!

DETECTION OF PERI-URBAN AND AGRICULTURAL EXPANSION 1990-2015

IN PAKSE, LAOS, USING DENSE TIME STACKS OF LANDSAT IMAGERY

by

Evan Applegate

A thesis submitted in partial fulfillment of the requirements

for the degree of

Master of Science
(Cartography and Geographic Information Systems)

at the

University of Wisconsin-Madison
2016
!

Abstract
Built-up areas in Southeast Asia have been expanding rapidly in the export-

oriented economies of Vietnam, China, and Thailand, but less attention has been paid to

cities in more rural nations such as Laos. The Laotian landscape has been shaped by both

internal migration to burgeoning urban areas and large-scale foreign agricultural

investment, but these trends have never been assessed together using remotely sensed

satellite data. This study presents results from efforts to characterize and quantify

expansion of built-up and agricultural land over a 35 km radius surrounding Pakse, the

third largest city in Laos, using a multi-date composite change detection approach applied

to dense time stacks of Landsat imagery. The method relies on a supervised, boosted

decision tree classification that exploits training data of stable/changed areas interpreted

from Google Earth and Landsat images spanning 1990 to 2015. The results show that the

decision tree approach provides a land cover change map with an overall accuracy of

92%, despite noisy and missing data, frequent cloud cover, and high land cover

heterogeneity across small areas. The results also show that, while built-up areas in

greater Pakse expanded 28% over 25 years, agricultural lands expanded 48% over the

same period. This work provides the most up-to-date land cover map of greater Pakse at

the highest spatial and temporal resolution available, and adds useful data to regional

environmental monitoring and sustainability efforts.

!

ii
Acknowledgments
!
This thesis work would not be possible without my diligent committee:

Annemarie Schneider, Robert Roth, and Tanya Buckingham. Professor Schneider

especially offered invaluable training, guidance, and encouragement during my entire

tenure at Madison and especially during the course of this work. None of the following

would have happened without her. I would also like to thank my distant correspondents

who provided data and information just because I sent a cold email: Juliet Lu, Matthew

Toro, Joost Foppes, and Jeremy Ferrand. Finally I would like to thank the UW Madison

Cartography Lab for their support, and feedback and friendship.

!

iii

Table of Contents
Abstract ........................................................................................................................................... i
Acknowledgments.......................................................................................................................... ii
Chapter 1. Introduction ................................................................................................................ 1
1.1: Land cover change in Southeast Asia: new urban and agricultural areas ................ 1
1.2: Research questions and goals ................................................................................... 5
1.3: Study area ................................................................................................................. 6
Chapter 2. Literature Review ...................................................................................................... 9
2.1 Urbanization and peri-urbanization trends in Southeast Asia ................................... 9
2.2: Urbanization and peri-urbanization trends in Laos ................................................ 11
2.3: Agricultural expansion trends in Southeast Asia ................................................... 14
2.4: Agricultural expansion trends in Laos ................................................................... 16
Chapter 3. Methods ..................................................................................................................... 20
3.1: Measuring land cover change using remotely sensed imagery .............................. 20
3.2: Using dense time stacks of Landsat imagery and a decision tree classifier to assess
greater Pakse ................................................................................................................. 25
3.3: Assessment of map accuracy ................................................................................. 31
3.4: Post-classification analyses .................................................................................... 32
Chapter 4. Results ....................................................................................................................... 33
4.1: Assessment of map accuracy ................................................................................. 33
4.2: Land cover change results ...................................................................................... 35
4.3: Spatial analyses to assess patterns of change ......................................................... 37
Chapter 5. Discussion .................................................................................................................. 38
5.1: Overview of classification results and spatial analysis .......................................... 38
5.2: Discussion of peri-urbanization results .................................................................. 39
5.3: Discussion of agricultural expansion results .......................................................... 40
5.4: Factors affecting map accuracy .............................................................................. 42
Chapter 6. Conclusion................................................................................................................. 45
6.1: Overview of research questions ............................................................................. 45
6.2: Significance of research ......................................................................................... 46
6.3: Future directions..................................................................................................... 48
Works Cited ................................................................................................................................. 50

Tables............................................................................................................................................ 61
Figures .......................................................................................................................................... 65
!

Chapter 1. Introduction

1.1: Land cover change in Southeast Asia: new urban and agricultural areas

Southeast Asia has witnessed massive land cover change over the past three

decades as a result of increased urbanization and an expansion of plantation agriculture

(Schneider 2003, Byerlee 2014). Driven by rapid economic growth beginning in the

1980s, East Asia’s urban land surface increased 22% from 2000-2010, while urban

populations grew 31% (Schneider et al. 2015). Most of this growth took place in the

region’s export manufacturing-oriented nations such as China and Vietnam, but even

relatively poor and rural Laos has experienced increased urbanization: urban land grew

37% between 2000-2010 (Schneider et al. 2015). Built-up areas in Laos’s two largest

cities, Vientiane and Savannaket, expanded at least 30% since 2005 (Sharifi et al. 2014,

Kimijiama et al. 2014). Today, 27% of Laotians live in urban areas and that figure is

expected to reach 38% by 2030 (Asian Development Bank 2011). While Laos’s rapid

urbanization -- defined here as both an increase in urban population and the expansion of

built-up areas -- is expected to continue, the country’s largest land cover changes have

resulted from agricultural concessions. Today, five percent of the nation’s land surface

area is leased to foreign businesses for agricultural, forestry, industrial, and mining

development (Global Witness 2013). These investments are usually sited in rural areas

but, as this study will show, they have greatly expanded in the peri-urban zone of a

regional capital (Schönweger et al. 2012).

!

2
While Southeast Asia’s enormous primary cities (e.g. Shanghai, Manila,

Bangkok) attract massive industrial and infrastructure investment, the region’s small and

mid-size cities attract the majority of rural migrants and, because there are more of these

cities, they affect a larger number of people and a greater proportion of land than their

megacity counterparts (Webster et al. 2014). Pakse, Laos’s third largest city and a

budding tourism and commercial hub near the Thai border, offers a useful case study for

urban land conversion in a heavily rural area that may be extensible to urban land cover

change studies in Cambodia, Vietnam, and Thailand (Asian Development Bank 2011).

It is likely that much of the projected urban growth will take the form of

converting formerly rural and agricultural land to residences, factories, or roads, a

process called “peri-urbanization” that has been extensively documented in the growing

metropolitan regions of Vietnam, Thailand, and China (Nguyen Thi et al. 2010, Gross et

al. 2014). While most urban growth in the region has been contiguous expansion from

urban cores, an increasing amount is taking place on the “peri-urban” fringe of cities

(Legates and Hudalah 2014). “Peri-urbanization” has many definitions across the fields

of urban planning and development studies, but broadly refers to the rapid, piecemeal,

and unplanned process by which rural areas on the outskirts of cities become more

“urban” socially, physically, and economically, often coincident with foreign direct

investment (Webster et al. 2002). For the purposes of this study, peri-urbanization is

defined as the unplanned and discontinuous growth of built-up areas up to 50 km from

the urban core, usually involving the conversion of agricultural or naturally vegetated

areas to roads, commercial structures, or residences (Simon 2008). Since 2000, Pakse’s
!

3
peri-urban zone has also hosted enormous forestry and agricultural projects, which

converted massive amounts of smallholder croplands and natural areas to monoculture

plantations (Global Witness 2013). Since peri-urban Pakse’s land cover change

trajectories cannot be adequately described by measuring urbanization or agricultural

expansion alone, both are explored in this work.

By 2030, 40% of the urban population added to Southeast Asia, or 200 million

people, will be in peri-urban areas (Webster 2002). Effective management can reduce the

harmful environmental and socio-economic effects of peri-urbanization, and satellite-

based assessments of East Asian peri-urbanization provide an invaluable tool to regional

officials and planners to support efficient growth and monitor progress (Gross et al. 2014,

Legates 2014). Recently, several remote sensing based studies of peri-urban areas have

been carried out across China, Thailand, Vietnam, Indonesia and the Philippines, but no

spatially and temporally detailed analyses have been conducted in Laos. This information

can inform local urban planning efforts, previously attempted in 1997 and 2011 (Asian

Development Bank 2011). Up-to-date land cover information is especially important in

Laos where government offices exercise significant control over land use and commercial

activity (Laungaramsri 2012).

In addition to peri-urbanization, large-scale agriculture is another major driver of

land cover change in Southeast Asia. Multinational companies contract with local farmers

in China, Vietnam, Thailand, Cambodia, and Laos, to produce cash crops, or they

establish and run large monoculture plantations on leased plots. Since the mid-1990s, oil

palm, rubber, cassava, eucalyptus, sugarcane, cashew, banana, and other forestry and
!

4
agricultural products have been grown in large plantations across Southeast Asia, often

established by multinational companies (Baird and Fox 2010, Global Witness 2013,

Byerlee 2014, Latsaphao 2016). While this agricultural investment provides revenues and

employment opportunities to host areas, it can have severe consequences for the local

environment and the livelihoods of subsistence farmers (Kenney-Lazar 2012).

Agricultural land cover change was also the primary driver of deforestation in the

Southeast Asia between 1990-2010, contributing to an 11.8% decrease in forest cover

that adversely affected biodiversity and carbon stocks (Stibig et al. 2014).

Since the mid 1990s official Laotian policies aimed at “turning land into capital”

have promoted foreign investment in mining, hydropower, light industry, and especially

agriculture (Manivong 2014). Provincial and national officials have leased vast tracts of

land to Vietnamese, Thai, and Chinese companies to plant coffee, tea, palm, sugarcane,

and rubber, with the latter representing the largest investments: 2,800 km2 of rubber trees

cover Laos, equivalent to 8% of the country’s arable land (Global Witness 2013). Fully

23% of leased land is located on “protected” forest areas nominally off-limits to

development, a contradiction to the Laotian government’s stated goal of restoring the

forest cover to 70% by 2020 (Lund 2010). Champasak province alone, of which Pakse is

the capital, contains more than 100 30-year agricultural land leases (Latsaphao 2016).

The true extent of the agricultural projects is unknown, as plantations regularly expand

beyond their permitted boundaries into neighboring forests and smallholder croplands

(Schönweger et al. 2012). Before the effect of Laotian plantation agriculture can be
!

5
accurately assessed, it is necessary to know its true extent, and accurate land cover maps

are a first step towards that goal.

1.2: Research questions and goals

Peri-urbanization and large-scale agricultural growth in Laos presents a unique set

of planning and environmental challenges, requiring accurate land cover change

assessments not only in the immediate surroundings of a city, but across the entire peri-

urban zone. Recently, several remote sensing-based studies of peri-urban areas have been

carried out across China, Thailand, and Vietnam, but no spatially- and temporally-

detailed analyses have been conducted in Southern Laos. Presented here is a land cover

change assessment of Laos’s third-largest city, Pakse, measuring the extent and spatial

pattern of new urban and agricultural land added in a 35 km radius from Pakse’s core

over three periods: 1990-2000, 2000-2006, and 2006-2015.

This work seeks to answer the following questions:

1) What was the amount of land converted to built-up and agricultural areas

between 1990-2015 in greater Pakse?

2) Is there a relationship between land conversion and distance from Pakse’s

center, or distance from roads?

This work is organized as follows. Chapter 2 reviews existing studies and their

conclusions. Chapter 3 describes the methods used to assess Pakse’s land cover changes

between 1990-2015. Chapter 4 presents the results and interpretation of the classification,

as well as an analysis of the study area’s spatial patterns of urban and agricultural growth.

Chapter 5 synthesizes the findings of this study and offers future directions for study.
!

6
This work makes a valuable contribution to Southeast Asian land cover change studies by

applying remotely sensed data and a semi-automated machine learning classifier to

characterize the extent and pattern of urban and agricultural land additions in Pakse,

Laos. The following methods and conclusions may be extensible to other nearby cities,

making Pakse a useful case study and template for land cover change science in the

region.

1.3: Study area

Pakse, Laos’s third largest city and capital of Champasak province in the

country’s southwest, spreads along the banks of the Mekong River near the Thai border.

Founded in 1908 as a French colonial outpost, its monsoon climate has a wet season

spanning May through October and a dry season from November to April. Upland forests

in the east reach elevations of 950m, hosting most of Laos’s coffee production (Toro

2012). Lowlands (elevation ~100m) to the north and west of Pakse are covered with rice

paddies and small croplands, fed by heavy rains that average 325mm a month during the

wet season. Natural vegetation includes dry evergreen forests on the hills, and mixed

deciduous forests, savannah, grasslands and the occasional wetland covering what has not

been converted to paddies (FIPD 2002).

The built-up area of the urban core is approximately 5 km in diameter and

surrounded by a patchwork of croplands composed of both smallholder plots and large

monoculture plantations. The city’s population grew from 20,000 residents in 1960 to

approximately 72,000 today, and is growing 5-6% yearly; in all, Pakse’s population is

expected to grow 50% by 2030 (Sisoulath et al. 2016, Asian Development Bank 2011).
!

7
Today, accelerating road development and small but growing local industries including

garment factories, construction, and agricultural processing give the region the second-

highest per-capita GDP in Laos, only trailing the capital city of Vientiane (Sisoulath et al.

2016). Pakse is the only urban district in Champasak province and one of the few Laotian

districts in general to experience net in-migration as populations from the southern

regions move to Pakse for work and education (Nolintha 2011).

The central government is courting investment in the city to transform Pakse into

a regional commercial hub, tourism destination, and a “bridge linking central and

northern regions and neighboring countries” (Ministry of Planning and Investment 2011,

Asian Development Bank 2011, Manivong 2014, Laine 2015). Investment in the area has

been growing: in 2015 eight Japanese and Lao-Japanese companies paid $5 million to

operate in a Pakse-Japan special economic zone, and another investment zone was

established in May 2016 to host a tourist resort and trade complex totaling $80 million in

investment (Vientiane Times 2015, 2016). The Laotian government’s National

Committee for Special and Specific Economic Zones has announced its intent to set up

three additional investment-friendly areas in Champasak by 2020 (National Committee

for Special and Specific Economic Zone Secretariat Office 2012).

This growth presents challenges to both the government’s planning efforts and the

financial and food security of its residents. As rice fields abutting the city are converted

to residential areas, factories, or cash crop plantations, residents have abandoned farming

for wage labor in Pakse or on plantation farms. As a result, agriculture’s share of the

regional economy has fallen from 40% in 2008 to 27.6% in 2014 (Asian Development
!

8
Bank 2011). While many welcome the new opportunities that urbanization brings, the

loss of agricultural land has affected tens of thousands of residents (Sharifi et al. 2014).

The city has an official urban plan to regulate development and prevent land degradation,

but it was last revised in 1997 and has not been followed due to informal development

and local officials granting building permits that contravene the plan (Asian Development

Bank 2011).

The area surrounding Pakse hosts dozens of land concessions, where private

companies lease government-owned land for business use (Figure 2) (Delang et al. 2013).

Most of these plots host monoculture agricultural and forestry plantations, where rubber,

eucalyptus, coffee, cassava, and rice are grown for export by Vietnamese, Thai, and

Chinese companies; of 31 concessions in the area, 26 are for agricultural products and 21

are foreign-owned (Government of Laos 2010). These concessions are an important part

of land cover change in the region, but their precise boundaries and extents are not public

information (Global Witness 2013).

This study assesses the pattern and extent of both peri-urbanization and

agricultural growth, as each affect the livelihoods of local residents; expanding built-up

areas on the fringes of Pakse have a “pull” effect on nearby rural dwellers, while more

and larger plantations displace villagers in a “push” towards the urban core (Kimijiama

and Nagai 2014). An accurate land cover map is necessary to assess these trends and

provide timely information towards the ends of inclusive, sustainable urban and

agricultural development.
!

9
Chapter 2. Literature Review

2.1 Urbanization and peri-urbanization trends in Southeast Asia

Peri-urbanization in the region is not a new phenomenon. 11th century Angkor

Wat experienced the same type of fragmented growth appearing in Southeast Asia today,

but it has accelerated since 1993 as the region’s major cities began to host large-scale

industry funded by foreign investment (Simon 2008). Indeed, growth in peri-urban zones

comprises a major share of the urban land added to East Asia over the past two decades

(Webster 2002, Legates and Hudalah 2014). In neighboring Vietnam, recent work has

shown that one-third of urban land added to Ho Chi Minh City from 1990-2012 appeared

more than 40 km from the urban core, and 57% of Hanoi’s urban land added over the

same period appeared 10-25 km from the city’s center (Kontgis et al. 2014, Nong et al.

2015). In addition, regional analyses of East Asia have revealed that 71% of new urban

surfaces 2000-2010 were added outside core municipal bounds (Schneider et al. 2015).

Peri-urban zones, containing both relatively rural and undeveloped land but high

concentrations of potential workers, are attractive “blank slates” for foreign and local

investors. The peri-urban regions of China’s Pearl River Delta received 70% of foreign

capital investment between 1980-1997, and 125 million jobs in peri-urban zones have

been added in China alone since 1978 (Hudalah 2007, Webster et al. 2014). The 1990s

brought foreign industrial and infrastructure investment that drove peri-urbanization in

Hanoi, Vietnam, and Bangkok, Thailand, converting rural croplands and forests to

factory complexes, railroads, expressways and residences (Keivani and Mattingly 2007).
!

10
Industrial peri-urbanization feeds economies but can hurt agricultural

productivity, local livelihoods, and the environment. Unconstrained industrial growth

around Bandung, Indonesia, caused farmland loss and lower yields, as well as increased

air and water pollution (Hudalah 2007). In 2008, the peri-urban Map Ta Phut Industrial

Estate on Thailand’s eastern seaboard, which hosts metalworks and chemical refineries,

damaged air quality and drew excessive water to the point that the government suspended

76 investment projects for one year (Webster et al. 2014). Industrial peri-urbanization in

China’s Hangzhou-Ningbo Corridor, which sprouted foreign-funded factories producing

everything from apparel to fiber optic cables, demolished productive agricultural land,

exacerbated acid rain and emitted large amounts of untreated wastewater into the

surrounding area (Webster 2002). Similar patterns of farm destruction and environmental

degradation began in the 1990s around Vietnam’s Ho Chi Minh City, Hanoi and the

greater Red River Delta (Kontgis et al. 2014, Nong et al. 2015). Laos, where agriculture

employs 71% of the workforce, may be especially vulnerable to the negative

environmental effects of peri-urban industrial growth (Simon 2008, World Bank 2010,

Nguyen Thi et al. 2010, Kontgis et al. 2014).

This type of growth has raised populations as migrants are attracted by the

opportunities and services that flow from new factories; on Thailand’s eastern seaboard,

which hosts dozens of factories and two high-capacity ports, the population has grown

four times faster in the peri-urban regions than in urban cores (Webster 2002). Peri-urban

growth can also follow informal residential growth on the boundaries between rural and

urban areas, as urban residents seek space or rural migrants seek domestic jobs, housing,
!

11
or education (Keivani and Mattingly 2007). Harm follows here as well; in Jakarta,

Indonesia, peri-urban residential development exacerbated economic segregation and

worsened flooding by encroaching onto a vital water catchment area (Firman 2004, Gross

et al. 2014, Hudalah et al. 2007).

Whatever the driver, East Asian peri-urbanization is a consequence of unplanned

development in areas of relatively weak land governance and high foreign investment,

causing landscape fragmentation that exacerbates transport problems, increases pollution,

and hurts agricultural productivity (Simon 2008). The negative effects of peri-

urbanization can be ameliorated by effective planning mechanisms implemented by a

centralized authority, but this is often impossible due to overlapping and/or weak regimes

of planning and control (Webster et al. 2014, Legates and Hudalah 2014). Peri-urban

informal growth, by definition outside urban core areas and thus also outside the

bailiwick of city-level authorities, does not fit into governance structures or jurisdictions.

This results in less attention from authorities, in turn causing uneven infrastructure

investment and inconsistent land use permitting (Hudalah et al. 2007). Indeed peri-

urbanization, by its very nature fragmented and discontinuous with existing services,

resists planning. Recent urban and regional growth plans in Laos and the Philippines

were unenforced by weak provincial and local authorities, and thus ignored by investors

and residents (Hudalah et al. 2007, Kritsanaphan and Sajor 2011, Gross et al. 2014).

2.2: Urbanization and peri-urbanization trends in Laos

The U.N. projects that 39% of Laos’s population lives in cities today, a figure that

may reach 51% by 2030, below the Southeast Asian average of 56% (U.N. Population
!

12
Division 2014). A regional-scale study found that urban land in the entire country

increased 37% from 2000 to 2010, giving Laos the largest annual growth rate in new

urban areas among East and Southeast Asian nations, though starting from a relatively

low base. Further study is needed, as that analysis used relatively coarse 250m resolution

imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS), and much

of the conversion to urban land takes place at scales far below 250m (Schneider 2015).

While Southeast Asian urban densities (defined here as the number of people per

unit area of built-up land) are increasing, 40% of new growth is projected to take place on

the peri-urban fringe of the region’s cities (Webster 2002, Schneider 2015). Peri-

urbanization’s extent and effects have been documented across cities in Vietnam,

Thailand, and China, but there are few urban land cover change studies for Laos and none

focusing on Pakse (Kritsanaphan and Sajor 2011, Schneider et al. 2012, Kontgis et al.

2014, Sharifi et al. 2014). Laos’s major urban areas have seen significant growth, but not

on the scale of its industrializing neighbors: urban land increased 34% between 1995-

2011 in the peri-urban zone of Laos’s capital and largest city, Vientiane, and in the core

areas of the second-largest city of Savannaket built-up areas jumped 165% from 2 km2 to

5.3 km2 between 1990-2013 (Kimijiama and Nagai 2014, Sharifi et al. 2014). Vientiane’s

peri-urban development followed four main patterns seen in other Southeast Asian cities:

discontinuous deforestation, development along roads, expansion from the center, and

annexation of surrounding villages into the core area (Sharifi et al. 2014).

Similar to its neighbors, a lack of enforcement and the fast pace of urban land

cover change limit the effectiveness of Laotian urban planning regimes (Legates and
!

13
Hudalah 2014, Webster et al. 2014). Two successive master plans for the capital city of

Vientiane, implemented in 2007 and 2010, were unsuccessful in either formalizing urban

land expansion or creating frameworks for future development; the 2007 plan was created

at too small a scale (1:60,000) for precise land use determinations, only included three

land use classifications, and was not followed by private developers, while the 2010 plan

was found to be “largely ignored” less than one year after approval (Lund 2010, Sharifi et

al. 2014). Pakse’s last master plan was implemented in 1997 and the results so far match

those of the capital: unplanned, informal conversion of agricultural, forested,

conservation, and drainage land has continued unabated (Asian Development Bank 2011,

2012).

Unlike urban areas in Thailand or Vietnam, Laos’s two largest cities of Vientiane

and Savannaket were not spurred into growth by export-oriented manufacturing; instead

mining, service, electrification and tourist projects spawned new built-up areas. Most of

this growth was largely fragmented, occurred irrespective of planning frameworks,

tended to follow roads and very often happened at the expense of cropland (Sharifi et al.

2014, Kimijiama et al. 2014). Though the drivers differ, peri-urbanization at the rural-

urban boundary has had similar effects in Laos as in neighboring countries: land

conversion of former smallholder agriculture at the edge of existing urban land has

accelerated deforestation, hurt agricultural productivity, and degraded natural resources.

This creates a self-reinforcing trend where former rural residents, for lack of better

options, move to cities for work after losing their subsistence farmland (Sharifi et al.

2014, Kimijiama and Nagai 2014).

!

14
2.3: Agricultural expansion trends in Southeast Asia

Between the late 1800s and mid 1960s, Southeast Asia’s colonial rulers

established large plantations to grow rubber, cassava, tea, oil palm, and other

commodities. As the colonial powers retreated, the plantations were nationalized and

most commodity production transitioned to smallholder systems. The late 20th century

brought a shift back to large-scale plantation: rising commodity prices and governments’

desire to exploit forest resources have driven a reduction in traditional agricultural

practices and an increase in plantation area across Southeast Asia, especially after

economic liberalization measures were carried out across the region in the 1990s. Today

Southeast Asia hosts 64% of the world’s oil palm plantations, 59% of all rubber

plantations, and 58% of all cassava plantations (Byerlee 2014). Southeast Asia is the only

region in the world where tree plantations make up a large portion of total agricultural

land, and these plantations have had a significant impact on the region’s forest cover,

biodiversity, and carbon stocks, as well as the welfare of small farmers (Gibbs et al.

2010).

Rubber, oil palm, tea, coffee, sugarcane, cassava, and other cash crop plantations

have been established at the expense of forest cover and protected areas: in Southeast

Asia as a whole, 56% of all land conversions to agricultural land between 1980-2000

took place in intact forests, and cash crop plantations have been assessed as the largest

driver of forest loss in the region (Gibbs et al. 2010, Stibig et al. 2014). In Malaysia,

rubber, cacao and oil palm plantations covered 80% of the nation’s cultivated land by
!

15
1990, threatening naturally forested and vegetated areas with adverse consequences for

the local environment (Härdter 1997).

Large-scale plantations tend to be more profitable than smaller ones; timber

plantations in Indonesia were found to be optimally profitable at 300-500 km2. Contract

farming agreements, where companies provide inputs and seeds to farmers in exchange

for products, are common in Thailand for eucalyptus and rubber but are less productive.

As a result, a large monoculture plantation run by a single entity is the preferred

arrangement for multinational companies establishing an agricultural project in Southeast

Asia (Hall 2003). These monoculture plantations are encouraged by the state as a more

“efficient” land use alternative to smallholders’ shifting/swidden agricultural practices,

i.e. when small fields are cleared with fire, cultivated for between one and three years,

and then left fallow for a longer period (Ichikawa 2007). Monoculture plantations bring

opportunities for local residents to clear land, care for crops and harvest and process

products in exchange for wages, but plantation expansion has removed many residents

from ancestral or traditional lands and has exacerbated poverty by cutting off access to

forest products (Souphonphacdy et al. 2012, Global Witness 2013).

The increase in cash crop cultivation constitutes one of the major land cover

change stories in Southeast Asia, with far-reaching consequences for environments and

people. By 2010, rubber alone covered at least 10,000 km2 across Laos, Thailand,

Vietnam, Cambodia, Myanmar, and South China (Li and Fox 2012). In Thailand between

1990-2008, approximately 1,100 km2 of forests were cleared for oil palm plantations, and

Indonesia’s palm oil plantings nearly doubled between 2000-2008 (Gibbs et al. 2010,
!

16
Thomas 2015). Today, Malaysia and Indonesia produce 85% of the world’s palm oil and

these plantations will likely expand; in 2015 the Indonesian president announced a plan

to add 15,000 km2 of new agricultural land by 2018 (Richards and Friess 2016).

2.4: Agricultural expansion trends in Laos

Due to the study area’s lack of updated land cover data and the opaque operations

of its industrial, mining, and agricultural land concessions, the precise extent and type of

land cover change in Laos is not known. However a review of existing literature shows

that land conversion to monoculture crops is the dominant land cover change trend in

both Laos in general and greater Pakse in particular (Ministry of Industry and Commerce

2009, Baird and Fox 2015).

Over the past twenty years large industrial estates have appeared on the periphery

of second-tier cities in Thailand, China, and Vietnam (Webster et al. 2014). Encouraged

by all levels of government, foreign manufacturers of computer parts, cars, and other

finished goods have converted a large amount of agricultural and forested land to urban

and built-up areas (Simon 2008). Laos’s relatively poor infrastructure and undeveloped

secondary sector preclude large industrial investments, but companies mainly based in

Thailand, China, and Vietnam have invested at least $1 billion since 2001 to produce

minerals, forestry, and food products in Laos (Ministry of Industry and Commerce 2009,

Vongkhamheng 2016). Rubber makes up the single largest product category produced

under concession, occupying 8% of all arable land and 26% of the 10,661 km2 leased to

foreign companies (Schönweger 2012, Vientiane Times 2013). Concessions vary in size

from 1.5 km2 to more than 150 km2, and the larger concessions have outsize effects: 9%
!

17
of leases cover 89% of total concession area. These massive investments have not been

transparently documented. There is no publicly available up-to-date listing of land

concessions or a map of their extents, and concession boundaries are often poorly defined

or ignored altogether in the course of land clearing and planting operations (Schönweger

et al. 2012).

Laotian officials view concessions as ways to attract investment, boost growth

and formal employment, and in the case of agricultural concessions, reduce the

“backwards” practice of shifting/swidden (Zurflueh 2013, Liao et al. 2015). Direct

revenues from concessions are a major incentive for local officials to grant them: a 190

km2 rubber project in Champasak was projected to yield $3.4 million in yearly taxes and

fees for the provincial government, a considerable amount in a province where an

agricultural laborer might earn $1,500 a year (Obein 2007, Portilla 2015).

French colonists first planted rubber in Laos in the early 1900s; eucalyptus and

teak projects followed in the late 1960s and 1970s. These plantations were relatively

small before the 1990s, but projects spanning dozens or hundreds of square kilometers

followed as the Laotian government began to allow investment in large-scale forest and

agricultural plantations (Phimmavong et al. 2009). Since Laos’s 1986 implementation of

the “New Economic Mechanism” to liberalize the economy and encourage trade,

provincial and central government bodies have granted decades-long land concessions to

private ventures (under Laotian law all land belongs to the state, so title is not transferred)

(Schönweger 2012, Souphonphacdy et al. 2012). In northern Laos, foreign agricultural

concessionaires often enter into contract arrangements with farmers, but in our study area
!

18
of Champasak province this arrangement is not common and land is typically cleared all

at once by the company (Portilla 2015).

During lease periods in southern Laos, investors clear forest, relocate villagers,

install monoculture plantations, and build processing centers to refine products for export

(Vongkhamheng et al. 2016). The provinces of Champasak and Salavan (comprising the

study area of 35 km surrounding Pakse) host many of these projects; indeed, 5% of land

(approximately 250 km2 of 4,899 km2) in the study area lies within one of approximately

30 concessions, and according to the most recent available data, 5% of Laos’s land

surface area is leased to businesses (Global Witness 2013, Messerli et al. 2015). Of the

study area’s documented land concessions as of 2010, 70% are foreign-owned plantations

raising vast tracts of rubber, coffee, sugarcane, cassava, eucalyptus, and livestock

(Ministry of Natural Resources and Environment 2011, Hirsch and Scurrah 2015). Recent

information on the number and size of concessions in the study area is not publicly

available, highlighting the importance of an up-to-date land cover assessment.

Large-scale agricultural projects have had a major impact on landscapes across

Laos, and it is necessary to understand their relationship to peri-urbanization around

Pakse. Concession-aires contribute to peri-urbanization directly by building roads,

factories, and other infrastructure, and indirectly by forcing the migration of residents

displaced by plantation agriculture (Baird 2010 and 2011, Dwyer 2007). While these

concessions have added a small amount of built-up land in the form of factories and

warehouses, the vast majority of land cover change in Laos has taken the form of
!

19
converting forested areas and village-scale agriculture to plantations. However, the true

extent of these plantations is not known (Zurflueh 2013).

The foreign projects in the study area have had negative effects on local residents,

as illustrated by the following example: in one concession located 15 km east of Pakse,

the Viet-Lao Rubber Joint Stock Company paid $65-$250 per planted hectare to

households whose land was cleared for the plantation, but paid no compensation for

fallow land, forest, or upland rice areas. Given that each household held 2-6 hectares of

land, this was a one-time compensation of $130-$1,500 (note that the average household

of eight consumes $900 of rice per year) (Portilla 2015). The plantations do offer jobs to

local villagers, a significant opportunity to land-poor residents, and after the plantation

was established most households sent at least one member to work there for $90-$125 per

month. However, older residents were not offered the chance to work and many

plantation employees complained of inconsistent pay and exposure to pesticides

(Zurflueh 2013, Portilla 2015).

It is uncertain if plantation expansion and its accompanying land conversion will

slow in coming years; in April 2016, the Champasak provincial government declared that

no more land would be granted to cultivate rice, coffee, rubber, cassava, or maize because

“there is no more land remaining” (Vientiane Times 2016). The existing concessions,

however, are leased for periods of at least 30 years, and are therefore likely to affect the

landscape and its residents for many years to come.

Smallholder coffee production in the Bolaven Plateau, 30 km east of Pakse, also

drives conversion to agricultural land and indirectly urbanizes the area. This is a much
!

20
more fragmented type of agricultural conversion than the large foreign-owned

monoculture plantations, but still occupies a significant amount of land area (Toro 2012).

The Bolaven Pleateau has some of the most ideal climatic, soil, and elevation conditions

for coffee production in mainland Southeast Asia and currently hosts 55% of the

country’s planted coffee area. Laotian coffee had been grown and exported since its

introduction in 1913, but a crash in the Brazilian coffee supply and the Laotian currency

devaluation that followed the Asian Financial Crisis of 1997 (subsequently making coffee

cheaper) made local coffee competitive in the global market. In 2015, Laos exported

20,700 metric tons of coffee, two and a half times more than in 1990. Coffee travels west

from the upland fields of the Bolaven Plateau to Pakse’s coffee processing centers before

continuing further west to Thailand for export by sea. This trade accelerates land cover

change as warehouses, roads, and processing centers are built to respond to demand,

which in turn attract workers who move into nearby residences (Toro 2012, International

Coffee Organization 2016). Coffee cultivation has been generally positive for the area,

giving Bolaven Pleateau residents the lowest poverty rate of all similarly-developed

regions in the country, though farmers often have to borrow capital at high interest rates

to maintain sufficient yields (Toro 2012, Thomas 2015).

Chapter 3. Methods

3.1: Measuring land cover change using remotely sensed imagery

Remote sensing data offers a cost-effective method for assessing land cover

change in general and peri-urbanization specifically. Satellite images assessed with

!

21
machine learning classifiers provide an expedient tool for evaluating urbanization and

land cover change patterns in Southeast Asia (Castrence et al. 2014, Nong et al. 2015).

Medium-resolution (30-250m spatial resolution) sensors including MODIS, Landsat, and

SPOT as well as high-resolution (1.5-5m) sensors such as RapidEye offer a relatively

inexpensive way to study remote areas with poor infrastructure (Dewan et al. 2010). The

United States’ Landsat series of sensors, which have provided free and continuous

medium-resolution global coverage since 1973 at a revisit rate of 16 days, have been

effective at characterizing change at Southeast Asia’s rural-urban interface. The area’s

small feature sizes, high variation in spectral response over agricultural lands, and

frequent cloud cover requires high-revisit sensors with sufficient resolution to

discriminate between agricultural, naturally vegetated, and urban land cover types

(Castrance et al. 2014). Landsat has been extensively used to study Southeast Asian land

cover change patterns (Dupuy 2012, Vongvisouk et al. 2016).

Satellite-derived land cover maps of Laos and Champasak province have been

produced both by the Laotian government and by researchers. The Laotian Forest

Inventory and Planning Division (FIPD) produced a nationwide vector-based land cover

map in 2002 with 24 classes and a minimum mapping unit of 40,000m2. However, the

map is coarse, outdated, and since it was created for forest management purposes, does

not focus on built-up areas or settlements (of which many are smaller than the minimum

mapping unit). A 2010 update was finished in 2012 (Ministry of Natural Resources and

Environment, 2012), but Thomas (2015) found that the map has too many quality issues

to be useable for land cover change analysis, and lacks up-to-date information on forest
!

22
extent and types. As part of a 2012 project assessing coffee production on the Bolaven

Pleateau (east of the study area), Toro classified the area using Landsat and Advanced

Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data and a maximum

likelihood classifier. He produced separate classifications for 1989, 2001, and 2008, but

accuracy assessments were not provided and the maps only cover the eastern half of

Pakse’s core (Toro 2012). These maps do not provide sufficient spatial or temporal detail

over the study area and lack detailed quality assurance metrics, highlighting the need for

a Pakse-specific land cover change assessment.

Detecting land cover change in the region presents unique challenges; conversion

to built-up land often happens at scales smaller than a sensor’s detection threshold, and

detection is further hampered by vegetation partially occluding structures that themselves

may be constructed using materials that are difficult to distinguish from surrounding

vegetation (Schneider et al. 2015). Peninsular Southeast Asia’s tropical monsoon climate

also significantly complicates remote sensing-based assessments: there are few cloud-free

images available, and clouds often completely obscure the landscape during the May-

October wet season (Sharifi et al. 2014, Kontgis 2014).

Dense time stacks, which combine several dozen or even hundreds of discrete

scenes over the same area but acquired on successive dates, can help overcome the

problem of endemic cloud cover by adding more data. Areas that are only seldom visible

through cloud cover can be assessed by compositing many images together; an area

consistently hidden by clouds will likely be obscured in any one image, but it is

extremely likely that the area will be visible in at least a few out of a hundred images.
!

23
The additional images, which provide multi-season views of stable and changed regions,

provide better results in complex landscapes than simpler change detection methods that

make use of just one cloud-free image per change period (Kontgis et al. 2014).

To overcome map quality issues in official maps, recent studies of land cover

change in Southeast Asia have used machine learning-based supervised classification

algorithms such as support vector machines, artificial neural networks, and decision trees

to determine land cover change trajectories. These classifiers make no a priori

assumptions about the distribution of the data (i.e. they are non-parametric), tolerate

noisy and/or incomplete data, and have been successfully applied to regional-scale

remote sensing problems in general and urban/peri-urban land cover change studies in

particular (Friedl et al. 2002, Schneider 2012, Castrence et al. 2014). Machine-learning

based classifiers such as decision trees have been shown to yield higher accuracies than

traditional statistical classification algorithms (e.g. maximum likelihood, minimum-

distance-to-means) and are now the method of choice for remote sensing-based land

change studies (Kontgis et al. 2014). They have been successfully deployed to study the

complex landscapes of Southeast Asia, where they have been used to assess 20 years of

land conversion across the peri-urban zones of Vietnam’s Ho Chi Minh City, the Red

River Delta, and the greater Mekong River Delta (Castrence et al. 2014, Kontgis et al.

2014, Nong et al. 2015, Schneider et al. 2015).

There are Laos-specific studies of land cover change using remotely sensed data,

but none have used decision trees to assess peri-urbanization and agricultural expansion,

none have focused on the greater Pakse area, and none reach the level of spatial and
!

24
temporal detail needed for the study area’s heterogeneous and fast-changing landscape.

The existing studies of Laotian peri-urbanization focused on the two largest cities, the

capital of Vientiane and the second-largest city of Savannaket, but neither used a machine

learning classification algorithm. They found that since the late 1970s, urban land cover

has expanded rapidly in these two cities, with already-developed Vientiane adding 34%

more urban land 1995-2011 within 10 km of the urban core and central Savannaket’s

urban land growing 31% from 2001-2011. Much of the growth was spread across roads

and isolated villages, not large industrial estates or residential areas as in Vietnam, China,

Indonesia, or the Philippines. These studies were hampered by heterogeneous landscapes

that include water, vegetation, and impervious surfaces commingled within Landsat’s

30x30 m pixels. Endemic wet season cloud cover also hurt classification accuracies

(Kimijiama and Nagai 2014, Sharifi et al. 2014).

Studies of agricultural land cover change have revealed the extent of cash crop

plantings across the region. A remote sensing-based assessment of areas in Thailand,

Cambodia, Vietnam and China using aerial photographs from the 1950s-1960s and an

unsupervised classification of mid-1990s Landsat imagery found new rubber, palm oil,

passion fruit, and Chinese cardamom plantations (Fox and Vogler 2005). 30m Landsat

data, 5m RapidEye imagery and an object-oriented classifier (which segments images

into spectrally-similar contiguous “objects” instead of classifying by each pixel) found

nearly 3,400 km2 of rubber plantations added between 1988-2010 in southwest China

alone (Chen et al. 2016). Landsat imagery and a machine learning random forest
!

25
classifier found large expansions of oil palm plantations between 2000-2012 across

Indonesia and Malaysia (Richards and Friess 2016).

As for detecting agricultural expansion in Laos specifically, an assessment of

Northern Laotian rubber plantations in the Luang Namtha province used 15m ASTER

imagery and an object-oriented classifier to find increased growth in plantations 2001-

2006, but difficulties in detecting immature plantations hindered accuracies (Hurni 2008).

A study of Laotian rubber plantation expansion in the area around Pakse found 137.3 km2

of potential rubber expansion between 2003-2012, but used only 10 Landsat 7 Enhanced

Thematic Mapper (ETM+) scenes and was hindered by a lack of data in the imagery; the

2003 failure of the ETM+ scan-line corrector (SLC) caused its imagery to be affected by

large striping artifacts, invalidating 22% of the data in each scene (Phompila et al. 2014).

These studies have found agricultural expansion beyond rubber plantations: in Laos’s

upland Bolaven Plateau (just east of the study region), a Landsat-based land cover change

assessment using a maximum likelihood classifier found 24.5% forest loss between 1989-

2008, of which most was attributable to new coffee fields (Toro 2012).

3.2: Using dense time stacks of Landsat imagery and a decision tree classifier to assess

greater Pakse

Building on past land cover change studies in Southeast Asia, this work relies on

a decision tree classifier to assess dense time stacks of Landsat imagery using a multi-

date composite change detection approach (Schneider 2012, Castrence et al. 2014,

Kontgis et al. 2014). Dense time stacks, in which dozens or even hundreds of successive

Landsat images are merged and assessed together, allow for long-term land cover change
!

26
measurement over wide regions (Huang et al. 2010). Dense time stacks ease

discrimination of land cover change by taking advantage of phenological (plant growth

cycle) patterns: agricultural, urban, and naturally vegetated areas have distinct cycles of

rising and falling reflectance in the near-infrared (NIR) and visible parts of the spectrum,

or, in the case of urban land, a conspicuous absence of this signal (Rogan and Chen

2004). Active croplands reflect a large amount of NIR energy to the sensor as the plants

grow and mature, then much less after they are harvested and the fields are cleared back

to bare soil. Healthy, well-watered natural vegetation also reflects NIR energy, but since

plants in the forest or grasslands are not harvested and replanted in a distinct cycle, they

are relatively easy to distinguish from agriculture by their temporal signature (Figure 3).

Surface reflectance data is the preferred input for remote sensing studies as it

partially corrects for atmospheric distortions that would introduce errors into downstream

analyses (Feng et al. 2012). Scenes spanning 1990-2015, each including three visible and

three infrared bands, were acquired from the USGS Earth Resources Observation and

Science Center Science Processing Architecture (ESPA). The data provided by ESPA

were processed using the Landsat Ecosystem Disturbance Adaptive Processing System

(LEDAPS), which corrected image distortions caused by the atmosphere (Masek et al.

2006). Although this radiometric correction was unnecessary as scenes were assessed

together and not individually, analysis of LEDAPS-corrected images provides less noisy

phenological signals that can boost change detection accuracies (Song et al. 2001,

Kontgis 2014). 102 scenes covering path/row 126/49 and 95 scenes covering path/row

126/50 were acquired; some 126/50 scenes had to be discarded due to data errors (Table
!

27
1). The scenes were assessed in their native projection of Universal Transverse Mercator

(UTM) Zone 48 based on the World Geodetic System 1984 datum (WGS84).

Any data collected after May 31, 2003, by the Landsat Enhanced Thematic

Mapper (ETM+) sensor is affected by striping artifacts due to the failure of the

instrument’s scan line corrector. The artifacts in affected scenes were masked and

excluded from the analysis. Though the scenes were generally un-obscured by clouds,

ESPA’s cloud mask was applied for each scene to set cloudy pixels to “no data,”

excluding them from analysis. The scenes were cropped to the study area of 35 km

around Pakse’s city center and combined into dense time stacks by footprint, yielding two

image stacks covering the two path/row footprints, 126/49 and 126/50 (Figure 1).

A classification scheme was chosen following the International Geosphere-

Biosphere Programme (IGBP) land cover criteria. Training samples were collected for

both “stable” land cover classes (i.e. areas that did not change over the study period) and

three change periods: 1990-2000, 2000-2006, 2006-2015. These periods were selected for

the availability of cloud-free imagery, to capture long-term land cover trajectories from

the beginning of Laos’s development spurt to the present day, and to align with the

change periods used in an ongoing NASA study of peri-urban change across mid-size

Southeast Asian cities. These uneven change periods were originally selected to coincide

with the latest available commune-level Vietnamese census data, and were retained in

this work for the sake of consistency and comparability (Asian Development Bank 2011,

Baird 2010, Kontgis et al. 2014, Hirsch and Scurrah 2015, Nong 2015).
!

28
Training pixels were selected in the ENVI software package by visual

interpretation of NIR composites for fall, spring, and summer Landsat scenes circa 1990,

2000, 2006, and 2015, as well as high resolution Google Earth imagery collected 2006-

2015. Sites were selected following the class criteria outlined in Table 2. Specifically,

“urban” pixels were selected based on the predominance of structures, roads, and other

impervious surfaces within the boundary of a single pixel; if more than half of a pixel

contained these features it was considered built-ip (Schneider 2009). 41 ground truth sites

collected in Laos during Toro 2012’s fieldwork were used to improve classification of the

vegetation-dominated change classes. Toro’s points included natural vegetation and

several different agricultural classes: cardamom, coffee, fruit trees, rubber, and rice.

Since these contained planting dates obtained from farmers, these were used to select

additional training sites for the agricultural change classes.

The classification was performed using the C5I implementation of Quinlan’s

decision tree algorithm (Figure 4) (Quinlan 1993, 1996). Support vector machines,

maximum likelihood, and artificial neural network classification algorithms were initially

considered for this project, but decision trees were likely to yield the best initial results in

Southeast Asia’s cloudy and heterogeneous landscape (Schneider 2012). A decision tree

iteratively splits a set of training data into increasingly homogenous groups using

decision rules (statistical tests), which are generated using a set of training data and

refined as the algorithm repeats its splits. After sufficiently homogenous groups are

generated using the training data (resulting in class labels at the terminal leaf nodes), the

decision rules are applied to the whole image to produce a classified map. “Boosting” is
!

29
used to enhance classification accuracy among nodes the algorithm had trouble with on

the first pass: on each successive run the classifier focuses on ambiguously-classified

nodes and adjusts its decision rules until final classes are maximally “pure” and separable

(Pal and Mather 20023, Schneider 2003). Like other machine learning classifiers

(artificial neural networks, support vector machines) decision trees are able to handle

noisy inputs, make no assumptions about the distribution of the data, and can handle

complex relationships between feature reflectance and final classes. They are also more

intuitive to use, transparent in the training data they make use of, and can arrive at the

same final classification from many different starting nodes (Schneider 2003).

Each iteration of the decision tree provides a probability estimate for each class at

each pixel, and a weighted vote is taken across all iterations to determine the final

classification. For this work the 1,493 training sites were separated by Landsat footprint

and given to the decision tree, which was “boosted" ten times (i.e. ten iterations of the

tree were generated) to produce a classification for each of the two dense time stacks. The

classification step was an iterative process: the fragmented landscape of southern Laos

landscape made it necessary to repeatedly assess the accuracy of the output, edit and add

training sites, then run the decision tree again before a satisfactory result was achieved (S.

Dupuy 2012, Liao et al. 2015).

The two resulting classifications were mosaicked together, and a series of post-

classification steps was performed to improve the accuracy of the final product. First, a

“sieve” spatial processing operation was used to remove “salt and pepper” artifacts

common to Landsat classifications (i.e. isolated, misclassified pixels). An eight-

!

30
connected sieve operation was performed in QGIS to remove areas below the minimum

mapping unit of 1,800m2 (2 pixels) (Dewan et al. 2010). Next, water areas were masked

to further refine the final output map. During the wet season the Mekong and Xe Dong

rivers inundate land that the classifier recognized as changed land as it switched from

vegetated to water and back again. Since measurement of this seasonal flooding was not

part of the study objectives, a water mask was manually digitized from the first cloud-

free image that followed the maximum water extent of the 2015 monsoon. This mask was

merged with the final output to label seasonally flooded areas as “stable water” regions

(Nong et al. 2015).

The third post-processing step included adding roads from non-imagery sources.

Since roads are an important cause and consequence of peri-urbanization in Southeast

Asia but their detection can be hampered by overhanging canopy and small feature sizes,

non-footpath dirt and paved roads were added from OpenStreetMap (OSM) and data

provided by the Laotian government (Hudalah et al. 2007, Ministry of Planning and

Construction 2010, Manivong 2014, Lainé 2015). The vector roads were projected to the

proper coordinate system, rasterized to one-pixel width (consistent with their appearance

on Landsat composites) and merged with the final output. These roads were added to the

stable urban class if they were visible on 1990 Landsat imagery, and added to one of the

three conversion-to-urban classes if they appeared on post-1990, post-2000, and post-

2006 imagery respectively. Roads that appeared in the vector data but not in Landsat or

Google imagery were discarded.

!

31
Finally, manual editing was performed to correct large, obvious errors and

increase the accuracy of the final classification. Since large portions of the final output

image were composed of monoculture plantations added after 2000, most of this editing

took place within these areas. As a starting point, agricultural concession point data

consisting of date, extent and type were acquired from the LaoDECIDE spatial

information project, a joint Lao-Swiss project to disseminate sociocultural and

geographic data in an online GIS (Messerli 2014). These data were used with Landsat

and Google Earth imagery to confirm the existence of concessions within the study area

and provide context for manual editing. Additional, more detailed concession boundaries

were digitized from a Global Witness report on rubber expansion, and land within those

boundaries was assigned to an agricultural change class coinciding with the date closest

to plantation establishment (Global Witness 2013).

3.3: Assessment of map accuracy

A stratified random accuracy assessment was performed by comparing the final

map to independently labeled “ground truth” sites. To prevent under-sampling of the

relatively sparse change classes, masks for each class were created and used separately as

boundaries for random truth site selection. Truth sites were randomly generated within

each class’s boundaries to reach 30-60 sites per class (Nong et al. 2015). Ground truth

sites were labeled based on visual interpretation of Landsat and very high resolution

Google Earth imagery. This process was repeated for the raw decision tree output until

classification accuracies reached 85%, and then after hand editing and road addition was

complete an assessment was performed on the final map using approximately 60 truth
!

32
sites per class. A confusion matrix was produced from these results to evaluate total map

accuracy and to document per-class omission and commission errors.

3.4: Post-classification analyses

Adapting techniques from existing studies of Southeast Asian land cover change,

a ring buffer and a road buffer analysis were performed to measure the distribution of

land conversion in the study area. To assess land cover change as a function of distance

from Pakse’s core (defined as the geographic center of the contiguous, higher-density

region visible on official maps and high-resolution satellite imagery), circular ring buffers

were generated at 5 km intervals from Pakse’s center and the share of developable

surface within each buffer that was converted to built-up or agricultural land was

calculated (Kontgis 2014). A separate road buffer study was conducted at several interval

distances from the OSM-derived roads to determine if there was a relationship between

road location and land cover change, similar to studies conducted in Vientiane, Laos, and

Taipei, Taiwan (Sharifi et al. 2014, Huang, Wang, Budd 2009). Buffers were created at

30m intervals up to 270m from the roads, and the share of developable surface within

each buffer that was converted to built-up or agricultural land was calculated. A summary

of the preceding steps can be seen in Figure 5.

!

33
Chapter 4. Results

4.1: Assessment of map accuracy

Figure 6 shows the classified map and Table 4 shows the full confusion matrix,

indicating which classes were confused with one another and the type of error. Note that

errors of omission for each class can be read in the columns, and the errors of omission in

the rows. Producer’s accuracy, which corresponds to the chance of an error of omission,

was 91.63%. The user’s accuracy, corresponding to the chance of an error of

commission, was about the same at 91.62%. The map’s overall accuracy was 91.63% and

the kappa coefficient was 0.90, indicating the classification was approximately 90%

better than one resulting from random chance. Stable classes had a minimum of 92%

accuracy, while 87% accuracy was the minimum for the change classes. Among change

classes the map most reliably classified land conversions to water and agriculture,

followed by conversions to built-up land.

Stable croplands were confused with other classes with similar spectral profiles,

e.g. dense, well-watered natural vegetation and lands converted to agriculture between

1990-2000. Stable water areas were easily distinguished due to water’s unique and

consistent low-reflectance signature. Stable urban and stable barren classes were

confused as they have a similar lack of phenological cycles and high overall reflectance.

Stable urban areas were confused with stable barren areas and early (1990-2000)

conversions to urban land, as these areas shared a high overall reflectance and were

especially mixed between vegetated and non-vegetated surfaces (e.g. bare soil, rock,

impervious surfaces, or structures). Stable agricultural areas, which often include bare-
!

34
soil fallow fields and flooded paddies, were confused with barren and early conversion-

to-water change classes with similar spectral signatures.

Trees, grasses, shrubs and crops closely commingled with structures and roads

caused relatively low accuracies for all urban change classes. Built-up areas in rural

Southeast Asia are difficult to distinguish from their surroundings; building materials

often match neighboring land cover types, and structures and roads are often too small to

distinguish using medium-resolution imagery such as Landsat (Castrence et al. 2014,

Schneider 2015). Classes with high NIR reflectance, i.e. natural vegetation, stable

agriculture, and conversions to agriculture, were confused with each other, likely because

of the similar spectral signatures of tropical vegetation and intensely cultivated rubber

plantations (Torbick et al. 2016). Conversion-to-water change class accuracies were

affected by the small size of most man-made ponds and co-location with woody

vegetation, which makes them harder to discriminate at 30m resolution.

There was mutual confusion between the 1990-2000 and 2000-2006 change

classes, but the 2006-2015 changes were more easily discriminated due to the availability

of high-resolution imagery and Toro 2012’s verification sites. The 1990-2000 and 2000-

2006 vegetation-to-urban change classes were consistently over-classified, i.e. they

included urban change pixels from other periods. These temporal errors were likely

caused by inconsistent training data, as built-up areas in the region have small feature

sizes and high-resolution imagery was not available before 2006. 1990-2000 conversions

to agriculture were consistently missed (i.e. under-classified) possibly due to land that

was cleared and not planted, but could be picked up as croplands converted in later
!

35
periods when the vegetation had grown enough to register as active agriculture to the

decision tree.

The well-documented challenges of finding built-up areas in Southeast Asia’s

heterogeneous and complex landscapes hurt classification accuracies for urban change

classes, but overall accuracy was good and the output accurately characterized the area’s

major land cover change story: new agricultural areas.

4.2: Land cover change results

The results show that naturally vegetated areas and stable agricultural areas

dominate the map, covering 65% and 18.5% of the area, respectively (Table 3, Figure 6).

Land conversion in the form of natural vegetation being replaced with built-up areas,

human-made ponds, or agricultural land covered 9.8% of the total study area, with the

majority of those changes being conversions to agricultural land (Figure 7). Agricultural

areas grew 435.3 km2 over 25 years, increasing the total cropland area by 48%. 27.2 km2

of built-up land was added between 1990-2015, bringing the study area’s urban land

cover to 124.7 km2 in 2015, a 28% increase (covering a total of 2.5% of the study area).

15.3 km2, or 0.3% of the total area, was converted from agriculture or vegetation to

artificial ponds or drainages. Change accelerated in the later periods, as average annual

additions of each type of land conversion increased successively from 1990-2000, 2000-

2006, and 2006-2015. For example, 65.9 km2, 132.8 km2, and 236.6 km2 of new

agricultural land was added during the three successive change periods, respectively

(Table 3, Figure 7a).

!

36
New agriculture comprised 90% of detected land cover change. Based on

information from local experts and visual interpretation of high resolution imagery, it is

likely that new agricultural areas (~50 km2) in the eastern highlands in the study area

were coffee cultivation that occurred during the 1990-2000 period. From 2000-2006,

agricultural change consisted of hundreds of square kilometers of rubber plantations

planted to the northeast, east, and southeast of Pakse; these areas effectively converted

naturally vegetated areas and small fields to monocultures (Figure 6). During 2006-2015

existing rubber plantations expanded into neighboring forest and village land, and expert

assessment of high-resolution satellite imagery shows that smallholder coconut palm,

cassava, sugarcane, and other plantings grew rapidly around villages in the northeast

corner of the study region (Foppes 2016).

Built-up areas increased 28% from 1990-2015, adding 27.2 km2 of urban land to

bring greater Pakse area’s 2015 total to 124.7 km2. This constitutes only a small portion

of the total study area, and is to be expected given the small size of the city and the area’s

low population. Pakse’s urbanization, while modest, appears to be accelerating: an annual

average of 0.9 km2 of urban land was added each year between 1990-2000, 1.1 km2 per

year from 2000-2006, and 1.3 km2 per year during 2006-2015.

Conversions to water were also small, as the results indicate that 15.3 km2 of

ponds were dug from natural vegetated areas and paddies between 1990-2015; these hold

irrigation water and can also be used to raise fish for sale or household consumption

(Nguyen Thi et al. 2010). These made up the smallest change class, covering 0.3% of the

study area.
!

37
4.3: Spatial analyses to assess patterns of change

The circular buffer analysis around Pakse’s center showed that 16% of all urban

areas as of 2015 were located within central Pakse, a region occupying only 2% of the

total study area. Urban land conversion steadily decreased as a function of distance from

Pakse, similar to the pattern seen in a previous study of Vientiane (Figure 8b, Sharifi et

al. 2014). The imagery and results suggest that a few industrial estates and agricultural

processing centers appeared after 2000 but most of the growth consisted of residences

added to tiny villages and roads cleared or expanded between them. Most agricultural

land was added between 15-25 km of Pakse’s center, likely due to rubber planting, as

most of the plantations added after 2000 fall within that range (Figure 8c). This may be

because cash crops generally require the transport and trade infrastructure of a city; a

plantation close enough to Pakse for convenient services but far enough to take advantage

of large contiguous tracts of land may be more efficiently sited than one closer or farther

from the city (Nahuelhual et al. 2012). The spike in agricultural land added 1990-2000 35

km from central Pakse coincides with a coffee plantation established after 2000 (Figure

8c).

Most built-up land was added within 60m of a road; urban land conversion across

all change periods tapered off as distance from a road increased, a common peri-urban

development pattern in Laos (Figure 8e, Kimijiama and Nagai 2014). Agricultural

conversion did not have a strong relationship to distance from a road: between 1990-2000

cropland conversion was unrelated to road distance, between 2000-2006 it was slightly

negatively correlated, and from 2006-2015 was slightly positively correlated (Figure 8f).
!

38
The degree of developable surface within each buffer converted to agriculture matched

the class shares over the entire study area: more was converted to agriculture between

2006-2015 than from 2000-2006 or 1990-2000.

Chapter 5. Discussion

5.1: Overview of classification results and spatial analysis

This work aims to provide a spatially and temporally-detailed assessment of the

land cover changes that occurred over the last three decades in the peri-urban region of

Pakse, Laos. The results show that built-up areas and agricultural lands grew 28% and

48%, respectively. The boosted decision tree approach using Landsat dense time stacks

detected 478 km2 of land conversion over 25 years, providing detailed class information

over four time points (1990, 2000, 2006, and 2015) with 91.6% overall accuracy. This

suggests the method was effective at capturing two important landscape changes in Laos:

development of built-up areas in peri-urban zones, and large-scale conversion of land to

agriculture, with the latter comprising 90% of the total land cover change in the study

area. The land cover change patterns explored in the previous section reinforce existing

studies: Laotian urban areas are small, rapidly growing, and tend to stick close to roads.

However, large agricultural and forest plantations take up much more land and are the

major land cover change story in Laos in general and the study area in particular (Global

Witness 2013, Hirsch and Scurrah 2015).

!

39
5.2: Discussion of peri-urbanization results

Pakse’s peri-urban growth of 28% over the 1990-2015 period is comparable to the

one other Laotian city with peri-urbanization data: built-up areas within 40 km of the

capital of Vientiane increased by 34% between 1995-2011 (Sharifi et al. 2014). Vientiane

and Pakse experienced similar in-fill urban growth in their core areas (e.g. structures

appearing in naturally vegetated areas between roads) while adding significant land along

roads, but Pakse did not undergo the patchy expansion and annexation of neighboring

villages that occurred in Vientiane (Sharifi et al. 2014). This is to be expected given that

Vientiane has more than five times as many residents as Pakse. Pakse’s core did not

greatly expand in the “spreading pancake” model of urban growth, but instead

settlements and commercial structures (agricultural processing centers, large buildings on

plantations) grew along roads; 38% of new built-up areas 1990-2015 appeared within

30m of a road. This is an established pattern in Champasak province, where an Asian

Development Bank study of road infrastructure projects showed that villagers tended to

build homes close to newly established roads, i.e. roads attracted built-up areas (Fujimura

and Ramesh 2010). In the aggregate, loss of natural vegetation was mostly due to

agricultural expansion, not urban growth as in Vientiane (Sharifi et al. 2014).

Within Laos, a review of primary and secondary sources reveals little of the

manufacturing investment that coincided with accelerating peri-urbanization in China’s

Pearl River Delta, Thailand’s eastern seaboard, and Vietnam’s Red River Delta. From

1989-2014, most of the $23.5 billion invested in Laos from abroad was in electricity

generation, mining and agriculture; only 8% was invested in industry (Ministry of

!

40
Planning and Investment Promotion Department 2015). While Laotian industrial

production has grown rapidly since the mid 1990s, recent data shows that industry

comprises only 31% of Laotian GDP, compared to 37% in Thailand and 43% in China

(World Bank 2014). Foreign investment may have driven land cover changes in the study

area, but it was investment in the primary sectors of agriculture, forestry, and

hydropower, not manufacturing capacity or residential development.

5.3: Discussion of agricultural expansion results

The results of this study confirm reports of rapidly increasing plantation

development by foreign companies, but at much finer scales than previous studies (Figure

9). Since 2000 Laotian, Vietnamese, Chinese, and Thai companies have leased large

tracts of land in greater Pakse; these concessions (totaling ~250 km2 in 2015) have been

planted primarily with rubber, cassava, coffee, eucalyptus, and rice (Figure 10, Kenney-

Lazar 2012, Delang et al. 2013, Vongvisouk et al. 2016). Between 2000-2015, 58.9% of

the agricultural land added to the study area appeared in large plantations first established

in 2004. The individual change classes reveal finer distinctions: 122.9 km2 of the

agricultural land added 2000-2006 (93% of the class total) lies inside land concessions

granted to Vietnam’s Dak-Lak Rubber Company and the Dau Tieng Viet-Lao Rubber

Joint Stock Company (Global Witness 2013).

The largest of the individual rubber concessions, belonging to the Viet-Lao

Rubber Joint Stock Company, covers more than 100 km2 of former forest, grassland, and

smallholder fields. These results highlight the intensity of plantation investment in the

region since 2000. After the first plantings, these plantations expanded further: 133.4 km2
!

41
were added to these plantations between 2006-2015, representing 56% of the total

agricultural land added within the study area during that period.

The 1990-2000 agricultural expansion results reveal a markedly different story:

75.7% of the period’s new cropland appeared in the Bolaven Plateau, the highland

eastern portion of the study area considered suitable for coffee. High-resolution Google

Earth imagery and Toro’s (2012) ground truth data point to extensive and mature shade-

grown coffee cultivation. Since Laotian coffee exports jumped 159% over the same

period and 58% of Laotian coffee is grown in this region, it is possible that coffee

planting increased rapidly in the eastern uplands of the study area from 1990-2000 (Lao

Coffee Association 2013, International Coffee Organization 2015). However, since

shade-grown coffee grows under the canopies of larger trees (effectively “hidden” from

satellite sensors) it is difficult to distinguish coffee plantings from natural forest cover

and thus plantation age cannot be confirmed using medium-resolution imagery alone

(Cordero-Sancho and Sader 2007). Further study is needed to isolate detailed crop

information (e.g. finer-grained coffee production figures, land use surveys) from each

change period to determine exactly when coffee plantings expanded in the region.

Unlike the post-2000 agricultural expansion that took the form of plantation

expansion, most of the agricultural expansion 1990-2000 occurred as new smallholder

plots 0.03-0.08 km2 in size. There are two large coffee plantations in the study area

totaling 4 km2 and several more plantations east of the study area’s boundary, but within

the study area coffee cultivation is much less consolidated than the plantations (Figure 6).
!

42
It is not possible to use this work to estimate forest loss due to agricultural

conversion. The pre-conversion land cover type for agricultural change classes lumped

together forests, savannahs, grasslands, wetlands, and shifting agricultural areas (the

latter certainly captured by mistake due to their spectral and temporal similarity to

naturally vegetated areas). In addition, since the agricultural change included only

naturally vegetated areas as initial land cover, this work does not measure the conversion

of smallholder croplands to plantation agriculture, i.e. agricultural consolidation.

5.4: Factors affecting map accuracy

Though map accuracy was high overall, the results were adversely affected by a

range of factors: few wet-season Landsat scenes, less information available to the

classifier when Landsat 7 SLC-off imagery comprised much of the available data, less

temporal coverage for high-resolution imagery, a paucity of “on the ground” information,

and issues related to training and verification site collection.

There was very little cloud-free Landsat data available during the wet seasons, a

problem endemic to land cover change analysis in tropical regions (Nong et al. 2015,

Kontgis et al. 2015). More wet season (May-October) data would help the decision tree

separate seasonal changes from permanent land conversion, especially for paddy rice;

wet-season flooding creates a very distinct drop in overall reflectance over those areas

that is paired with a rebound in NIR reflectance as the water drains from the paddies

(Kontgis et al. 2015). Also across the study area’s two footprints, 58% of the 2003-2013

scenes were recorded by Landsat 7 after the SLC failure; this led to “no data” gaps in

20% of the pixels in each affected image. While the Landsat SLC-off data is still useful
!

43
for characterizing land cover change in this period, there is definitely a lower signal to

noise ratio, making it more difficult for the classifier to effectively locate areas similar to

the training sites.

High-resolution imagery was not available prior to 2006, hampering training and

verification site selection for the 1990 and 2000 periods. For those years, sites had to be

picked using only the spectral and temporal information available at 30m resolution. At

that scale, the area’s complex landscape and small feature sizes make it difficult to

distinguish stable from changed areas. In addition, the only “ground truth” information

used in this study came from Toro (2012), which was limited to agricultural conversions

and did not confirm new built-up areas.

Training and verification site selection was hampered by additional factors, as

well. No detailed cadastral, planning, or land use data was available to confirm historical

land cover types. Spectral confusion was common across classes: land cover types in the

region often include a variety of materials with unique signatures, e.g. urban land cover

in Southeast Asia includes both dirt and asphalt roads (Jensen 2006). It was impossible to

distinguish crop intensities (e.g. single- vs. double-cropped rice) for agricultural training

sites due to the lack of agricultural census data, high-resolution imagery, and field

surveys (Nong et al. 2015). These factors hindering training and verification site selection

are not unique to Pakse; rural Southeast Asia’s difficult terrain, relative inaccessibility,

and monsoon climate have affected almost all previous studies of land cover change in

the region (Nong et al. 2015). The preceding factors and the well-documented challenges
!

44
of using Landsat to distinguish small croplands and built-up areas harmed the accuracy of

this work’s land cover assessment (Pham and Yamaguchi 2011, Schneider 2012).

Many of the errors in the maps occurred across areas that changed (Table 4).

Often the classifier was able to capture land conversion, but the timing of the conversion

was incorrect (e.g. confusion between the conversion to urban 2000-2006 and conversion

to urban 2006-2015 classes). This result was likely caused by inconsistent spectral and

temporal profiles of the classified areas: nearby vegetation rapidly occluded roads,

homes, and other built-up surfaces soon after land was cleared for construction, creating

pixels with such inconsistent temporal signals that even “stable” pixels appeared changed

and changed pixels were sorted into the wrong time periods. Mixed pixels, exacerbated

by the region’s small feature sizes, are a common source of error in land cover change

studies in tropical regions (Kontgis et al. 2014, Nong et al. 2015). This is clearly the case

in this study, as shown in the confusion between the 1990-2000 and 2000-2006

conversion to agriculture classes.

Finally, one-way transitions were assumed for all change classes in this study. For

urban pixels this is almost always the case and is likely a reasonable assumption for

Pakse (Lambin and Geist 2008). However, the presence of shifting (swidden) agriculture

in the area, where naturally vegetated areas are cleared for planting and then allowed to

regrow naturally after harvest, presents a major challenge for change detection: tropical

swidden plots are often irregularly-shaped, only occupy a few Landsat pixels, and have

highly variable and complex land cover change patterns that may cycle between forest,

soil, burned areas, and planted crops from year to year. This yields unpredictable
!

45
temporal and spectral signals that significantly complicate remote sensing-based

assessments. Landsat imagery has been used to measure the expansion of shifting

agriculture in Vietnam and northern Laos, but the studies required hundreds of field

survey samples and high-resolution imagery for validation (Schmidt-Vogt et al. 2009,

Liao et al. 2015).

Chapter 6. Conclusion

6.1: Overview of research questions

This work characterized land cover change trajectories between 1990-2015 in the

peri-urban region of Pakse, Laos. The first research question focused on “…the amount

of land converted to built-up and agricultural areas in greater Pakse”; this work showed

that 27.2 km2 of urban land and 435.3 km2 of agricultural land were added over the study

period. In all, the amount of new urban land was dwarfed by the amount of new cropland,

with the bulk of the latter established by foreign agribusiness enterprises. While urban

areas grew 28% in the 4,900 km2 study area comprising greater Pakse, agricultural areas

grew 48%; today, fully 5% of the land surface area in Pakse’s peri-urban zone consists of

monoculture plantations growing rubber, cassava, palm, coffee, and other agricultural

products. Agricultural land conversion was the major land cover change story in the

region.

Two Vietnamese rubber companies, The Dak Lak Rubber Company and Dau

Tieng Viet-Lao Rubber Joint Stock Company, were responsible for 52% of all land

conversion in the study area. While the plantations operated by these companies have
!

46
been documented before, detailed information on their extent is not available. This

creates a need for accurate and expedient monitoring of land use and land cover change, a

need that this work shows can be met by remotely sensed imagery and semi-automated

classifiers (Fan et al. 2015, Chen et al. 2016).

Turning to the second research question, “is there a relationship between land

conversion and distance from Pakse’s center, or distance from roads?”, this work showed

that new built-up areas tend to appear closer to Pakse’s core and along roads. This aligns

with current studies on peri-urbanization patterns in Southeast Asia in general and Laos

in particular (Kontigs et al. 2014, Sharifi et al. 2014, Nong et al. 2015). As for new

agricultural areas, there was no evidence of a strong relationship between distance from

central Pakse and the degree of expansion. Agricultural areas tended to grow at a middle

distance 15-25km from Pakse’s core, but it’s not possible to separate “proximity to city”

from the multitude of factors that affect plantation siting (soil quality, labor availability,

land tenure regime, elevation, etc.). There was also no strong relationship between the

location of new croplands and their distance from roads, though this requires further

study; research has found that roads (or often, one main road) are correlated with

increased oil palm and rubber plantation development across Malaysia and China, though

the causal direction has not been established (Ichikawa 2007, Zhou and Thomson 2014).

6.2: Significance of research

This work adds an important spatial-quantitative aspect to land use studies in

Southern Laos, updating existing work on Pakse’s urban land extent. Accurate

assessments of regional land cover trajectories are necessary for sustainable urbanization
!

47
that preserves productive forests, farmland, local biodiversity, and water quality (Kontgis

2014). The work may also be useful for development; the Laotian government’s goal of

turning Pakse into a regional commercial and tourism hub depends on adding structures

and roads in a way that eases the movement of goods and people while preserving the

local environment, a process that can be aided by accurate land cover maps (Rabe et al.

2007, Asian Development Bank 2011). This type of work is also necessary for planning

and assessment towards the government’s goal of restoring the country to 70% forest

cover by 2020 (Lund 2010).

Further study is needed to assess the long-term biophysical effects of the region’s

monoculture plantations on local biodiversity, carbon stocks, and soil quality, as well as

their impacts on local populations (Baird 2010). Harm can follow when these projects fail

or overextend, making knowledge of their true extents even more valuable. Over the past

decade in Laos, for example, low yields in eucalyptus plantations and large coffee

projects caused poverty to increase as contract farmers had to repay loans taken out to

plant cash crops (Asian Development Bank 2005, Schönweger and Messerli 2015).

Swings in commodity demand (especially for rubber) have caused Laotian state

economists to caution against expanding plantation capacity beyond what the export

market can support (Asian Development Bank 2012, Schönweger and Messerli 2015,

Vientiane Times 2013).

Large-scale plantation agriculture is likely to continue even though Champasak’s

government declared in April 2016 that no new concessions would be granted (Vientiane

Times 2016). Existing concessions are under decades-long leases and there is precedent
!

48
for ignoring these kinds of bans: in 2007 the Lao Prime Minister announced an indefinite

moratorium on tree plantations and mining, but it was never fully enforced and land deals

have continued apace. It is also likely that rubber, which makes up the largest plantation

crop in the study area, will continue to be a lucrative export: global rubber demand is

projected to increase 30% by 2020, incentivizing more production (Zurflueh 2013). This

trend highlights the importance of remote sensing-based assessments of Laotian

plantation agriculture, since it provides an open, public, spatially- and temporally-

detailed record of their expansion where official records are opaque and/or incomplete.

6.3: Future directions

Future work should exploit additional medium-resolution imagery, e.g. from the

ASTER or Sentinel-2 sensors, or apply other machine learning algorithms such as

support vector machines or neural networks; support vector machines especially have

been successfully used in Southeast Asian land cover change assessments and would be

extensible to southern Laos (Kontgis et al. 2014, Castrence et al. 2014, Nong et al. 2015).

In addition, medium resolution RADAR sensors such as PALSAR and Sentinel-1 can

peer through clouds to detect land cover change and have been used to measure rubber

plantation extent in Myanmar (Torbick et al. 2016). Though it is difficult to differentiate

forest crops such as rubber from other trees using medium-resolution imagery, high-

resolution data (e.g. RapidEye) and object-based classification can help make more

precise land use determinations by distinguishing between types of plantation crops (S.

Dupuy 2012, Avtar et al. 2012, Toro 2012). Since forest loss and subsistence farmer

displacement are important topics in Southeast Asian land cover change science, future
!

49
studies should exploit higher-resolution imagery and land use surveys to create finer

divisions of land cover classes; this would make it possible to quantify forest loss,

smallholder displacement, and land consolidation to monoculture plantations (Kenney-

Lazar 2012).

Fusing survey, census, cadastral, and other sub-district level information with

satellite data has enhanced land cover change studies in Vietnam by providing more

detailed trend assessment and land use information, but the most recent demographic and

land use information in Laos is only available at the coarser district level. Village-level

agricultural data is available, but was last collected in 2011 (Nong et al. 2015, Sisoulath

et al. 2016). Field surveys and household questionnaires are highly desirable adjuncts to

remotely sensed data that increase accuracy and strengthen the findings of land cover

change studies, but they are expensive and time consuming (Vongvisouk et al. 2016).

Future studies could include more extensive spatial analysis of peri-urban and

agricultural expansion patterns, e.g. measuring new road additions vs. distance from city,

a temporal study of road additions by change period vs. when plantations were

established, or a point-to-point distance analysis of road networks vs. new urban areas.

Peri-urbanization and plantation development present unique challenges in

Southeast Asia, spurring economic growth while displacing villagers, accelerating

environmental degradation, and destroying carbon stocks (Baird 2010, Fox and Castella

2013, Gross et al. 2014). Accurate and expediently-produced land cover trajectories offer

a powerful tool to researchers and environmental advocates to measure these trends, and

may be potentially used by local stakeholders to minimize the impacts of land conversion
!

50
while maximizing economic gains to a broad strata of residents (Hudalah et al. 2007,

Sharifi et al. 2014). More of this work is needed in Laos, where few urbanization studies

have been conducted and where land cover change is projected to accelerate in the

coming decades as cities grow and more land is cleared for export-oriented agriculture.

Sustainable development in Southeast Asia in general and Laos in particular depends on

timely and precise land cover assessments. This study, characterizing Pakse’s land cover

change trajectories at the highest spatial and temporal resolution yet, takes a step in that

direction.

Works Cited
Ahrends, Antje, Peter M. Hollingsworth, Alan D. Ziegler, Jefferson M. Fox, Huafang
Chen, Yufang Su, and Jianchu Xu. 2015. “Current Trends of Rubber Plantation
Expansion May Threaten Biodiversity and Livelihoods.” Global Environmental
Change 34 (September): 48–58. doi:10.1016/j.gloenvcha.2015.06.002.
Asian Development Bank, (first). 2011. “Urban Development Strategy: Pakse Urban
Environmental Improvement Project.”
http://www.adb.org/sites/default/files/linked-documents/43316-012-lao-oth-
03.pdf.
Asian Development Bank. 2012. “Project Administration Manual: Pakse Urban
Environmental Improvement Project.”
http://www.adb.org/projects/documents/pakse-urban-environmental-
improvement-project-pam.
Avtar, Ram, Wataru Takeuchi, and Haruo Sawada. 2012. “Monitoring of Biophysical
Parameters of Cashew Plants in Cambodia Using ALOS/PALSAR Data.”
Environmental Monitoring and Assessment 185 (2): 2023–37.
doi:10.1007/s10661-012-2685-y.
Baird, Ian G. 2010. “Land, Rubber and People: Rapid Agrarian Changes and Responses
in Southern Laos.” Journal of Lao Studies 1 (1): 1–47.
———. 2011. “Turning Land into Capital, Turning People into Labor: Primitive
Accumulation and the Arrival of Large-Scale Economic Land Concessions in the
Lao People’s Democratic Republic.” New Proposals: Journal of Marxism and
Interdisciplinary Inquiry 5 (1): 10–26.
Baird, Ian G., and Jefferson Fox. 2015. “How Land Concessions Affect Places
Elsewhere: Telecoupling, Political Ecology, and Large-Scale Plantations in
!

51
Southern Laos and Northeastern Cambodia.” Land 4 (2): 436–53.
doi:10.3390/land4020436.
Byerlee, Derek. 2014. “The Fall and Rise Again of Plantations in Tropical Asia: History
Repeated?” Land 3 (3): 574–97. doi:10.3390/land3030574.
Castrence, Miguel, Duong H. Nong, Chinh C. Tran, Luisa Young, and Jefferson Fox.
2014. “Mapping Urban Transitions Using Multi-Temporal Landsat and DMSP-
OLS Night-Time Lights Imagery of the Red River Delta in Vietnam.” Land 3
(1): 148–66. doi:10.3390/land3010148.
Castrence, Miguel, Duong H. Nong, Chinh C. Tran, Luisa Young, and Jefferson Fox.
2014. “Mapping Urban Transitions Using Multi-Temporal Landsat and DMSP-
OLS Night-Time Lights Imagery of the Red River Delta in Vietnam.” Land 3
(1): 148–66. doi:10.3390/land3010148.
Chang, Chaoyi. 2013. “Understanding Spatial And Temporal Patterns Of Urban
Expansion In Western China During The Post-Reform Era.” University of
Wisconsin-Madison.
Chen, Bangqian, Xiangping Li, Xiangming Xiao, Bin Zhao, Jinwei Dong, Weili Kou,
Yuanwei Qin, et al. 2016. “Mapping Tropical Forests and Deciduous Rubber
Plantations in Hainan Island, China by Integrating PALSAR 25-M and Multi-
Temporal Landsat Images.” International Journal of Applied Earth Observation
and Geoinformation 50 (August): 117–30. doi:10.1016/j.jag.2016.03.011.
Chen, Huafang, Zhuang-Fang Yi, Dietrich Schmidt-Vogt, Antje Ahrends, Philip
Beckschäfer, Christoph Kleinn, Sailesh Ranjitkar, and Jianchu Xu. 2016.
“Pushing the Limits: The Pattern and Dynamics of Rubber Monoculture
Expansion in Xishuangbanna, SW China.” PLoS ONE 11 (2).
doi:10.1371/journal.pone.0150062.
Cordero-Sancho, S., and S. A. Sader. 2007. “Spectral Analysis and Classification
Accuracy of Coffee Crops Using Landsat and a Topographic-Environmental
Model.” International Journal of Remote Sensing 28 (7): 1577–93.
doi:10.1080/01431160600887680.
Delang, Claudio O, Matthew Toro, and Marieke Charlet-Phommachanh. 2013. “Coffee,
Mines and Dams: Conflicts over Land in the Bolaven Plateau, Southern Lao
PDR: Coffee, Mines and Dams.” The Geographical Journal 179 (2): 150–64.
doi:10.1111/j.1475-4959.2012.00481.x.
Department, ADB Independent Evaluation. 2005. Champasack Road Improvement
Project in the Lao People’s Democratic Republic (Loan 1369-LAO[SF]). ADB
Independent Evaluation Department.
https://www.adb.org/documents/champasack-road-improvement-project-lao-
peoples-democratic-republic-loan-1369-laosf.
Dewan, Ashraf M., Yasushi Yamaguchi, and Md Ziaur Rahman. 2010. “Dynamics of
Land Use/Cover Changes and the Analysis of Landscape Fragmentation in
Dhaka Metropolitan, Bangladesh.” GeoJournal 77 (3): 315–30.
doi:10.1007/s10708-010-9399-x.
!

52
“Development Strategy for Special and Specific Economic Zone (SEZ) in the Lao PDR,
2011 - 2020.” 2012. National Committee for Special and Specific Economic
Zones Secretariat Office.
Ducourtieux, Olivier. 2006. “Is the Diversity of Shifting Cultivation Held in High
Enough Esteem in Lao PDR?” Moussons. Recherche En Sciences Humaines Sur
l’Asie Du Sud-Est, no. 9–10 (December): 61–86. doi:10.4000/moussons.1887.
Dwyer, M. 2007. Turning Land into Capital: A Review of Recent Research on Land
Concessions for Investment in Lao PDR. Land Issues Working Group,
Vientiane, Laos.
Eco-Business. 2016. “Laos Bans New Mining Projects in a Polluted Province.” Eco-
Business. Accessed October 27. http://www.eco,business.com/news/laos,
bans,new,mining,projects,in,a,polluted,province/.
“Ecosystems in the Greater Mekong: Past Trends, Current Status, Possible Futures.”
2013. World Wildlife Fund.
Fan, Hui, Xiaohua Fu, Zheng Zhang, and Qiong Wu. 2015. “Phenology-Based
Vegetation Index Differencing for Mapping of Rubber Plantations Using Landsat
OLI Data.” Remote Sensing 7 (5): 6041–58. doi:10.3390/rs70506041.
“FDI-DDI by Sector.” 2015. Ministry of Planning and Investment Investment Promotion
Department. http://www.investlaos.gov.la/index.php/resources/statistics.
Feng, Min, Chengquan Huang, Joseph O. Sexton, Saurabh Channan, Raghuram
Narasimhan, and John R. Townshend. 2012. “An Approach for Quickly Labeling
Land Cover Types for Multiple Epochs at Globally Selected Locations.” In ,
6203–6. IEEE. doi:10.1109/IGARSS.2012.6352674.
Firman, Tommy. 2004. “New Town Development in Jakarta Metropolitan Region: A
Perspective of Spatial Segregation.” Habitat International 28 (3): 349–68.
doi:10.1016/S0197-3975(03)00037-7.
Fortunel, Frédéric. 2007. “Le plateau des Boloven et la culture du café, entre division
interne et intégration régionale.” L’Espace géographique Tome 36 (3): 215–28.
Fox, Jefferson, and Jean-Christophe Castella. 2013. “Expansion of Rubber (Hevea
Brasiliensis) in Mainland Southeast Asia: What Are the Prospects for
Smallholders?” The Journal of Peasant Studies 40 (1): 155–70.
doi:10.1080/03066150.2012.750605.
Fox, Jefferson, and John B. Vogler. 2005. “Land-Use and Land-Cover Change in
Montane Mainland Southeast Asia.” Environmental Management 36 (3): 394–
403. doi:10.1007/s00267-003-0288-7.
Friedl, M. A, D. K McIver, J. C. F Hodges, X. Y Zhang, D Muchoney, A. H Strahler, C.
E Woodcock, et al. 2002. “Global Land Cover Mapping from MODIS:
Algorithms and Early Results.” Remote Sensing of Environment, The Moderate
Resolution Imaging Spectroradiometer (MODIS): a new generation of Land
Surface Monitoring, 83 (1–2): 287–302. doi:10.1016/S0034-4257(02)00078-0.
Friedl, M. A., and C. E. Brodley. 1997. “Decision Tree Classification of Land Cover
from Remotely Sensed Data.” Remote Sensing of Environment 61 (3): 399–409.
doi:10.1016/S0034-4257(97)00049-7.
!

53
Fujimura, Manabu, and Adhikari Ramesh. 2010. “Critical Evaluation of Cross-Border
Infrastructure Projects in Asia.” Asian Development Bank.
Gibbs, H. K., A. S. Ruesch, F. Achard, M. K. Clayton, P. Holmgren, N. Ramankutty, and
J. A. Foley. 2010. “Tropical Forests Were the Primary Sources of New
Agricultural Land in the 1980s and 1990s.” Proceedings of the National
Academy of Sciences 107 (38): 16732–37. doi:10.1073/pnas.0910275107.
Global Witness. 2013. “Rubber Barons.”
https://www.globalwitness.org/en/campaigns/land-deals/rubberbarons/.
Gross, Jill Simone, Lin Ye, and Richard Legates. 2014. “Asia and the Pacific Rim: The
New Peri-Urbanization and Urban Theory.” Journal of Urban Affairs 36 (s1):
309–14. doi:10.1111/juaf.12109.
Hall, Derek. 2003. “The International Political Ecology of Industrial Shrimp Aquaculture
and Industrial Plantation Forestry in Southeast Asia.” Journal of Southeast Asian
Studies, June. /core/journals/journal-of-southeast-asian-studies/article/div-
classtitlethe-international-political-ecology-of-industrial-shrimp-aquaculture-
and-industrial-plantation-forestry-in-southeast-
asiadiv/A8855AEDF14137F468CDA793F1A8CBE8.
Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina,
D. Thau, et al. 2013. “High-Resolution Global Maps of 21st-Century Forest
Cover Change.” Science 342 (6160): 850–53. doi:10.1126/science.1244693.
Härdter, Rolf, Woo Yin Chow, and Ooi Soo Hock. 1997. “Intensive Plantation Cropping,
a Source of Sustainable Food and Energy Production in the Tropical Rain Forest
Areas in Southeast Asia.” Forest Ecology and Management, Agroforestry and
Land use Change in Industrialized Nations, 91 (1): 93–102. doi:10.1016/S0378-
1127(96)03880-7.
Hirsch, Philip, and Natalia Scurrah. 2015. “The Political Economy of Land Governance
in Lao PDR.” http://mrlg.org/wp,
content/uploads/2015/12/Political_Economy_of_Land_Governance_in_Lao_
PDR_FA.pdf.
“Historical Data on the Global Coffee Trade.” 2015. International Coffee Organization.
http://www.ico.org/new_historical.asp?section=Statistics.
Huang, Chengquan, Samuel N. Goward, Jeffrey G. Masek, Nancy Thomas, Zhiliang Zhu,
and James E. Vogelmann. 2010. “An Automated Approach for Reconstructing
Recent Forest Disturbance History Using Dense Landsat Time Series Stacks.”
Remote Sensing of Environment 114 (1): 183–98. doi:10.1016/j.rse.2009.08.017.
Hudalah, Delik, Haryo Winarso, and Johan Woltjer. 2007. “Peri-Urbanisation in East
Asia: A New Challenge for Planning?” International Development Planning
Review 29 (4): 503–19. doi:10.3828/idpr.29.4.4.
Hurni, Kaspar. 2008. “Rubber in Laos: Detection of Actual and Assessment of Potential
Plantations in the Lao PDR Using GIS and Remote Sensing Technologies.”
University of Berne.
https://www.researchgate.net/publication/307169845_Rubber_in_Laos_Detectio
n_of_actual_and_assessment_of_potential_plantations_in_the_Lao_PDR_using_
GIS_and_remote_sensing_technologies.
!

54
Ichikawa, Masahiro. 2007. “Degradation and Loss of Forest Land and Land-Use Changes
in Sarawak, East Malaysia: A Study of Native Land Use by the Iban.” Ecological
Research 22 (3): 403–13. doi:10.1007/s11284-007-0365-0.
“Industry, Value Added (% of GDP).” 2014. World Bank.
http://data.worldbank.org/indicator/NV.IND.TOTL.ZS.
Jensen, John R. 2006. Remote Sensing of the Environment: An Earth Resource
Perspective. 2 edition. Upper Saddle River, NJ: Pearson.
Keivani, Ramin, and Michael Mattingly. 2007. “The Interface of Globalization and
Peripheral Land in the Cities of the South: Implications for Urban Governance
and Local Economic Development.” International Journal of Urban and Regional
Research 31 (2): 459–74. doi:10.1111/j.1468-2427.2007.00718.x.
Kenney-Lazar, Miles. 2012. “Plantation Rubber, Land Grabbing and Social-Property
Transformation in Southern Laos.” The Journal of Peasant Studies 39 (3–4):
1017–37. doi:10.1080/03066150.2012.674942.
Kimijiama, S., and M. Nagai. 2014. “Study for Urbanization Corresponding to Socio-
Economic Activities in Savannaket, Laos Using Satellite Remote Sensing.” IOP
Conference Series: Earth and Environmental Science 20 (1): 12005.
doi:10.1088/1755-1315/20/1/012005.
Kontgis, Caitlin, Annemarie Schneider, and Mutlu Ozdogan. 2015. “Mapping Rice Paddy
Extent and Intensification in the Vietnamese Mekong River Delta with Dense
Time Stacks of Landsat Data.” Remote Sensing of Environment 169
(November): 255–69. doi:10.1016/j.rse.2015.08.004.
Kontgis, Caitlin, Annemarie Schneider, Jefferson Fox, Sumeet Saksena, James H.
Spencer, and Miguel Castrence. 2014. “Monitoring Peri-Urbanization in the
Greater Ho Chi Minh City Metropolitan Area.” Applied Geography 53
(September): 377–88. doi:10.1016/j.apgeog.2014.06.029.
Kritsanaphan, Amorn, and Edsel Sajor. 2011. “Intermediaries and Informal Interactions
in Decentralised Environmental Management in Peri-Urban Bangkok.”
International Development Planning Review 33 (3): 247–72.
doi:10.3828/idpr.2011.11.
Lainé, Elsa. 2015. “Urban Development and New Actors in Lao PDR in the Context of
Regionalization: Case Studies of two Border Towns,” September.
doi:urn:doi:10.4000/moussons.3258.
Lambin, Eric F., and Helmut J. Geist. 2008. Land-Use and Land-Cover Change: Local
Processes and Global Impacts. Springer Science & Business Media.
Land Issues Working Group. 2012. “Cases from the Field.”
http://www.laolandissues.org/case-studies/.
Lao PDR Ministry of Agriculture and Forestry. 2010a. “Agricultural Master Plan 2011 to
2015.”
———. 2010b. “Strategy for Agricultural Development 2011 to 2020.”
Lao Population and Housing Census. 2015. Results of Population and Housing Census
2015.
!

55
Latsaphao, Khonesavanh. 2016. “No More Land Concessions in Champassak.” Vientiane
Times, April 7. http://www.farmlandgrab.org/post/view/25971-no-more-land-
concessions-in-champassak.
———. 2016b. “Korean Company Looking to Crack Open Cashew Nut Market.”
Vientiane Times, August 23.
Laungaramsri, Pinkaew, and Pornpana Kuaycharoen. 2009. “Research Evaluation of
Economic, Social, and Ecological Implications of the Programme for
Commercial Tree Plantations: Case Study of Rubber in the South of Laos PDR.”
Centre for Research and Information on Land and Natural Resources, National
Land Management Authority, Office of Prime Minister, Lao PDR.
Laungaramsri, Pinkaew. 2012. “Frontier Capitalism and the Expansion of Rubber
Plantations in Southern Laos.” Journal of Southeast Asian Studies 43 (3): 463–
77. doi:10.1017/S0022463412000343.
Legates, Richard, and Delik Hudalah. 2014. “Peri-Urban Planning for Developing East
Asia: Learning from Chengdu, China and Yogyakarta/Kartamantul, Indonesia.”
Journal of Urban Affairs 36 (s1): 334–53. doi:10.1111/juaf.12106.
Leinenkugel, Patrick, Michel L. Wolters, Natascha Oppelt, and Claudia Kuenzer. 2015.
“Tree Cover and Forest Cover Dynamics in the Mekong Basin from 2001 to
2011.” Remote Sensing of Environment 158 (March): 376–92.
doi:10.1016/j.rse.2014.10.021.
Li, Xia, and Anthony Gar-On Yeh. 2004. “Analyzing Spatial Restructuring of Land Use
Patterns in a Fast Growing Region Using Remote Sensing and GIS.” Landscape
and Urban Planning 69 (4): 335–54. doi:10.1016/j.landurbplan.2003.10.033.
Li, Zhe, and Jefferson M. Fox. 2012. “Mapping Rubber Tree Growth in Mainland
Southeast Asia Using Time-Series MODIS 250 M NDVI and Statistical Data.”
Applied Geography 32 (2): 420–32. doi:10.1016/j.apgeog.2011.06.018.
Liao, Chenhua, Zhiming Feng, Peng Li, and Jinghua Zhang. 2015. “Monitoring the
Spatio-Temporal Dynamics of Swidden Agriculture and Fallow Vegetation
Recovery Using Landsat Imagery in Northern Laos.” Journal of Geographical
Sciences 25 (10): 1218–34. doi:10.1007/s11442-015-1229-0.
Luck, Matthew, and Jianguo Wu. 2002. “A Gradient Analysis of Urban Landscape
Pattern: A Case Study from the Phoenix Metropolitan Region, Arizona, USA.”
Landscape Ecology 17 (4): 327–39. doi:10.1023/A:1020512723753.
Lund, Christian. 2010. “Urbanization and Land Conversion in Vientiane, Lao PDR.”
Manivong, Vongpaphane. 2014. “Agrarian Transition in Lowland Southern Laos:
Implications for Rural Livelihoods.”
http://espace.library.uq.edu.au/view/UQ:344279.
Masek, J. G., E. F. Vermote, N. E. Saleous, R. Wolfe, F. G. Hall, K. F. Huemmrich, Feng
Gao, J. Kutler, and Teng-Kui Lim. 2006. “A Landsat Surface Reflectance
Dataset for North America, 1990-2000.” IEEE Geoscience and Remote Sensing
Letters 3 (1): 68–72. doi:10.1109/LGRS.2005.857030.
McIver, D. K., and M. A. Friedl. 2001. “Estimating Pixel-Scale Land Cover
Classification Confidence Using Nonparametric Machine Learning Methods.”
!

56
IEEE Transactions on Geoscience and Remote Sensing 39 (9): 1959–68.
doi:10.1109/36.951086.
Mertes, C. M., A. Schneider, D. Sulla-Menashe, A. J. Tatem, and B. Tan. 2015.
“Detecting Change in Urban Areas at Continental Scales with MODIS Data.”
Remote Sensing of Environment 158 (March): 331–47.
doi:10.1016/j.rse.2014.09.023.
Messerli, Peter, Amaury Peeters, Oliver Schoenweger, Vong Nanhthavong, and Andreas
Heinimann. 2015. “Marginal Lands or Marginal People? Analysing Key
Processes Determining the Outcomes of Large-Scale Land Acquisitions in Lao
PDR and Cambodia.” International Development Policy | Revue Internationale
de Politique de Développement, no. 6.1 (September). doi:10.4000/poldev.2037.
Messerli, Peter. 2014. “The Lao DECIDE Info Project.”
https://cdeweb4.unibe.ch/pages/project/2/23/the-lao-decide-info-project.aspx.
Ministry of Industry and Commerce. 2009. Scoping Study on Cross-Border Agribusiness
in Lao PDR: Focus on Champasak Province!: Trade Development Facility. Trade
Development Facility.
Ministry of Planning and Investment. 2015. “The Eighth Five-Year National Socio-
Economic Development Plan (2016-2020).”
Nahuelhual, Laura, Alejandra Carmona, Antonio Lara, Cristian Echeverría, and Mauro E.
González. 2012. “Land-Cover Change to Forest Plantations: Proximate Causes
and Implications for the Landscape in South-Central Chile.” Landscape and
Urban Planning 107 (1): 12–20. doi:10.1016/j.landurbplan.2012.04.006.
Nguyen Thi, Dien, Philippe Lebailly, and Ton Vu Dinh. 2010. “Land Conversion to
Industrialization and Its Impacts on Household Food Security in Red River
Delta, Vietnam.” http://orbi.ulg.ac.be/handle/2268/72522.
Nolintha, Vanthana. 2011. “Cities, SEZs and Connectivity in Major Provinces of Laos.”
Nong, Duong H., Jefferson Fox, Tomoaki Miura, and Sumeet Saksena. 2015. “Built-up
Area Change Analysis in Hanoi Using Support Vector Machine Classification of
Landsat Multi-Temporal Image Stacks and Population Data.” Land 4 (4): 1213–
31. doi:10.3390/land4041213.
Nong, Duong H., Jefferson Fox, Tomoaki Miura, and Sumeet Saksena. 2015. “Built-up
Area Change Analysis in Hanoi Using Support Vector Machine Classification of
Landsat Multi-Temporal Image Stacks and Population Data.” Land 4 (4): 1213–
31. doi:10.3390/land4041213.
Obein, F. 2007. “Industrial Rubber Plantation of the Viet-Lao Rubber Company,
Bachiang District, Champasack Province: Assessment of the Environmental and
Social Impacts Created by the VLRC Industrial Rubber Plantation and Proposed
Environmental and Social Plans. Produced for Agence Francaise de
Développement.” Earth Systems Lao.
Otukei, J. R., and T. Blaschke. 2010. “Land Cover Change Assessment Using Decision
Trees, Support Vector Machines and Maximum Likelihood Classification
Algorithms.” International Journal of Applied Earth Observation and
Geoinformation, Supplement Issue on “Remote Sensing for Africa – A Special
Collection from the African Association for Remote Sensing of the Environment
!

57
(AARSE),” 12, Supplement 1 (February): S27–31.
doi:10.1016/j.jag.2009.11.002.
“OSM WikiProject Laos.” 2016. July 29.
http://wiki.openstreetmap.org/wiki/WikiProject_Laos#Road_status.
Pal, Mahesh, and Paul M Mather. 2003. “An Assessment of the Effectiveness of Decision
Tree Methods for Land Cover Classification.” Remote Sensing of Environment
86 (4): 554–65. doi:10.1016/S0034-4257(03)00132-9.
Pham, Hai Minh, and Yasushi Yamaguchi. 2011. “Urban Growth and Change Analysis
Using Remote Sensing and Spatial Metrics from 1975 to 2003 for Hanoi,
Vietnam.” International Journal of Remote Sensing 32 (7): 1901–15.
doi:10.1080/01431161003639652.
Phimmavong, S., B. Ozarska, S. Midgley, and R. Keenan. 2009. “Forest and Plantation
Development in Laos: History, Development and Impact for Rural
Communities.” International Forestry Review 11 (4): 501–13.
doi:10.1505/ifor.11.4.501.
Phompila, Chittana, Megan Lewis, Kenneth Clarke, and Bertram Ostendorf. 2014.
“Monitoring Expansion of Plantations in Lao Tropical Forests Using Landsat
Time Series.” In , 9260:92601M–92601M–11. doi:10.1117/12.2068283.
Portilla, Gilda Sentíes. 2015. “Land Concessions and Rural Youth in Southern Laos.” In .
http://www.iss.nl/research/research_programmes/political_economy_of_resource
s_environment_and_population_per/networks/land_deal_politics_ldpi/conferenc
es/land_grabbing_perspectives_from_east_and_southeast_asia/.
Quinlan, J. R. 1996. “Bagging, Boosting, and C4.5.” In In Proceedings of the Thirteenth
National Conference on Artificial Intelligence, 725–730. AAAI Press.
Quinlan, J. Ross. 1993. C4.5: Programs for Machine Learning. San Francisco, CA, USA:
Morgan Kaufmann Publishers Inc.
Rabé, Paul, Thenekham Thongbonh, and Vongdeuane Vongsiharath. 2007. “Study on
Urban Land Management and Planning in Lao PDR.” Lao-German Land Policy
Development Project (German Contribution to the Lao Land Titling Project II in
Lao PDR).
Rennenberg, Nils. n.d. Assessment of Land Use Changes in Mukdaham and Nakhon
Phanom Provinces (NE Thailand) by Means of Remote Sensing. diplom.de.
Richards, Daniel R., and Daniel A. Friess. 2016. “Rates and Drivers of Mangrove
Deforestation in Southeast Asia, 2000-2012.” Proceedings of the National
Academy of Sciences of the United States of America 113 (2): 344–49.
doi:10.1073/pnas.1510272113.
Rogan, John, and DongMei Chen. 2004. “Remote Sensing Technology for Mapping and
Monitoring Land-Cover and Land-Use Change.” Progress in Planning 61 (4):
301–25. doi:10.1016/S0305-9006(03)00066-7.
S. Dupuy, A, V. Herbreteau, T. Feyfant, S. Morand. 2012. “Land-Cover Dynamics in
Southeast Asia: Contribution of Object-Oriented Techniques for Change
Detection.” In 4th International Conference on GEographic Object-Based Image
Analysis (GEOBIA 2012).
!

58
Schmidt-Vogt, Dietrich, Stephen J. Leisz, Ole Mertz, Andreas Heinimann, Thiha Thiha,
Peter Messerli, Michael Epprecht, et al. 2009. “An Assessment of Trends in the
Extent of Swidden in Southeast Asia.” Human Ecology 37 (3): 269–80.
Schneider, A., C. M. Mertes, A. J. Tatem, B. Tan, D. Sulla-Menashe, S. J. Graves, N. N.
Patel, et al. 2015. “A New Urban Landscape in East–Southeast Asia, 2000–
2010.” Environmental Research Letters 10 (3): 34002. doi:10.1088/1748-
9326/10/3/034002.
Schneider, A., M. A. Friedl, and D. Potere. 2009. “A New Map of Global Urban Extent
from MODIS Satellite Data.” Environmental Research Letters 4 (4): 44003.
doi:10.1088/1748-9326/4/4/044003.
Schneider, A., M.A. Friedl, and C.E. Woodcock. 2003. “Mapping Urban Areas by Fusing
Multiple Sources of Coarse Resolution Remotely Sensed Data.” In Geoscience
and Remote Sensing Symposium, 2003. IGARSS ’03. Proceedings. 2003 IEEE
International, 4:2623–25 vol.4. doi:10.1109/IGARSS.2003.1294530.
Schneider, Annemarie. 2012. “Monitoring Land Cover Change in Urban and Peri-Urban
Areas Using Dense Time Stacks of Landsat Satellite Data and a Data Mining
Approach.” Remote Sensing of Environment 124 (September): 689–704.
doi:10.1016/j.rse.2012.06.006.
Schönweger, O., A. Heinimann, M. Epprecht, J. Lu, and P. Thalongsengchanh. 2012.
“Concessions and Leases in the Lao PDR: Taking Stock of Land Investments.”
Centre for Development and Environment (CDE), University of Bern, Bern and
Vientiane.
Schönweger, Oliver, and Peter Messerli. 2015. “Land Acquisition, Investment, and
Development in the Lao Coffee Sector: Successes and Failures.” Critical Asian
Studies 47 (1): 94–122. doi:10.1080/14672715.2015.997095.
“Sector Assistance Program Evaluation for the Agriculture and Natural Resources Sector
in the Lao People’s Democratic Republic.” 2005. Asian Development Bank.
Seto, K. C., C. E. Woodcock, C. Song, X. Huang, J. Lu, and R. K. Kaufmann. 2002.
“Monitoring Land-Use Change in the Pearl River Delta Using Landsat TM.”
International Journal of Remote Sensing 23 (10): 1985–2004.
doi:10.1080/01431160110075532.
Seto, Karen C., Michail Fragkias, Burak Güneralp, and Michael K. Reilly. 2011. “A
Meta-Analysis of Global Urban Land Expansion.” PLOS ONE 6 (8): e23777.
doi:10.1371/journal.pone.0023777.
Sharifi, Ayyoob, Yoshihiro Chiba, Kohei Okamoto, Satoshi Yokoyama, and Akito
Murayama. 2014. “Can Master Planning Control and Regulate Urban Growth in
Vientiane, Laos?” Landscape and Urban Planning 131 (November): 1–13.
doi:10.1016/j.landurbplan.2014.07.014.
Simon, David. 2008. “Urban Environments: Issues on the Peri-Urban Fringe.” Annual
Review of Environment and Resources 33 (1): 167–85.
doi:10.1146/annurev.environ.33.021407.093240.
Sisoulath, Vilaysouk, Michael Epprecht, Obert Pimhidzai, and Harold Coulombe. 2016.
“Where Are the Poor? Lao PDR 2015 Census-Based Poverty Map!: Province and
District Level Results.” 106899. The World Bank.
!

59
http://documents.worldbank.org/curated/en/477381468415961977/Where-are-
the-poor-Lao-PDR-2015-census-based-poverty-map-province-and-district-level-
results.
Son, Nguyen-Thanh, Chi-Farn Chen, Cheng-Ru Chen, Huynh-Ngoc Duc, and Ly-Yu
Chang. 2013. “A Phenology-Based Classification of Time-Series MODIS Data
for Rice Crop Monitoring in Mekong Delta, Vietnam.” Remote Sensing 6 (1):
135–56. doi:10.3390/rs6010135.
Song, Conghe, Curtis E. Woodcock, Karen C. Seto, Mary Pax Lenney, and Scott A.
Macomber. 2001. “Classification and Change Detection Using Landsat TM
Data: When and How to Correct Atmospheric Effects?” Remote Sensing of
Environment 75 (2): 230–244.
Souphonphacdy, Daovinh, Mitsuyasu Yabe, Goshi Sato, , and . 2012.
Impact of Rubber Concession on Rural Livelihood in Champasack Province,
Lao PDR. https://core.ac.uk/display/37855219.
Steele, Brian M. 2005. “Maximum Posterior Probability Estimators of Map Accuracy.”
Remote Sensing of Environment 99 (3): 254–70. doi:10.1016/j.rse.2005.09.001.
Steele, Brian M. 2005. “Maximum Posterior Probability Estimators of Map Accuracy.”
Remote Sensing of Environment 99 (3): 254–70. doi:10.1016/j.rse.2005.09.001.
Stibig, H.-J., F. Achard, S. Carboni, R. Raši, and J. Miettinen. 2014. “Change in Tropical
Forest Cover of Southeast Asia from 1990 to 2010.” Biogeosciences 11 (2): 247–
58. doi:10.5194/bg-11-247-2014.
Thai National Mekong Committee Knowledge Center. 2000. “Overview and Socio-
Economic Development Plans, Lao PDR.”
http://www.tnmckc.org/upload/document/bdp/2/2.2/National_Policy/Soc
io,economic%20lao%20pdr2.pdf.
“The Seventh Five-Year National Socio-Economic Development Plan (2011-2015).”
2011. Ministry of Planning and Investment.
Thomas, Ian. 2015. “Drivers of Forest Change in the Greater Mekong Subregion - Lao
PDR Country Report.”
Torbick, Nathan, Lindsay Ledoux, William Salas, and Meng Zhao. 2016. “Regional
Mapping of Plantation Extent Using Multisensor Imagery.” Remote Sensing 8
(3). doi:10.3390/rs8030236.
Toro, Matthew. 2012. “Coffee Markets, Smallholder Credit, and Landscape Change in
the Bolaven Plateau Region, Laos.” Open Access Theses, May.
http://scholarlyrepository.miami.edu/oa_theses/333.
“Traditional Production Area: The Bolaven Plateau.” 2013. Lao Coffee Association.
http://www.laocoffeeassociation.org/index.php/en/about-lao-coffee/production-
areas.
United Nations, Department of Economic and Social Affairs, and Population Division.
2014. World Urbanization Prospects: The 2014 Revision.
Vientiane Times. 2013. “Economists Warn Laos on Rubber,” October 3.
http://www.laolandissues.org/2013/10/03/economists-warn-laos-on-rubber/.
———. 2014. “Champassak Coffee Exports Soar,” July 17.
!

60
———. 2015. “Laos-Japan Economic Zone to Benefit Local Community,” December 30.
———. 2016a. “Govt Aims to Restore Forest Cover to 70 Percent by 2020,” November
22.
———. 2016b. “Growth of Agriculture Wilting under Challenges,” November 24.
———. 2016c. “Govt Approves Development of Champa Nakhone SEZ,” May 12.
Vongkhamheng, Chansack, Jianhua Zhou, Mukete Beckline, and Sythud Phimmachanh.
2016. “Socioeconomic and Ecological Impact Analysis of Rubber Cultivation in
Southeast Asia.” OALib 3 (1): 1–11. doi:10.4236/oalib.1102339.
Vongvisouk, Thoumthone, Rikke Brandt Broegaard, Ole Mertz, and Sithong
Thongmanivong. 2016. “Rush for Cash Crops and Forest Protection: Neither
Land Sparing nor Land Sharing.” Land Use Policy 55 (September): 182–92.
doi:10.1016/j.landusepol.2016.04.001.
Webster, Douglas, Jianming Cai, and Larissa Muller. 2014. “The New Face of Peri-
Urbanization in East Asia: Modern Production Zones, Middle-Class Lifesytles,
and Rising Expectations.” Journal of Urban Affairs 36 (s1): 315–33.
doi:10.1111/juaf.12104.
Webster. 2002. On the Edge: Shaping the Future of Peri-Urban East Asia. Asia/Pacific
Research Center.
Zhou, Adrian, and Elspeth Thomson. 2009. “The Development of Biofuels in Asia.”
Applied Energy, Bio-fuels in Asia, 86, Supplement 1 (November): S11–20.
doi:10.1016/j.apenergy.2009.04.028.
Zurflueh, Joel. 2013. “Vietnamese Rubber Investments in the South of the Lao PDR.”
University of Berne.
!

Table 1: Landsat TM, ETM+ and OLI surface reflectance scenes used as inputs to the decision tree. 102 scenes from path/row
126/49 and 95 scenes from path/row 126/50 were used.

61
!

Table 2: The 14 land cover and land cover change classes (based on Seto 2002, Schneider 2012), criteria for class inclusion,
and number of training exemplars used during the supervised classification procedure.

62
!

Table 3: Results of the decision tree classification for each stable and changed land cover class. Conversion to agriculture
made up the largest change class: 435 km2 of agricultural land was added 1990-2015, most of it large-scale plantation
agriculture. 27 km2 of urban land was added, which includes residential areas, roads and commercial structures.

63
 !

Reference data

1990 to
2000: 1990 to 2000 to 2000 to 2006 to 2006 to User’s
natural 2000: 1990 to 2006: 2006: 2000 to 2015: 2015: 2006 to accuracy
vegetation natural 2000: natural natural 2006: natural natural 2015:
Stable / vegetation/ vegetation vegetation/ vegetation/ vegetation vegetation/ vegetation/ vegetation (commis
natural Stable Stable Stable Stable agriculture agriculture to agriculture agriculture to agriculture agriculture to sion
vegetation agriculture water urban barren to water to urban agriculture to water to urban agriculture to water to urban agriculture Total error)
Stable natural vegetation 54 1 1 1 1 Reading this table left to right: 58 93%
The “stable natural vegetation” mistakenly included four pixels belonging
Stable agriculture 1 55 1 to the “stable agriculture,” “stable water,” “stable barren” and “1990 to 57 96%
2000: vegetation to agriculture” classes. That’s an error of commission
Stable water 1 55 1 1 58 95%

Stable urban 54 2 1 1 1 59 92%

Stable barren 1 2 55 1 1 60 92%

Decision tree output

1990 to 2000: natural

vegetation/agriculture to water 1 1 50 2 54 93%

1990 to 2000: natural

vegetation/agriculture to urban 3 1 52 2 1 59 88%

1990 to 2000: vegetation to agriculture 3 1 55 1 2 62 89%

2000 to 2006: natural

vegetation/agriculture to water 1 1 52 1 55 95%

2000 to 2006: natural

vegetation/agriculture to urban Reading this table top to bottom: 4 52 3 59 88%
“Stable agriculture” should have included
2000 to 2006: vegetation to agriculture four more pixels of stable agricultural areas, 2 57 1 2 62 92%
but these were left out, instead assigned to
2006 to 2015: natural “stable natural vegetation” and “1990 to 2 2 52 56 91%
vegetation/agriculture to water 2000: natural vegetation to agriculture.”
2006 to 2015: natural That’s an error of omission.
vegetation/agriculture to urban
2 3 53 1 59 90%

2006 to 2015: vegetation to agriculture 2 4 55 61 90%

Total 58 59 58 59 59 53 60 61 57 59 62 54 60 60 819

Producer’s accuracy
(omission error)
93% 93% 95% 92% 93% 94% 87% 90% 91% 88% 92% 94% 88% 92% 91.6%

Producer’s accuracy average: 91.6% Overall accuracy ↑

User’s accuracy average: 91.6%
Kappa coefficient: 0.9024

Table 4: Confusion matrix comparing a sample of test sites to the final map of land cover change to assess map accuracy.
Producer’s, user’s and overall accuracies were all approximately 92%. The map had a kappa coefficient of 0.9024, indicating
the classification was 90% better than one resulting from random chance.

64
!

65

Figure 1: The Pakse study area and the Landsat scene footprints used in this analysis.
!

66

Figure 2: The extent of the study area, encompassing a 35km radius around central
Pakse. Elevation, roads, drainages, and available information on land concessions are
also indicated.
!

67

Figure 3: Simplified NIR reflectance trends for different land cover types and
conversions. Croplands have a distinct cycle of rising and falling near-infrared
reflectance as plants are sown and harvested. Urban cover still shows a slight cycle as
vegetation is commingled among the structures, roads, and other impervious surfaces
that comprise an urban area. Conversions to urban or water show a large drop in overall
reflectance, with water losing the phenological signal altogether. Most of the paddies in
the study area contain single-cropped rain-fed rice (Manivong 2014). The illustrative
reflectance trends pictured here were adapted and simplified from Son et al (2013).
!

68

Figure 4: Schematic decision tree. Training data is used to generate a series of tests that
are applied to the dataset, recursively splitting it into increasingly homogeneous classes.
After the classes are satisfactorily homogenous and separable, the rules are applied to
the entire image.
!

69

Figure 5: Work flow used to generate the land cover map.

!

70
(a)

(b) (c) (d)

Figure 6: The final land cover map 1990-2015 for the greater Pakse area for (a) the full
study area, as well as zoom boxes illustrating (b) areas that changed to water, (c) areas
converted to urban and built-up areas, and (d) areas converted to agriculture.
 !

(a) (b) 71

Figure 7: Summary of results showing (a) land cover change by time period and (b) land
cover types by total share of study region. Note that natural vegetation and barren land
cover types (66% of the study area) were excluded to better highlight the change classes.
!

(a) (b) (c)

(d) (e) (f)

Figure 8: Results of the buffer analyses measuring land cover conversion vs. distance from the Pakse core (a) and from roads
(d). The results are shown for new urban land (b, e) and for new agricultural land (c, f). Note that developable land refers to
any land that is not in a water or stable urban class.

72
!

Figure 9: Land cover

comparisons between
Landsat near-infrared
composites (bright red
means high near-
infrared reflectance),
high-resolution imagery
from Google Earth, and
the classification result.
All images are at the
same scale. (a-f) shows
conversions to
agriculture 1990-2006,
(g-l) shows conversions
to urban land 1990-
2006, (m-q) shows
conversions to
agriculture 2006-2015
with additional high-
resolution imagery, and
(r-v) shows conversions
to urban land 2006-2015
with additional high-
resolution imagery.

73
!

Figure 10: Views of plantations in the

greater Pakse area captured from
Landsat, Google Earth, and on the
ground. Photos are representative only
and were not taken at the sites imaged
by the satellite.

74