A SYSTEMS APPROACH TO ANALYZE HOUSEHOLD VULNERABILITY TO FOOD

INSECURITY IN RURAL SOUTHERN MALI USING A SPATIALLY-EXPLICIT

INTEGRATED SOCIAL AND BIOPHYSICAL MODEL

By

Rajiv Paudel

A DISSERTATION

Michigan State University

in partial fulﬁllment of the requirements

Submitted to

for the degree of

Geography – Doctor of Philosophy

2020

ABSTRACT

A SYSTEMS APPROACH TO ANALYZE HOUSEHOLD VULNERABILITY TO FOOD

INSECURITY IN RURAL SOUTHERN MALI USING A SPATIALLY-EXPLICIT

INTEGRATED SOCIAL AND BIOPHYSICAL MODEL

By

Rajiv Paudel

Mali is expected to be profoundly impacted by climate change. Its mid-century temperature could
increase by 2◦C, which could have detrimental impacts on crop production. Besides, northern
Mali is currently highly arid and is not suitable for agriculture. The South has the responsibility to
feed the national population. However, the region is experiencing rapid population growth. Mali is
likely to struggle to feed its growing population under climate change. Food production and food
security in the South have a wider inﬂuence on the national food supply. Food security lies at the
interface of biophysical, climatic, and socio-economic systems and demands a systemic approach
for evaluation. Using a biophysical crop model with an agent-based model of household food
systems, we capture the dynamics within and across multiple associated systems. We focus in the
South and use multidisciplinary tools to explore the trajectories of household food security under
climate change and population growth in the region.

The dissertation is organized into three research papers. Paper 1, entitled “A Largely Unsu-
pervised Domain-Independent Qualitative Data Extraction Approach for Empirical Agent-based
Model Development.”, focuses on exploring household food systems and identifying actors of
household food security and their behaviors. Chieﬂy, we aim to extract the information needed
to develop an ABM of household food systems. We apply largely automatic eﬃcient approaches
for information extraction from contextually rich qualitative ﬁeld narratives. Using a combination
of semantics and syntactic Natural Language Processing, we identify actors (agents) of household
food security, their properties, and actions and interactions responsible for household food supply.
The data extraction is primarily unsupervised and, apart from being eﬃcient, it controls manual
manipulation and bias introduction in the model development. We use the extracted information

for developing a contextual model of household food security. Finally, we subject the model to
stakeholder evaluation for credibility and validity.

Paper 2, entitled “Analyzing household vulnerability to food insecurity in rural southern Mali
- a coupled biophysical and social model approach.”, combines a biophysical process-based crop
model with an ABM of household food system to analyze household vulnerability to food insecurity
in southern Mali. We use the Systems Approach to Land Use Sustainability (SALUS) as the crop
model to simulate the cultivation of maize, millet, and sorghum in the region. While SALUS
provides information on food production, ABM simulates interactions for food access. We measure
household vulnerability using Food Security Vulnerability Index (FSVI), a coping-based index,
that evaluates households’ vulnerability to food insecurity by assessing the mechanisms used by
the households to address household food scarcity. Running a business-as-usual scenario, deﬁned
by low access to input, high population growth, and a high emission climate change scenario
at Representative Concentration Pathway (RCP) 8.5 level, we ﬁnd that maize and sorghum lose
their productivity signiﬁcantly under future climate change. Besides, the region sees a signiﬁcant
increase in its food insecure population. Around 80% of the regional households are at risk of food
insecurity by mid-century.

In paper 3, entitled “Global or Local? Eﬀects of policy interventions on household food security
in Koutiala, southern Mali: A coupled biophysical and social systems approach.”, we evaluate the
eﬀectiveness of selected global and local level policy intervention in promoting household food
security in the South. The model recommends local level interventions that include improved access
to input and lower population growth for a prompt and signiﬁcant increase in food security in the
region. More than 60% of the regional households could be food secure under the combinations
of the interventions. However, the global level intervention that consists of lowering emission
to RCP4.5 level does not have signiﬁcant impacts on local household food production and food
security.

Copyright by
RAJIV PAUDEL
2020

ACKNOWLEDGEMENTS

I would like to express my deepest gratitude to my major advisor Dr. Arika Ligmann-Zielinska
for giving me the opportunity and providing me guidance throughout the doctoral program. The
completion of my dissertation would not have been possible without her constant support and
mentorship. I would also like to express my deepest appreciation to my guidance committee, Dr.
Bruno Basso, Dr. Laura Schmitt Olabisi, and Dr. Leo Zulu for oﬀering encouragement, guidance,
and disciplinary insights throughout the program. I am very grateful to GEO faculty and staﬀ,
especially Dr. Alan Arbogast and Dr. Raechel Portelli, for their invaluable guidance and support. I
must also thank Ms. Sharon Ruggles, who helped me navigate the complex administrative processes
at MSU. Special thanks to Dr. Amadou Sidibé and Ms. Kadiatou Touré for supporting me with
the ﬁeld works in Mali. Also, I am extremely grateful to Mr. Brian Baer and the Basso lab for the
support they provided me with the crop model.

I should not forget to thank many fellow graduate students with whom I shared this journey.
Special thanks to BJ Baule, Jonah White, and Naﬁseh Haghtalab for helping me get through the
sedentary basement life. I am blessed to have an incredible community of friends, especially Dr.
Ujjwal Karki, Dr. Shital Poudyal, Anurag Dawadi, Dr. Udita Sanga, and Subechchya Sharma, who
oﬀered support during this journey.

Finally, I would like to express my deepest and sincere gratitude to my family for their continuous
and unparalleled love, help, and support. I cannot thank my wife, Sunita, enough for the sacriﬁces
she had to make for me to succeed. I want to thank my parents, whose love and guidance are with
me in whatever I pursue. I would also like to thank my brother Roshan Paudel and sister-in-law
Dr. Misti Ames Paudel for their relentless support and guidance. Last but not least, I would like to
thank Yuvan and Trisha for cheering me up every single day.

v

TABLE OF CONTENTS

.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

1.1 Background .
.

LIST OF TABLES .
LIST OF FIGURES .
CHAPTER 1

.
.
.
.
INTRODUCTION .
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
ix
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.

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

1
1
7

. .
.
.

BIBLIOGRAPHY .
CHAPTER 2 A LARGELY UNSUPERVISED DOMAIN-INDEPENDENT QUALITA-
TIVE DATA EXTRACTION APPROACH FOR EMPIRICAL AGENT-
BASED MODEL DEVELOPMENT.

2.1
Introduction .
2.2 Background .

2.3 Using NLP for largely unsupervised ABM development . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . 10
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 The use of qualitative data for ABM conceptualization . . . . . . . . . . . 14
2.2.2
Formulating ABM for model development . . . . . . . . . . . . . . . . . . 16
. 18
2.3.1 Unsupervised data processing and extraction . . . . . . . . . . . . . . . . . 20
2.3.1.1 Data preprocessing (cleaning and normalization) . . . . . . . . . 20
2.3.1.2 Data volume reduction . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1.3
Tagging and Information Extraction . . . . . . . . . . . . . . . . 21
Supervised contextualization and evaluation . . . . . . . . . . . . . . . . . 23

2.3.2

.
.

.
.

.
.

.
.

2.4 Development and evaluation of ABM of household food security using the pro-

.

.

.

.

.

.

.
.

.
.

.
.

.
.

posed framework .

APPENDICES .
APPENDIX A

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Limitations and future direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.6 Conclusion .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
.

. .
.
.
HOUSEHOLD-LEVEL ACTION AS DESCRIBED BY ELDERLY
MALES .
. 34
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
INDIVIDUAL ACTIONS AS DESCRIBED BY ELDERLY MALES 35
. 36
AN EXCERPT FROM THE FIELD NARRATIVES . . . . . . . .
. 37
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

APPENDIX B
APPENDIX C
BIBLIOGRAPHY .
CHAPTER 3 ANALYZING HOUSEHOLD VULNERABILITY TO FOOD INSECU-
RITY IN RURAL SOUTHERN MALI - A COUPLED BIOPHYSICAL
AND SOCIAL MODEL APPROACH.

. . . . . . . . . . . . . . . . . . . . 46
3.1
Introduction and Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Measuring household vulnerability to food insecurity . . . . . . . . . . . . . . . . 48
3.2.1
. . . . . . . . . . . . . . . . . 48
3.2.2 Measuring vulnerability to food insecurity . . . . . . . . . . . . . . . . . . 50
3.2.3 Coupled biophysical and social food security analysis . . . . . . . . . . . . 53
3.2.3.1 Crop models for estimating food production . . . . . . . . . . . . 54

Food security: deﬁnition and measurement

.

.

.

.

.

.

.

vi

3.3 Results and Discussions .
3.4 Limitations and future development
3.5 Conclusion .
.

3.2.3.2
3.2.3.3 Model parameterization, setup, and evaluation:

Integrating ABM for food security analysis . . . . . . . . . . . . 56
. . . . . . . . . . 58
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
. . . . . . . . . . . . . . . . . . . . . . . . . 69
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
.

.
.
PICTORIAL CONCEPTUAL DIAGRAM USED FOR STAKE-
HOLDER EVALUATION . . . . . . . . . . . . . . . . . . . . . . . 74
ABM DESCRIPTION IN ODD FORMAT . . . . . . . . . . . . . . 75
. 85
.

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

. .
.
.

.
.

.
.

.
.

APPENDICES .
APPENDIX D

.

.

.

.

.

.

.

.

.

.
.
.

.
.
.

.
.
.

4.3.2

4.3.1

Introduction .

APPENDIX E
BIBLIOGRAPHY .
CHAPTER 4 GLOBAL OR LOCAL? EFFECTS OF POLICY INTERVENTIONS ON
HOUSEHOLD FOOD SECURITY IN KOUTIALA, SOUTHERN MALI:
A COUPLED BIOPHYSICAL AND SOCIAL SYSTEMS APPROACH.

. .
4.1
4.2 The coupled model
4.3 Scenario analysis

. . 94
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Intervention at the global level
. . . . . . . . . . . . . . . . . . . . . . . . 99
4.3.1.1 Climate change mitigation (CCM) . . . . . . . . . . . . . . . . . 99
Interventions at the local level
. . . . . . . . . . . . . . . . . . . . . . . . 99
Lower population growth (LPG) . . . . . . . . . . . . . . . . . . 99
4.3.2.1
4.3.2.2
Improved access to agricultural inputs . . . . . . . . . . . . . . . 100
. 101
Intervention at global level
. . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.4.1.1 Climate change mitigation . . . . . . . . . . . . . . . . . . . . . 102
Interventions at local/regional levels . . . . . . . . . . . . . . . . . . . . . 102
Lower Population Growth . . . . . . . . . . . . . . . . . . . . . 102
4.4.2.1
Improved access to agricultural inputs . . . . . . . . . . . . . . . 104
4.4.2.2
4.5 Limitations
.
. .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
.
4.6 Conclusion .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
BIBLIOGRAPHY .
.
.
.
CHAPTER 5 CONCLUDING REMARKS . . . . . . . . . . . . . . . . . . . . . . . . . . 117

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

4.4 Results and discussion .

. .
.
.
.
.

. .
.
.

4.4.2

4.4.1

. .

.

.

.

vii

LIST OF TABLES

Table 2.1: An excerpt of action and decision verbs . . . . . . . . . . . . . . . . . . . . . . 25

Table 3.1: Food Security Vulnerability Index (FSVI) . . . . . . . . . . . . . . . . . . . . . 53

Table 3.2: Koutiala crop yields (reported) . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Table 3.3: SALUS Parameters .

Table 3.4: ABM Parameters .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

Table E.1: State variables used in the ABM . . . . . . . . . . . . . . . . . . . . . . . . . . 76

Table E.2: SALUS Parameters .

Table E.3: ABM Parameters .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Table 4.1: Combinations of intervention scenarios

. . . . . . . . . . . . . . . . . . . . . . 101

viii

LIST OF FIGURES

Figure 1.1: Bio-Climatic Zones of Mali . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 1.2: Koutiala: the study area . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.

2

3

Figure 2.1: Largely unsupervised information extraction for ABM development . . . . . . . 22

Figure 2.2: Mental Model of household food security . . . . . . . . . . . . . . . . . . . .

. 24

Figure 2.3: Candidate agents before and after using the external database . . . . . . . . . . 26

Figure 2.4: UML class diagram of agents of household food security . . . . . . . . . . . . 27

Figure 2.5: Conceptual model of household food security in Koutiala, Mali

. . . . . . . .

. 29

Figure A.1: Household-level action as described by elderly males

. . . . . . . . . . . . . . 34

Figure B.1: Individual actions as described by elderly males . . . . . . . . . . . . . . . .

. 35

Figure C.1: An excerpt from the ﬁeld narratives . . . . . . . . . . . . . . . . . . . . . . . . 36

Figure 3.1: The pillars of food security . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 49

Figure 3.2: Space of Vulnerability .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

Figure 3.3: Household responses to food shortages . . . . . . . . . . . . . . . . . . . . .

. 52

Figure 3.4: An integrated systems approach to food security analysis

. . . . . . . . . . . . 54

Figure 3.5: A systems approach in SALUS model . . . . . . . . . . . . . . . . . . . . . . . 56

Figure 3.6: Structure of Crop-ABM coupled Model . . . . . . . . . . . . . . . . . . . . .

. 60

Figure 3.7: Future climate under RCP 8.5 scenario . . . . . . . . . . . . . . . . . . . . . . 64

Figure 3.8: SALUS predicted yields (kg/ha) for business-as-usual scenario . . . . . . . . . 65

Figure 3.9: Spatial distribution of crop productivity in Koutiala . . . . . . . . . . . . . . . 66

Figure 3.10: Maize productivity decline from 2010 to 2050 . . . . . . . . . . . . . . . . . . 68

Figure 3.11: Regional food production vs. demand . . . . . . . . . . . . . . . . . . . . . .

. 68

ix

Figure 3.12: Trajectory of food security . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Figure 3.13: Spatial distribution of vulnerability to food insecurity . . . . . . . . . . . . .

. 70

Figure 3.14: Vulnerability to food insecurity by household types

. . . . . . . . . . . . . . . 71

Figure D.1: Pictorial conceptual diagram used for stakeholder evaluation . . . . . . . . .

. 74

Figure E.1: ABM ﬂow diagram .

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

. 78

Figure 4.1: Climate predictions under BAU (RCP 8.5) and CCM (RCP 4.5) . . . . . . . . . 103

Figure 4.2: Crop yield (Kg/ha) under BAU (RCP 8.5) and CCM (RCP 4.5)

. . . . . . . . . 105

Figure 4.3: Household food security under diﬀerent scenarios . . . . . . . . . . . . . . . . 106

Figure 4.4: Crop yield (Kg/ha) under low and high access to input . . . . . . . . . . . . . . 108

Figure 4.5: Food insecurity by household types . . . . . . . . . . . . . . . . . . . . . . . . 110

x

CHAPTER 1

INTRODUCTION

1.1 Background

Unlike many countries in the world, Mali could not seize the beneﬁts of the Green Revolution.
The development and application of improved technology and farm management, that were nec-
essary for a successful Green Revolution, were severely inadequate (Otsuka and Muraoka 2017).
As a result, farmers in the region could not intensify their agriculture (Brown, Hintermann, and
Higgins 2009), and their productivity remained stagnant for decades (Djurfeldt et al. 2005). The
current farming practices in Mali are still very traditional. Conventional hand tools, low-yielding
cultivars, and poor application of inputs are quite ubiquitous (Brown, Hintermann, and Higgins
2009). Consequently, its crop productivity remains amongst the lowest in the world. The future of
Malian agriculture does not look promising either. The country is expected to be profoundly im-
pacted by climate change. Climate models used by the Intergovernmental Panel on Climate Change
(IPCC) unanimously predict a warmer climate in Mali (Traore et al. 2013). High temperature
has detrimental eﬀects on crop productivity (Asseng et al. 2017; Basso, Kendall, and Hyndman
2013). As a result, the region may see a further decline in its crop production. With some models
predicting increased rainfall, Mali can expect to see improved production in the future (Patricola
and Cook 2010; Sultan and Gaetani 2016). However, the increased frequency of erratic rainfall
and a warmer climate expected under climate change may nullify such beneﬁts (Sivakumar 1988).
Mali has four broad bio-climatic zones (Figure 1). The northern half of Mali is highly arid
and unsuitable for agriculture. People living in the North are generally nomadic; their livelihood
depends on trans-Saharan trades. They rely entirely on food imports. The country becomes wetter
and more productive in the South. The Sudano-Guinean zone receives more than 1000mm of
annual rainfall (Traore et al. 2013). Due to the relative abundance of fertile land, people in the
South are more involved in agricultural activities. The region is the primary exporter of food in

1

Figure 1.1: Bio-Climatic Zones of Mali. Source: https://www.climatelinks.org/

the nation. However, the regional farmers are predominantly poor and into subsistence farming.
Intensiﬁcation in the region is highly limited due to inadequate access to technology and resources.
Furthermore, soil degradation is omnipresent (Adger et al. 2001) in the region.

Mali is experiencing rapid population growth (Farvacque-Vitkovic et al. 2007; Dijk, Bruijn, and
Beek 2004). With the current state of farming in the South, the country will face diﬃculties feeding
its growing population. The South has a daunting task of feeding its people while maintaining
food supply to the nation. Household food security here has broader implications for the national

2

Figure 1.2: Koutiala: the study area

food supply. The purpose of this study is to explore household vulnerability to food insecurity
from climate change and population growth in rural southern Mali.
In this study, we explore
the trajectories of household food security in Koutiala, a region in rural southern Mali (Figure
2). Koutiala makes a signiﬁcant contribution to the national food supply and is regarded as the
breadbasket of the country (Bingen 1998). Using a systems approach, we look at household
food systems in the region. Additionally, combining multidisciplinary approaches, we explore
their trajectories under climate change and population growth and evaluate the eﬃcacies of selected
policy interventions in promoting food security in the region. Household food security in rural Mali
mainly relies on household food production and access. Food production responds to biophysical
and climatic factors and farm management. Food access, on the other hand, is aﬀected by resource
availability and social norms and practices. In other words, food security lies at the intersection

3

of biophysical, climatic, and socio-economic systems and, therefore, demands a systems approach
for its analysis (Jones et al. 2017). In this study, we integrate a bio-physical crop model with an
agent-based model (ABM) developed using ﬁeld narratives of household food systems to evaluate
vulnerability to food insecurity. The crop model incorporates the dynamics within biophysical
and climatic systems for estimating household food production while the ABM explores social
interactions contributing to household food access.

Vulnerability to food insecurity is governed by complex interactions between the associated
dynamic systems. Vulnerability measurement needs to capture the dynamics of the complex systems
(Krishnamurthy, Lewis, and Choularton 2014). In this study, we measure household vulnerabilities
using a coping-based Food Security Vulnerability Index (FSVI). FSVI evaluates vulnerability by
exploring household coping mechanisms. Households start to cope when faced with food scarcity
at home. The options they choose generally reﬂect their current severity and future vulnerability
(D. Maxwell, Caldwell, and Langworthy 2008; D. G. Maxwell 1996; D. Maxwell, Watkins, et al.
2003). By incorporating FSVI into the ABM, we are able to capture the dynamics and interactions
occurring within and across the associated systems.

Although food security is deﬁned in terms of food availability, access, utilization, and stability,
its measurement is often limited to food production (availability) or access.
In this study, we
try to capture all four dimensions for measuring food security in our analysis. We evaluate
food availability and access through analyzing household food production, social interactions, and
resource availability. Similarly, we measure utilization using household calorie suﬃciency. The
simulations we perform provide information on the temporal stability of household food security.
This dissertation is organized into three papers, each with speciﬁed research questions. The ﬁrst
paper delves into household food systems in Koutiala. The second paper explores the trajectories of
household food security under climate change and population growth, and the third paper evaluates
selected policy intervention. The following research questions guide this study:

Research Question 1: What are the actors and factors of household food security? How do

household actions and interactions aﬀect their food security?

4

Research Question 2: How do dynamic interactions across socio-economic, climatic, and
bio-physical systems spatially and temporally aﬀect household vulnerabilities to food insecurity?
Research Question 3: What are the global and local level management recommendations for

eﬀective policy interventions?

The development of ABM needs information on agents, their properties, actions, and inter-
actions. In Paper one, we use qualitative ﬁeld narratives to identify the actors (agents) of food
security, their attributes, and actions/interactions responsible for maintaining food security at the
households in Koutiala. Qualitative data, primarily narratives, is an excellent source of informa-
tion. The narratives contain rich contextual information for systems representation essential for
model development. However, qualitative data processing and extraction is complex, and time and
resource consuming. ABM developers often avoid extracting model components from qualitative
data. Models are usually developed ad-hoc, based on modelers’ subjective interpretation of target
systems. The resulting models often have questionable quality and reliability. In this paper, we
use Natural Language Processing tools to extract ABM components from qualitative narratives.
The process is largely unsupervised and, therefore, highly eﬃcient. Further, it limits manual
manipulation, reducing opportunities for human bias introduction during model development.

In paper two, we integrate a biophysical crop model with an ABM of household food systems
to analyze household vulnerability to food insecurity. We simulate a ‘business-as-usual’ scenario
based on the current level of farm management, population growth, and expected climate change.
The ABM receives information on household crop production from the crop model. Household
food suﬃciency is determined by calculating food availability and demand at the household level.
The model treats food suﬃcient households as food secure. Food insuﬃcient households look for
coping alternatives, and the ABM provides FSVI indices to the households based on the severity
indicated by alternatives used by the households.

In paper three, we evaluate the eﬀects of global and local level policy interventions on house-
hold food security in the study area. We simulate alternative futures combining diﬀerent farm
management, population growth, and climate change interventions. We explore the trajectories

5

of household food security under combinations of interventions. Additionally, we compare the
outputs with the ‘business-as-usual’ scenario from paper two to assess the eﬀectiveness of the
interventions.

6

BIBLIOGRAPHY

7

BIBLIOGRAPHY

Adger, W Neil et al. (2001). “Advancing a political ecology of global environmental discourses”.

In: Development and change 32.4, pp. 681–715. issn: 0012-155X.

Asseng, S. et al. (2017). “Hot spots of wheat yield decline with rising temperatures”. In: Glob
Chang Biol 23.6, pp. 2464–2472. issn: 1365-2486 (Electronic) 1354-1013 (Linking). doi:
10.1111/gcb.13530. url: https://www.ncbi.nlm.nih.gov/pubmed/27860004.

Basso, Bruno, Anthony D Kendall, and David W Hyndman (2013). “The future of agriculture over
the Ogallala Aquifer: Solutions to grow crops more eﬃciently with limited water”. In: Earth’s
Future 1.1, pp. 39–41. issn: 2328-4277.

Bingen, R James (1998). “Cotton, democracy and development in Mali”. In: The Journal of Modern

African Studies 36.2, pp. 265–285. issn: 0022-278X.

Brown, Molly E, Beat Hintermann, and Nathaniel Higgins (2009). “Markets, climate change, and
food security in West Africa”. In: Environ. Sci. Technol. 43.21, pp. 8016–8020. issn: 0013-
936X.

Dijk, Han van, Mirjam de Bruijn, and Wouter van Beek (2004). “Pathways to mitigate climate
variability and climate change in Mali: The districts of Douentza and Koutiala compared”. In:
The Impact of Climate Change on Drylands. Springer, pp. 173–206.

Djurfeldt, Göran et al. (2005). The African food crisis: lessons from the Asian Green Revolution.

Cabi. isbn: 1845930975.

Farvacque-Vitkovic, Catherine et al. (2007). “Development of the cities of Mali: Challenges and

priorities”. In:

Jones, J. W. et al. (2017). “Brief history of agricultural systems modeling”. In: Agric Syst 155,
pp. 240–254. issn: 0308-521X (Print) 0308-521X (Linking). doi: 10.1016/j.agsy.2016.
05.014. url: https://www.ncbi.nlm.nih.gov/pubmed/28701816.

Krishnamurthy, PK, K Lewis, and RJ Choularton (2014). “A methodological framework for rapidly
assessing the impacts of climate risk on national-level food security through a vulnerability
index”. In: Global Environmental Change 25, pp. 121–132. issn: 0959-3780.

Maxwell, Daniel G (1996). “Measuring food insecurity: the frequency and severity of “coping

strategies””. In: Food policy 21.3, pp. 291–303. issn: 0306-9192.

Maxwell, Daniel, Richard Caldwell, and Mark Langworthy (2008). “Measuring food insecurity:
Can an indicator based on localized coping behaviors be used to compare across contexts?” In:
Food Policy 33.6, pp. 533–540. issn: 0306-9192.

8

Maxwell, Daniel, Ben Watkins, et al. (2003). “The coping strategies index: A tool for rapidly
measuring food security and the impact of food aid programs in emergencies”. In: Nairobi:
CARE Eastern and Central Africa Regional Management Unit and the World Food Programme
Vulnerability Assessment and Mapping Unit.

Otsuka, Keijiro and Rie Muraoka (2017). “A green revolution for sub-Saharan Africa: Past failures
and future prospects”. In: Journal of African Economies 26.suppl1, pp. i73–i98. issn: 0963-
8024.

Patricola, Christina M and Kerry H Cook (2010). “Northern African climate at the end of the
twenty-ﬁrst century: an integrated application of regional and global climate models”. In:
Climate Dynamics 35.1, pp. 193–212. issn: 0930-7575. url: https://link.springer.
com/content/pdf/10.1007%2Fs00382-009-0623-7.pdf.

Sivakumar, MVK (1988). “Predicting rainy season potential from the onset of rains in Southern
Sahelian and Sudanian climatic zones of West Africa”. In: Agricultural and Forest Meteorology
42.4, pp. 295–305. issn: 0168-1923.

Sultan, B. and M. Gaetani (2016). “Agriculture in West Africa in the Twenty-First Century: Climate
Change and Impacts Scenarios, and Potential for Adaptation”. In: Front Plant Sci 7, p. 1262.
issn: 1664-462X (Print) 1664-462X (Linking). doi: 10 . 3389 / fpls . 2016 . 01262. url:
https://www.ncbi.nlm.nih.gov/pubmed/27625660.

Traore, Bouba et al. (2013). “Eﬀects of climate variability and climate change on crop production

in southern Mali”. In: European Journal of Agronomy 49, pp. 115–125. issn: 1161-0301.

9

CHAPTER 2

A LARGELY UNSUPERVISED DOMAIN-INDEPENDENT QUALITATIVE DATA

EXTRACTION APPROACH FOR EMPIRICAL AGENT-BASED MODEL

DEVELOPMENT.

Agent-based model (ABM) development needs information on system components and interactions.
Qualitative narratives contain contextually rich system information immensely useful for ABM
conceptualization. However, traditional qualitative data extraction is usually manual, complex, and
time and resource consuming. As a result, modelers often shy away from using qualitative data or
follow unorthodox ad-hoc approaches for ABM development. That could yield models diﬃcult to
validate and replicate. Besides, manual data extraction is generally bias-prone and may produce
models of questionable quality and reliability.
In this study, we present a largely-unsupervised
qualitative data extraction for ABM development. Using a combination of semantic and syntactic
Natural Language Processing tools, our methodology extracts information on system agents, their
attributes, and actions and interactions. In addition to expediting information extraction for ABM,
the largely-unsupervised approach also minimizes biases arising from modelers’ preconceptions
about target systems. To make it usable on large unstructured datasets, we also introduce multiple
automatic and manual noise-reduction stages. We demonstrate the approach by developing a
conceptual ABM of household food security in rural Mali. The data for the model contains a large
set of unstructured qualitative ﬁeld narratives. The data extraction is swift and mainly automatic
and devoid of human manipulation. We contextualize the model manually using the extracted
information. For added credibility and validity, we also put the conceptual model to stakeholder
evaluation.

2.1 Introduction

Qualitative data provides thick contextual information (Miles 1979; Rich and Ginsburg 1999;
Watkins 2012) that can support reliable quantitative socioecological systems (SES) model devel-
opment. Qualitative data analysis explores systems components, their complex relationships and

10

behavior (Kemp-Benedict 2004; Rich and Ginsburg 1999; Watkins 2012), and provides a struc-
tured framework that can guide the formulation of quantitative SES models (Ackermann, Eden,
and Williams 1997; Coyle 2000; Forbus and Falkenhainer 1990; Wolstenholme 1999). However,
qualitative research is complex, time, and resource-consuming (Miles 1979; Watkins 2012). Data
analysis usually involves keyword-based data extraction and evaluation that requires multiple coders
to reduce biases. Moreover, model development using qualitative data requires multiple, lengthy,
and expensive interactions with stakeholders (Polhill, Sutherland, and Gotts 2010), which adds
to its inconvenience. Consequently, quantitative modelers often shy away from using qualitative
data for their model development. Often, modelers either entirely skip qualitative data analysis or
use unorthodox approaches for framework development. This usually fails to capture the complex
dynamics of target systems and produces inaccurate and unreliable outputs (Landrum and Garza
2015).

The development in the information technology sector has substantially increased access to
qualitative data over the past few decades. Harvesting large, credible data is crucial for reliable
model development.
Increased access to voluminous data presents a challenge, as well as an
opportunity for model developers (B. Runck 2018). However, qualitative data analysis has always
been a hard nut to crack for SES modelers. Most of the existing qualitative data analyses are
highly supervised (i.e., done mainly by humans) and, hence, bias-prone and ineﬃcient for large
datasets. An approach that can eﬃciently analyze and extract qualitative data for developing
credible, reliable and valid SES models is urgently required. This study presents a methodology
that uses an eﬃcient largely unsupervised qualitative data extraction method for credible Agent-
based Model (ABM) development using Natural Language Processing (NLP) toolkits. As the
process of information extraction is unsupervised and does not require manual interventions, it
reduces potential for subjectivity and biases in the model development. ABM, one of the widely
used SES models, requires information on socioecological agents, their attributes, and interactions
for its development.

The paper is organized as follows. Section 2.2 describes the importance of qualitative data

11

and outlines some of the existing approaches for translating qualitative data to empirical ABM.
In section 2.3, we present a largely unsupervised domain-independent qualitative data processing
and extraction using NLP for ABM development. We use the approach to develope an ABM of
household food security in Koutiala Mali, which is reported in section 2.4. Section 2.5 outlines
some of the limitations and recommendations for future research. Finally, Section 2.6 concludes
our paper.

2.2 Background

ABM is useful for understanding phenomena that emerge from nonlinear interactions of au-
tonomous and heterogeneous constituents of socioecological systems (An et al. 2005). ABM is
a bottom-up approach; interactions occurring at micro-level produce complex and emergent phe-
nomena at a macro (higher) level. To simulate such behaviors, ABM requires micro-scale data.
However, obtaining data at the required scale is not always feasible. ABM had enjoyed a prosperous
era of nonempirical models that showcased the model’s eﬀectiveness in exploring complex sys-
tems (Janssen and Ostrom 2006). Using synthetic data, the nonempirical abstract models simulate
simple system behaviors to generate and test complex systems theories (O’Sullivan et al. 2016).
Models like Conway’s Game of Life and Shelling’s Segregation are some of the ﬁrst and renowned
nonempirical ABMs, that despite lacking real-world data, simulate simple behaviors to adequately
explain system complexity and emergence (Axelrod 1997).

As the access to micro-level data becomes easier, modelers are increasingly using empirical
data for more realistic system representation and simulation (Janssen and Ostrom 2006; O’Sullivan
et al. 2016; Robinson et al. 2007). These empirical models use quantitative as well as qualitative
data to evaluate systems theories. Quantitative data is primarily useful for parameterizing and
running simulation. Additionally, quantitative model outputs are also used for model veriﬁcation
and validation. Qualitative data, on the other hand, ﬁnds its uses at various stages of the model cycle
(Seidl 2014). Apart from the routine tasks of identifying systems constituents and behaviors for
model development, qualitative data also supports the model structure and output representations

12

(Grimm, U. Berger, et al. 2010; Ligmann-Zielinska, Siebers, et al. 2020; B. Müller et al. 2014).
Qualitative model representations facilitate communication for learning and model evaluation and
replication.

Several researchers believe that the qualitative/quantitative dichotomy is artiﬁcial and inaccurate
(Gorard 2002; Kemp-Benedict 2004; Rich and Ginsburg 1999). They argue that both approaches
have similar philosophies - both aim to advance scientiﬁc knowledge through veriﬁed tools and
methods. Scientiﬁc rigor and integrity of theoretical knowledge are central to both paradigms
(Gorard 2002; Rich and Ginsburg 1999). Although their inductive and deductive approaches are
diﬀerent, data used for the analyses are predominantly nonbinary. Numbers can be textualized,
and textual data can be counted (Gorard 2002). However, pro-qualitative researchers challenge
such ideas. They believe that the qualitative and quantitative approaches have distinct philosophies,
characteristics, and goals (Leininger 1992). They ﬁnd quantitative research to be highly rigid
and context-deprived (Rich and Ginsburg 1999; Watkins 2012). On the contrary, the quantitative
domain ﬁnds qualitative research vague, bias-prone, and without scientiﬁc rigor (Gorard 2002).

Such ﬁnger-pointing and blame game is omnipresent in the scientiﬁc community, especially in
the post-Kuhn era (Niglas 2000). Moreover, there is a broader dissatisfaction amongst qualitative
researchers as quantitative research is deemed more important in the scientiﬁc community and is
prioritized for funding and publication. Improperly mixed methods create more issues than they
typically address (Leininger 1992). However, SES models can naturally embrace both approaches.
ABM, speciﬁcally, uses both deductive and inductive approaches to generate and test systems
theories. Epstein (2006) promotes such ABM quality as a generative social science, while Axelrod
(1997) calls it a third way of doing science.

Mixed-methods are gradually getting traction in the wider modeling community (Forbus and
Falkenhainer 1990; Gorard 2002; Kemp-Benedict 2004; J. D. A. Millington and Wainwright 2016;
Wolstenholme 1999). Speciﬁcally, qualitative analysis is used at the beginning of research to
explore system dynamics that inform the development of structured frameworks for quantitative
model formulation (Coyle 2000; Forbus and Falkenhainer 1990). Without such frameworks,

13

quantitative modelers often get bogged down by systems complexities. Besides, qualitative research
also lays down model assumptions explicitly that solidify the credibility and reliability of the
resulting models (Ackermann, Eden, and Williams 1997; Coyle 2000; Forbus and Falkenhainer
1990; Wolstenholme 1999). As a result, such frameworks are frequent in the quantitative models
across various disciplines including engineering (Forbus and Falkenhainer 1990), transportation
management (Ackermann, Eden, and Williams 1997), public health (Rich and Ginsburg 1999),
geography (Ligmann-Zielinska and Jankowski 2010), and other socioecological systems (Schmitt
Olabisi et al. 2010).

Some quantitative models have predeﬁned structures for model development. System Dynamics,
for instance, uses Causal Loop Diagrams as qualitative tools (A. Ford and F. A. Ford 1999). Causal
Loop Diagrams elucidate systems components, their interrelationships, and feedbacks that can be
used for learning and developing quantitative System Dynamics models. ABM, however, does not
have a predeﬁned structure for model development; models are mostly based on ad-hoc structures.
Structural validation is crucial for ABM (Heath, Ciarallo, and Hill 2012). However, such ad-hoc
development is problematic for model validation and replication (Grimm, Augusiak, et al. 2014).
Besides, similar systems can have diﬀerent structures that make learning through models extremely
diﬃcult (O’Sullivan et al. 2016). Hence, there is a need for a standard structure for the development
of credible, reliable, and educational ABMs.

2.2.1 The use of qualitative data for ABM conceptualization

Social and cognitive theories often form the basis for translating qualitative data to empirical
ABM (Becu et al. 2005). Since gathering and organizing micro-level qualitative data is diﬃcult,
using theories streamlines data management and analysis for model development. Theories like
Belief-Desire-Intension (Balke and Gilbert 2014), Theory of Planned Behavior (Edwards-Jones
2006), Utility Maximization (Doscher et al. 2014), and Consumat (Janssen and Jager 1999) are
typical for deﬁning behavioral rules in ABM. Since social behavior is complex and challenging to
comprehend, the use of social and cognition theories helps in determining the expected behavior

14

of the system.

Another school of thought bases model development on stakeholder cognition. Rather than
relying mainly on social theories, this approach focuses on extracting empirical information about
system components and behaviors. Nicolas Becu et al. (2003) identiﬁed two approaches for
using stakeholder information in model building – a modeling view and a transfer view. The
modeling view requires direct involvement of stakeholders in model development. Participatory
or companion modeling, as well as role-playing games, are some of the conventional approaches
of eliciting stakeholder knowledge for model development (Robinson et al. 2007). Stakeholders
usually develop model structures in real-time, while some modelers prefer to process stakeholder’s
information after the discussions. For instance, Bharwani (2006) employs computer technologies to
post-process stakeholder responses to develop a rule-induction algorithm for her ABM. Stakeholders
are assumed to be the experts of their systems, and using their knowledge in model building
makes model valid and reliable. However, this approach also has its limitations. Working with
stakeholders can be complicated and may require specialized skills. Stakeholder bias (Seidl 2014)
is also a common problem. Relying entirely on stakeholders for model development, on the other
hand, can completely hijack the initially intended model objectives. As social systems are ‘socially
constructed,’ conﬂicting views on the systems are also likely to occur. Collating the conﬂicting and
contrasting views in one representation may not always be possible. That may require developing
multiple models of the same system to reﬂect diﬀerent perspectives (Nicolas Becu et al. 2003).
However, Participatory Modeling (Voinov and Bousquet 2010) has been continually developed to
acknowledge and minimize various biases that may be introduced by stakeholders involvement in
model building (Sterling et al. 2019; Voinov, Kolagani, et al. 2016).

Stakeholder involvement is not always feasible; for instance, when modeling remote or historical
events. In such cases, a transfer view is preferred. The transfer view requires textual information,
usually in the form of narratives or interviews. Modelers use knowledge engineers and information
elicitation tools for information extraction. Thus, translating empirical textual data into agent
architecture is diﬃcult and requires concrete algorithms and structures (Seidl 2014). Edmonds

15

(2015) formulated a context, scope, and narrative element structure for eliciting model elements
from textual data. Modelers ﬁrst explore the context of the narratives and then identify potential
context-speciﬁc scopes. Once context and scopes are identiﬁed, according to Edmonds (2015),
identifying narrative elements becomes easy.

Many ABM modelers have formulated structures for organizing qualitative data for model
development. For instance, Ghorbani, Schrauwen, and Dijkema (2013) used Ostrom’s Institu-
tional Analysis and Development (IAD) framework for managing qualitative data in their Modeling
Agent-system based on Institutions Analysis (MAIA). MAIA comprises ﬁve structures: collective,
constitutional, physical, operational, and evaluative. Information on agents is populated in collec-
tive structure while behavior rules and environment go in constitutional and physical structures,
respectively. A similar example can also be found in Gilbert and Terna (2000). They presented
an Environment-Rule-Agent structure for organizing qualitative data for ABM development. Like-
wise, the MameLuke framework developed by Huigen (2004) is also a highly popular approach. It
uses Action-in-Context (De Groot 1992) for translating real-life stories into computerized ABM.

2.2.2 Formulating ABM for model development

ABMs are essentially computer tools that have software agents and environments. The high-level
model descriptions elicited from qualitative data needs to be translated to low-level computer
languages. Frequently, modelers are not computer programmers. That often prompts modelers
to represent their models to programmers. The widely used ABM representations are either in
the form of natural languages (e.g., ODD (Grimm, Railsback, et al. 2020)), Graphics (e.g., UML
(Bersini 2012)), or formal language (e.g., pseudocodes, ontologies). B. Müller et al. (2014)
compared these approaches and found that UML representation has broader applicability in coding,
validating, communicating, and replicating models. UML is a standardized graphical representation
of software development (Bersini 2012). It has a set of well-deﬁned class and activity diagrams
that are eﬀective in representing the inner workings of ABMs (Collins et al. 2015). Although there
are other forms of graphical ABM representations like Petri Nets (Bakam et al. 2000), Conceptual

16

Model for Simulation (Heath, Ciarallo, and Hill 2012) and sequence and activity diagrams (Gilbert
2004), UML is natural in representing ABM. Due to which, Miller and Page (2009) expected UML
to be the default lingua franca of ABM. Recently, ABMs are increasingly being represented using
UML (Li et al. 2015; Serrano and Iglesias 2016). Furthermore, ABM speciﬁc UMLs like Agent
UML (AUML) (Bauer, J. P. Müller, and Odell 2001) that captures the adaptive nature of agents
and Agent Modeling Language (AML) (Trencansky and Cervenka 2005) are also being developed
(Subburaj and Urban 2018).

ABM has a close resemblance to Object-Oriented Programming (OOP) (An 2012; Dearden
In OOP, objects belong to classes; they have
and Wilson 2012; Gilbert 2004; Simoes 2012).
attributes as well as methods and relationships. Likewise, in ABM, agents have attributes and
actions/interactions. Code proﬁcient modelers often develop their ABMs using object-oriented
programs (T. Berger 2001; Standish 2008). OOP ﬂows naturally and is easy to understand (Bersini
2012; Torrens 2010); hence, translating ABM to OOP makes coding much easier (Crooks and
Heppenstall 2012). However, ABM is not necessarily an OOP. Many ABMs have been developed
using non-OOP NetLogo (Krejci et al. 2016; J. D. Millington 2012). NetLogo eﬀectively caters to
modelers who have an elementary coding experience (Tisue and Wilensky 2004).

UML is widely accepted for representing OOP (Bauer, J. P. Müller, and Odell 2001). Trans-
ferring narratives to an object-oriented framework and representing it through UML is useful for
communicating and developing ABM. Many software engineers are exploring ways to extract infor-
mation from qualitative data directly to object-oriented frameworks. Similarly, database designers
are also trying to auto-populate entity-relationship diagrams from textual information (Harmain
and Gaizauskas 2000). They primarily exploit the syntactic structure of sentences for information
extraction. The syntactic analysis usually treats the subject of a sentence as a class (an entity for a
database) and the main verb as a method (relationship for a database).

Software engineers have been exploring various supervised and unsupervised approaches for
In supervised approaches, syntactical patterns are deﬁned (Clark et al.
information extraction.
2012), and text is manually scanned for such patterns. In unsupervised information extraction, the

17

machine does the pattern matching. NLP (Loper and Bird 2002; Manning, Surdeanu, et al. 2014)
toolkits are increasingly used for unsupervised pattern matching and information extraction (Fraga
et al. 2017; Salloum et al. 2018). Tools like lemmatizing, tokenizing, stemming, and part of speech
tagging (Bird, Klein, and Loper 2009) are useful for syntactic information extraction. These tools
can normalize texts and identify subjects and main verbs from their sentences.

Supervised approaches are reliable but slow. On the contrary, the faster unsupervised approaches
are diﬃcult and prone to errors, mostly due to the word sense ambiguation (Al-Safadi 2009). A
purely syntactical analysis cannot capture the nuanced meaning of texts, which is often the culprit
of the problem. Recently, pattern matching also involves semantic analysis. External databases
of hierarchically-structured words (e.g., WordNet, VerbNet) (Navigli 2009), as well as machine
learning tools, are increasingly used for understanding semantics for reduced word sense ambiguity
(Husain and Khanum 2016; Orkphol and Yang 2019).

2.3 Using NLP for largely unsupervised ABM development

NLP is instrumental in eﬃciently analyzing, tagging, and extracting information from broad
textual data (Liddy 2001; Sun, Luo, and Chen 2017). The ability to convert highly unstructured
texts to structured information through predominantly unsupervised approaches is one of the main
advantages of NLP in qualitative data analysis. NLP eﬃciently analyzes intertextual relationships
using syntactic and semantics algorithms. Approaches like word co-occurrence statistics and
sentiment analysis (Nasukawa and Yi 2003) are highly useful for domain modeling (Clark et al.
2012), and for exploring contextual and behavioral information from textual data. Similarly,
its eﬃciency in pattern matching for information extraction is particularly important for model
development. Although being present for decades in object-oriented programming and database
development (Harris 1978; Lees 1970), NLP has a minimal footprint in ABM development.

The introduction of NLP in ABM development is very recent. Runk (B. Runck 2018; B. C.
Runck et al. 2019) used NLP to model human cognition through web embedding. Word embeddings
are contextually analyzed vector representations of texts. They place closely related texts next to

18

each other. Placing agents with similar worldviews together supports theorizing agent decision-
making. Although the approach helps in developing agent decision-making, it is not yet well
equipped for developing a comprehensive agent architecture.

Padilla, Shuttleworth, and O’Brien (2019) applied NLP in conjunction with machine learning to
create an ABM structure from unstructured textual data. Texts are translated to the agent-attribute-
rule framework. They deﬁned agents as nouns (e.g., person, place) that perform some actions, and
attributes as words that represent some variables. Similarly, the sentences that contain agents or
attributes together with action verbs are considered rules. The primary goal of their approach is to
create an ABM structure mainly for communicating the model to non-modelers.

As with machine learning approaches in general, the Padilla el al. approach requires a large
amount of training data. They used ten highly concise formally written ABM descriptions from
published journals as their training datasets. Another limitation, according to the researchers, is the
lack of precise distinctions between agents and attributes - attributes could also be nouns that might
confuse the machine. Using repeated training with large data is found to be eﬀective in increasing
the accuracy of the agent-attribute-rule detection. However, the model frequently resulted in under
and over predictions.

Our study proposes and tests a largely unsupervised domain-independent approach for de-
veloping ABM structures from unstructured, informal narratives using Python-based semantic
and syntactic NLP tools Figure 2.1. The method primarily uses syntactic NLP approaches for
information extraction directly to the widely accepted OOP framework (i.e., agents, attributes,
actions/interactions), and creates UML for its representation. Since the approach is not based
on machine learning, it does not require large training data. The semantic analysis is limited
to the use of external static datasets like WordNet (https://wordnet.princeton.edu/) and VerbNet
(https://verbs.colorado.edu/verbnet/). The process is mainly unsupervised and involves the follow-
ing:

• Unsupervised data processing and extraction

19

– Data preprocessing (cleaning and normalization)

– Data volume reduction

– Tagging and information extraction

• Supervised contextualization and evaluation

– UML/Model conceptualization

– Model evaluation

2.3.1 Unsupervised data processing and extraction

2.3.1.1 Data preprocessing (cleaning and normalization)

Qualitative data analysis is computationally expensive. Unstructured narratives often contain
redundant or inﬂected texts that can bog down NLP analysis. Hence, removing non-informative
contents from large textual data is highly recommended at the start of the analysis. NLP is well
equipped with stop words removal tools that can eﬀectively remove redundant texts. Similarly,
tools like stemming and lemmatizing are useful for normalizing texts to their base forms (Manning,
Raghavan, and Schütze 2010).

2.3.1.2 Data volume reduction

Data volume reduction can tremendously speed up NLP analyses. Traditional volume reduction
approaches usually contain highly supervised keyword-based methods. Data analysts use predeﬁned
keywords to select and extract sentences perceived to be relevant (Namey et al. 2008). Keyword
identiﬁcation generally requires a priori knowledge on the system and is often bias-prone. We
recommend a domain-independent unsupervised Term Frequency Inverse Document Frequency
(TFIDF) approach (Ramos 2003) that eliminates the requirement for manual keyword identiﬁcation.
The approach provides weightage to individual words based on their uniqueness and machine-
perceived importance. The TFIDF diﬀerentiates between important words and common words by

20

comparing their frequency in individual documents and across the entire body of texts. Sentences
that have high cumulative TFIDF scores are perceived to have higher importance. Given a document
collection D, a word w, and an individual document d  D, TFIDF can be deﬁned as,

, ∗ (||\ , )
where fw, d equals the number of times w appears in d, |D| is the size of the corpus, and fw, D

equals the number of documents in which w appears in D (Ramos 2003).

2.3.1.3 Tagging and Information Extraction

Once the preprocessed data is reduced, tagging for agents, attributes, and actions/interactions can
occur. We propose the following approaches for tagging agent architecture:

Candidate agents: Following the conventional approaches in database design and object-
oriented programming, we propose the subjects of sentences to be identiﬁed as candidate agents.
For instance, the farmer in ‘The farmer grows cotton’ can be a candidate agent. NLP has well-
developed tools like part-of-speech tagger and named-entity tagger that can be used to detect
subjects of sentences.

Candidate actions: The main verbs of sentences can become candidate actions. The main
verbs need to have candidate agents as the subject of the sentences. In the sentence ‘The farmer
grows cotton,’ the farmer is a candidate agent and the subject of the sentence; grows is the main
verb and, hence, a candidate action.

Candidate attributes: Attributes are properties inherent to the agents. Sentences containing
candidate agents as subjects and be or have verbs as their primary (non-auxiliary) verbs provide
attribute information. E.g. The farmer is a member of a cooperative. The farmer has 10 ha
of land. Additionally, the use of possessive words also indicates attributes. E.g., the cow in the
sentence ‘My cow is very small’ is an attribute of ‘I’.

Candidate interactions: Main verbs indicating relationships between two candidate agents are
identiﬁed as interactions. Hence the sentences containing two or more candidate agents provide
information on interactions. E.g., The Government trains the farmers.

21

Figure 2.1: Largely unsupervised information extraction for ABM development, TFIDF: Term
Frequency Inverse Document Frequency

22

Since the data tagging is strictly unsupervised, false positives are likely to occur. The algorithm
can overpredict agents, as the subjects of all the sentences are treated as candidate agents. In ABM,
however, agents are deﬁned as autonomous actors – they act and make decisions. We propose to
use a hard-coded list of action (e.g., eat, grow, walk) and decision (e.g., choose, decide, think)
verbs to ﬁlter agents from the list of candidate agents. Only the candidate agents that use both types
of verbs qualify as agents. Candidate agents that do not use both verbs are categorized as entities
that may be subjected to manual evaluation. Similarly, people use diﬀerent terminologies that are
semantically similar. We recommend using external databases like WordNet to group semantically
similar terminologies.

2.3.2 Supervised contextualization and evaluation

The unsupervised analysis reduces data volume and translates unstructured narratives to the agent-
action-attribute structure. However, since the process is unsupervised, noise can still percolate to
the outputs. Additionally, the outputs need to be contextualized. We suggest performing a series
of supervised output ﬁltration, followed by manual contextualization and validation.

The domain-independent unsupervised analysis extracts individual sentences that can some-
times be ambiguous or domain irrelevant. Hence the output should be ﬁltered based on the level
of ambiguity and domain relevancy. Once output ﬁltration is done, contextual structures can be
developed and validated with domain experts and stakeholders.

2.4 Development and evaluation of ABM of household food security using

the proposed framework

We applied the above approach to develop a structural ABM of household food security using
unstructured ﬁeld narratives. The data contains 42 semi-structured interviews collected from
diﬀerent members (young and old; male and female) of farm households in Koutiala, Southern Mali.
The interviews were initially conducted for developing mental models of household food security
in the region (Rivers III et al. 2017), see, for example, Figure 2.2. The mental model development

23

Figure 2.2: Mental Model of household food security (Rivers III et al. 2017)

24

Table 2.1: An excerpt of action and decision verbs

Decision Verbs Action Verbs
adhere
advise
approve
assess
choose
comply
consult
decide
determine
discourage
educate
encourage
expect
favor
follow
guide
instruct
learn
obey
oblige
plan
prefer
propose
provoke
recommend
refuse
regulate
reject
reprimand
suggest
withdraw

abandon
accelerate
accept
access
accompany
accord
achieve
acquaint
acquire
add
address
adjust
adopt
advertise
aﬀect
aﬀord
agree
aim
alert
allow
analyze
apply
appoint
appreciate
argue
arrange
arrive
assemble
assist
associate
attach

25

Figure 2.3: Candidate agents before and after using the external database

followed the lengthy conventional qualitative data analysis approach that used multiple coders and
keyword-based sentence extraction. That inspired us to develop a more eﬃcient alternative data
processing and extraction approach for ABM development.

For this study, we used Python 3.7 programming language (https://www.python.org/) along with
several NLP libraries (e.g., scikit-learn, nltk, spacy, textacy) to perform data reduction, tagging,
extraction, and structuration. We grouped the narratives by the member types (i.e., elder male,
younger male, elder female, and younger female) and analyzed the grouped narratives collectively.
After preprocessing the narratives using NLTK tools, we used the scikit-learn TﬁdfVectorizer to
reduce the volume of qualitative data. Textacy was primarily used for identifying candidate agents,
actions, and attributes. Additionally, Textacy extract (textacy.extract.semistructured_statements)
was useful in converting sentences to structured outputs. We manually ﬁltered the unsupervised
outputs based on their domain relevancy and ambiguity. The ﬁnal outputs were then visualized
and conceptualized using Gephi (https://gephi.org/) and Lucid Chart (https://www.lucidchart.com/)
platforms.

As expected, the unsupervised tagging overpredicted the agents. Subjects that do not make a
decision were also identiﬁed as candidate agents. We created an external database of action/decision
verbs (Table 2.1) that somewhat addressed the problem. There was more than a 60% reduction in
the number of agents after the external database was used (e.g., Figure 2.3). The initially identiﬁed
candidates, such as porridge, food, cereal, or farm, were discarded by the ﬁltration process. We
also used external WordNet database to group semantically similar actions.

Next, we developed UML class diagrams (Figure 2.4) and contextual diagram (Figure 2.5) using
the extracted information. We found that diﬀerent members of households support household food
security diﬀerently (see Appendix A; Appendix B). Male members of the households are generally

26

Figure 2.4: UML class diagram of agents of household food security

27

involved in farming. They grow cereal crops, vegetables and are also into cash-cropping. Women
look after household works and also support men in the farm. Households generally consume the
food they produce. In case of food shortages, households seek help from their fellow villagers or buy
food from the market. They use money obtained from cash-cropping for buying food. Additionally,
women in the households are involved in small businesses that can support food purchases. Some
households may need to rely on oﬀ-farm jobs or sell their livestock to buy food. There are also
other organizations and credit agencies that provide the households with credits and supports.

28

Figure 2.5: Conceptual model of household food security in Koutiala, Mali (EM: Elderly male, YM: Younger male, EF: Elderly female,
YF: Younger female)

29

We simpliﬁed the contextual model (see Appendix D) and brought it to the stakeholders
(interviewees) for validation. The stakeholders received the structure well and acknowledged that it
included all the principal dynamics of the household food system. They, however, pointed out that
the contextualized structure did not provide the dynamics of the government and non-government
actors. Since the input data only contained interviews from farm households, we failed to capture
the dynamics occurring outside of the households. In addition to representing the dynamics of
the household food system, the structure also revealed data gaps. More information needs to be
gathered on the government and non-government actors of household food security.

2.5 Limitations and future direction

The proposed unsupervised information extraction picked individual sentences based on their
cumulative TFIDF weights. Some of the individually extracted sentences lacked contextuality
and were ambiguous. To add context and reduce this ambiguity, we propose to use neighboring
sentences at the unsupervised data extraction and processing phase (Figure 2.1). We hypothesize
that extracting a tuple of preceding and trailing sentences along with the identiﬁed sentence,
could provide vital contextual information. Similarly, some of the extracted sentences contained
pronouns. Without the information on preceding sentences, these pronouns were impossible to
resolve. Extracting the preceding sentence should help in resolving their references. NLP also
has a coreference resolution tool that automatically replaces pronouns with their referenced nouns.
However, the tool is still in development. We found that it generated too many errors that would
require manual checks. We had to proceed without using the tool, and the pronouns identiﬁed as
agents were ignored.

We only collected information on agents, attributes, and actions/interactions for ABM. However,
ABM also requires information on agent decision-making. Although the use of social and behavioral
theories in deﬁning agent decision-making is predominant, empirically derived decision-making
frameworks are context-speciﬁc and, therefore, more desirable. We realize that some sentences are
particularly useful in deriving agent decision-making. Speciﬁcally, the conditional sentences like ‘if

30

it rains, we plant maize,’ and compound sentences like ‘when production is low, we buy food from
the market’ can reveal decision-making. Harvesting these sentences with semantics and machine
learning approaches can open up new avenues for formulating empirically-based decision-making
rules for ABM.

Information obtained in the outputs is limited by the information contained in the input. We
noticed that agent tagging, after using the action and decision verbs, underpredicted agents. Entities
like ‘father’ and ‘the government’ should also be identiﬁed as agents of this particular system.
However, since subjects did not use both types of verbs (action and decision) in the provided
narratives, some information was missed out. Additionally, stakeholders pointed out that our
model structure did not include the dynamics of the governmental and non-governmental actors.
It prompts a need for a careful analysis of entities that failed to qualify as agents for data gaps.
Furthermore, the narratives (see Appendix C)went through diﬀerent stages of translations (from
local dialect to French and English) that could have corrupted some of their original meanings.

2.6 Conclusion

Complexities, ambiguities, and diﬃculties in data processing often discourage ABM devel-
opers from using qualitative data for model development. This prevents modelers from using
rich-contextual information about their target systems. ABMs are usually developed using ad-hoc
approaches, potentially producing models that lack credibility and reliability.
In this paper, we
introduced a systematical approach for ABM development from unstructured qualitative narra-
tives using NLP. The proposed methodology contained largely unsupervised, domain-independent,
eﬃcient, and bias-controlled data processing and extraction for ABM. We demonstrated its ef-
fectiveness by developing an ABM of household food security from large open-ended qualitative
ﬁeld narratives. Additionally, we outlined some of the signiﬁcant limitations of the approach and
recommended improvements for future development.

We initially aimed to develop an eﬃcient and bias-free, completely unsupervised information
extraction for conceptualizing an ABM. After preliminary algorithm development, we realized that,

31

with the current level of NLP capabilities, entirely unsupervised data processing and conceptual-
ization is not realistic. We decided to use manual ﬁltration and contextualization that potentially
introduced subjectivity and biases in model development. To address this deﬁciency, we performed
a stakeholder validation to check for subjectivity and biases.

Although we could not fully develop a completely unsupervised approach, we successfully
managed to reduce subjectivity and biases through limiting manipulation in data extraction. Data
processing and extraction are fully unsupervised, and manual inputs are only required towards the
end of model development. That eﬀectively limits the opportunities for introducing human bias in
model development. Furthermore, the unsupervised approach is much faster compared to manual
coding. The NLP development community is highly active, and hopefully, these limitations will
soon be resolved, making semantic and syntactic NLP more eﬀective for unsupervised information
extraction and model conceptualization.

32

APPENDICES

33

APPENDIX A

HOUSEHOLD-LEVEL ACTION AS DESCRIBED BY ELDERLY MALES

Figure A.1: Household-level action as described by elderly males

34

APPENDIX B

INDIVIDUAL ACTIONS AS DESCRIBED BY ELDERLY MALES

Figure B.1: Individual actions as described by elderly males

35

APPENDIX C

AN EXCERPT FROM THE FIELD NARRATIVES

Figure C.1: An excerpt from the ﬁeld narratives (Note: Conversation took place in a local dialect
that was later transcribed into English by a local translator)

36

BIBLIOGRAPHY

37

BIBLIOGRAPHY

Ackermann, Fran, Colin Eden, and Terry Williams (1997). “Modeling for litigation: Mixing quali-

tative and quantitative approaches”. In: Interfaces 27.2, pp. 48–65. issn: 0092-2102.

An, Li (2012). “Modeling human decisions in coupled human and natural systems: review of agent-
based models”. In: Ecological Modelling 229, pp. 25–36. issn: 0304-3800. url: http://ac.
els-cdn.com.proxy1.cl.msu.edu/S0304380011003802/1-s2.0-S0304380011003802-
main.pdf?_tid=3d7e25ce- 90d1- 11e7- 9c02- 00000aab0f26&acdnat=1504461581_
8c1ad049c5ac7f41cdb53de217138485.

An, Li et al. (2005). “Exploring Complexity in a Human–Environment System: An Agent-Based
Spatial Model for Multidisciplinary and Multiscale Integration”. In: Annals of the association
of American geographers 95.1, pp. 54–79. issn: 1467-8306.

Axelrod, Robert (1997). “Advancing the art of simulation in the social sciences”. In: Simulating

social phenomena. Springer, pp. 21–40.

Bakam, Innocent et al. (2000). “Formalization of a spatialized multiagent model using coloured
petri nets for the study of an hunting management system”. In: International Workshop on
Formal Approaches to Agent-Based Systems. Springer, pp. 123–132.

Balke, Tina and Nigel Gilbert (2014). “How Do Agents Make Decisions? A Survey”. In: Journal
of Artiﬁcial Societies and Social Simulation 17.4, p. 13. issn: 1460-7425. doi: 10.18564/
jasss.2687. url: http://jasss.soc.surrey.ac.uk/17/4/13.html.

Bauer, Bernhard, Jörg P Müller, and James Odell (2001). “Agent UML: A formalism for specifying
multiagent software systems”. In: International journal of software engineering and knowledge
engineering 11.03, pp. 207–230. issn: 0218-1940.

Becu, Nicolas et al. (2003). “A methodology for eliciting and modelling stakeholders’ representa-

tions with agent based modelling”. In: Multi-Agent-Based Simulation III, pp. 131–148.

Becu, N et al. (2005). “A methodology for identifying and for-malizing farmers’ representations of
watershed management: a case study from northern Thailand”. In: Companion Modeling and
Multi-Agent Systems for Integrated Natural Resource Management in Asia, p. 41.

Berger, Thomas (2001). “Agent-based spatial models applied to agriculture: a simulation tool for
technology diﬀusion, resource use changes and policy analysis”. In: Agricultural economics
25.2-3, pp. 245–260. issn: 1574-0862.

Bersini, Hugues (2012). “UML for ABM”. In: Journal of Artiﬁcial Societies and Social Simulation

15.1, p. 9. issn: 1460-7425. url: http://jasss.soc.surrey.ac.uk/15/1/9.html.

38

Bharwani, Sukaina (2006). “Understanding complex behavior and decision making using ethno-
graphic knowledge elicitation tools (KnETs)”. In: Social Science Computer Review 24.1, pp. 78–
105. issn: 0894-4393. url: http : / / journals . sagepub . com / doi / pdf / 10 . 1177 /
0894439305282346.

Bird, Steven, Ewan Klein, and Edward Loper (2009). Natural language processing with Python:

analyzing text with the natural language toolkit. " O’Reilly Media, Inc.". isbn: 0596555717.

Clark, Malcolm et al. (2012). “Automatically structuring domain knowledge from text: An overview
of current research”. In: Information Processing & Management 48.3, pp. 552–568. issn: 0306-
4573.

Collins, Andrew et al. (2015). “A Call to Arms: Standards for Agent-Based Modeling and Simu-
lation”. In: Journal of Artiﬁcial Societies and Social Simulation 18.3, p. 12. issn: 1460-7425.
url: http://jasss.soc.surrey.ac.uk/18/3/12.html.

Coyle, Geoﬀ (2000). “Qualitative and quantitative modelling in system dynamics: some research
questions”. In: System Dynamics Review: The Journal of the System Dynamics Society 16.3,
pp. 225–244. issn: 0883-7066.

Crooks, Andrew T and Alison J Heppenstall (2012). “Introduction to agent-based modelling”. In:

Agent-based models of geographical systems. Springer, pp. 85–105.

De Groot, Wouter T (1992). Environmental science theory: concepts and methods in a one-world,

problem-oriented paradigm. Elsevier. isbn: 0080875114.

Dearden, Joel and Alan Wilson (2012). “The relationship of dynamic entropy maximising and
agent-based approaches in urban modelling”. In: Agent-based models of geographical systems.
Springer, pp. 705–720.

Doscher, C et al. (2014). “An Agent-Based Model of Tourist Movements in New Zealand”. In:

Empirical Agent-Based Modelling-Challenges and Solutions. Springer, pp. 39–51.

Edmonds, Bruce (2015). “A context-and scope-sensitive analysis of narrative data to aid the spec-
iﬁcation of agent behaviour”. In: Journal of Artiﬁcial Societies and Social Simulation 18.1,
p. 17.

Edwards-Jones, Gareth (2006). “Modelling farmer decision-making: concepts, progress and chal-

lenges”. In: Animal science 82.6, pp. 783–790. issn: 1748-748X.

Epstein, Joshua M (2006). Generative social science: Studies in agent-based computational mod-

eling. Princeton University Press. isbn: 0691125473.

Forbus, Kenneth D and Brian Falkenhainer (1990). “Self-Explanatory Simulations: An Integration

of Qualitative and Quantitative Knowledge”. In: AAAI, pp. 380–387.

39

Ford, Andrew and Frederick Andrew Ford (1999). Modeling the environment: an introduction to

system dynamics models of environmental systems. Island press. isbn: 1559636017.

Fraga, Anabel et al. (2017). “Extraction of Patterns Using NLP: Genetic Deafness”. In: SEKE,

pp. 428–431.

Ghorbani, A, N Schrauwen, and GPJ Dijkema (2013). “Using Ethnographic Information to Con-
ceptualize Agent-based Models”. In: European Social Simulation Association. url: http :
//cfpm.org/qual2rule/essa20130_submission_80.pdf.

Gilbert, Nigel (2004). “Agent-based social simulation: dealing with complexity”. In: The Complex

Systems Network of Excellence 9.25, pp. 1–14.

Gilbert, Nigel and Pietro Terna (2000). “How to build and use agent-based models in social science”.

In: Mind & Society 1.1, pp. 57–72. issn: 1593-7879.

Gorard, Stephen (2002). “Can we overcome the methodological schism? Four models for combining
qualitative and quantitative evidence”. In: Research Papers in Education Policy and Practice
17.4, pp. 345–361. issn: 0267-1522.

Grimm, Volker, Jacqueline Augusiak, et al. (2014). “Towards better modelling and decision support:
documenting model development, testing, and analysis using TRACE”. In: Ecological modelling
280, pp. 129–139. issn: 0304-3800.

Grimm, Volker, Uta Berger, et al. (2010). “The ODD protocol: a review and ﬁrst update”. In:

Ecological modelling 221.23, pp. 2760–2768. issn: 0304-3800.

Grimm, Volker, Steven F. Railsback, et al. (2020). “The ODD Protocol for Describing Agent-Based
and Other Simulation Models: A Second Update to Improve Clarity, Replication, and Structural
Realism”. In: Journal of Artiﬁcial Societies and Social Simulation 23.2, p. 7. issn: 1460-7425.
doi: 10.18564/jasss.4259. url: http://jasss.soc.surrey.ac.uk/23/2/7.html.

Harmain, Harmain Mohamed and R Gaizauskas (2000). “CM-Builder: an automated NL-based
CASE tool”. In: Proceedings ASE 2000. Fifteenth IEEE International Conference on Automated
Software Engineering. IEEE, pp. 45–53. isbn: 0769507107.

Harris, Larry R (1978). “The ROBOT System: Natural language processing applied to data base

query”. In: Proceedings of the 1978 annual conference, pp. 165–172.

Heath, Brian L, Frank W Ciarallo, and Raymond R Hill (2012). “Validation in the agent-based
modelling paradigm: problems and a solution”. In: International Journal of Simulation and
Process Modelling 7.4, pp. 229–239. issn: 1740-2123.

40

Huigen, Marco GA (2004). “First principles of the MameLuke multi-actor modelling framework
for land use change, illustrated with a Philippine case study”. In: Journal of Environmental
Management 72.1, pp. 5–21. issn: 0301-4797.

Husain, Mohd Shahid and M Akheela Khanum (2016). “Word Sense Disambiguation in Software
Requirement Speciﬁcations Using WordNet and Association Mining Rule”. In: Proceedings
of the Second International Conference on Information and Communication Technology for
Competitive Strategies, pp. 1–4.

Janssen, Marco and Wander Jager (1999). “An integrated approach to simulating behavioural
processes: A case study of the lock-in of consumption patterns”. In: Journal of Artiﬁcial
Societies and Social Simulation 2.2, pp. 21–35.

Janssen, Marco and Elinor Ostrom (2006). “Empirically based, agent-based models”. In: Ecology

and Society 11.2. issn: 1708-3087.

Kemp-Benedict, Eric (2004). “From Narrative to Number: A Role for Quantitative Models in

Scenario analysis”. In: International Congress on Environmental Modelling and Software.

Krejci, Caroline C et al. (2016). “Analysis of food hub commerce and participation using agent-
based modeling: integrating ﬁnancial and social drivers”. In: Human factors 58.1, pp. 58–
79. issn: 0018-7208. url: http : / / journals . sagepub . com / doi / pdf / 10 . 1177 /
0018720815621173.

Landrum, Brittany and Gilbert Garza (2015). “Mending fences: Deﬁning the domains and ap-
proaches of quantitative and qualitative research”. In: Qualitative psychology 2.2, p. 199. issn:
2326-3598.

Lees, B (1970). “Artiﬁcial intelligence education for software engineers”. In: WIT Transactions on

Information and Communication Technologies 12. issn: 1853123854.

Leininger, Madeleine (1992). “Current issues, problems, and trends to advance qualitative paradig-
matic research methods for the future”. In: Qualitative Health Research 2.4, pp. 392–415. issn:
1049-7323.

Li, C et al. (2015). “Designing an agent-based model of SMEs to assess ﬂood response strategies

and resilience”. In: International Journal of Bioengineering and Life Sciences 9.1, pp. 7–12.

Liddy, Elizabeth D (2001). “Natural language processing”. In: Encyclopedia of Library and Infor-

mation Science. NY: Marcel Decker, Inc.

Ligmann-Zielinska, Arika and Piotr Jankowski (2010). “Exploring normative scenarios of land use
development decisions with an agent-based simulation laboratory”. In: Computers, Environment
and Urban Systems 34.5, pp. 409–423. issn: 0198-9715.

41

Ligmann-Zielinska, Arika, Peer-Olaf Siebers, et al. (2020). “‘One Size Does Not Fit All’: A
Roadmap of Purpose-Driven Mixed-Method Pathways for Sensitivity Analysis of Agent-Based
Models”. In: JASSS-THE JOURNAL OF ARTIFICIAL SOCIETIES AND SOCIAL SIMULA-
TION 23.1. issn: 1460-7425.

Loper, Edward and Steven Bird (2002). “NLTK: the natural language toolkit”. In: Proceedings of
the ACL-02 Workshop on Eﬀective Tools and Methodologies for Teaching Natural Language
Processing and Computational Linguistics, pp. 63–70.

Manning, Christopher, Prabhakar Raghavan, and Hinrich Schütze (2010). “Introduction to infor-

mation retrieval”. In: Natural Language Engineering 16.1, pp. 100–103.

Manning, Christopher, Mihai Surdeanu, et al. (2014). “The Stanford CoreNLP natural language
processing toolkit”. In: Proceedings of 52nd annual meeting of the association for computational
linguistics: system demonstrations, pp. 55–60.

Miles, Matthew B (1979). “Qualitative data as an attractive nuisance: The problem of analysis”. In:

Administrative science quarterly 24.4, pp. 590–601. issn: 0001-8392.

Miller, John H and Scott E Page (2009). Complex adaptive systems: An introduction to computa-

tional models of social life. Princeton university press. isbn: 1400835526.

Millington, James D. A. and John Wainwright (2016). “Mixed qualitative-simulation methods”. In:
Progress in Human Geography 41.1, pp. 68–88. issn: 0309-1325 1477-0288. doi: 10.1177/
0309132515627021. url: %3CGo%20to%20ISI%3E://WOS:000396830200004%20http:
//journals.sagepub.com/doi/pdf/10.1177/0309132515627021.

Millington, James DA (2012). “Using social psychology theory for modelling farmer decision-
making”. In: International Environmental Modelling and Software Society (iEMSs) 2012 In-
ternational Congress on Environmental Modelling and Software. Managing Resources of a
Limited Planet: Pathways and Visions under Uncertainty. Ed. by S. Lange R. Seppelt A. Voinov,
pp. 2485–2492.

Müller, Birgit et al. (2014). “Standardised and transparent model descriptions for agent-based
models: Current status and prospects”. In: Environmental Modelling & Software 55, pp. 156–
163. issn: 1364-8152. url: https://www-sciencedirect-com.proxy2.cl.msu.edu/
science/article/pii/S1364815214000395.

Namey, Emily et al. (2008). “Data reduction techniques for large qualitative data sets”. In: Handbook

for team-based qualitative research 2.1, pp. 137–161.

Nasukawa, Tetsuya and Jeonghee Yi (2003). “Sentiment analysis: Capturing favorability using
natural language processing”. In: Proceedings of the 2nd international conference on Knowledge
capture. ACM, pp. 70–77. isbn: 1581135831.

42

Navigli, Roberto (2009). “Word sense disambiguation: A survey”. In: ACM computing surveys

(CSUR) 41.2, pp. 1–69. issn: 0360-0300.

Niglas, Katrin (2000). “Quantitative and qualitative inquiry in educational research: is there a
paradigmatic diﬀerence between them? Paper given at ECER99, Lahti, 22-25 September 1999”.
In: Education Line.

O’Sullivan, D. et al. (2016). “Strategic directions for agent-based modeling: avoiding the YAAWN
syndrome”. In: J Land Use Sci 11.2, pp. 177–187. issn: 1747-423X (Print) 1747-423X (Link-
ing). doi: 10.1080/1747423X.2015.1030463. url: https://www.ncbi.nlm.nih.gov/
pubmed/27158257.

Orkphol, Korawit and Wu Yang (2019). “Word sense disambiguation using cosine similarity col-

laborates with Word2vec and WordNet”. In: Future Internet 11.5, p. 114.

Padilla, Jose J, David Shuttleworth, and Kevin O’Brien (2019). “Agent-Based Model Character-
ization Using Natural Language Processing”. In: 2019 Winter Simulation Conference (WSC).
IEEE, pp. 560–571. isbn: 1728132835.

Polhill, J Gary, Lee-Ann Sutherland, and Nicholas M Gotts (2010). “Using qualitative evidence
to enhance an agent-based modelling system for studying land use change”. In: Journal of
Artiﬁcial Societies and Social Simulation 13.2, p. 10.

Ramos, Juan (2003). “Using tf-idf to determine word relevance in document queries”. In: Pro-
ceedings of the ﬁrst instructional conference on machine learning. Vol. 242. Piscataway, NJ,
pp. 133–142.

Rich, Michael and Kenneth R Ginsburg (1999). “The reason and rhyme of qualitative research:
why, when, and how to use qualitative methods in the study of adolescent health”. In: Journal
of Adolescent health. issn: 1879-1972.

Rivers III, Louie et al. (2017). “Mental models of food security in rural Mali”. In: Environment

Systems and Decisions, pp. 1–19. issn: 2194-5403.

Robinson, Derek T. et al. (2007). “Comparison of empirical methods for building agent-based
models in land use science”. In: Journal of Land Use Science 2.1, pp. 31–55. issn: 1747-423X
1747-4248. doi: 10.1080/17474230701201349.

Runck, Bryan (2018). “GeoComputational Approaches to Evaluate the Impacts of Communication

on Decision-making in Agriculture”. Thesis.

Runck, Bryan C et al. (2019). “Using word embeddings to generate data-driven human agent
decision-making from natural language”. In: GeoInformatica 23.2, pp. 221–242. issn: 1384-
6175.

43

Al-Safadi, Lilac AE (2009). “Natural Language Processing for Conceptual Modeling”. In: JDCTA

3.3, pp. 47–59.

Salloum, Said A et al. (2018). “Using text mining techniques for extracting information from re-
search articles”. In: Intelligent natural language processing: Trends and Applications. Springer,
pp. 373–397.

Schmitt Olabisi, Laura K et al. (2010). “Using scenario visioning and participatory system dynamics
modeling to investigate the future: Lessons from Minnesota 2050”. In: Sustainability 2.8,
pp. 2686–2706.

Seidl, Roman (2014). “Social scientists, qualitative data, and agent-based modeling”. In: Social

Simulation Conference.

Serrano, Emilio and Carlos A Iglesias (2016). “Validating viral marketing strategies in Twitter via
agent-based social simulation”. In: Expert Systems with Applications 50, pp. 140–150. issn:
0957-4174.

Simoes, Joana A (2012). “An agent-based/network approach to spatial epidemics”. In: Agent-based

models of geographical systems. Springer, pp. 591–610.

Standish, Russell K. (2008). “Going Stupid with EcoLab”. In: Simulation 84.12, pp. 611–618. issn:
0037-5497 1741-3133. doi: 10.1177/0037549708097146. url: http://sim.sagepub.
com/cgi/content/abstract/84/12/611.

Sterling, Eleanor J et al. (2019). “Try, try again: Lessons learned from success and failure in

participatory modeling”. In: Elementa: Science of the Anthropocene 7.1.

Subburaj, Vinitha Hannah and Joseph E Urban (2018). “Specifying security requirements in multi-
agent systems using the descartes-agent speciﬁcation language and AUML”. In: Information
technology for management: Emerging research and applications. Springer, pp. 93–111.

Sun, Shiliang, Chen Luo, and Junyu Chen (2017). “A review of natural language processing
techniques for opinion mining systems”. In: Information Fusion 36, pp. 10–25. issn: 15662535.
doi: 10 . 1016 / j . inffus . 2016 . 10 . 004. url: http : / / www . sciencedirect . com /
science/article/pii/S1566253516301117.

Tisue, Seth and Uri Wilensky (2004). “NetLogo: Design and implementation of a multi-agent

modeling environment”. In: Proceedings of agent. Vol. 2004, pp. 7–9.

Torrens, Paul M (2010). “Geography and computational social science”. In: GeoJournal 75.2,
pp. 133–148. issn: 0343-2521. url: https://link.springer.com/content/pdf/10.
1007%2Fs10708-010-9361-y.pdf.

44

Trencansky, Ivan and Radovan Cervenka (2005). “Agent Modeling Language (AML): A compre-

hensive approach to modeling MAS”. In: Informatica 29.4. issn: 1854-3871.

Voinov, Alexey and Francois Bousquet (2010). “Modelling with stakeholders”. In: Environmental

Modelling & Software 25.11, pp. 1268–1281. issn: 1364-8152.

Voinov, Alexey, Nagesh Kolagani, et al. (2016). “Modelling with stakeholders–next generation”.

In: Environmental Modelling & Software 77, pp. 196–220. issn: 1364-8152.

Watkins, D. C. (2012). “Qualitative research: the importance of conducting research that doesn’t
"count"”. In: Health Promot Pract 13.2, pp. 153–8. issn: 1524-8399 (Print) 1524-8399 (Link-
ing). doi: 10 . 1177 / 1524839912437370. url: https : / / www . ncbi . nlm . nih . gov /
pubmed/22382490.

Wolstenholme, Eric F (1999). “Qualitative vs quantitative modelling: the evolving balance”. In:

Journal of the Operational Research Society 50.4, pp. 422–428. issn: 0160-5682.

45

CHAPTER 3

ANALYZING HOUSEHOLD VULNERABILITY TO FOOD INSECURITY IN RURAL
SOUTHERN MALI - A COUPLED BIOPHYSICAL AND SOCIAL MODEL APPROACH.

Household food security is dependent on, among other things, food production and access. How-
ever, food security discourse is dominated by the productionist arguments, and the role of food
access is often ignored. Food security lies at the interface of biophysical, climatic, and socioe-
conomic systems and requires a systems approach for evaluation.
In this study, we integrate a
biophysical crop model with an agent-based model (ABM) of household food systems to analyze
household vulnerability to food insecurity in rural southern Mali. The crop model analyzes the
eﬀects of biophysical and climatic environments on household food production, while the ABM
evaluates the eﬀects of socioeconomic systems on household food access. We measure household
vulnerability to food insecurity using a coping-based food security vulnerability Index (FSVI).
The FSVI reﬂects households’ current severity and also indicates their future vulnerability to food
insecurity. Vulnerability to food insecurity results from complex and dynamic interactions between
biophysical and social systems. By integrating the FSVI into the model, we capture the interplay
between the dynamic systems for the vulnerability analysis. We simulate a ‘business-as-usual’
scenario and ﬁnd that around 80% of the regional households are likely to be food insecure by
2050. Large families and those living in the North are more susceptible to future food insecurity.
However, the severity of food insecurity is higher especially in some smaller families.

3.1 Introduction and Background

Despite realizing the importance of food access for decades (Sen 1981), food availability still
dominates the food security discourse. Regional food availability is often used as a sole indicator
of food security (Pinstrup-Andersen 2009), and policies are ﬁxated on increasing food production.
However, food security is a multicausal issue that varies across households. Multiple factors,
including lack of resources, socio-cultural norms, and policies, often dictate households’ access to

46

available food. As food access is non-homogenous, regional food availability does not adequately
reﬂect household food security. A comprehensive food security analysis, therefore, needs evaluating
heterogeneous food availability and access at the household level.

In this study, we develop a spatial agent-based model (ABM) of a household food system
coupled with a process-based crop model to analyze household vulnerability to food insecurity in
rural Southern Mali. The crop model examines the eﬀects of biophysical and climatic factors on
household food production. At the same time, the ABM explores the inﬂuence of socioeconomic
factors on adequacy and stability of household food access. Social vulnerability is often deﬁned in
terms of risk exposure, coping capacity, and recovery potential (Bohle, Downing, and Watts 1994;
Watts and Bohle 1993). We deﬁne vulnerability at the interface of biophysical and socioeconomic
systems and measure it using a coping-based food security vulnerability index (FSVI). The FSVI
measures the current severity of food shortages at the households as well as provides an outlook on
their future vulnerability.

Malian agriculture had a disappointing past; it could not seize the beneﬁts of the Green
Revolution (Otsuka and Muraoka 2017). The promoted technologies and farm management were
largely inadequate for improved crop production (Djurfeldt et al. 2005). The regional farmers could
not intensify their agriculture, and their productivity remained stagnant for decades (M. E. Brown,
Hintermann, and Higgins 2009). On the other hand, due to rapid population growth, food demand
kept on rising. Farmers had to rely on extensiﬁcation to meet the expanding food demand. Also, due
to poor management, there was widespread soil degradation (Ruelland, Levavasseur, and Tribotté
2010) that further deteriorated the agricultural productivity.

The outlook of Malian agriculture looks bleak. The country is expected to be profoundly
impacted by future climate change. All of the General Circulation Models (GCMs) used by the
Intergovernmental Panel for Climate Change (IPCC) predict a warmer Mali (Traore et al. 2013).
Since crop yield generally declines with increased temperature (Bassu et al. 2014), Mali is likely to
see reduced crop production under climate change. Some climate models foresee increased rainfall
in the future (Patricola and Cook 2010; Sultan and Gaetani 2016) that can arguably beneﬁt crop

47

production. However, an increased frequency of erratic rainfall predicted under climate change,
together with warmer weather, is expected to nullify such beneﬁts (Sivakumar 1988).

The northern half of Mali lies in the Saharan region. It is highly arid and not suitable for agricul-
ture. However, southern Mali receives higher rainfall and is more productive. Koutiala (Figure 1.2),
our study area, lies in the Sudano-Guinean region in the South and receives 600-1400mm of annual
rainfall (Falconnier et al. 2015). Agriculture is a prime activity here. Maize, millet, and sorghum
are the major staple crops grown in the region. However, farming is predominantly traditional. The
Head of the household, usually the oldest male member, makes most of the farm-related decisions
(Rivers III et al. 2017). Besides, farmers are mostly poor and perform subsistence farming, and
their access to resources is highly limited.

Koutiala houses a large cotton industry that supports local farmers with logistics and training.
Since cotton production technologies also help with crop production (Bingen 1998; Cooper and
West 2017), local farmers also cultivate cotton to receive such beneﬁts (Dembele et al. 2017). In a
way, the cotton industry plays a vital role in the growth of regional agriculture. Food produced here
often gets exported to other parts of the country, including the northern region. Since the area plays
a critical role in addressing broader food security, it is considered a breadbasket of Mali (Bingen
1998). Ironically, Koutiala has one of the largest food-poor and malnourished populations in the
country (Cooper and West 2017; Eozenou, Madani, and Swinkels 2013) that highlights existing
disparities in food access at a granular level. Nevertheless, local policy discussions focus on
improving food production (Olabisi et al. 2018); discourse on food access remains largely dormant.

3.2 Measuring household vulnerability to food insecurity

3.2.1 Food security: deﬁnition and measurement

Food security is dominated by ‘productionist’ arguments (Clapp 2014; Fouilleux, Bricas, and
Alpha 2017) where food production has been emphasized for achieving food security. Traditionally
regional food production used to be the indicator of food security. However, with globalization
and commercialization, food transport got easier, and food-deprived regions like Northern Mali

48

could become equally food secure as the South (Cooper and West 2017; Eozenou, Madani, and
Swinkels 2013; Staatz, D’agostino, and Sundberg 1990). This brought food access into the center
of food security deﬁnition. Recently, food security is increasingly analyzed at more granular levels.
Households’ or individuals’ access to suﬃcient, stable, and diverse food supply becomes the central
theme of current food security deﬁnitions. Food security is increasingly deﬁned in terms of its
three pillars and a foundation (Figure 3.1): food availability, access, utilization, and stability (FAO
2008).

Figure 3.1: The pillars of food security

Since food production indicates food availability, its measurement is generally straightforward.

49

However, measuring food access and utilization is challenging. Due to the lack of suitable indicators,
food access and utilization are usually measured using proxies. For instance, household assets are
often used as a proxy for food access. Wealthy households are assumed to have better access to food
and, hence, poorer families are considered vulnerable (S. Maxwell and Smith 1992). Utilization,
on the other hand, is often measured using anthropometric indices (Scaramozzino 2006). These
indicators are usually expensive to collect and prone to errors (Bickel et al. 2000; Scaramozzino
2006). Subjective experiences with food and hunger (Bickel et al. 2000; Hoddinott 1999) are also
increasingly used to analyze food security. Although they are easier to collect, subjective indicators
are bias-prone and often unreliable for making future projections.

3.2.2 Measuring vulnerability to food insecurity

The notion of vulnerability diﬀers across disciplines but is often considered as synonymous to
poverty (Chambers 1989; Scaramozzino 2006). However, Chambers (1989) dichotomizes poverty
and vulnerability - the ﬁrst indicates a lack of resources, while the latter is a defenseless state.
Speciﬁc groups of people like the elderly, pregnant women, and children are increasingly vulnerable
to stresses regardless of their economic conditions (Kelly and Adger 2000). Furthermore, there are
external and internal sides of vulnerability. The external side indicates shock, and the internal side
of vulnerability is a defenseless state (Chambers 1989). Watts and Bohle (1993) extend this idea
and deﬁne vulnerability in terms of exposure to risk (external), the capacity to cope (internal), and
the ability to recover (internal). Risk exposure does not necessarily indicate a vulnerable state, as
long as one can cope with it. However, coping can impede future recovery potential.

Based on Watts and Bohle’s approach, we deﬁne vulnerability to food insecurity at the interface
of biophysical and socioeconomic systems (Figure 3.2). The biophysical environment generally
oﬀset food availability by aﬀecting household food production, while socioeconomic factors inﬂu-
ence households’ access to food. Households look for options to feed their family when exposed to
food shortages. These measures (coping strategies) can vary by geography, culture, and severity but
usually start with easily accessible low-impact approaches such as meal rationing and in-kind help

50

Figure 3.2: Space of Vulnerability (Adapted from Watts and Bohle, 1993)

from relatives or neighbors (S. Maxwell and Smith 1992). However, as the severity of food scarcity
increases, households need to resort to more drastic measures, including the sale of household as-
sets and migration (Figure 3.3). Such approaches generally put a strain on domestic resources that
can aﬀect households’ ability to cope with future food scarcity (D. G. Maxwell 1996; S. Maxwell
and Smith 1992).

Vulnerability assessment provides early warning and helps with target identiﬁcation for relief
planning and emergency management (Flanagan et al. 2011; Füssel and Klein 2006; Riely 2000).
Disaster management literature uses several vulnerability indices for social (Flanagan et al. 2011),
economic (Guillaumont 2009), and climatic (Pandey and Jha 2012) stresses. However, indices
speciﬁc to food security are highly limited. Existing food security vulnerability indices often work
at macro-levels (e.g., Burg 2008; Krishnamurthy, Lewis, and Choularton 2014) and fail to represent
household level vulnerabilities. Moreover, these indices focus on current vulnerabilities. Future
predictions generally use trend extrapolations that are unreliable as they ignore the dynamics of
associated systems (Burg 2008). Since vulnerability is dynamic and complex, it demands an index
that captures the dynamics of the complex systems (Riely 2000).

51

Figure 3.3: Household responses to food shortages (based on FGD and Maxwell 1992)

Despite recognizing food security as a function of food availability, access, utilization, and
stability, literature still lacks an index that simultaneously checks all the dimensions of food
security. Besides, vulnerability to food insecurity results from complex and dynamic interactions
between biophysical and social systems and is diﬃcult to quantify (Krishnamurthy, Lewis, and
Choularton 2014). The FSVI, used in the study, is a simple coping-based index that provides
vulnerability information based on the coping mechanisms households use to address their food
scarcities. Furthermore, we integrate it into the ABM to eﬀectively capture systems complexities
needed to explore future vulnerabilities.

The FSVI has four severity1 levels (Table 3.1).

It evaluates food availability by analyzing
household food production and food access by examining their interactions. Similarly, it checks
1Since current severity also aﬀects future vulnerability, we use severity and vulnerability inter-

changeably in the paper.

52

Table 3.1: Food Security Vulnerability Index (FSVI)

Index
FSVI I

FSVI II

FSVI III

FSVI IV

Description
Food secure - households have suﬃcient food
production/availability to feed their family.
Slightly food insecure – households face slight
shortages that can be covered by rationing or
in-kind help from neighbors and relatives.
Moderately food insecure – households need to
purchase food. They might have to sell their
assets or go for oﬀ-farm jobs.
Severely food insecure – households deplete all
their domestic resources. Members gone for
oﬀ-farm jobs decide to migrate, creating work-
force shortages at home and further exacerbating
household’s food insecurity.

household calorie suﬃciency for evaluating utilization. The longitudinal ABM simulations, in turn,
provide information on the stability of household food supply. The highlighted terms constitute
a comprehensive approach to food security (Figure 3.1). Maxwell and the team (D. Maxwell,
Caldwell, and Langworthy 2008; D. G. Maxwell 1996; D. Maxwell, Watkins, et al. 2003) are big
proponents of coping-based indices for measuring food security. According to them, these indices
work as ‘current’ as well as ‘leading’ indicators; i.e., beside reﬂecting the current severity of food
scarcity, the indices also indicate households’ susceptibility to future vulnerabilities.

3.2.3 Coupled biophysical and social food security analysis

Biophysical and social sciences are highly active but often disjoint in deﬁning food security.
Biophysical crop science emphasizes promoting crop productivity for food security (Gebbers and
Adamchuk 2010). Social science, on the other hand, attributes food insecurity to unjust resource
distribution and access (Jenkins and Scanlan 2001). Household food security requires suﬃcient food
supply supported through stable production and access. Food production is generally inﬂuenced by
biophysical and climatic conditions and farm management. Food access, on the other hand, relies
on resource availability controlled by prevailing socioeconomic policies and practices. Besides,

53

household access to resources also aﬀects farm management.

Hence, food security lies at the intersection of biophysical, climatic, and socioeconomic systems,
and requires a broader systems approach for its analysis (Hammond and Dube 2012; J. W. Jones et
al. 2017). In this study, we integrate SALUS (Systems Approach to Land Use Sustainability) (Basso,
Ritchie, et al. 2006), a bio-physical process-based crop model, with an ABM for a comprehensive
household food security analysis. By using the crop model that explores the dynamics of crop
production coupled with the ABM of household food system, we aim to capture the interplay
between biophysical, climatic, and socioeconomic environments responsible for household food
availability and access (Figure 3.4).

Figure 3.4: An integrated systems approach to food security analysis

3.2.3.1 Crop models for estimating food production

Farmers regularly monitor the growth and development of their crops at diﬀerent stages of crop cycle
to get the real-time evaluation of yield and management. These estimations are often subjective,
restricted to small spatial and temporal scales, and are not expandable or transferable (i.e., the
same approach cannot be used at diﬀerent locations). Crop modelers, on the other hand, develop
statistical and process-based models that can forecast yields based on empirical biophysical and

54

climatic observations. These models can predict harvests at a much larger extent and are often
transferable (Basso and Liu 2019).

Statistical crop models extrapolate past trends to predict future yields. These models are
simple and have low data requirements but usually need high-quality data for reliable estimations.
Recent advancements in technologies like remote sensing and machine learning have increased the
reliability of statistical models (Basso and Liu 2019). However, these models do not capture the
complex dynamics of agricultural sub-systems and cannot explain ‘out of sample’ scenarios (Basso
and Liu 2019; J. W. Jones et al. 2017). Unlike statistical approaches, process-based models capture
systems complexities, making them useful in exploring plausible future scenarios. Process-based
crop models are based on plant responses to physical and bio-climatic variables and are suitable
alternatives to statistical models for long-term yield estimations.

The process-based models use agroecological and climatic indicators to evaluate crop biomass.
They are modular (Asseng et al. 2014) and contain a series of mathematical equations representing
the growth and development of crops. Some of their key processes include phenology, biomass
accumulation, and water and nutrient uptake (Asseng et al. 2014). These models are highly
useful for evaluating management practices, yield gap, nutrient cycling, environmental impacts,
and climate change mitigation and adaptation (Asseng et al. 2014; Basso and Ritchie 2015; Basso,
Ritchie, et al. 2006; Basso, Sartori, et al. 2012; Liu and Basso 2017).

SALUS uses a systemic approach (Figure 3.5) to simulate daily plant growth and soil conditions.
It consists of three major components – a crop growth module, soil organic matter and nutrient
cycling module, and soil water balance and temperature module. The model has been used
extensively for simulating crop productivity (Basso and Ritchie 2015; Basso, Sartori, et al. 2012;
Liu and Basso 2017) and evaluating farm management and their environmental impacts (Basso,
Ritchie, et al. 2006; Basso, Sartori, et al. 2012). However, SALUS has a large data requirement.
It generally requires information on physical and chemical composition at each layer of soil, daily
weather conditions, crop genetics, and crop management. As a result, using SALUS in developing
African countries, where detailed quality data is usually scarce, is often challenging (Liu and Basso

55

2017).

Figure 3.5: A systems approach in SALUS model (Source: Basso, Ritchie, et al. 2006)

A key criticism of crop sciences is their total ﬁxation on crop production, ignoring the social
dimensions of food security (James W Jones et al. 2017). Furthermore, the crop models generally
treat farmers as rational ‘homo-economicus’. They supposedly have the knowledge and resources
to optimally exploit the productivity of their farms completely. However, farmers are neither
uniformly capable and knowledgeable; nor do all want to maximize productivity. Hence, one needs
to consider household heterogeneity and bounded rationality for a reliable food security analysis.

3.2.3.2

Integrating ABM for food security analysis

ABM is useful for understanding phenomena that emerge from complex and nonlinear interactions
of autonomous as well as heterogeneous social constituents of a system (An et al. 2005). ABM
can adequately consider population heterogeneity and bounded rationality, and is highly useful
in analyzing food security (An et al. 2005; Bithell, Brasington, and Richards 2008; Kennedy
2012). It has been widely used in evaluating land management (Ligmann-Zielinska 2009) and its

56

repercussion on food production (T. Berger and Troost 2014; Magliocca, D. G. Brown, and Ellis
2013; Polhill, Sutherland, and Gotts 2010). Since ABM can handle social interactions well, it can
naturally analyze food access (Koh, Reno, and Hyder 2019).

In Mali, rural households are mostly subsistence farmers. Thus, capturing food production is
essential for rural food security analysis. However, ABM, being primarily a model of the social
system, cannot capture crop dynamics. Coupling a process-based crop model with an ABM of
household food system incorporates the dynamics required for an eﬀective and comprehensive
food security analysis. However, literature severely lacks integrated crop-ABM models for food
security analysis. Most of the integrated models focus on the impacts of policies and practices
on land and farm management and their eﬀects on household food production and economy. For
instance, the People and Landscape Model (Matthews 2006) integrates a process-based crop model
with an ABM to analyze crop-nutrient management. Marohn et al. (2013) used Land Use Change
Impact Assessment (LUCIA), a process-based crop model, with a Multi-Agent System to evaluate
soil conservation strategies in Vietnam. Similarly, Schreinemachers and Thomas Berger (2011)
integrated a Mathematical Programming-based Multi-Agent System with biophysical models to
examine the eﬀects of agricultural technologies, market dynamics, environmental change, and
policies on household and agroecological resources. Earlier research by them (Schreinemachers,
Thomas Berger, and Aune 2007) used ABM with the Tropical Soil Fertility Calculator (TSPC) to
simulate the eﬀect of crop productivity on household economy.

A more recent integrated food security analysis, done by Wossen and Thomas Berger (2015),
used a crop model with an ABM to analyze the impacts of climate and price ﬂuctuations on
household food availability and access. However, their research does not check the stability aspect
of food security as it lacks long-term climate and market projections. Recently, Dobbie et al. (2018)
looked at all four dimensions of food security for community food security analysis using crop and
agent-based models. They integrate climate change scenarios in their study to simulate future
crop production. However, they do not use a process-based crop model and represent crop yield
as a function of input and water availability. Although crop productivity is profoundly aﬀected

57

by temperature, their model does not capture such dynamics. Additionally, they used RCP 2.6, a
‘highly stringent’ climate change scenario, which only simulates a best-case future. Our research
addresses these gaps by capturing all four dimensions of food security by integrating a process-
based crop model with an ABM. Additionally, we use a high emission RCP 8.5 climate change
scenario that helps with risk estimation and planning interventions.

3.2.3.3 Model parameterization, setup, and evaluation:

Agents in the ABM represent farm households in Koutiala. The households are represented in
the landscape by vector polygons (representing households’ land parcels) of diﬀerent shapes and
sizes. We develop a simple GIS model for creating the distributions of land parcels. GIS places the
parcels randomly in and around arable land. Additionally, the GIS model is set to place the land
parcels denser and smaller within urban areas reﬂecting their distribution in reality (N’Danikou
et al. 2017). We divide households into the following four categories based on the size of their land
parcels (Falconnier et al. 2015):

1. High Resource Endowed – Large Herds (LRE_LH) households

2. High Resource Endowed (HRE) households

3. Medium Resource Endowed (MRE) households

4. Low Resource Endowed (LRE) households

Koutiala has only one growing season; households partition their farmland to cultivate diﬀerent
crops. We use SALUS to simulate the cultivation of maize, millet, and sorghum. At the beginning
of a model run, the ABM randomly selects one set of land parcels generated by the GIS model.
The ABM then receives the productivity estimates from SALUS and interpolates the values to the
land parcels. The model then calculates household food production and evaluates food suﬃciency
(Figure 6). Food suﬃcient families are labeled as food secure households (FSVI I), and food insuf-
ﬁcient households look for coping alternatives beginning with in-kind help from close neighbors

58

or relatives. The families whose food deﬁciency is addressed by in-kind support get labeled as
slightly food insecure (FSVI II) and the households that need to purchase food are moderately food
insecure (FSVI III). Households either sell their livestock or go for oﬀ-farm jobs to purchase food.
The households that deplete their internal resources (i.e., livestock and workforce) are labeled as
severely food insecure (FSVI IV).

For the model simulations, We deﬁne a ‘business-as-usual’ scenario based on the current trends
in climate change, population growth, and farmers’ access to input. We select a high emission
climate change scenario (Representative Concentration Pathway – RCP 8.5) (Riahi et al. 2011).
Similarly, the population growth rate for Koutiala is set at 5% per annum (Farvacque-Vitkovic
et al. 2007; Dijk, Bruijn, and Beek 2004). During our focus group discussions, the local farmers
highlighted the issue of the lack of access to agricultural input. Farmers usually produce manure
for their ﬁelds using household and agricultural residues. Based on the discussions, we set the
input application at 1T/ha of manure. We use the Met Oﬃce Hadley Centre ESM (HadGEM2-ES)
(C. Jones et al. 2011) climate change model to generate future weather conditions for the crop
model. Daily estimated weather data is obtained from Marksim DSSAT weather ﬁle generator
(http://gismap.ciat.cgiar.org/MarkSimGCM). Although the region lies within a uniform Sudano-
Guinean agroecological zone, to account for local variability in weather (albeit small), we divide
our study area into four quadrants and use their centroids as weather stations.

Each model simulation runs for 40 years (from 2010 to 2050) with one-year timesteps. SALUS
is run for the entire Koutiala landscape containing 690,000 grids of 250m spatial resolution. SALUS
outputs are validated using reported ﬁeld data (Table 3.2). ABM is generated from stakeholder vali-
dated ABM structure (Appendix D). We applied standard veriﬁcation including methods such as unit
tests, extreme tests, assertions, and debugging (An et al. 2005; Gilbert 2019). We run 1000 ABM
simulations for the business-as-usual scenario. The ABM is developed in Python 3.7 programming
language (https://www.python.org/) using multiple libraries, including Pandas, GeoPandas,
and Rasterio. ABM parallelization is achieved through the concurrent.futures library. Since both
crop and agent-based models are extensive and resource-intensive, we used High Performance Com-

59

Figure 3.6: Structure of Crop-ABM coupled Model

60

Table 3.2: Koutiala crop yields (reported) (Source: Cellule de Planiﬁcation et de Statistique, Mali)

Year Millet Yield (Kg/ha) Sorghum Yield (Kg/ha) Maize Yield (Kg/ha)
2003
1800
1804
2004
1773
2005
2006
1896
2040
2007
2000
2008
2009
2300
2500
2010
2285
2011
2012
2214
2118
2013
2475
2014
2015
2896
2440
2016

1059
994
989
991
1095
1150
1100
1100
710
1000
955
1106
1099
1066

958
921
983
979
1000
1100
1000
1000
892
900
979
1050
1046
1025

puting (HPC) at Michigan State University for the simulations. Detailed information on the ABM
is presented in the ODD format (V. Grimm et al. 2006; Volker Grimm et al. 2010) in Appendix E,
and the details on SALUS can be found at https://basso.ees.msu.edu/salus/index.html.
Additionally, itemized model parameters are presented in Table 3.3 and Table 3.4..

61

Source
International Soil Reference and Information Center
(http://www.isric.org)

Marksim DSSAT weather ﬁle generator (http://
gismap.ciat.cgiar.org/MarkSimGCM)

Focus Group Discussion/ Expert Advice/ (Soumaré
et al. 2002)

Table 3.3: SALUS Parameters

SALUS Parameterization
Soil

Parameters

Physical Properties

Bulk density, Clay content,
Coarse fragments, Sand, Silt

Chemical Properties Cation exchange capacity, Nitro-

gen, Soil Organic Carbon, pH

Weather/Climate

Parameters

Weather data

Future Climate

Tmax, Tmin, Rainfall, Solar ra-
diation
RCP 8.5 (HadGEM2-ES)

Farm Management Parameters

Resolution

250m

250m

Resolution

Daily

-

Value

Crops

Input

Maize, Millet, Sorghum

Barnyard Compost

20%, 40%, 40% of total
land respectively
1T/ha

62

Table 3.4: ABM Parameters

ABM Parameterization
Total Farm Households (No of
Agents)
HH typologies

High Resource Endowed - Large
Herdsize (HRE_LH)
High Resource Endowed (HRE)
Medium Resource Endowed
(MRE)
Low Resource Endowed (LRE)
Food supply
Calorie demand

Calorie supply

Sale of 1 cattle provides

Food Production
Workforce available
Worker-days required (Crops)
Worker-days required (Cotton)
Food Access
Neighborhood distance

Neighborhood support

Distribution

13%

28%
40%

19%
Values
2400 per person

per

3100
kg
maize; 3550 per kg
sorghum/millet
1285 kg of grain

Values
1/2*hh_size
63
149
Distribution
Uniform
2000,3000,4000
(meters)
Discrete (0%, 5%,
10%, 20%) % of
surplus

9000

Land_size:
mean (std)
16.6 (8.33)

11.8 (6)
7.5 (2.4)

3.2 (2)

HH_Size:
mean (std)
45.3 (18)

27.3 (6.6)
12.6 (1.7)

7.9 (4.9)

Livestk:
mean (std)
5.7 (1.5)

2.7 (1.6)
4.1 (6.5)

1.3 (1)

Source
Census (2009)

(Falconnier et al. 2015)

FAO
(http://www.fao.org/)

Derived from ICRISAT,
2017

(Ruben et al. 1997)

based on Focus Group
Discussion

P(0.1,0.5,0.3,0.1)

63

Figure 3.7: Future climate under RCP 8.5 scenario

3.3 Results and Discussions

The climate model predicts a warmer Koutiala; its mid-century temperature could rise by
more than 2◦C (Figure 3.7). Temperature increase accelerates plant phenological development and
shortens the growing season. This can lead to reduced biomass accumulation (S. Asseng et al.
2017; Basso, Kendall, and Hyndman 2013). Bassu et al. (2014) suggested that with every ◦C
increase in temperature maize productivity could decline by 0.5T/ha. Since Koutiala is expected to
see an increase in future temperature, crop productivity is likely to decline. The region experiences
no signiﬁcant changes in future rainfall patterns. However, increased temperature causes increased
evapotranspiration (Roudier et al. 2011) that can limit water availability to plants. Our SALUS
modeling produced results where maize and sorghum suﬀered from water stress, and their future
yields declined by more than 40% (Figure 3.8). Since millet is more water-eﬃcient than maize and
sorghum (Saxena et al. 2018; Schlenker and Lobell 2010), it showed no water stress in the model.
Millet productivity remained relatively stable, with a mere 2% decline in its future yields. We also
found that southern Koutiala was more productive than the north (Figure 3.9). However, the region
would signiﬁcantly lose its productivity in the future (Figure 3.10).

64

Figure 3.8: SALUS predicted yields (kg/ha) for business-as-usual scenario (dashed lines represent
ﬁeld reported data)

65

Figure 3.9: Spatial distribution of crop productivity in Koutiala

66

The crops we simulated belong to C4 plant categories. These plants are highly resistant to
heat and drought (Roudier et al. 2011). Still, SALUS predicted a signiﬁcant decline in their future
productivity, especially for maize and sorghum. Farm management plays a vital role in plant water
use eﬃciency (Basso, Kendall, and Hyndman 2013; Ritchie and Basso 2008). The lack of improved
technologies and farm management predominant in the region were not eﬀective in reducing crop
water stress. Additionally, these plants are less sensitive to atmospheric CO2 (Roudier et al. 2011;
Schlenker and Lobell 2010). Although RCP 8.5 has a high level of CO2, its beneﬁts on the
productivity of the crops have been minuscule.

Koutiala is witnessing population growth higher than the national average (Farvacque-Vitkovic
et al. 2007; Dijk, Bruijn, and Beek 2004). At this rate, the region is likely to see a signiﬁcant
increase in its baseline population by 2050.
Increased population, together with reduced crop
productivity, substantially oﬀsets regional food balance. We found that the regional cumulative
food demand would surpass its food production by 2025 (Figure 3.11). This can severely impact the
region’s food export, jeopardizing its status as the breadbasket of the country. However, the ABM
revealed that household food insecurity, deﬁned by using the FSVI, existed well before 2025. It
further signiﬁes the inadequacy of regional food availability in reﬂecting household food security.
Similarly, we found that almost 80% of the households were likely to be food insecure by 2050
(Figure 3.12). For the subsistence farmer in the region, household food production is critical. The
households in the less productive north were found to be more susceptible to food insecurity than
their counterparts in the South (Figure 3.13). Additionally, population dynamics play an important
role in household food security. Population growth increases food demand to a level that some
households cannot easily meet. The model revealed a higher number of FSVI II than FSVI III
households at the beginning of the simulation, indicating that the food shortages were small with
a smaller population. However, it was later superseded by FSVI III households that needed food
purchases to feed the family. The model identiﬁed larger households (HRE_LH and HRE) to be
more dependent upon food purchases (Figure 3.14)

However, population growth is not all doom and gloom. Traditionally, rural households in

67

Figure 3.10: Maize productivity decline from 2010 to 2050 (kg/ha)

Figure 3.11: Regional food production vs. demand

Koutiala are large and contain extended family members living and working together. Since
farming is mainly manual, a large family usually means the availability of a large agricultural
workforce. Older generation in the region still prefers the traditionally large households (Rivers III
et al. 2017). The large households often have suﬃcient workforce for farm and oﬀ-farm jobs.
Although they were found to be prone to food insecurity in the study, the large households usually
had resources – in the form of workforce and livestock - to address the problem. This prevents them

68

from being severely food insecure (FSVI IV). However, smaller households have highly limited
livestock and workforce for farm and oﬀ-farm jobs. Any depletion of domestic resources could
easily expose them to severe forms of food insecurity (Figure 3.14).

Figure 3.12: Trajectory of food security

3.4 Limitations and future development

There is only one-way communication between SALUS and the ABM; the ABM uses outputs
from the crop model as its inputs. Due to this loose coupling, incorporating some dynamics are
challenging (Parker, Hessl, and Davis 2008). Households learn from their experience, and farming
evolves over time. Our model considers farm management uniform across time and space. Learning
and adaptation, some of the most critical characteristics of ABM, are not incorporated in the model.
Future updates to the model should integrate such dynamics through tighter integration between
SALUS and ABM. This enables the crop model to receive farming information from ABM in
real-time.

Similarly, post-harvest crop loss and loss due to pests and diseases are not considered in the
model. Households consume fruits, vegetables, and animal products. However, non-cereal food is

69

Figure 3.13: Spatial distribution of vulnerability to food insecurity

not included in the calorie calculation. Additionally, calorie requirements and access are considered
uniform across household members.

Almost all farm households in Koutiala cultivate cotton. In our model, we did not explicitly
incorporate non-farm income and income from cash-crops. According to Cooper and West (2017),
only a small portion of such income is used for food purchases. Our model considers households
having minor food deﬁcits (less than 5% of total food demand) as food secure, assuming that such
income will cover these shortages.

70

Figure 3.14: Vulnerability to food insecurity by household types

71

Furthermore, the model assumes a constant population growth. In reality, resource availability
dictates growth (A. Ford and F. A. Ford 1999). The population might not be able to grow
continuously at the current rate under food scarcity. Similarly, the ABM agents are farm households
only; non-farm households and other stakeholders are not considered in the model. Other agents of
the food security system like government/non-government agencies, credit agencies, and extension
workers should also be included in the updated model.

3.5 Conclusion

Productionist arguments dominate food security discourse. Food production is, indeed, essential
for food security. However, adequate food access is also necessary, especially for attaining food
security at the household level. Household food security analysis, therefore, primarily needs to look
at both household food production and access. We integrate a crop model with an ABM of household
food systems for a comprehensive evaluation of household food security. The coupled model allows
us to explore food security at the intersection of biophysical, climatic, and socioeconomic systems.
Additionally, the model has the granularity, heterogeneity, and complexities needed to eﬀectively
analyze future food security at the household level.

Households usually ﬁnd ways to address food scarcity at home. The measure they use reﬂects
not only the current severity but also their future vulnerability. Food security is inherently dynamic
and non-binary. The use of FSVI integrated with ABM allows us to explore the dynamics and
nuances of household food security. We ﬁnd that the majority of regional households will be
food insecure in the future. Large families, especially, are more susceptible to food insecurity.
In contrast, smaller households are prone to higher severity of food insecurity. Also, the spatial
feature of the coupled model highlights the areas prone to food insecurity. Such information has
broad policy implications, especially for target identiﬁcation and intervention.

72

APPENDICES

73

PICTORIAL CONCEPTUAL DIAGRAM USED FOR STAKEHOLDER EVALUATION

APPENDIX D

Figure D.1: Pictorial conceptual diagram used for stakeholder evaluation

74

APPENDIX E

ABM DESCRIPTION IN ODD FORMAT

The model description follows the ODD (Overview, Design concepts, Details) protocol for describ-
ing individual- and agent-based models.

E.1 Purpose

The ABM explores the eﬀect of biophysical and socioeconomic factors on household vulnera-
bility to food insecurity. For a comprehensive food security analysis, the model looks at household
food availability, access, utilization, and stability. It is coupled with a process-based crop model
that provides the information on household food productivity. Using the information, the ABM
evaluates household food production, checks for suﬃciency, and looks for coping options during
food shortages. Household vulnerability is measured by evaluating recovery potential of the options
used for coping food shortages.

E.2 Entities, state variables, and scales

Agents: Agents represents rural farm households in Koutiala district of Southern Mali. They
are represented in the model by vector polygons (i.e., their farmland) of diﬀerent shapes and sizes
distributed randomly across the landscape. Households work towards securing food for their family.
There are the following four types of households (Falconnier et al. 2015):

• High Resource Endowed – Large Herd size (HRE_LH) households

• High Resource Endowed (HRE) households

• Medium Resource Endowed (MRE) households

• Low Resource Endowed (LRE) households

The agents contain following state variables:

75

Entity

Sub-group Variable name Description

Table E.1: State variables used in the ABM

size of the households
number of the livestock
household land size
neighborhood distance
portion of food saved for supporting the neighbors
Food Security Vulnerability Index

Possible Values Units/Notes
45.3 (18)
5.7 (1.5)
16.6 (8.3)
2000,3000,4000 meters
0,0.5,0.10,0.20 %
I - IV

>4 in a household
>= 0
Ha / (2-100 ha)

Household

HRE_LH

HRE

MRE

LRE

HH_size
Livestk
Land_ha
Neigh_dist
Neigh_support
FSVI

HH_size
Livestk
Land_ha
Neigh_dist
Neigh_support
FSVI

HH_size
Livestk
Land_ha
Neigh_dist
Neigh_support
FSVI

HH_size
Livestk
Land_ha
Neigh_dist
Neigh_support
FSVI

size of the households
number of the livestock
household land size
neighborhood distance
portion of food saved for supporting the neighbors
Food Security Vulnerability Index

>4 in a household
>= 0
Ha / (2-100 ha)

27.3 (6.6)
2.7 (1.6)
11.8 (6.2)
2000,3000,4000 meters
0,0.5,0.10,0.20 %
I - IV

size of the households
number of the livestock
household land size
neighborhood distance
portion of food saved for supporting the neighbors
Food Security Vulnerability Index

>4 in a household
>= 0
Ha / (2-100 ha)

12.6 (1.7)
4.1 (6.5)
7.5 (2.4)
2000,3000,4000 meters
0,0.5,0.10,0.20 %
I - IV

size of the households
number of the livestock
household land size
neighborhood distance
portion of food saved for supporting the neighbors
Food Security Vulnerability Index

>4 in a household
>= 0
Ha / (2-100)

7.9 (4.9)
1.3 (1)
3.2 (2)
2000,3000,4000 meters
0,0.5,0.10,0.20 %
I - IV

76

Spatial Units: One household owns only one land parcel. The land parcels are of various sizes
ranging from 2 ha to 100 ha. The crop model simulates biophysical and climatic conditions to
estimate crop productivity of the land parcels.

Temporal Units: The model is run for 41 years (from 2010 to 2050) and one timestep represents

one year.

Environment: Since crop production inﬂuence households’ interactions and behavior, crop

yield represents model environment.

E.3 Process overview and scheduling

Every time step, the model runs in a predeﬁned order (Figure 1). First, a deﬁned number of land
parcels representing the households are placed randomly in space. Model then calculates cultivable
area and food production for each household. Households with adequate food supply are identiﬁed
as food secure. Households with food shortages look for options to secure food for the family.

get_cultivable_area: The model ﬁrst estimates cultivable area of each household. Since farming
is predominantly manual, depending upon the farm workers at home, cultivable land may be smaller
than total farmland. Famers have one crop cycle a year; they divide the cultivable area for maize,
millet, and sorghum cultivation.

get_status: Once the ABM calculates household food production, it converts the available food
to calorie supply and checks for its suﬃciency. Households with positive calorie suﬃciency (calorie
suﬃciency = calorie supply – calorie demand) are tagged as food secure (FSVI I) households.

get_inkind: Food insecure households seek for in-kind support from their neighbors. Every
household is provided with a predeﬁned neighborhood size. Neighbors with surplus food contribute
a small portion from their food storage.

use_cattle: Households with more severe food shortages sell cattle to buy food. However,

households do not sell all their livestock as they need some for farming.

oﬀ_farm: Households also send some family members to oﬀ-farm jobs to be able to purchase
food. Some members decide to migrate that creates labor shortages at home. As households do

77

Figure E.1: ABM ﬂow diagram

78

not want to lose their farm workers, oﬀ-farm job is used as the last available option.

E.4 Design concept

Basic principles: Vulnerability is deﬁned by exposure to risk, capacity to cope, and ability
to recover (Watts and Bohle 1993). Households starts to cope when faced with food shortages.
Coping mechanisms usually start with high accessible low impact options. Households gradually
move to low accessible high impact options as severity of food shortages increases. Such options
may hinder households capacity to cope in the future. The ABM checks household food suﬃciency,
looks for available coping options and measures the vulnerability based on the recovery potential.
Emergence: The spatial and temporal distribution of household vulnerability to food insecurity
is an emergent phenomenon that results from households’ interactions with each other and their
environment.

Adaptation: During food shortages, households ﬁrst start with low-impact coping options like
skipping or rationing meals and getting in-kind help from relatives and neighbors, and slowly move
to high-impact alternatives.

Objectives: The overall objective of the households is to secure enough food for the family

without impairing their ability to ﬁght future vulnerabilities.

Interaction: Households interact with each other and the environment. Environment (i.e., crop
productivity) inﬂuences household food suﬃciency that aﬀects household behavior. Households
interact with other households for food (for in-kind support). Households also go for oﬀ-farm jobs
and to market to buy food or sell livestock.

Stochasticity: Stochasticity is incorporated at various places in the model. First, the land
parcels’ shape, size, and the distribution over the landscape are random. Additionally, the model
populates agents’ attributes from the predeﬁned probability distribution (Table 1).

Observation: The model measures household vulnerability to food insecurity using Food
Security Vulnerability Index (FSVI). Households with FSVI I are self-suﬃcient and food secure.
FSVI II households face slight food shortages that can be addressed through meal rationing or

79

in-kind help. FSVI III households need to purchase food. They might need to go for oﬀ-farm jobs
or sell their livestock to be able to purchase food. FSVI IV households neither have suﬃcient food
production nor they have resources to purchase food. The model shows the temporal distribution of
household vulnerability to food insecurity. Since this is coupled with a process-based crop model,
we also see the impacts of climate change on household food production and security. Additionally,
as the model is spatial in nature, we can also see the distribution of household vulnerabilities across
space.

E.5 Initialization

The model consists of 9000 agents, amongst them 1170 are HRE_LH, 2520 are HRE, 1710
are LRE, and 3600 are MRE. Land parcels representing the agents are randomly placed within
the cultivable area of the landscape. We use ‘Thiessen Polygon’ feature in ArcGIS to generate the
distributions of land parcels. We have above 2000 realizations of GIS model and ABM picks one
land parcel distribution (at random) at the start of a run. ABM divides households’ cultivable areas
to parcels for maize (20%), millet (40%), and sorghum (40%). The model receives productivity
information from the crop model. Other attributes are populated from the provided distribution
(Table 2).

80

Source
International Soil Reference and Information Center
(http://www.isric.org)

Marksim DSSAT weather ﬁle generator (http://
gismap.ciat.cgiar.org/MarkSimGCM)

Focus Group Discussion/ Expert Advice/ (Soumaré
et al. 2002)

Table E.2: SALUS Parameters

SALUS Parameterization
Soil

Parameters

Physical Properties

Bulk density, Clay content,
Coarse fragments, Sand, Silt

Chemical Properties Cation exchange capacity, Nitro-

gen, Soil Organic Carbon, pH

Weather/Climate

Parameters

Weather data

Future Climate

Tmax, Tmin, Rainfall, Solar ra-
diation
RCP 8.5 (HadGEM2-ES)

Farm Management Parameters

Resolution

250m

250m

Resolution

Daily

-

Value

Crops

Input

Maize, Millet, Sorghum

Barnyard Compost

20%, 40%, 40% of total
land respectively
1T/ha

81

Table E.3: ABM Parameters

ABM Parameterization
Total Farm Households (No of
Agents)
HH typologies

High Resource Endowed - Large
Herdsize (HRE_LH)
High Resource Endowed (HRE)
Medium Resource Endowed
(MRE)
Low Resource Endowed (LRE)
Food supply
Calorie demand

Calorie supply

Sale of 1 cattle provides

Food Production
Workforce available
Worker-days required (Crops)
Worker-days required (Cotton)
Food Access
Neighborhood distance

Neighborhood support

Distribution

13%

28%
40%

19%
Values
2400 per person

per

3100
kg
maize; 3550 per kg
sorghum/millet
1285 kg of grain

Values
1/2*hh_size
63
149
Distribution
Uniform
2000,3000,4000
(meters)
Discrete (0%, 5%,
10%, 20%) % of
surplus

9000

Land_size:
mean (std)
16.6 (8.33)

11.8 (6)
7.5 (2.4)

3.2 (2)

HH_Size:
mean (std)
45.3 (18)

27.3 (6.6)
12.6 (1.7)

7.9 (4.9)

Livestk:
mean (std)
5.7 (1.5)

2.7 (1.6)
4.1 (6.5)

1.3 (1)

Source
Census (2009)

(Falconnier et al. 2015)

FAO
(http://www.fao.org/)

Derived from ICRISAT,
2017

(Ruben et al. 1997)

based on Focus Group
Discussion

P(0.1,0.5,0.3,0.1)

82

E.6 Input data

The ABM is coupled with SALUS - a process-based crop model. SALUS requires soil, climate
crop, and farm management information to predict crop productivity (Basso, Ritchie, et al. 2006).
SALUS is provided with information on soil biophysical and chemical properties, climate, and farm
management. Additionally, we use RCP 8.5 to represent future climate scenario. Farm information
is obtained from focus group discussions in the ﬁeld. SALUS provides estimated crop productivity
for each land parcel for every year until 2050.

E.7 Submodels

get_cultivable_area: Since households are essentially into manual farming, available farm labor
at home determines the areas that a household can cultivate . According to Ruben et al. (1997), only
½ of household members can be considered as farm labors. They also suggest that cultivating 1 ha
of cereal crop requires 63 man-days in the region. Assuming a farm labor can work for 120 days
(a crop season) per year, we estimate cultivable_area of the household by: Man-days available = ½
household * 120 Man-days required = 63 cultivable_area = man-days available/man-days required
get_status: This simply calculates household food suﬃciency. Households with positive caloric
suﬃciency (calorie suﬃciency = calorie supply – calorie demand) are categorized as food suﬃcient.
get_inkind: Households with food shortages initially try low-impact options for securing food.
If a household has a low food shortage (less than 5% of total food demand), it skips/rations to
address the problem. However, if it suﬀers from larger food shortages, it seeks for in-kind help
from neighbors or relatives. Every household has deﬁned a neighborhood area (walkable distance
- ranging from 2 to 4 km). It looks for help from food secure households within the neighborhood
area. Neighbors usually donate a small portion (0-20%) from their surplus food.

use_cattle: When neighbors’ support is not enough, households buy food from market. Since
most of the households are cash-constrained, they might have to sell their livestock to be able to
purchase food. Every time step, a household sells one of their livestock (cattle) to purchase food.
They will keep at least 2 livestock at home for farm work. Using ICRISAT data (Badolo 2017), we

83

estimate the average price of 1 kg crop (140 FCFA ) and unit price for adult cattle (180,000 FCFA).
Using the information, we estimate the amount of food that can be purchased by selling one cattle.
We also assume a uniform inﬂation across commodities in the future (i.e., prices for cattle and food
increases/decreases at the same rates).

oﬀ_farm: In addition to sale of livestock, households also send their members for oﬀ-farm jobs.
We assume that the oﬀ-farm jobs result in food suﬃciency at home. However, based on the focus
group discussion, we learned that around 15% of members going for oﬀ-farm jobs do not return.
Such out-migration results in lower food demand, but also decreases farm labor at home.

84

BIBLIOGRAPHY

85

BIBLIOGRAPHY

An, Li et al. (2005). “Exploring Complexity in a Human–Environment System: An Agent-Based
Spatial Model for Multidisciplinary and Multiscale Integration”. In: Annals of the association
of American geographers 95.1, pp. 54–79. issn: 1467-8306.

Asseng, S. et al. (2017). “Hot spots of wheat yield decline with rising temperatures”. In: Glob
Chang Biol 23.6, pp. 2464–2472. issn: 1365-2486 (Electronic) 1354-1013 (Linking). doi:
10.1111/gcb.13530. url: https://www.ncbi.nlm.nih.gov/pubmed/27860004.

Asseng, S et al. (2014). “Simulation modeling: applications in cropping systems”. In: issn:

0080931391.

Badolo, Felix (2017). “Monitoring of market prices of agricultural products in Bougouni and

Koutiala, Mali”. In:

Basso, Bruno, Anthony D Kendall, and David W Hyndman (2013). “The future of agriculture over
the Ogallala Aquifer: Solutions to grow crops more eﬃciently with limited water”. In: Earth’s
Future 1.1, pp. 39–41. issn: 2328-4277.

Basso, Bruno and Lin Liu (2019). “Seasonal crop yield forecast: Methods, applications, and

accuracies”. In: Advances in Agronomy. Vol. 154. Elsevier, pp. 201–255. isbn: 0065-2113.

Basso, Bruno and Joe T Ritchie (2015). “Simulating crop growth and biogeochemical ﬂuxes
in response to land management using the SALUS model”. In: The ecology of agricultural
landscapes: long-term research on the path to sustainability. Oxford University Press, New
York, NY USA, pp. 252–274.

Basso, Bruno, Joe T Ritchie, et al. (2006). “Simulation of tillage systems impact on soil biophysical
properties using the SALUS model”. In: Italian Journal of Agronomy, pp. 677–688. issn:
2039-6805.

Basso, Bruno, Luigi Sartori, et al. (2012). “Environmental and economic evaluation of N fertilizer
rates in a maize crop in Italy: A spatial and temporal analysis using crop models”. In: Biosystems
Engineering 113.2, pp. 103–111. issn: 1537-5110.

Bassu, Simona et al. (2014). “How do various maize crop models vary in their responses to
climate change factors?” In: Global change biology 20.7, pp. 2301–2320. issn: 1365-2486.
url: https://onlinelibrary.wiley.com/doi/abs/10.1111/gcb.12520.

Berger, T. and C. Troost (2014). “Agent-based Modelling of Climate Adaptation and Mitigation
Options in Agriculture”. In: Journal of Agricultural Economics 65.2, pp. 323–348. issn: 0021-
857x. doi: 10.1111/1477-9552.12045.

86

Bickel, G et al. (2000). Guide to measuring household food security. Washington, DC: US Depart-

ment of Agriculture. Journal Article.

Bingen, R James (1998). “Cotton, democracy and development in Mali”. In: The Journal of Modern

African Studies 36.2, pp. 265–285. issn: 0022-278X.

Bithell, Mike, James Brasington, and Keith Richards (2008). “Discrete-element, individual-based
and agent-based models: Tools for interdisciplinary enquiry in geography?” In: Geoforum 39.2,
pp. 625–642. issn: 0016-7185. url: http://ac.els- cdn.com.proxy2.cl.msu.edu/
S0016718506001667/1-s2.0-S0016718506001667-main.pdf?_tid=f981ca02-98a4-
11e7-a689-00000aacb361&acdnat=1505322178_c67dc0ade85c1b60e56ab89bf49b58a7.

Bohle, Hans G, Thomas E Downing, and Michael J Watts (1994). “Climate change and social
vulnerability: toward a sociology and geography of food insecurity”. In: Global environmental
change 4.1, pp. 37–48. issn: 0959-3780.

Brown, Molly E, Beat Hintermann, and Nathaniel Higgins (2009). “Markets, climate change, and
food security in West Africa”. In: Environ. Sci. Technol. 43.21, pp. 8016–8020. issn: 0013-
936X.

Burg, Jericho (2008). “Measuring populations’ vulnerabilities for famine and food security inter-
ventions: the case of Ethiopia’s Chronic Vulnerability Index”. In: Disasters 32.4, pp. 609–630.
issn: 0361-3666.

Chambers, Robert (1989). “Editorial introduction: vulnerability, coping and policy”. In: issn:

1759-5436.

Clapp, Jennifer (2014). “Food security and food sovereignty: Getting past the binary”. In: Dialogues

in Human Geography 4.2, pp. 206–211. issn: 2043-8206.

Cooper, M. W. and C. T. West (2017). “Unraveling the Sikasso Paradox: Agricultural Change
and Malnutrition in Sikasso, Mali”. In: Ecol Food Nutr 56.2, pp. 101–123. issn: 1543-5237
(Electronic) 0367-0244 (Linking). doi: 10.1080/03670244.2016.1263947. url: https:
//www.ncbi.nlm.nih.gov/pubmed/27976918.

Dembele, Bandiougou et al. (2017). “Dynamics and adaptation of agricultural farming systems in
the boost of cotton cropping in southern Mali”. In: African Journal of Agricultural Research
12.18, pp. 1552–1569. issn: 1991-637X.

Dijk, Han van, Mirjam de Bruijn, and Wouter van Beek (2004). “Pathways to mitigate climate
variability and climate change in Mali: The districts of Douentza and Koutiala compared”. In:
The Impact of Climate Change on Drylands. Springer, pp. 173–206.

Djurfeldt, Göran et al. (2005). The African food crisis: lessons from the Asian Green Revolution.

Cabi. isbn: 1845930975.

87

Dobbie, Samantha et al. (2018). “Agent-based modelling to assess community food security and
sustainable livelihoods”. In: Journal of Artiﬁcial Societies and Social Simulation 21.1, pp. 1–23.

Eozenou, Patrick, Dorsati Madani, and Rob Swinkels (2013). Poverty, malnutrition and vulnera-

bility in Mali. The World Bank. isbn: 1813-9450.

Falconnier, Gatien N et al. (2015). “Understanding farm trajectories and development pathways:
Two decades of change in southern Mali”. In: Agricultural Systems 139, pp. 210–222. issn:
0308-521X.

FAO, An (2008). “An introduction to the basic concepts of food security”. In: FAO, Rome, Italy.

Farvacque-Vitkovic, Catherine et al. (2007). “Development of the cities of Mali: Challenges and

priorities”. In:

Flanagan, Barry E et al. (2011). “A social vulnerability index for disaster management”. In: Journal

of homeland security and emergency management 8.1. issn: 2194-6361.

Ford, Andrew and Frederick Andrew Ford (1999). Modeling the environment: an introduction to

system dynamics models of environmental systems. Island press. isbn: 1559636017.

Fouilleux, Eve, Nicolas Bricas, and Arlène Alpha (2017). “‘Feeding 9 billion people’: Global food
security debates and the productionist trap”. In: Journal of European Public Policy 24.11,
pp. 1658–1677. issn: 1350-1763.

Füssel, Hans-Martin and Richard JT Klein (2006). “Climate change vulnerability assessments: an
evolution of conceptual thinking”. In: Climatic change 75.3, pp. 301–329. issn: 0165-0009.
url: https://link.springer.com/article/10.1007%2Fs10584-006-0329-3.

Gebbers, Robin and Viacheslav I Adamchuk (2010). “Precision agriculture and food security”. In:

Science 327.5967, pp. 828–831. issn: 0036-8075.

Gilbert, Nigel (2019). Agent-based models. Vol. 153. Sage Publications, Incorporated. isbn:

1506355595.

Grimm, V. et al. (2006). “A standard protocol for describing individual-based and agent-based
models”. In: Ecological Modelling 198.1-2, pp. 115–126. issn: 0304-3800. doi: 10.1016/j.
ecolmodel.2006.04.023. url: %3CGo%20to%20ISI%3E://WOS:000240823100008.

Grimm, Volker et al. (2010). “The ODD protocol: a review and ﬁrst update”. In: Ecological

modelling 221.23, pp. 2760–2768. issn: 0304-3800.

Guillaumont, Patrick (2009). “An economic vulnerability index: its design and use for international

development policy”. In: Oxford Development Studies 37.3, pp. 193–228. issn: 1360-0818.

88

Hammond, R. A. and L. Dube (2012). “A systems science perspective and transdisciplinary models
for food and nutrition security”. In: Proc Natl Acad Sci U S A 109.31, pp. 12356–63. issn:
1091-6490 (Electronic) 0027-8424 (Linking). doi: 10.1073/pnas.0913003109. url: http:
//www.ncbi.nlm.nih.gov/pubmed/22826247.

Hoddinott, John (1999). Choosing outcome indicators of household food security. International

Food Policy Research Institute Washington, DC.

Jenkins, J. C. and S. J. Scanlan (2001). “Food security in less developed countries, 1970 to
1990”. In: American Sociological Review 66.5, pp. 718–744. issn: 0003-1224. doi: Doi10.
2307 / 3088955. url: %3CGo % 20to % 20ISI % 3E : / / WOS : 000172062400005 % 20http :
//www.jstor.org/stable/3088955.

Jones, CDea et al. (2011). “The HadGEM2-ES implementation of CMIP5 centennial simulations”.

In: Geoscientiﬁc Model Development 4, pp. 543–570.

Jones, J. W. et al. (2017). “Brief history of agricultural systems modeling”. In: Agric Syst 155,
pp. 240–254. issn: 0308-521X (Print) 0308-521X (Linking). doi: 10.1016/j.agsy.2016.
05.014. url: https://www.ncbi.nlm.nih.gov/pubmed/28701816.

Jones, James W et al. (2017). “Toward a new generation of agricultural system data, models,
and knowledge products: State of agricultural systems science”. In: Agricultural systems 155,
pp. 269–288. issn: 0308-521X. url: https://www.ncbi.nlm.nih.gov/pmc/articles/
PMC5485672/pdf/main.pdf.

Kelly, P Mick and W Neil Adger (2000). “Theory and practice in assessing vulnerability to climate
change andFacilitating adaptation”. In: Climatic change 47.4, pp. 325–352. issn: 0165-0009.
url: https://link.springer.com/article/10.1023%2FA%3A1005627828199.

Kennedy, William G (2012). “Modelling human behaviour in agent-based models”. In: Agent-based

models of geographical systems. Springer, pp. 167–179.

Koh, Keumseok, Rebecca Reno, and Ayaz Hyder (2019). “Examining disparities in food accessi-
bility among households in Columbus, Ohio: an agent-based model”. In: Food Security 11.2,
pp. 317–331. issn: 1876-4517.

Krishnamurthy, PK, K Lewis, and RJ Choularton (2014). “A methodological framework for rapidly
assessing the impacts of climate risk on national-level food security through a vulnerability
index”. In: Global Environmental Change 25, pp. 121–132. issn: 0959-3780.

Ligmann-Zielinska, Arika (2009). “The impact of risk-taking attitudes on a land use pattern: an
agent-based model of residential development”. In: Journal of Land Use Science 4.4, pp. 215–
232. issn: 1747-423X.

89

Liu, Lin and Bruno Basso (2017). “Spatial evaluation of maize yield in Malawi”. In: Agricultural

Systems 157, pp. 185–192. issn: 0308-521X.

Magliocca, Nicholas R, Daniel G Brown, and Erle C Ellis (2013). “Exploring agricultural livelihood
transitions with an agent-based virtual laboratory: global forces to local decision-making”. In:
PLoS One 8.9, e73241. issn: 1932-6203. url: https://www.ncbi.nlm.nih.gov/pmc/
articles/PMC3764159/pdf/pone.0073241.pdf.

Marohn, Carsten et al. (2013). “A software coupling approach to assess low-cost soil conservation
strategies for highland agriculture in Vietnam”. In: Environmental Modelling & Software 45,
pp. 116–128. issn: 1364-8152.

Matthews, Robin (2006). “The People and Landscape Model (PALM): Towards full integration of
human decision-making and biophysical simulation models”. In: Ecological Modelling 194.4,
pp. 329–343. issn: 0304-3800.

Maxwell, Daniel G (1996). “Measuring food insecurity: the frequency and severity of “coping

strategies””. In: Food policy 21.3, pp. 291–303. issn: 0306-9192.

Maxwell, Daniel, Richard Caldwell, and Mark Langworthy (2008). “Measuring food insecurity:
Can an indicator based on localized coping behaviors be used to compare across contexts?” In:
Food Policy 33.6, pp. 533–540. issn: 0306-9192.

Maxwell, Daniel, Ben Watkins, et al. (2003). “The coping strategies index: A tool for rapidly
measuring food security and the impact of food aid programs in emergencies”. In: Nairobi:
CARE Eastern and Central Africa Regional Management Unit and the World Food Programme
Vulnerability Assessment and Mapping Unit.

Maxwell, Simon and Marisol Smith (1992). “Household food security: a conceptual review”. In:
Household Food Security: concepts, indicators, measurements. Edited by S. Maxwell and T.
Frankenberger. Rome and New York: IFAD and UNICEF.

N’Danikou, Sognigbe et al. (2017). “Foraging Is Determinant to Improve Smallholders’ Food

Security in Rural Areas in Mali, West Africa”. In: Sustainability 9.11, p. 2074.

Olabisi, Laura Schmitt et al. (2018). “Using participatory modeling processes to identify sources
of climate risk in West Africa”. In: Environment Systems and Decisions 38.1, pp. 23–32. issn:
2194-5403.

Otsuka, Keijiro and Rie Muraoka (2017). “A green revolution for sub-Saharan Africa: Past failures
and future prospects”. In: Journal of African Economies 26.suppl1, pp. i73–i98. issn: 0963-
8024.

90

Pandey, Rajiv and ShashidharKumar Jha (2012). “Climate vulnerability index-measure of climate
change vulnerability to communities: a case of rural Lower Himalaya, India”. In: Mitigation
and Adaptation Strategies for Global Change 17.5, pp. 487–506. issn: 1381-2386.

Parker, Dawn C, Amy Hessl, and Sarah C Davis (2008). “Complexity, land-use modeling, and
the human dimension: Fundamental challenges for mapping unknown outcome spaces”. In:
Geoforum 39.2, pp. 789–804. issn: 0016-7185.

Patricola, Christina M and Kerry H Cook (2010). “Northern African climate at the end of the
twenty-ﬁrst century: an integrated application of regional and global climate models”. In:
Climate Dynamics 35.1, pp. 193–212. issn: 0930-7575. url: https://link.springer.
com/content/pdf/10.1007%2Fs00382-009-0623-7.pdf.

Pinstrup-Andersen, Per (2009). “Food security: deﬁnition and measurement”. In: Food security 1.1,

pp. 5–7. issn: 1876-4517.

Polhill, J Gary, Lee-Ann Sutherland, and Nicholas M Gotts (2010). “Using qualitative evidence
to enhance an agent-based modelling system for studying land use change”. In: Journal of
Artiﬁcial Societies and Social Simulation 13.2, p. 10.

Riahi, Keywan et al. (2011). “RCP 8.5—A scenario of comparatively high greenhouse gas emis-

sions”. In: Climatic Change 109.1-2, p. 33. issn: 0165-0009.

Riely, Frank (2000). “A comparison of vulnerability analysis methods and rationale for their use in

diﬀerent contexts”. In: A FIVIMS synthesis document.

Ritchie, Joe T and Bruno Basso (2008). “Water use eﬃciency is not constant when crop water
supply is adequate or ﬁxed: The role of agronomic management”. In: European Journal of
Agronomy 28.3, pp. 273–281. issn: 1161-0301.

Rivers III, Louie et al. (2017). “Mental models of food security in rural Mali”. In: Environment

Systems and Decisions, pp. 1–19. issn: 2194-5403.

Roudier, Philippe et al. (2011). “The impact of future climate change on West African crop yields:
What does the recent literature say?” In: Global Environmental Change 21.3, pp. 1073–1083.
issn: 0959-3780.

Ruben, R. et al. (1997). “The impact of agrarian policies on sustainable land use”. In: Applications
of Systems Approaches at the Farm and Regional Levels Volume 1: Proceedings of the Second
International Symposium on Systems Approaches for Agricultural Development, held at IRRI,
Los Baños, Philippines, 6–8 December 1995. Ed. by P. S. Teng et al. Dordrecht: Springer
Netherlands, pp. 65–82. isbn: 978-94-011-5416-1. doi: 10.1007/978-94-011-5416-1_6.
url: https://doi.org/10.1007/978-94-011-5416-1_6.

91

Ruelland, Denis, Florent Levavasseur, and Antoine Tribotté (2010). “Patterns and dynamics of
land-cover changes since the 1960s over three experimental areas in Mali”. In: International
Journal of Applied Earth Observation and Geoinformation 12, S11–S17. issn: 0303-2434.

Saxena, Rachit et al. (2018). “Millets for food security in the context of climate change: A review”.

In: Sustainability 10.7, p. 2228.

Scaramozzino, Pasquale (2006). “Measuring vulnerability to food insecurity”. In: Agricultural
and Development Economics Division of the Food and Agriculture Organization of the United
Nations (FAO-ESA).

Schlenker, Wolfram and David B Lobell (2010). “Robust negative impacts of climate change on

African agriculture”. In: Environmental Research Letters 5.1, p. 014010. issn: 1748-9326.

Schreinemachers, Pepijn and Thomas Berger (2011). “An agent-based simulation model of hu-
man–environment interactions in agricultural systems”. In: Environmental Modelling & Soft-
ware 26.7, pp. 845–859. issn: 1364-8152. url: https : / / www . sciencedirect . com /
science/article/pii/S1364815211000314?via%3Dihub.

Schreinemachers, Pepijn, Thomas Berger, and Jens B Aune (2007). “Simulating soil fertility and
poverty dynamics in Uganda: A bio-economic multi-agent systems approach”. In: Ecological
economics 64.2, pp. 387–401. issn: 0921-8009.

Sen, Amartya (1981). “Ingredients of famine analysis: availability and entitlements”. In: The

quarterly journal of economics 96.3, pp. 433–464. issn: 1531-4650.

Sivakumar, MVK (1988). “Predicting rainy season potential from the onset of rains in Southern
Sahelian and Sudanian climatic zones of West Africa”. In: Agricultural and Forest Meteorology
42.4, pp. 295–305. issn: 0168-1923.

Soumaré, M. et al. (2002). “Chemical characteristics of Malian and Belgian solid waste composts”.
In: Bioresource Technology 81.2, pp. 97–101. issn: 0960-8524. doi: https://doi.org/10.
1016/S0960- 8524(01)00125- 0. url: http://www.sciencedirect.com/science/
article/pii/S0960852401001250.

Staatz, John M, Victoire C D’agostino, and Shelly Sundberg (1990). “Measuring food security
in Africa: conceptual, empirical, and policy issues”. In: American Journal of Agricultural
Economics 72.5, pp. 1311–1317. issn: 0002-9092.

Sultan, B. and M. Gaetani (2016). “Agriculture in West Africa in the Twenty-First Century: Climate
Change and Impacts Scenarios, and Potential for Adaptation”. In: Front Plant Sci 7, p. 1262.
issn: 1664-462X (Print) 1664-462X (Linking). doi: 10 . 3389 / fpls . 2016 . 01262. url:
https://www.ncbi.nlm.nih.gov/pubmed/27625660.

92

Traore, Bouba et al. (2013). “Eﬀects of climate variability and climate change on crop production

in southern Mali”. In: European Journal of Agronomy 49, pp. 115–125. issn: 1161-0301.

Watts, Michael J and Hans G Bohle (1993). “The space of vulnerability: the causal structure of

hunger and famine”. In: Progress in human geography 17.1, pp. 43–67. issn: 0309-1325.

Wossen, Tesfamicheal and Thomas Berger (2015). “Climate variability, food security and poverty:
Agent-based assessment of policy options for farm households in Northern Ghana”. In: En-
vironmental Science & Policy 47, pp. 95–107. issn: 1462-9011. doi: https://doi.org/
10.1016/j.envsci.2014.11.009. url: http://www.sciencedirect.com/science/
article/pii/S1462901114002226%20https://www.sciencedirect.com/science/
article/pii/S1462901114002226?via%3Dihub.

93

CHAPTER 4

GLOBAL OR LOCAL? EFFECTS OF POLICY INTERVENTIONS ON HOUSEHOLD
FOOD SECURITY IN KOUTIALA, SOUTHERN MALI: A COUPLED BIOPHYSICAL

AND SOCIAL SYSTEMS APPROACH.

Northern Mali is highly arid and not suitable for agriculture. The South is mainly responsible for
supplying food to the nation. However, climate change and rapid population growth in the South
are threatening local food production and food security. Regional food demand is likely to surpass
its production very soon. Besides, the majority of households in the South are likely to be food
insecure in the near future. This can tremendously aﬀect the national food supply. The South
needs to act soon to maintain its food security and the food supply to the nation. In this study,
we evaluate the eﬀectiveness of global and local level interventions in addressing household food
security in the region. We integrate a biophysical process-based crop model with an agent-based
model of household food systems to explore the trajectories of household food security under policy
interventions. We ﬁnd that the local level interventions that include improved access to agriculture
inputs and reduced population growth are highly eﬀective in promoting household food security
in the region. At least 60% of the households could remain food secure by mid-century under the
local level interventions. On the other hand, global level intervention that includes climate change
mitigation to RCP4.5 levels does not show immediate eﬀects on food production and food security
in the region.

4.1 Introduction

Mali is expected to be profoundly impacted by climate change. Climate models used by
the Intergovernmental Panel for Climate Change (IPCC) unanimously agree that the country will
experience warmer weather in the future (Traore et al. 2013). High temperature shortens the plant’s
phenological cycle and reduces biomass accumulation (S. Asseng et al. 2017; Basso, Kendall, and
Hyndman 2013) that can have a signiﬁcant impact on regional crop production. Maize, millet, and
sorghum are the major staple crops grown in the region. Since these crops are C4 plants, they are

94

less likely to be aﬀected by increased temperature (Roudier et al. 2011). However, every degree
rise in temperature can decrease maize productivity by 0.5T/ha (Bassu et al. 2014). Yields of other
C4 crops are also expected to decline under climate change (Schlenker and Lobell 2010; Sultan
et al. 2013). Since productivity in the region is currently very low, further loss in yields could be
detrimental to the regional food supply. Although Mali may see an increased rainfall in the future,
its impact on crop production is expected to be marginal (Sivakumar 1988).

Northern Mali is highly arid and not suitable for agriculture. Being more productive, the South
has the responsibility to feed the entire national population. Koutiala, our study area, lies in the
South and receives 600-1400mm of rainfall annually (refer to Figure 1.1) (Falconnier et al. 2015).
Farming is the prime activity here. Food grown locally gets exported to the North and other parts
of the nation. Since the region plays a signiﬁcant role in the national food supply, it is considered
the breadbasket of Mali (Bingen 1998). However, with climate change, the region is expected to
lose its productivity signiﬁcantly. Our study in the previous paper ﬁnds that the mid-century maize
and sorghum yields could fall by 40%. Additionally, with the current level of population growth,
Koutiala is likely to see a massive surge in food insecure population in the future. Around 80% of
the households will be at risk of food insecurity by 2050. Moreover, regional food demand is likely
to surpass its food production by 2025. This could signiﬁcantly impact food supply to the nation,
potentially jeopardizing the breadbasket status of the region. Mali needs to act soon to promote
food security in the region in order to maintain the national food supply.

Food security lies at the interface of biophysical, climatic, and socio-economic systems (Ham-
mond and Dube 2012). Besides, household food security is determined by household food access
and production. In this study, we couple a biophysical process-based crop model with an agent-
based model (ABM) of household food systems to explore the trajectories of household food
security in Koutiala. Using the coupled model, we test the eﬃciency of global and local level
interventions in promoting food production and security in the region. The crop model integrates
the dynamics of biophysical and climatic systems to simulate household food production, and the
ABM explores prevailing socio-economic systems to evaluate household food access. The study

95

has signiﬁcant policy implications in that, it explores future trajectories of household food security
and evaluates the eﬃcacies of selected interventions.

The process-based crop and the agent-based models are systems models (J. W. Jones et al. 2017;
Parker et al. 2003) and ideal tools for performing scenario analyses (Boero and Squazzoni 2005).
Unlike the models based on statistical relationships, systems models capture the interplay within
and across the associated dynamic systems (Basso and Liu 2019). Such complexity is essential for
eﬀective scenario analyses. ABM can simulate human decisions and interactions with great realism
and, hence, has been widely used as a virtual laboratory for social experimentations (Ligmann-
Zielinska and Jankowski 2007; Parker et al. 2003).
It allows modelers to explore micro-level
interactions responsible for macro events. Even simple ABMs can explain complex phenomena
(Epstein 2006). Conway’s Game of Life and Shelling’s Segregation are prime examples of simple
models explaining the emergence of macro-phenomena from micro-interactions (Axelrod 1997).
Complex and more realistic ABMs need empirically-based sophisticated algorithms for decision
and interaction. These models allow modelers to explore complex systems, perform modiﬁcations,
and evaluate alternative outcomes.

Farm ﬁeld experimentations may not always be feasible due to spatio-temporal and economic
limitations. Crop models provide cost-eﬀective, practical, and eﬃcient options to carry out agri-
cultural investigations (J. W. Jones et al. 2017). Process-based crop models have interconnected
modules containing a series of mathematical equations representing crop growth, soil nutrient
cycling, soil water balance, and temperature (Basso, Ritchie, et al. 2006). As the changes in
one module are reﬂected in others, these models allow researchers to evaluate the implications of
modiﬁcation in one system on others. Such analyses allow farmers and policymakers to identify
sustainable and eﬀective management options (Basso and Ritchie 2015). For example, Basso,
Sartori, et al. (2012), and Liu and Basso (2017) used crop models to evaluate changes in Nitrogen
application on crop production and environment. S. Asseng et al. (2017) used multiple crop mod-
els to check the eﬀects of changes in atmospheric temperature on wheat yield. Similarly, Basso,
Ritchie, et al. (2006) used a process-based crop model to evaluate the eﬀects of tillage diﬀerences

96

on soil conditions and crop yields. García-Vila and Fereres (2012), on the other hand, used a crop
model, in combination with an economic model, to evaluate sustainable water management at the
farm level.

The coupled model has added advantages over the decoupled crop and agent-based models.
Crop models do not consider the social side of farming and crop production (Antle et al. 2017;
James W Jones et al. 2017). Similarly, ABMs cannot incorporate crop growth and dynamics.
Our integrated model adds crop dynamics to the representation of a social system. This allows
modelers to explore human interactions and their implications on farming and crop production.
Literature contains only a few examples integrating crop models with ABM for analyzing the
eﬀects of farm management on household food production (Wossen and Berger 2015), economy
(Schreinemachers and Berger 2011; Schreinemachers, Berger, and Aune 2007), and environmental
management (Marohn et al. 2013). Crop modelers are increasingly aware of such limitations
demanding more integration with social systems (James W Jones et al. 2017).

4.2 The coupled model

We develop the ABM of household food systems using information from ﬁeld interviews and
focus group discussions. The ﬁeld interviews are useful for developing the ABM structure, while
the focus group discussions help us formulate decision making in the model. Agents in the ABM
represent farm households from Koutiala. We have 9000 agents in the models represented in the
landscape by their land parcels. We develop a GIS model to generate distributions of land parcels
as vector polygons of diﬀerent shapes and sizes. Since the agents are farmers, the land parcels are
placed in arable land. Additionally, land parcels are smaller and densely distributed in and around
urban areas reﬂecting their actual spatial distribution (N’Danikou et al. 2017). Households are
classiﬁed into the following four groups based on their land sizes (Falconnier et al. 2015):

1. High Resource Endowed – Large Herds (LRE_LH) households

2. High Resource Endowed (HRE) households

97

3. Medium Resource Endowed (MRE) households

4. Low Resource Endowed (LRE) households

We select the Systems Approach to Land Use Sustainability (SALUS) model as our crop model.
SALUS is a process-based crop model that requires information on daily weather conditions,
physical and chemical properties at each layer of soil, and information on farm management to
simulate daily plant growth and development. Detailed information on SALUS can be found at
https://basso.ees.msu.edu/salus/index.html. We use SALUS to simulate household cultivation of
maize, millet, and sorghum in the region. The model structure, and SALUS and ABM parameters
are presented in Figure 3.6 and Table 3.3 and Table 3.4 of Chapter 3, respectively. Additionally, we
present detailed information on ABM in ODD format (V. Grimm et al. 2006; Volker Grimm et al.
2010) in Appendix E.

At the beginning of a model run, the ABM randomly selects a set of land parcel distributions
generated by the GIS model.
It then receives crop productivity from SALUS and interpolates
the values to land parcels. ABM calculates household food suﬃciency based on food production
and household food demand. We develop a Food Security Vulnerability Index (FSVI) to measure
household food security. FSVI is a coping-based vulnerability index that measures the severity of
food security based on applied coping mechanisms by the households. Households generally use
diﬀerent coping mechanisms when faced with food scarcity. Such mechanisms range from low
impact alternatives such as meal rationing or neighborhood support to high impact alternatives like
sale of household assets, oﬀ-farm jobs, and migration. Their application reﬂects the severity of
households’ food scarcity and their future vulnerability (D. Maxwell, Caldwell, and Langworthy
2008; D. G. Maxwell 1996; D. Maxwell, Watkins, et al. 2003). The FSVI has four severity levels
(Table 3.1).

4.3 Scenario analysis

Farmers in the region increasingly worry about low soil fertility and crop productivity (Olabisi
et al. 2018; Rivers III et al. 2017). Poor access to input and farm management are major factors

98

responsible for productivity decline. Farmers are also concerned about sporadic rainfall in the
region (Rivers III et al. 2017). Climate change was one of the signiﬁcant issues local farmers raised
during our focus group discussions. Besides, the rapidly growing population is increasing regional
food demand. Farmers demand a future where they have increased access to input and water
availability for increased food production (Olabisi et al. 2018) to feed their household population.
Based on farmers’ concerns and observed issues in the region, we develop the following three
interventions grouped by their spatial scale.

4.3.1

Intervention at the global level

4.3.1.1 Climate change mitigation (CCM)

Since Mali is likely to be profoundly aﬀected by climate change, it is an essential issue for regional
food security and food production. As seen in the previous paper (Figure 3.7), under Representative
Concentration Pathway (RCP) 8.5, the mid-century temperature could rise by 2◦C. That can have
detrimental impacts on regional food production. Although future rainfall largely remains the
same, the increased temperature can in turn increase evapotranspiration that can exacerbate the
regional water problem. Malian alone cannot contribute to climate change mitigation as it requires
a global eﬀort. With the realization that climate change is currently unavoidable, we consider a less
severe climate change, i.e., RCP 4.5, under which, emission of greenhouse gases is curbed. This
can positively contribute to regional food production by lowering the temperature and its resulting
evapotranspiration.

4.3.2

Interventions at the local level

4.3.2.1 Lower population growth (LPG)

Population growth rate in Koutiala, currently at 5% per annum, is higher than the national average.
Households in the region are usually big, containing extended family living together and working
in collective farms (Rivers III et al. 2017). Since farming is mainly manual, large households are

99

generally preferred. For locals, a large family means more people to work in the ﬁeld and go for oﬀ-
farm jobs. Remittances from oﬀ-farm jobs generally support household food purchases (Eozenou,
Madani, and Swinkels 2013; Rivers III et al. 2017). Since the large family is considered beneﬁcial,
population growth is often a non-issue here. However, the increase in population increases food
demand, and for food constrained households having a large family might not always be beneﬁcial.
On the other hand, small families have a small agricultural workforce that can negatively impact
household food production. It is crucial to explore the tradeoﬀs of population growth on household
food production and demand. For this scenario, we set the population growth equal to the national
average of 3% per annum.

4.3.2.2

Improved access to agricultural inputs

Koutiala has a large cotton industry that provides logistic supports in the form of credits, training,
and agricultural inputs to cotton farmers in the region. The technologies and inputs used for cotton
also help with crop production. Hence, to become its beneﬁciaries, most of the households in the
region choose cotton production (Dembele et al. 2017). Households allocate a small portion of their
farms for cotton plantation. However, support from the cotton industry is not usually suﬃcient for
the entire farms. Since most of the farmers are food poor, they severely lack resources to purchase
agricultural input. Households usually prepare manure using household and farm residues, but its
application is often insuﬃcient (Traore et al. 2013).

Under this scenario, we increase the application of input and check its impact on household

food production and food security. We test the following two sub-scenarios:

• Mixed access to input (MAI): Since resourceful households usually have better access to inputs
than resource-scarce households, HRE_LH and HRE households apply 5T/ha of manure in
their ﬁelds. In contrast, households with low endowments (i.e., MRE and LRE households)
apply 1T/ha.

• High access to input (HAI): All households have high access to inputs; all apply 5T/ha of

100

Table 4.1: Combinations of intervention scenarios

Scenario ID

Assumptions

Access to Input

Business-as-usual
Climate Change Mitigation
High Access to Input
Low Population Growth
Climate Change Mitigation + Low
Population Growth
High Access to Input + Climate
Change Mitigation

Low (1T/ha)
Low (1T/ha)
High (5T/ha)
Low (1T/ha)
Low (1T/ha)

High (5T/ha)

High (5T/ha)

Population
Growth

Climate
Change
RCP8.5 5%
RCP4.5 5%
RCP8.5 5%
RCP8.5 3%
RCP4.5 3%

RCP4.5 5%

RCP4.5 3%

BAU
CCM
HAI
LPG
CCM+LPG

HAI+CCM

HAI+LPG

MAI+LPG

HAI+CCM+LPG High Access to Input + Climate
Change Mitigation + Low Popula-
tion Growth
High Access to Input + Low Popu-
lation Growth
Mixed Access to Input
Mixed Access to Input + Climate
Change Mitigation
Mixed Access to Input + Low Pop-
ulation Growth

MAI
MAI+CCM

MAI+CCM+LPGMixed Access to Input + Climate
Change Mitigation + Low Popula-
tion Growth

High (1T/ha)

RCP8.5 3%

Mixed (5T/ha, 1T/ha) RCP8.5 5%
Mixed (5T/ha, 1T/ha) RCP4.5 5%

Mixed (5T/ha, 1T/ha) RCP8.5 3%

Mixed (5T/ha, 1T/ha) RCP4.5 3%

manure in their ﬁelds.

4.4 Results and discussion

In paper 2, we simulated a business-as-usual (BAU) scenario using high emission climate
change (RCP 8.5), low access to input (1T/ha) and high population growth rate (5%). In this study,
we ran multiple simulations for diﬀerent combinations of scenarios (Table 1) and evaluated their
impacts on household food production and security. Each simulation was run for 40 years (from
2010 to 2050) with a one-year timestep. For each scenario, we ran 50 simulations. We used High
Performance Computing (HPC) at Michigan State University for the simulations. In this section,
we present only the signiﬁcant outputs of the above scenarios.

101

4.4.1

Intervention at global level

4.4.1.1 Climate change mitigation

The future climate under this scenario is highly similar to that of BAU (Figure 4.1). Although there
are visible interannual variabilities, the overall rainfall pattern remains the same. However, the
rise in temperature is lower under the RCP 4.5 climate change scenario. Since high temperature
is detrimental to crop productivity, we expected to see an increase in yields under milder RCP 4.5
climate. Surprisingly, the crop model did not show signiﬁcant variations in crop yields (Figure 4.2).
In addition to the lower temperature, RCP 4.5 has lower ambient CO2. CO2 in the atmosphere acts
as a natural fertilizer (Senthold Asseng et al. 2013) and plays a signiﬁcant role in crop productivity,
especially in dry climates (Asseng et al. 2004). Additionally, the increased level of ambient CO2
reduces evapotranspiration (Asseng et al. 2004). Hence, we believe that the lower level of CO2
under CCM largely neutralizes the beneﬁts of lower temperatures in the atmosphere. Since the
productivity did not change, household food security under CCM (Figure 4.3:1b) remained similar
to that of BAU (Figure 4.3:1d).

4.4.2

Interventions at local/regional levels

4.4.2.1 Lower Population Growth

The coupled model showed a signiﬁcant impact of LPG on household food security. Compared to
BAU, which had a 5% annual population growth, the LPG intervention increased mid-century food
security by 40% (Figure 4.3: 1c). Also, the severity of food security was reduced - the number
of moderately food insecure (FSVI III) households signiﬁcantly declined under this intervention.
FSVI III households either became food secure (FSVI I) or slightly food insecure (FSVI II). We
expected to see the impact of decreased farm labor under lower population growth on household
food production and food security. However, the model suggested that such impacts were miniscule.

102

Figure 4.1: Climate predictions under BAU (RCP 8.5) and CCM (RCP 4.5) (RCP: Representative
Concentration Pathway)

103

4.4.2.2

Improved access to agricultural inputs

Crop productivity increased substantially with increased application of inputs (Figure 4.4). That
contributed positively to household food security in the region. Under the MAI scenario, mid-
century food security rose to almost 40% (Figure 4.3:3d) compared to 20% in BAU. Under this
scenario, only large endowed households (i.e., HRE_LH and HRE households), that constitute
40% of total farm households in Koutiala, have increased access to inputs. However, when all
the households get equal and high access to inputs (HAI), there was a drastic improvement in
household food security. More than 55% of the households would be food secure by 2050 under
this scenario(Figure 4.3: 2d).

104

Figure 4.2: Crop yield (Kg/ha) under BAU (RCP 8.5) and CCM (RCP 4.5)

105

Figure 4.3: Household food security under diﬀerent scenarios (LPG: Lower Population Growth, CCM: Climate Change Mitigation, HAI:
High Access to Input, MAI: Mixed Access to Input)

106

In sum, the intervention at the global level was not eﬀective in promoting mid-century household
food security. However, trajectories of crop yields under RCP 4.5 and RCP8.5 are expected to change
signiﬁcantly beyond 2050 (Levis et al. 2018). This can have a very diﬀerent impact on household
food security in the region. Note that analysis beyond 2050 is not included in our study.

Both local level interventions are highly eﬀective in addressing future food insecurity. However,
diﬀerent interventions aﬀect households diﬀerently. Although LPG and HAI were equally eﬀective,
LPG has a slight edge in reducing the severity of food security, especially of smaller households
(Figure 4.5). A combination of interventions that includes improved access to input and lower
population growth, can signiﬁcantly increase future food security. When HAI and LPG are
combined, mid-century household food security can increase by up to 90%. Such an increase
in regional food security can have signiﬁcant impacts on local and national food supply.

The region contains many mildly food insecure (FSVI II) households. These households are
supported by formal and informal cooperation between local community members. Households
with surplus food generally support neighbors that are in need. Koutiala has various groups and
cooperatives that promote supports amongst community members. They usually act as social
safety-nets that absorb the impacts of food scarcity in the locality. They allow sharing of beneﬁts
amongst the community members. The impact of such cooperation is clearly visible in the MAI
scenario. Although households having low endowment could not access high inputs, they could
still reduce their food insecurity by 20% (Figure 4.5). These smaller households beneﬁted from the
improved production in their neighborhood, especially in large endowed households.

4.5 Limitations

The empirical results reported above should be considered in light of several limitations.
There are inter and intra-model variabilities in climate change projections. Crop models are
highly sensitive to climatic parameters (Senthold Asseng et al. 2013). However, for the sake of
simplicity and parsimony, we used only one observation from one climate model.

Furthermore, due to the loose coupling between the crop and the agent-based models, we

107

Figure 4.4: Crop yield (Kg/ha) under low and high access to input

108

assumed a uniform farm management across time and space. We also did not include learning and
adaptation in the model. Such limitations could have resulted in overprediction of the food insecure
households. Farmers regularly modify their farming practices; they apply improved technologies
and crop varieties improving crop productivity and food security. Furthermore, resource availability
may aﬀect population growth. A constant population growth, assumed in the study, may not be
possible under household food scarcity.

A desire for large households is engraved in the local culture due to their apparent beneﬁts on
farm and oﬀ-farm jobs. This raises a question of practicality of low population growth interventions
in the local context. Consequently, the feasibility of improved and uniform access to agricultural
inputs for all farm households in a developing country like Mali requires further study.

109

Figure 4.5: Food insecurity by household types

110

4.6 Conclusion

As northern Mali is unsuitable for agricultural farming, the South has the responsibility to feed
the nation. So far, it has been able to export local production to other regions. Since Mali is
expected to be impacted by climate change, its crop productivity is likely to decline. Moreover,
the South is observing a rapid increase in its population. Local food demand will rise while the
production declines. As a result, future without interventions could lead to massive food insecurity
in the region.

Southern Mali has a daunting task of providing food to the locals while maintaining food
supply to the nation. There is an urgency to start interventions to promote food security in the
region. Based on our simulations, we conclude that local level interventions, such as improving
farmers’ access to input and controlling population growth, are eﬀective in promoting household
food security. In contrast, the global intervention that includes climate change mitigation to the
RCP4.5 level does not have similar success in maintaining food production and food security in the
region.

The coupled model incorporates the dynamics within and across biophysical, climatic, and
social systems. As food security lies at the intersection of these systems, the coupled model is an
appropriate tool for its analysis. Moreover, the interconnectedness between biophysical and social
systems allows for exploring the impacts of changes in one system on others. Such characteristic
is essential for evaluating policy interventions. Using the coupled model for scenario analyses has
broader policy implications as it provides valuable information on the potential eﬃcacies of the
interventions in the modeled system.

111

BIBLIOGRAPHY

112

BIBLIOGRAPHY

Antle, John M et al. (2017). “Towards a new generation of agricultural system data, models and
knowledge products: Design and improvement”. In: Agricultural systems 155, pp. 255–268.
issn: 0308-521X. url: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5485644/
pdf/main.pdf.

Asseng, S. et al. (2017). “Hot spots of wheat yield decline with rising temperatures”. In: Glob
Chang Biol 23.6, pp. 2464–2472. issn: 1365-2486 (Electronic) 1354-1013 (Linking). doi:
10.1111/gcb.13530. url: https://www.ncbi.nlm.nih.gov/pubmed/27860004.

Asseng, Senthold et al. (2013). “Uncertainty in simulating wheat yields under climate change”. In:

Nature Climate Change 3.9, p. 827. issn: 1758-6798.

Asseng, S et al. (2004). “Simulated wheat growth aﬀected by rising temperature, increased water
deﬁcit and elevated atmospheric CO2”. In: Field Crops Research 85.2-3, pp. 85–102. issn:
0378-4290.

Axelrod, Robert (1997). “Advancing the art of simulation in the social sciences”. In: Simulating

social phenomena. Springer, pp. 21–40.

Basso, Bruno, Anthony D Kendall, and David W Hyndman (2013). “The future of agriculture over
the Ogallala Aquifer: Solutions to grow crops more eﬃciently with limited water”. In: Earth’s
Future 1.1, pp. 39–41. issn: 2328-4277.

Basso, Bruno and Lin Liu (2019). “Seasonal crop yield forecast: Methods, applications, and

accuracies”. In: Advances in Agronomy. Vol. 154. Elsevier, pp. 201–255. isbn: 0065-2113.

Basso, Bruno and Joe T Ritchie (2015). “Simulating crop growth and biogeochemical ﬂuxes
in response to land management using the SALUS model”. In: The ecology of agricultural
landscapes: long-term research on the path to sustainability. Oxford University Press, New
York, NY USA, pp. 252–274.

Basso, Bruno, Joe T Ritchie, et al. (2006). “Simulation of tillage systems impact on soil biophysical
properties using the SALUS model”. In: Italian Journal of Agronomy, pp. 677–688. issn:
2039-6805.

Basso, Bruno, Luigi Sartori, et al. (2012). “Environmental and economic evaluation of N fertilizer
rates in a maize crop in Italy: A spatial and temporal analysis using crop models”. In: Biosystems
Engineering 113.2, pp. 103–111. issn: 1537-5110.

Bassu, Simona et al. (2014). “How do various maize crop models vary in their responses to
climate change factors?” In: Global change biology 20.7, pp. 2301–2320. issn: 1365-2486.
url: https://onlinelibrary.wiley.com/doi/abs/10.1111/gcb.12520.

113

Bingen, R James (1998). “Cotton, democracy and development in Mali”. In: The Journal of Modern

African Studies 36.2, pp. 265–285. issn: 0022-278X.

Boero, Riccardo and Flaminio Squazzoni (2005). “Does empirical embeddedness matter? Method-
ological issues on agent-based models for analytical social science”. In: Journal of Artiﬁcial
Societies and Social Simulation 8.4.

Dembele, Bandiougou et al. (2017). “Dynamics and adaptation of agricultural farming systems in
the boost of cotton cropping in southern Mali”. In: African Journal of Agricultural Research
12.18, pp. 1552–1569. issn: 1991-637X.

Eozenou, Patrick, Dorsati Madani, and Rob Swinkels (2013). Poverty, malnutrition and vulnera-

bility in Mali. The World Bank. isbn: 1813-9450.

Epstein, Joshua M (2006). Generative social science: Studies in agent-based computational mod-

eling. Princeton University Press. isbn: 0691125473.

Falconnier, Gatien N et al. (2015). “Understanding farm trajectories and development pathways:
Two decades of change in southern Mali”. In: Agricultural Systems 139, pp. 210–222. issn:
0308-521X.

García-Vila, Margarita and Elías Fereres (2012). “Combining the simulation crop model AquaCrop
with an economic model for the optimization of irrigation management at farm level”. In:
European Journal of Agronomy 36.1, pp. 21–31. issn: 1161-0301.

Grimm, V. et al. (2006). “A standard protocol for describing individual-based and agent-based
models”. In: Ecological Modelling 198.1-2, pp. 115–126. issn: 0304-3800. doi: 10.1016/j.
ecolmodel.2006.04.023. url: %3CGo%20to%20ISI%3E://WOS:000240823100008.

Grimm, Volker et al. (2010). “The ODD protocol: a review and ﬁrst update”. In: Ecological

modelling 221.23, pp. 2760–2768. issn: 0304-3800.

Hammond, R. A. and L. Dube (2012). “A systems science perspective and transdisciplinary models
for food and nutrition security”. In: Proc Natl Acad Sci U S A 109.31, pp. 12356–63. issn:
1091-6490 (Electronic) 0027-8424 (Linking). doi: 10.1073/pnas.0913003109. url: http:
//www.ncbi.nlm.nih.gov/pubmed/22826247.

Jones, J. W. et al. (2017). “Brief history of agricultural systems modeling”. In: Agric Syst 155,
pp. 240–254. issn: 0308-521X (Print) 0308-521X (Linking). doi: 10.1016/j.agsy.2016.
05.014. url: https://www.ncbi.nlm.nih.gov/pubmed/28701816.

Jones, James W et al. (2017). “Toward a new generation of agricultural system data, models,
and knowledge products: State of agricultural systems science”. In: Agricultural systems 155,
pp. 269–288. issn: 0308-521X. url: https://www.ncbi.nlm.nih.gov/pmc/articles/
PMC5485672/pdf/main.pdf.

114

Levis, Samuel et al. (2018). “CLMcrop yields and water requirements: avoided impacts by choosing

RCP 4.5 over 8.5”. In: Climatic Change 146.3-4, pp. 501–515. issn: 0165-0009.

Ligmann-Zielinska, Arika and Piotr Jankowski (2007). “Agent-based models as laboratories for
spatially explicit planning policies”. In: Environment and Planning B: Planning and Design
34.2, pp. 316–335. issn: 0265-8135.

Liu, Lin and Bruno Basso (2017). “Spatial evaluation of maize yield in Malawi”. In: Agricultural

Systems 157, pp. 185–192. issn: 0308-521X.

Marohn, Carsten et al. (2013). “A software coupling approach to assess low-cost soil conservation
strategies for highland agriculture in Vietnam”. In: Environmental Modelling & Software 45,
pp. 116–128. issn: 1364-8152.

Maxwell, Daniel G (1996). “Measuring food insecurity: the frequency and severity of “coping

strategies””. In: Food policy 21.3, pp. 291–303. issn: 0306-9192.

Maxwell, Daniel, Richard Caldwell, and Mark Langworthy (2008). “Measuring food insecurity:
Can an indicator based on localized coping behaviors be used to compare across contexts?” In:
Food Policy 33.6, pp. 533–540. issn: 0306-9192.

Maxwell, Daniel, Ben Watkins, et al. (2003). “The coping strategies index: A tool for rapidly
measuring food security and the impact of food aid programs in emergencies”. In: Nairobi:
CARE Eastern and Central Africa Regional Management Unit and the World Food Programme
Vulnerability Assessment and Mapping Unit.

N’Danikou, Sognigbe et al. (2017). “Foraging Is Determinant to Improve Smallholders’ Food

Security in Rural Areas in Mali, West Africa”. In: Sustainability 9.11, p. 2074.

Olabisi, Laura Schmitt et al. (2018). “Using participatory modeling processes to identify sources
of climate risk in West Africa”. In: Environment Systems and Decisions 38.1, pp. 23–32. issn:
2194-5403.

Parker, Dawn C et al. (2003). “Multi-agent systems for the simulation of land-use and land-cover
change: A review”. In: Annals of the association of American Geographers 93.2, pp. 314–337.
issn: 1467-8306.

Rivers III, Louie et al. (2017). “Mental models of food security in rural Mali”. In: Environment

Systems and Decisions, pp. 1–19. issn: 2194-5403.

Roudier, Philippe et al. (2011). “The impact of future climate change on West African crop yields:
What does the recent literature say?” In: Global Environmental Change 21.3, pp. 1073–1083.
issn: 0959-3780.

115

Schlenker, Wolfram and David B Lobell (2010). “Robust negative impacts of climate change on

African agriculture”. In: Environmental Research Letters 5.1, p. 014010. issn: 1748-9326.

Schreinemachers, Pepijn and Thomas Berger (2011). “An agent-based simulation model of hu-
man–environment interactions in agricultural systems”. In: Environmental Modelling & Soft-
ware 26.7, pp. 845–859. issn: 1364-8152. url: https : / / www . sciencedirect . com /
science/article/pii/S1364815211000314?via%3Dihub.

Schreinemachers, Pepijn, Thomas Berger, and Jens B Aune (2007). “Simulating soil fertility and
poverty dynamics in Uganda: A bio-economic multi-agent systems approach”. In: Ecological
economics 64.2, pp. 387–401. issn: 0921-8009.

Sivakumar, MVK (1988). “Predicting rainy season potential from the onset of rains in Southern
Sahelian and Sudanian climatic zones of West Africa”. In: Agricultural and Forest Meteorology
42.4, pp. 295–305. issn: 0168-1923.

Sultan, Benjamin et al. (2013). “Assessing climate change impacts on sorghum and millet yields in
the Sudanian and Sahelian savannas of West Africa”. In: Environmental Research Letters 8.1,
p. 014040. issn: 1748-9326.

Traore, Bouba et al. (2013). “Eﬀects of climate variability and climate change on crop production

in southern Mali”. In: European Journal of Agronomy 49, pp. 115–125. issn: 1161-0301.

Wossen, Tesfamicheal and Thomas Berger (2015). “Climate variability, food security and poverty:
Agent-based assessment of policy options for farm households in Northern Ghana”. In: En-
vironmental Science & Policy 47, pp. 95–107. issn: 1462-9011. doi: https://doi.org/
10.1016/j.envsci.2014.11.009. url: http://www.sciencedirect.com/science/
article/pii/S1462901114002226%20https://www.sciencedirect.com/science/
article/pii/S1462901114002226?via%3Dihub.

116

CHAPTER 5

CONCLUDING REMARKS

Koutiala, a region in the South of Mali, is currently considered the breadbasket of the country. The
trajectories of national food supply depend heavily on food production and food security in Koutiala.
In this study, we analyzed the vulnerability of households to food insecurity in Koutiala. Mali
is expected to be profoundly impacted by climate change. Under a high emission climate change
scenario, which is very likely without a global intervention, its mid-century temperature could rise
by 2◦C. This can have detrimental impacts on national crop production. Furthermore, southern
Mali is mainly responsible for providing food to the nation as the North is highly unsuitable for
agriculture. However, farming in the South has its own set of problems. Poor access to inputs and
farm management has substantially lowered crop productivity in the region. With climate change,
the region may see a further decline in its agricultural productivity. Moreover, Mali is experiencing
rapid population growth. With reduced production, the country will struggle to feed its growing
population.

Food security crisscrosses biophysical, climatic, and socio-economic systems. Household food
security, particularly in rural areas, depends on household food production and access. We used
a systems approach for deﬁning and analyzing household food security and employed multidisci-
plinary tools to capture the dynamics of associated systems. Through combining an ABM, a model
of the social system that analyzes social interactions for food access, with a biophysical crop model
simulating crop production, we essentially captured the basic dynamics responsible for household
food security. We used Food Security Vulnerability Index (FSVI), a coping-based index, to measure
vulnerability to future food insecurity at the household level. Furthermore, we incorporated all four
dimensions of household food security in our analysis. Household food production was used to
obtain information on food availability. We assessed household food access by analyzing resource
availability and social interactions. Similarly, calorie suﬃciency was used to estimate household
food utilization, and temporal stability was represented through simulations.

117

Agent-based model development needs information on systems actors and their interactions. In
chapter two, we aimed to identify actors (agents) of household food security in Koutiala, together
with their characteristics, and their interactions required for the development of agent-based model
(ABM). Since qualitative data, especially in the narrative form, contains rich contextual systems
information, we decided to use ﬁeld narratives to extract ABM components. Qualitative data
extraction is usually tricky and often blamed for being subjective and bias-prone. Modelers often
skip extracting model components from narratives and develop models based on their subjective
interpretation of the target systems. This can produce models with questionable credibility.
In
this chapter, we used Natural Language Processing (NLP) tools for enhancing and expediting
qualitative data extraction for model development. Using a combination of semantic and syntactic
tools, we used an unsupervised approach to extract information on agents and their attributes and
interactions. We identiﬁed subjects of sentences as agents, while root verbs were characterized as
actions. Similarly, we obtained attribute information from sentences containing possessive nouns
and be or have as their root verbs. We contextualized the extracted information manually and asked
the stakeholders for evaluation. Since the approach is mainly unsupervised, especially during the
initial information extraction, it controls opportunities for subjectivity and biases during model
development.

In chapter three, we explored the future trajectory of household food security in Koutiala. We
developed an ABM of household food security and integrated it with (Systems Approach to Land
Use Sustainability) SALUS, a process-based biophysical crop model. Using the crop model, we
simulated household cultivation of maize, millet, and sorghum – the three major cereal crops in the
region. For the simulations, we deﬁned a ‘business-as-usual’ scenario based on the current state
of regional farm management, population growth, and global climate change. The ABM received
productivity information from SALUS and calculated household food availability and demand.
Households with suﬃcient food were tagged as food secure. We simulated the interactions for
food access by food-deprived households and provided FSVI values to the households based on
their interactions. Under status-quo, the model predicted a massive reduction in crop productivity;

118

maize and sorghum productivity, in particular, could decline by almost 40%. We also found that
almost 80% of the Koutiala population is likely to be food insecure by 2050. Large households are
particularly vulnerable to food insecurity, and smaller houses are more likely to suﬀer from severe
forms of food insecurity.

A staggering level of food insecurity in one of the most productive regions of the nation is
troublesome. Mali needs to start promoting food production and food security in the South. In
chapter four, we formulated three global and local level interventions based on observed issues
and local public demand. We simulated climate change mitigation to the RCP4.5 level as a global
level intervention. For the local level, we formulated scenarios with low population growth and
a high level of input application. The consequences of climate change mitigation appeared to be
highly similar to the Business-as-usual scenario. Even though the intervention represents lower
temperature, its beneﬁts on regional food production and food security were noninﬂuential. We
believe that the lower level of atmospheric CO2, predicted under RCP4.5, canceled the expected
beneﬁts of a lower temperature. As a result, household food security would remain unaﬀected
under the global level intervention, at least until 2050. However, both local level interventions were
highly eﬀective in showing immediate impacts on household food security in the region. More
than 60% of the households in Koutiala are likely to remain food secure under any of the local
interventions. Besides, a combination of the local level interventions could substantially increase
food secure households, which can signiﬁcantly promote food supply to the nation.

In sum, Mali relies on the South for food, and it needs to act soon to maintain the food supply
to the nation. Without any interventions, the majority of households in the South will soon be food
insecure potentially disrupting the national food supply. Our simulations suggest that Mali should
immediately focus on local level interventions to promote regional food security and national food
supply.

There are ﬁve major technical and methodological limitations in this study to be addressed in
future research. First, we only extracted individual sentences during NLP data extraction, which
resulted in increased ambiguity and loss of contextual information. We encourage extracting a set

119

containing sentences before and after the sentence identiﬁed for extraction. This can help preserve
contextual information and reduce ambiguity. Additionally, extracting a set instead of individual
sentences can help resolving pronouns used in the sentences. Second, the NLP data extraction
was largely limited to syntactic analysis. We focused mainly on extracting agents, attributes,
and interactions using syntactic approaches. Future research should increase the use of semantic
analysis for understanding decision-making in the system. Compound and conditional sentences,
we believe, could provide the key to decisions in the system. The third limitation is the failure
to identify other actors of household food security. Since our data source contains information
collected from household members, information on other actors of household food security was
left out. This prompts for careful consideration of all candidate agents identiﬁed during the NLP
analysis.

The fourth major shortcoming is related to the loose coupling between the ABM and the crop
model. Due to models high resource requirements, we opted for a loose coupling between the
models. This essentially prevented us from introducing some crucial dynamics in the analysis.
Although the crop model captures the heterogeneity in weather and soil conditions, we kept the
farm management uniform across time and space. Farmers learn from experience and make
modiﬁcations in their farming approaches. Such adaptation could not be incorporated due to the
loose coupling. A more integrated model can update farm management input to the crop model in
real-time. The ﬁnal limitation involves the exclusion of cash crops and non-farm incomes in model
simulations. The regional farmers are involved in cash cropping and cottage industries, and their
income helps in food purchases. Future research should include all types of income contributing
to food security at the household level.

With all the limitations, the data extraction process is quick, structured, and eﬀective.

It
can become a suitable substitute for lengthy and subjective traditional data extraction for model
development. This can help ABM developers in identifying model components and interactions
needed for credible and reliable model development. Similarly, the coupled model is highly useful
for comprehensive household food security analysis. The model eﬀectively integrates dynamics

120

across multiple systems responsible for household food security. Besides, the coping-based index
is simple and incorporates multiple dimensions of food security. Overall, the model we presented is
an eﬃcient and cost-eﬀective laboratory for exploring expected and alternative futures of household
food security for evaluating eﬀective intervention strategies.

121