend

far

beyond environmen
tal and health issues and it has been
acknowledged that sprawl can contribute to an increasingly uneven and socially
fragmented city.
Development in the urban periphery often requires huge investments in
roads, sewers, schools and other public infrastructu
re. As new communities are built, new
infrastructure needs to expand

to accommodate them (Benedict and McMahon, 2002)
.
This growth is believed to reduce the incentive for development of land closer to city
centres, contributing to the decay of downtown are
as. As a population moves towards the
suburbs industry seeks out these outlying communities and the relocation of jobs (often
termed ‘job sprawl’) can exacerbate economic disparities within a city (Squires, 2002).
16

Although the process of job sprawl has bee
n debated, a study by Kneebone (2009)
highlighted that 95 out of 98 metropolitan areas in the US saw a decrease in the share of
jobs in the traditional city centre. Sprawl is also believed to lead to a decrease in
community engagement as suburban living is

often characterised by a longer commute and
segregated and homogenous neighbourhoods. Putnam (2000) estimates that each
additional 10 minutes of driving time predicts a 10% decline in civic involvement. Suburban
development patterns also tend to lead to e
conomic stratification


many developments
are built to specific price ranges which leads to a segregation and

whilst

this creates income
homogeneity wit
hin neighbourhoods, it intensifies income inequality across the

wider
metropolitan area (Frumkin 2002).

Collectively, these types of trends suggest that certain
features of sprawl can create an increasingly socially fragmented and uneven city.


As highlighted above, there are a number of problems associated with urban sprawl. There
has been a response to sp
rawl, which has included the ‘Smart Growth’ movement in the
United States.
Smart Growth strives to protect farmland and open space, revitali
s
e
neighbo
u
rhoods, and provide more transportation choic
es (Geller, 2003). Many places in
the U.S. have adopted smar
t growth initiatives, such as the state of Maryland, which in
1997 provided incentives for brownfield site development whilst denying similar subsidies
for projects outside of the target area.


In
Denver,

another city in which the

metropolitan

area has gro
wn
rapidly
, a light rail system has been built which connects the suburban area
from all directions to the downtown area, and a free shuttle bus runs the length of the
downtown spine (Geller, 2003). One of the most controversial measures in combatting
spra
wl is an Urban Growth Boundary
, which

is a zoning tool that slows

urban growth by
banning development in designated areas on the urban fringe (similar to the Greenbelt
adopted in the United Kingdom). In 1979, Portland, Oregon created an urban growth
bounda
ry. Whilst effective in limiting sprawl it has been argued that this created a number
of negative effects, including an increase in housing costs
. Geller (2003) notes that whilst
17

there is some debate over opinions on smart growth’s efficacy, it is becoming

part of the
urban landscape as the debate over the negative impacts created by urban sprawl
continue.

It is hugely important that urban growth is monitored to develop strategies for sustainable
development and

to

improve the livelihood of a city (Yang
,

20
0
2
).
The ability to monitor

urban land cover changes is highly desirable by local communities and

pol
icy decision
makers alike. The increased

availa
bility and improved quality of
multi
-
temporal remote
sensing

data offers opportunities to monitor urban land

cover

changes a
nd urban sprawl in
a timely and
cost
-
effective way.

3
.
2

Remote Sensing

Remote sensing is the process of obtaining information about an object or area through the
analysis of data acquired by a device that is not in contact with the o
b
je
ct or area under
investigation. Remotely collected data can take many forms, including electromagnetic
energy and sensors to detect this are operated from airborne and spaceborne platforms to
assist in inventorying, mapping and monitoring earth resources a
nd land cover
(Lillesand
and Kiefer, 2000)
.

Remote sensing has been increasingly used to monitor urban land cover change, with the
Landsat data series being widely utilised (Masek et al 2000, McMahon et al 2002, Ji et al
2006, Yuan et al, 2005). Remote se
nsing in urban areas present many challenges as these
areas are heterogeneous

and most urban image pixels at the resolution of Landsat

comprise a mix of different surfaces.
Mixed pixels are problematic for mapping using
conventional

classification methods
because most algorithms are

predicated on the
assumption of spectral homogeneity within a

particular land cover; therefore, the urban
mosaic can

result in high rates of misclassification between urban and other

land cover
classes

(Martinuzzi et al, 2007).
It is for this reason that a detailed differentiation and
classification of different land covers and uses is not easy in urban environments (Griffiths
18

et al, 2010). Improvements in spatial and spectral sensor resolutions in the last few years
have led to
further research in classifying urban areas and narrower wavelength ranges
allow the extraction of detailed signatures to better discriminate objects on the ground
(Yang
,

200
2
). Whilst not being of a high enough resolution for this, the benefit of Landsat
is
in the now 40 year archive of data it provides. It is well suited for a synoptic view of urban
development and has been widely used to measure urban form and changes from natural
to impervious surfaces


which mark a loss of open spaces to urbanisation.

Landsat has
been used to study the extent of the loss of green spaces in urban areas; Raifee et al (2009)
provide a study into the changes and extent of green spaces in Mashad, Iran between 1987
and 2006 and McMahon et al (2002) assess the change in green

space over time in two
cities in Idaho, USA. There have been a large number of studies using a time
-
series of
Landsat data to measure the change in urban land
-
cover over time, which are presented in
the table below. Many of these are focused on growing US

cities.

Table 1: Urban sprawl studies utilising a time
-
series of Landsat data

Author

Date

Study Area

Focus of study

Yang

2002

Atlanta

Urban sprawl in Atlanta metropolitan region
from 1973
-

1999

Yuan et al

2005

Minnesota

Change in Urban Land
-
cover in
Twin Cities 7
county metropolitan area from 1986
-

2002

Ji et al

2005

Kansas

Trends and patterns in urban land cover
from 1972 to 2001 in

the

Kansas City
metropolitan area

Masek et al

2000

Washington DC

Urban land changes from 1973 to 1996 and
links with

socio
-
economic situation

Xian and
Crane

2003

Tampa

Increase in impervious land in Tampa
between 1991 and 2002 and prediction of
19

future growth using a cellular automata
model

Martinuzzi et
al

2007

Puerto Rico

Urban sprawl mapping in San Juan
metropolitan

area

Catalán et al

2008

Barcelona

Urban growth in Barcelona between 1993


2000 慮T 業p慣琠on 污lT
-
u獥猠on 瑨攠u牢慮
p敲楰U敲e

V
á
捬慶
í
欠 慮T
副g慮

2009

佬Omouc 剥杩on,
䍺散栠剥pub汩c

䍨慮g敳e楮 u牢慮 污lT cov敲eb整睥敮 1991
-

2001


周敲攠i猠con獩V敲敤e
to be a difference in US style sprawl and European ‘compact’ urban
form
,

however Landsat has proved effective in identifying changes in both (Huang et al,
2007).

When utilising a time
-
series of satellite imagery, change detection methods are employed
to a
ssess changes in the land cover between the different dates of imagery. There are a
variety of methods that can be employed to detect land cover changes from remote
sensing data
and it is acknowledged that there is no consensus as to a single method that i
s
uni
versally applicable (Yang
,

200
2
).

However the methods can be summarised into two
broad categories


those that detect changes and assign change (pre
-
classification) and
those that first assign classes and then detect change (post
-
classification). Thes
e will now
be discussed in more detail.

3
.
3

Change Detection

There are many applications for change detection analysis which may range from short
term phenomena such as

the

measurement of flood water to longer term phenomena such
as urban fringe developmen
t (Lillesand and Kiefer
,

2000). Using multi
-
date satellite imagery
to detect land cover changes dates back to the early 1970’s. It is important that imagery
20

data used for change detection is acquired by the same (or very similar) sensor and be
recorded
using the same spatial resolution, spectral bands and viewing geometry. It is also
important to consider images that have

near

anniversary dates in order to minimise
seasonal variability in vegetation and sun angle. A variety of change detection methods
h
ave been employed in studies which generally fall into the pre and post classification
methods noted above. The pre
-
classification techniques apply various algorithms

to detect
change between images. These

includ
e

image differencing which was employed by R
idd
and Liu (1998) in the Salt
L
ake Valley, image ratioing, used by Prakash and Gupta
(
1998
)

to
detect land use changes in the Jharia coalfield in India and
neural networks
,
used by Dai
and Khorram (1999) to detect changes in land cover in Wilmington, Nort
h Carolina from
Landsat imagery
.

Other pre classification methods include
Principal Components Analysis,

Change Vector Analysis and

vegetation index differencing and

all of these methods can be
applied to a single or multiple spectral band
s

and directly to

multiple dates of satellite
imagery to generate ‘‘change’’ vs. ‘‘no
-
change’’ maps (Yuan et al, 2005). While the pre
-
classification methods are able to detect areas of change between images, a draw
-
back of
these methods
is

that the type of change is not sp
ecified.

Some studies have proposed
change detection techniques for monitoring urban growth changes by using the
Normalised Difference Vegetation Index (NDVI). The NDVI is the normalised difference
between near
-
infrared and visible reflectance so can be di
rectly related to the amount of
biomass within a pixel (Masek et al, 2000). Urban growth in an area replaces open spaces
(higher NDVI) with impervious built up land (low NDVI), so sudden decreases in NDVI have
been shown to reflect urban development. Studi
es by Howarth and Boasson (1983) in
Hamilton, Ontario and Masek et al 2000 in Washington DC have found this to be true. It has
been noted that using the NDVI differencing method alone can confuse urban growth
where agricultural areas are included due to th
e effects of crop rotation over a period of
time. Griffiths (1998) has suggested filtering NDVI maps to remove agricultural noise.

21

Post classification methods detect land cover change by comparing independently
produced classifications of images from
different dates. Through the application of this
technique a detailed matrix of ‘from and to’ changes can be built up (Rafiee et al, 2009) and
the classification of each

date of imagery builds a his
torical series that can be more
easily
understood and also

used for

applications other than change
detection. In addition, post
-
classification comparison minimis
es the problems caused by varia
tion in sensors and
atmospheric
conditions, as well as vegetation phenolo
gy between different dates as the

data from diffe
rent dates are separately classified

(Zhou et al, 2008)
. An issue with post
c
lassification methods is that of

error propagation


the technique is dependent on the
accuracy of each classified map

and a poor classification can lead to uncertainty
propagate
d through the change analysis.

Despite this, post classification has been widely
utilised in change detection of urban environments in the US, with studies including Kansas
(Ji et al
,

2006), Minnesota (Yuan et al
,

2005) and Baltimore (Zhou et al
,

2008).

3
.
4

Land Use/Cover Change Models

Models are used as a way of providing a simplified representation of a complex system
(Chang, 2010). The use of models in the context of Land Use/Cover Change (LUCC) opens
up opportunities to understand the factors and intera
ctions behind land cover change and
can serve as useful tools in projecting future environmental and economic impacts of land
use change (Alig, 1986). Attempts have been made to develop spatially explicit models to
predict urban land cover change, includin
g the SLEUTH cellular automata model which has
been used in studies in Tampa and also Baltimore
-
Washington DC (Xian and Crane, 2005,
Goetz et al, 2003). In recent years, agent based models have been increasingly inves
tigated,
with these models focu
sing on
human actions in defining landscape transitions. Parker et al
(2003) describe the use of agent based models in studying LUCC, no
ting a number of
applications for

these in the urban environment. Ligtenberg, Bregt, and van Lammeren
(2001) look at modelling s
patial planning in the Netherlands and Torrens (2001)
inspects
residential location dynamics.


22

In terms of predicting urban sprawl specifically, various modelling approaches have been
suggested, from landscape metrics (Sudhira et al, 2004), cellular
automata (Xian and Crane,
2005) to neural networks. Artificial neural networks are powerful tools that use a machine
learning approach to quantify and model complex behaviour and patterns. These have
been applied in Grand Traverse Bay, Michigan and also Do
ngguan, China (Pijanowski et al,
2001), (Li

and Yeh,

2002). The common theme amongst the various different model types is
the choice of ‘predictor variables’


which are the elements which are interacting to drive
the change. Given the complexity of urban
sprawl a number of predictor variables have
been used, with varying degrees of success. Two common variables which are amongst
others used by Xian and Crane in the study in Tampa and Sudhira et al in Mangalore, India
include:



Current urban extent/previous
urban sprawl



Distance to roads

Depending on the complexity of the model used,

a number of other variables

can be
included. Past studies have included factors such as, terrain and slope, density of urban
centre, distance from lakes and annual population gro
wth rates.

For
this study, the NDVI differencing change detection method has been employed as
described in section 3.3. However, to produce a future prediction of land cover change
IDRISI’s Land Change Modeler (LCM)

was employed. I will now introduce this

and some of
the features in modelling changes in land cover using the Multi
-
Layer Perceptron Neural
Network.

3
.
5

IDRISI Land Change Modeler

IDRISI’s Land Change Modeler

(LCM)

provides a

unique

t
ool to analyse and predict land
cover change. It has its root
s in biodiversity and the problem of habitat loss and
conservation, although also has a
pplications in any type of land
cover transformation,
including that of urban growth. It allows a user to analyse patterns and trends in past
land
cover by change detect
ion methods and also provides a

modelling and prediction
23

environment to create future landscape scenarios. This is accomplished by an ability to
integrate user defined driver
s

of change to a model that will impact upon the change.

The
three major parts of
the model used in this study were tools to



Analyse

past

land

cover change



Model

the potential for land transitions



Predict

the course of change into the future

These option
s

are all separated into a set of tabs within the model which provide the user
with
a

set of operations that should

be followed in sequence (Eastman, 2009)
.

3.6 Multi
-
Layer Perceptron Neural Network


The Land Change Modeler uses

la
nd transition information and

variable
s

that might d
rive
or explain such change and

create
s

a

transition pot
ential

th
e likelihood that

land
cover

will
change

in the

future

(Clark Labs, 2009).
Each transition is model
l
ed with either Logistic
Regression or a

Multi
-
Layer Perceptron neural network
. The recommended model for use
in the LCM is the Multi
-
Layer Perceptr
on (MLP) neural network as this can calculate
multiple transitions at a time, whereas the logistic regression can only model a single
transition at a time. The MLP was selected for this study.

Neural networks are predictive models based loosely on the acti
on of biological neurons
(DTREG, 2012). The unit of an artificial neural network is a neuron which receives inputs
and calculates outputs on the basis of a simple function. A simplified view of an artificial
neural network is as below:


24


Figure 2
:

Neural Network diagram. Source neuroph.sourceforge.net


Information is stored in the weights of the connections between neurons. These weights
among

neurons change according to a process of learning or adaptation, using a process or
trial and err
or to observe relationships in data. Rosenblatt (1958) is credited with
developing one of the first artificial neural networks when he created his ‘‘perceptron’’,
which consisted of a single node, receiving weighted inputs and categorising the results
acco
rding to a defined rule.

(Pijanowski et al, 2001)

The multi
-
layer perceptron (MLP) neural network is one of the most widely used Neural
Networks. The MLP consists of three layers: input, hidden, and output and is able to
solve
problems which are not linea
rly separable
. The neural network algorithms calculate
weights for the values

and nodes
and

introduce

the input in a feed forward manner, which
propagates through the hidden layer and the output layer
(Pijanowski et al, 2001)
. The
signals disseminate acros
s the nodes and are modified by the weights between each
connection. The receiving node sums the weighted inputs from all of the nodes connected
to it from the previous layer. The output of this node is then computed as the function of
its input and the da
ta moves forward from node to node with multiple weighted
summations occurring before reaching the output layer.
This type of network is trained
with the back
-
propagation learning algorithm. This algorithm randomly selects the initial
Hidden Layer

25

weights, then compare
s the calculated output for a given observation

with the expected
output for that observation.


3.7 Predicting Change


Markov Chain


After the transition potentials have been created the Land Change Modeler provides tools
for a dynamic land

cover change prediction.
After specifying an end date, the quantity of
change in each transition can either be modelled through a Markov Chain analysis.

A
M
arkov chain

is a process w
ith a fin
ite number of states in which

the probability of being in
a parti
cular state at step
n + 1

depends only on the state occupied at step
n

and not on the
sequence of events that preceded it.
The
MARKOV
module is used in the LCM to calculate
the transition probability matrix

from two input maps and use
s

this to create a pro
jection
of the transition potentials into the future.





26

4
. Study Area

The study comprises the urbanised area of Austin, Texas, as delimited by the US Census
Bureau in 2010. Austin is located in central Texas, which is in
the
American South West.


Figure 3: Location of Austin in the USA and Texas

Austin is the state capital of Te
xas and the urban area delineated

by the US Census bureau
has an area of 1365 km
2
. It is predominantly located within Travis county, though also
covers parts of Hays, Willia
mson and a very small part of Caldwell county. Austin is the
thirteenth most populous city in the USA, with a population of 820,611 as at 2011 (US
Census Bureau, 2011). The Austin urban area is shown below, along with the primary and
secondary roads in the

region.

27


Figure
4
: Austin, Texas. The primary
/secondary

roads downloaded from the US Census bureau are
present


Austin presents a very interesting area for studying urban growth. The Sierra Club, an
influential American environmental organisation, ranked

Austin highly amongst sprawl
threatened cities in the US in a 1998 report (Sierra Club 1998) and noted that from 1990
-
1998 on average 1000 people per month had moved to the Austin area. The city has
continued to develop rapidly thanks to a bouyant local e
conomy and between 2000 and
2006 was the third fastest growing large city in the US (
CNN Money
, 2007). The city is
28

considered a major centre for high technology and a number of global companies have
operations within the city, including Apple, Oracle and G
oogle, amongst many others.
Urban sprawl has been acknowledged to be an issue and the amount of land occupied by
Austin’s urban area has continued to rise. The map below, produced by the city of Austin
GIS team, displays Austin’s urbanised area from 1970


2004 and shows a steady
expansion.


Figure 5: Austin’s urban area expansion from 1970


2004. Source: City of Austin GIS Team

29

I
n 1998 Smart Growth planning was initiated by Mayor Kirk Watson in order to combat
sprawl and make Austin a more environmentall
y conscious city, along the lines of

a

cit
y

like
Portland, Oregon (Summerville, 2011). Tax incentives and expedited building approvals
have been made to encourage developments in the Central Business District

of Austin and
whilst this has created a large

amount of residential development within the
urban core of
the city
, the thousands of new units being created there

only represent a tiny percentage

of

the

total residential units created

on the urban periphery (City of Austin, 2012)
. There
simply are ver
y few land availability constraints in t
he territory surrounding the city, so

urban sprawl
remains a concern in

Austin
.



30

5.

Methodology

This chapter provides an overview of the methodology employed in order to answer the
aims and objectives set out. The chapter is broadly spilt into two parts


1) methods for
change detection analysis of the satellite
imagery and 2) methods for applying the Multi
-
layer Perceptron Neural Network model to the change detection results to firstly create a
map for 2010 which could be va
lidated against the actual land
cover change and secondly
produce a prediction for the fut
ure.



Figure 6: Generalised work flow for creation of change detection maps

DOWNLOAD

Landsat Data


US Census files

Reference Material


IMAGE PRE
-
PROCESSING

Geo
-
correction

Projection

Clip images to Area
of interest


Creation of False
colour composites


Creation of NDVI maps

1988

1995

2003

2010

Δ
NM噉
䕰N捨c
1

Δ
NM噉
䕰N捨c
2

Δ
NM噉
䕰N捨c
3

Minus
Agricultural
Variability

INCREASE/DECREASE
IN NDVI

3x3 Modal
Filter

&
Accuracy
Assessment

URBAN
GROWTH

Spatial
Analysis

31

5
.
1

Data Sources

5.1.1



Landsat Imagery

The materials for this study relied completely on what was available on the internet.
Satellite imagery was downloade
d from the United States Geological Survey’s “Earth
Explorer” website
http://earthexplorer.usgs.gov/

which provides over 16,000 Landsat
images that are freely available for download. Landsat 5 was launched in 1984 and this
(along with Landsat 4) marked an introduction of the Thematic Mapper (TM) sensor. This
provided a significant improvement over the pr
evious Landsat missions as the spatial
resolution was increased to 30 metres and three additional spectral bands were added, two
in shortwave infrared and one in thermal infrared. (Williams et al
,

2006). Landsat 5 has an
altitude of 705km which provides a
ground swath width of 185 kilometres. The temporal
resolution is 16 days. Remarkably, Landsat 5 has significantly exceeded its designed life
expectancy and as of March 2012 celebrated 28 years in space, 25 years beyond the
original 3 year mission that was
planned.

Due in part to its significant temporal archive, the Landsat
-
5 TM dataset was utilised for
this study and four Landsat scenes were chosen. This provided the core data that was used
in this study. Images from mi
d
-
July to mid
-
August were selected

a
s these can provide
better land cover detection in the fast vegetation growth season (Ji et al, 2006) and the
importance of choosing near anniversary dates was taken into account to avoid seasonal
differences in vegetation skewing the NDVI results. Images
were also selected that
displayed zero cloud cover over the study area. All data were received in the UTM
projection Zone 14N, using the World Geodetic System 1984 datum.






32

Table 2: Landsat Imagery used in the study

Imagery

Path

Row

Date

Source

Landsat
-
5 TM

27

39

25
th

July 1988

USGS Earth Explorer website

Landsat
-
5 TM

27

39

13
th

July 1995

USGS Earth Explorer website

Landsat
-
5 TM

27

39

4
th

August 2003

USGS Earth Explorer website

Landsat
-
5 TM

27

39

23
rd

August 2010

USGS Earth Explorer website


5.1.2


US Census Bureau shapefiles

A number of other pieces of data were also used in this study. The US Census TIGER files for
2010 were used to download county information for Texas, the US urban areas (clipped to
the Austin area), census tracts for Aus
tin and primary and secondary road data (US Census
Bureau, 2010). This data was all downloaded in shapefile format and
was in the Global
Coordinate System North American Datum of 1983 (GCS NAD83).

Table 3: US census bureau shapefiles

Data

Type

Source

Texas counties

Shapefile

US Census Bureau

US Urban Areas

Shapefile

US Census Bureau

Texas census tracts

Shapefile

US Census Bureau

Primary and secondary roads

Shapefile

US Census Bureau


5.1.
3

Reference data

Various data was also downloaded from the City of Austin GIS datasets website
ftp://ftp.ci.austin.tx.us/GIS
-
Data/Regional/coa_gis.html

in order to help with accuracy
assessment for th
e study. This excellent resource provided a number of pieces of data
crucial to the study, including aerial photography which aided in classification and accuracy
assessment as well as shapefiles for the City of Austin parks and land use information. The
p
arks layer was used in the accuracy assessment stage and the land use information played
33

a critical role in utilising the NDVI differencing method, as this gave information on the
parts of the study area that were classified as agricultural. Using this, it

was possible to
filter out the agricultural signal in the study area, which can otherwise cause significant
noise when using this method for classifying urban growth. All of the data downloaded
from the Austin GIS website was in the Texas Central State Pl
ane projection, using the
North American Datum of 1983.

Table 4:
City of Austin GIS Data, used as

reference data

for the study

Data

Type

Source

Year 2000 aerial photography

Infrared 24 cm resolution

City of Austin GIS Data Sets

Year 2009 aerial
photography

True colour 12cm
resolution

City of Austin GIS Data Sets

Austin parks (2005)

Shapefile

City of Austin GIS Data Sets

Austin land
-
use information
2003

Shapefile

City of Austin GIS Data Sets

Austin land
-
use information
2006

Shapefile

City of
Austin GIS Data Sets


5
.
2 Image Pre
-
processing

Geometric registration is crucial f
or producing spatially correct

change mapping through
time. The Landsat data downloaded from the USGS Earth Explorer site had already been
accurately rectified and geo
-
refer
enced to the UTM projection zone 14N, WGS84 datum.
The supporting vector data was also registered to this projection in ArcGIS10 using the
Project tool. Once all of the data was in the same projection it was loaded into Idrisi Taiga.
Although all seven ban
ds of the Landsat imagery were downloaded from the USGS Earth
Explorer website, just bands 1


4 of the Landsat data were used for this study. A summary
of these bands can be found in the table below:

34


Table 5: Landsat bands summary information

Band

Bandwi
dth
(μm)

Spatial resolution
(metres)

1 (visible blue)

0.45


0.52

30

2 (visible green)

0.52


0.60

30

3 (visible red)

0.63


0.69

30

4 (near infrared)

0.76


0.90

30


Once loaded into Idrisi Taiga the Landsat imagery was converted from a
G
eotiff to the Idrisi
raster format. An area of interest boundary was generated using the Austin urban area
shapefile, which had been clipped from the US urban areas national file in ArcGIS and then
loaded into Idrisi.


Figure
7
: Austin urban area, as de
lineated by the US Census bureau, 2010

35

The Landsat imagery was clipped to the area of interest boundary using the
WINDOW

module. A series of false colour composite images were produced for the four years. In
displaying a colour composite image, three prima
ry colours (red, green and blue) are used.
When the three colours are combined in various proportions, they produce different
colours in the visible spectrum. Associating each spectral band to a separate primary colour
results in a colour composite image (
CRISP 2001). Bands 4, 3 and 2 were assigned to the
red, green and blue colour guns respectively. In this image, vegetation appears in shades of
red whilst urban areas appear blue. These provide an excellent contrast between natural
and man
-
made areas and w
ere very useful

in distinguishing between land
cover types.

36


Figure
8
: Austin false colour composite for 1988

37


Figure 9
: Austin false colour composite for 1995




38


Figure
10
: Austin false colour composite for 2003


39


Figure 1
1
: Austin false colour
composite for 2010

5
.
3 Change Detection

As a map highlighting urban expansion had already been created by the City of Austin GIS
team (
see figure 5
)

I decided against performing a supervised classification of the data and
potentially duplicating this work. Instead,

t
o detect changes in land cover
,

the NDVI
differencing method was employed
. Bands 3 (visible red) and 4 (near infrared
-

NIR) of the
40

Lands
at data were used and

the NDVI for each image was calculated.

The NDVI exploits the
circumstance that vegetation has a much higher reflectance in the near
-
infrared region
compared to the visible red region and it is calculated by using a simple red and nea
r
-
infrared ratio:

NDVI


=









NDVI results are calculated from
-
1 (low/no vegetation) to 1 (high vegetation). Maps for all
of the years were created in Idrisi using this method, by running the
VEGINDEX

module,
specifying bands 3 (re
d) and 4 (infrared) and selecting the NDVI as the index type.



Figure 1
2
: VEGINDEX module in Idrisi


NDVI selected


41



Figure 1
3
: Austin NDVI map for 1988



Figure 1
4
: Austin NDVI map for 19
95




Figure 1
5
: Austin NDVI map for
2003




Figure 1
6: Austin NDVI map for 2010


N

Austin

NDVI

-

1988

NDVI index

N

Austin NDVI
-

1995

N

N

Austin NDVI
-

2003

Austin NDVI
-

2010

NDVI index

NDVI index

NDVI index

42

Agricultural variability has been shown to be
an issue in past studies using the NDVI
differencing method (Griffith 1998, Masek et al 2000), so the NDVI maps were filtered
through the Austin land
-
use

information

and agricultural areas were discounted from the
analysis. This was achieved in Idrisi by p
erforming an overlay and reclassing the appropriate
agricultural areas to zero. This was an important step in ensuring change detected did not
include highly variable

agricultural

areas which would have impacted significantly on the
results.
Figure 17

belo
w

shows the agricultural areas within the Austin urban area
boundary.


Figure 1
7
: Austin agricultural areas
, identified from the Austin land use data set

43


After the agricultural areas were removed
,

the change detection was performed.
NDVI
images were
overlaid and the earlier image NDVI values were

subtracted,

producing a map
of
changes in
NDVI

(ΔNDVI)

in which positive values represent ‘
greening’ (increased
vegetation) and negative values represent ‘browning’ (decreased vegetation).



Table 6
:

Change
detection map calculations

Source Maps

Calculation

Change Map

1988 and 1995

1995 NDVI


1988 NDVI = Change 1

Epoch 1

1995 and 2003

2003 NDVI


1995 NDVI = Change 2

Epoch 2

2003 and 2010

2010 NDVI


2003 NDVI = Change 3

Epoch 3



A threshold ΔNDVI value

needed to be decided upon to successfully capture true changes
in NDVI between the two images from noise. It was decided upon a change threshold of
two standard deviations from the mean to represent a true change in NDVI value. If the
threshold chosen is
too small, then any noise such as sensor fluctuation or slight
differences in ground reflectance will be viewed as land cover change. Alternatively, if the
threshold chosen is too large then true changes may not be identified. Visual checks were
undertaken

to ensure that two standard deviations did indeed capture true changes and
after being satisfied that this was a sensible threshold the NDVI differencing image values
were reclassified into three classes: vegetation decrease if a pixel value was lower tha
n the
low
-
end threshold, vegetation increase if the pixel was higher than the high
-
end threshold
and a no change class which contained all values between the thresholds. Full details
regarding this method are available in the appendix (see appendix A). As
a last step, the
change

map
s were

passed through a
3x3
m
odal

fi
lter to remove

any isolated pixels, which
as Masek et al 2000 note are mostly associated with registration errors. The change
detection maps for the three time periods are as presented below.

44



Figure 1
8
: Change detection map for epoch 1
, between 1988 and 1995


45


Figure 19
: Change detection map for epoch
2, between 1995 and 2003


46


Figure
20
: Change detection map for epoch
3, between 2003 and 2010

5
.
4 Validation and Accuracy Assessment

Map
validation exercises were undertaken to assess the accuracy of the NDVI differencing
method. This composed of two stages. Stage 1 involved e
stimating

the amount of
47

background s
peckle that were errors of commission. By using the false colour composites,
aer
ial photography and the Austin parks layers I was able to identify areas that have
remained as green spaces for the period from 1988
-

2010
. Within these regions
all
decrease in NDVI

pixels

present

could be shown to r
epresent errors of commission
, as
there

should be no change.

Figure
21

below illustrates this method of checking for speckle:


Figure 2
1
: Accuracy assessment for errors of commission


snapshot of epoch 2. The blue circles
highlight areas in which a decline in NDVI was experienced, but the
area remained as a green
space
, so these could be seen to be errors of commission.


48

The areas for parks were converted to a raster and reclassed to have a value of 1. The
urban growth for the epoch was also given a value of 1 and all other areas had a valu
e of 0.
The raster calculator was then utilised to multiply these together, with the areas that
returned a value of 1 showing an error of commission.
These areas were
summ
ed

to give
an estimate of the frequency of speckle within the growth map for each cha
nge epoch.


Figure 2
2
: Raster calculator to assess errors of commission
.

The second accuracy assessment exercise was with relation to areas experiencing an
increase in NDVI values. This was completed by visual assessment of the false colour
composites, ae
rial imagery and land use information, which contained a category of
‘undeveloped’ land. Similarly to the parks layer, this was overlaid to assess the extent of
increases in NDVI against this. As highlighted below, this was a factor and it seemed the
spect
ral signal for undeveloped land would change between the years which had a bearing
on the NDVI results obtained.

49


Figure 2
3
: Accuracy assessment for
increase in NDVI



snapshot of epoch 2. The blue circles
highlight areas in which a
n

increase in NDVI wa
s experienced in ‘undeveloped’ land.

The red
circles are places where NDVI increased but the land use is not undeveloped.


5
.
5 Spatial Analysis

5.5.1


Location and trends of
u
rban
s
prawl

After completing the validation exercises the general trends and d
ynamics of urban growth
were analysed
. A feature of urban sprawl is growth outwards, away from the traditional
50

city centre. Using a Euclidean distance calculation in ArcGIS 10 the distance from the Austin
CBD was calculated. The growth from the three epoch

maps were overlaid onto this
Euclidean distance surface and each pixel displaying growth had a value of 1, which was
multiplied by the distance value in the raster calculator. Using this method, it was possible
to identify the minimum, maximum and mean di
stance of urban growth for each epoch
from the CBD.


Figure 24
: Calculating the distance of urban growth away from the CBD

Another feature of urban sprawl is growth along highways. Using the same method as
above, distances of urban growth for each epoch
from roads was calculated.

As growth away from the city centre and nearby roads are two features highly associated
with urban sprawl these were plotted graphically in order to explore them further. A
correlation test was undertaken for distance to roads us
ing the Pearson’s product moment
correlation and a
S
tudent’s t
-
test performe
d to test the significance of any

relationships
found.

51

The census tracts for Austin were also analysed to identify general areas in which the
greatest urban change has taken place.

The full breakdown of all 304 census tracts is
available in the appendix. The 10 census tracts which experienced the most urban growth
(as a percentage of their total size) for each change epoch were analysed to identify any
spatial patterns, which are p
r
esented in the results section.

5.5.2


Link to the population

As a final stage in the change detection analysis the population of Austin over the period of
the study was considered in order to assess how the patterns observed in the change
detection map
s linked with the population in the city.


5
.
6 Land Cover Change Modelling

In the second part of this study, the NDVI differencing maps were used as an input into the
Idrisi Land Change Modeler to produce a predicted NDVI map for 2015. The first step
however, involved

using the first two change in NDVI maps to create

a prediction for the
third epoch, as this could then be validated against the actual land cover map for that time
to assess the model’s performance.

52

F
igure

25
: Generalised workflow for the future urban growth prediction map

As the emphasis for the study was on urban growth, just the areas of ‘NDVI dec
rease’ and
‘No Change (Austin)’ were used as inputs into the model to reduce the number of
ou
tcomes possible. Since there were

onl
y two land cover classes there we
re four possible
outcomes for each pixel



Austin with no change (Current land use persistence)



Decrease in NDVI to Austin (vegetation regrowth)



Austin to decrease in NDVI (urban growth)

Validation of map
with actual epoch
1
-
3 change map

Selection of explanatory variables

1)

Distance from past sprawl

2)

Distance from roads

53



Area of decreased NDVI with no change (urban growth persistence).

The change results for epoch 1 (1988
-
1995) and epoch 1
-

epoch 2 (1988
-
2003) were input
into the c
hange analysis tab to highlight the change in the NDVI between the two epochs.



Figure 26: Change in urban growth between epoch 1 and epoch 2

Between the two epochs a decrease in NDVI was witnessed over a 177km
2

area.

5
.
7 Transition Potentials

The
next step is the selection of transition sub
-
models, which allow a user to specify the
type of transition they are interested in (Clark Labs, 2009). The transition from Austin to a
decrease in NDVI was selected to be modelled in this study as the interest
is in urban
growth over time.

5.7.1


Explanatory variables

After selecting the transition sub
-
model it was required to select explanatory variables,
which are the factors driving the change. Urban sprawl has been acknowledged to be the
54

outcome of a number

of complex factors, though attempts have been made to model this
previously (Parker et al, 2003). Although a number of different factors have been input into
models in past research, variables have frequently included factors related to



Previous sprawl an
d



Distance to roads

As the variables above have been regularly used as inputs into other models and it was
possible to get data for them, these were selected as input
s

into the model. Importantly,
LCM differentiates between static variables, those that do not change over time and
dynamic variables (i.e.

proximity to distu
rbed areas). The choice of dynamic or static
impacts upon the transition achieved. For this study, d
istance to roads was selected as a
static variable as I was only considering

established

primary and secondary roads from the
US census which are less likely to vary over time. However, distance from previous sprawl
needed to be calculated as a dynamic var
iable as this variable will change over time. A
distance to roads map was created by converting the Austin roads vector layer from the US
Census Bureau to a raster in Idrisi using the
RASTERVECTOR

module. Once in a raster
format the Idrisi
DISTANCE

module
was run, which produced a map of distance from the
feature of interest.

55



Figure 27


Distance from roads calculated in Idrisi

In order to feed the previous sprawl in as an explanatory variable the decrease in NDVI in
epoch 1
was used as an input to create a Euclidean distance from disturbance layer.

Distance
from
roads
(km)

56


Figure 2
8



Distance from
previous sprawl

calculated in Idrisi

These maps could now be selected within LCM to test their respective explanatory power
in driving change. The test is based on contingency table analysis and the quantitative
measure of association used is Cramer’s V. A high Cramer’s V indicates that the p
otential
explanatory value of the variable is good, but this doesn’t guarantee a strong performance
Distance
from
previous
disturbance

(km)

57

(Clark Labs, 2009). Values of 0.4 or higher are regarded as good while anything over 0.15 is
useful. The explanatory variables used were selected and tested

with the results as below
:

Table 7
: Cramer’s V results for the two selected explanatory variables

Explanatory Variable

Cramer’s V

Distance from roads

0.1327

Distance from previous sprawl

0.4463


The distance from previous disturbance showed a high
Cramer’s V value, which suggests
that this variable is strongly associated with change. The value for distance from roads was
perhaps lower than would have been expected. However, given how roads have been
discussed as a factor in sprawl and have been used

in models in the past it was decided to
keep this variable in the model.

5.7.2


Multi
-
Layer Perceptron Neural Network

After the sub
-
model ha
d

been selected and explanatory variables input and tested, the
choice of model needed to be decided upon. The MLP

is the recommended option within
the Land Change Modeler program and this option was selected for use in this study.

The MLP launches in an automatic training mode.
When
the model is executed the MLP
creates a
random sample of cells that have experienced

a

transition fr
om Austin to a
decrease in NDVI as well as an
additional set of random samples for pixels

which persisted.
Thus the neural network is fed with

examples of the two cases, one
transition class and one
persistent class.
Using the samples the M
LP develops a multivariate function that can
predict transition potential based on the values at any location for the two explanatory
variables, by using half of the samples to train and the other half to test how well it is
doing. The model builds a neura
l network between the explanatory variables (distance
from sprawl and distance from roads) and the transition and persistence classes. This web
of connections between the neurons is applied as a set of weights which structure the
58

multivariate function
.
As
it analyses the pixels in the training data the model gauges error
and
adjust
s weights. As it continues to train it gets better and the accuracy and precision
can improve. These are measured by the accuracy rate and

Root Mean Square

(
RMS
)

value
in the mode
l.

The transition from Austin to decrease in NDVI

sub model was run in order to create the
tr
ansition potential map. The MLP
a
chieved an accuracy rate of 94.0
6 % and a RMS value of
0.2
3. The model was run using the default training parameters, as advised
by Clark Labs.



Figure 29: MLP Neural Network sub model

run

Once the model had successfully run and achieved the desired number of iterations to train
the data and
required
accuracy rate, the transition potential map was created. This map
marks the probability that a pixel will transition from Austin to a decrease in NDVI.

59


Figure 30: transition potential map


from Austin to a decrease in NDVI

It is evident from the map ab
ove that sprawl is expected across the study area. The impact
of distance from roads cannot be clearly seen and in the transition the distance from
previous sprawl has appeared to have come out as the stronger factor, as suggested by the
higher Cramer’s V
value.

The creation of a transition potential is the first stage in developing a prediction and has a
large bearing on the success of this. The next step involved feeding the transition potential
into the change prediction tab in LCM and using a Markov ch
ain to model the transition in
60

land
-
cover. The

MARKOV

module uses the earlier and later land
-
cover maps along with the
specified date for the prediction and identifies how much land would transition to a
decrease in NDVI based on the transition potentials
into the future. The Land Change
Modeler can then create a predicted future land cover map. An end date of 2010 was
selected so that the resulting map could be compared to the actual land cover map for that
period (epoch 3). A model of change is generated

and this will be discussed in the results,
along with the validation exercise that was undertaken to test how accurate this predicted
map was. Finally, a predicted map for 2015 was created.



61

6.

Results and Discussion

This chapter provides a description
of the analysis and results that were carried out for this
study. These are broken down into:



Change detection analysis map validation and accuracy assessment



General trends and dynamics of urban growth



Comparison of change to Austin population trend



Gener
ation of a predicted land cover map for 2010



Validation of predicted map for 2010



Prediction for 2015

6
.
1

Map validation and accuracy assessment

Figure
31

displays a composite map of the changes across the three epochs. This is an
overlay of
figures 1
8
,
19

and 2
0

and gives an indication of the change from 1988 to 2010.

62


Figure 31
: Composite change in NDVI, from 1988
-

2010


The study area size is 1365 square kilometres (527 square miles). The changes in NDVI were
calculated for each of the epochs
separately and are presented as below:


63

Table 8
: Amount of NDVI change across the study area for the three change epochs

Epoch

Increase in NDVI
(sq km)

No Change (sq km)

Decrease in NDVI
(sq km)

1988


1995

70.98

1246.38

47.64

1995


2003

10.44

1177.43

177.13

2003
-

2010

14.56

1223.36

127.09


Omitting the no change figures

for clarity

the results for change in the table above can be
seen graphically

below
:


Figure 32:
Change in NDVI graphs for the three

change epochs


The graph above shows a significant change between the first and second epochs, with a
marked decrease in NDVI experienced. The results for epoch 1 are somewhat surprising,
since it may be expected that an overall increase in NDVI would not

have been experienced
given that the city has been growing. The
m
ap validation exercises undertaken to assess
the accuracy of the NDVI differencing method composed of the two stages discussed in
70.98

10.44

14.56

47.64

177.13

127.09

0
20
40
60
80
100
120
140
160
180
200
1988


1995

1995


2003

2003 - 2010
Increase in NDVI (sq km)
Decrease in NDVI (sq km)
Change in NDVI values across the three epochs

64

section 5.4. The amount of
speckle

found in the Austin parks

and green areas was found to
be fairly low.

A
s a percentage of

the total image area these were as below:



Table 9:

Accuracy assessment for decrease in NDVI