An agentbased model for analyzing land use dynamics in response to farmer behaviour and...

Seediscussions,stats,andauthorprofilesforthispublicationat:http://www.researchgate.net/publication/257015480

Anagent-basedmodelforanalyzinglandusedynamicsinresponsetofarmerbehaviourandenvironmentalchangeinthePampangadelta(Philippines)

ARTICLEinAGRICULTUREECOSYSTEMS&ENVIRONMENT·OCTOBER2012

ImpactFactor:3.4·DOI:10.1016/j.agee.2012.07.016

CITATIONS

3AUTHORS:

FrançoisMialhe

UniversitéLumiereLyon2

17PUBLICATIONS34CITATIONS

SEEPROFILE

NicolasBecu

FrenchNationalCentreforScientificResea…

39PUBLICATIONS248CITATIONS

SEEPROFILE

YanniGunnell

UniversitéLumiereLyon2

98PUBLICATIONS1,169CITATIONS

SEEPROFILE

Availablefrom:YanniGunnell

Retrievedon:09October2015

(This is a sample cover image for this issue. The actual cover is not yet available at this time.)

This article appeared in a journal published by Elsevier. The attached

copy is furnished to the author for internal non-commercial research

and education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling or

licensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of the

article (e.g. in Word or Tex form) to their personal website or

institutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies are

encouraged to visit:

http://www.elsevier.com/copyright

Author's personal copy

Agriculture, Ecosystems and Environment 161 (2012) 55– 69

Contents lists available at SciVerse ScienceDirect

Agriculture, Ecosystems and Environment

jo ur n al homepage: www.elsev ier .com/ lo cate /agee

An agentbased model for analyzing land use dynamics in response to farmerbehaviour and environmental change in the Pampanga delta (Philippines)

F. Mialhe a,b,∗, N. Becu c, Y. Gunnelld

a Department of Geography, Université Paris Diderot 7, CNRS UMR 8586, 2 rue Valette, 75005 Paris, Franceb Department of Geography, Facultés Universitaires Notre Dame de la Paix, 61 rue de Bruxelles, 5000 Namur, Belgiumc Department of Geography, Université Paris PanthéonSorbonne 1, CNRS UMR 8586, 2 rue Valette, 75005 Paris, Franced Department of Geography, Université de Lyon, CNRS UMR 5600, 86 rue Pasteur, 69365 Lyon Cedex 07, France

a r t i c l e i n f o

Article history:

Received 21 November 2011

Received in revised form 14 May 2012

Accepted 18 July 2012

Keywords:

Agentbased modelling

Aquaculture

Landuse change

Pampanga delta

Farmer behaviour

a b s t r a c t

Agentbased models (ABMs) are increasingly employed to understand land use change in agro

ecosystems. Here we use an ABM named CHANOS to capture how a range of variables influences

decisionmaking processes among farmers with respect to their choice of cropping system, and to analyze

the resulting changes in land use patterns. The model is experimental but is empirically based and nour

ished by field data acquired in the Pampanga delta, Philippines, where rice cropping and aquaculture have

been competing over the last 40 years at the expense of natural habitats. Among the variables we include

agent behavioural profiles but also forcing factors relevant to the natural, economic and political settings

of the system: e.g. continuous (deltaic land subsidence) and discrete (typhoon events) environmental

processes, external market forces, and changes in governmentdriven agricultural policies. Assessing the

relative weights of these factors was performed through a detailed analysis of decisional outcomes. The

farmers fall into three behavioural categories: rational, collective minded and boundedly rational. Like

wise, four different environmental dynamics are driven respectively by no deltaic subsidence, steady

subsidence, accelerating subsidence, and subsidence punctuated by additional external variables such as

listed above. Twelve scenarios were elaborated by combining the agent behaviour algorithms with the

environmental dynamics. Results reveal three categories of landuse change: an extension of paddy over

natural habitat, of aquaculture over natural habitat and paddy, and a succession of periods alternating

between paddy and aquaculture. Several indicators show that the rational agents are the most reactive

and adaptive to environmental changes.

Collectiveminded agents act independently from environmental changes. Their ability to cope with

change is limited and adaptations take longer to propagate. Boundedly rational agents reveal adaptive

capacities but are less reactive than rational agent. CHANOS thus provides a dynamic tool for understand

ing the social fabric and behavioural processes behind land use change.

1. Introduction

Landuse change (LUC) can trigger different environmental

changes such as biodiversity loss, local and regional climate change,

soil degradation and enhanced carbon emissions (Huigen, 2004;

Lambin et al., 2001; Meyfroidt et al., 2010; Lambin and Meyfroidt,

2011). Societies are affected by these changes through the erosion

of ecosystem services, food shortages, a higher vulnerability to nat

ural hazards, and may more generally experience a reduction in

wellbeing (Turner II, 2010).

∗ Corresponding author. Current address: Department of Geography, Facultés

Universitaires Notre Dame de la Paix, 61 rue de Bruxelles, 5000 Namur, Belgium.

Tel.: +32 485363424.

Email address: francois.mialhe@fundp.ac.be (F. Mialhe).

A major tool for understanding and forecasting LUC is computer

modelling. Among modelling methods, the appeal of agentbased

models (ABMs) is that these allow the exploration of interactions

between micro and macrolevel structures, e.g. the farmstead

level and its wider environment (Huigen, 2004). Simple rules and

specific process–response interactions between social and natural

elements can lead to emergent and often unpredictable spatial pat

terns or time delays (Bandini et al., 2009). Thus, the aim of many

ABMs is to study how microlevel processes affect macrolevel

outcomes within socialecological systems (SES) that are com

plex, unpredictable, adaptive and often evolve in a nonlinear way

(Berkes et al., 2003; Jager et al., 2000; Matthews et al., 2007).

Because LUC emerges from a wide set of decisions taken at different

scales, ABMs are tools well suited to the study of LUC.

Existing ABMs dedicated to LUC are concerned with the explo

ration of the causes and the consequences of LUC, e.g. on landscape

http://dx.doi.org/10.1016/j.agee.2012.07.016

56 F. Mialhe et al. / Agriculture, Ecosystems and Environment 161 (2012) 55– 69

patterns or biodiversity (Verburg et al., 2004; Veldkamp and

Verburg, 2004). In agrarian systems, farmers are central decision

makers. Thus an important component of LUC models when applied

to agrarian systems is the decisionmaking process of the farmer,

or agent. Modelling this behavioural component can be based on

behavioural theory and/or on empirical data (Smajgl et al., 2011;

Macal and North, 2010). Jager et al. (2000) have developed the so

called consumat approach, in which the behaviour of people when

facing a consumer decision is not merely motivated by economic

optimization but relies on multidimensional optimizations. Drawn

from theories in social psychology, this is based on the assumption

that people do not always seek to optimize outcomes outside of

open competitive situations (Jager et al., 2000; Janssen and Ostrom,

2006). Another assumption is that people learn from one another’s

behaviours, particularly when they are beset by uncertainty (Jager

et al., 2000).

For farmers, decisions about cropping and farming systems are

both influenced by internal (e.g. cognitive capacities and social

capital) and by external drivers (e.g. presence or absence of govern

ment subsidies, variations in global markets or the environment)

(AcostaMichlik and Espaldon, 2008). Decisions result from a com

bination of profits and satisfaction seeking, showing that farmers

are not perfectly rational decision makers (Polhill et al., 2010).

Reaching a rational decision would be further complicated by vari

ables such as access to useful resources or to information (Lambin

et al., 2001; AcostaMichlik and Espaldon, 2008). Nonmonetary

factors, such as cultural barriers, can also contribute to modify the

strategies on which decisions are based (Polhill et al., 2010).

This paper aims to assess the influence of both internal and

external variables on decisions made by farmers regarding their

cropping systems and how these decisions generate LUC. It also

seeks to improve our understanding of the social fabric behind

land use change processes by examining a range of decisionrelated

social and economic outcomes that hinge on adaptive mechanisms.

The real environment on which the model is based is located in

the Pampanga river delta (Philippines), where in recent decades a

transition from agriculture to aquaculture and a conversion from

natural habitat to aquaculture have occurred. However, the ABM

we present here, called CHANOS (Changement d’Occupation du Sol),

relies both on empirical and on experimental data. The use of exper

imental data was necessary because obtaining accurate data about

environmental processes and systematic information from farmers

concerning landuse decision processes is not a simple task (Smajgl

et al., 2011; Bakker and van Doorn, 2009).

2. Study area

The study area is a portion of the Pampanga delta located in

Pampanga province, Luzon Island, Philippines (Fig. 1). The climate

is tropical with a monsoon season from May to October, which is

also the typhoon season. ENSO events (El Nino Southern Oscillation)

affect the area by lengthening the dry season. Natural subsidence

of the delta provokes the diffusion of seawater to inland areas. Its

recent acceleration is due to human actions such as groundwater

mining, floodplain constriction as a result of urban planning, the

reduction of river discharge because of flood control devices such

as dams, and the construction of fish ponds that have desiccated the

moistureretentive upper layers of the soil (Rodolpho and Siringan,

2006). Between 1991 and 2001, subsidence rates varied spatially

between 2 and 8 cm year−1 (Soria et al., 2005).

The two municipalities of Masantol and Macabebe, with a pop

ulation of ∼120,000, encompass the study area. Until the 1970s,

the main land use and land cover categories were rice, aquacul

ture, and natural habitat (open marshland and mangrove or nipa

swamp). Since then, aquaculture has expanded dramatically over

all other types of land use, thus becoming the dominant form of land

use within a period of 40 years. In the entire Pampanga delta, the

total area under aquaculture increased from 3900 ha to 33,000 ha

between 1965 and 2008 (+595% in the case of Macabebe, +817% in

the case of Masantol: Mialhe, 2010). Several aquaculture farming

systems coexist today, the main ones being extensive polyculture

(shrimp, crab, tilapia, milkfish) and semiintensive monoculture

(tilapia or milkfish). Fish farms range from small (1 ha and less)

to largescale (>100 ha), with a variety of tenure arrangements

(Mialhe, 2010).

In addition to tectonic subsidence in the delta, farmers have

faced seasonal events such as brutal changes in salinity, difficulties

in collecting molluscs and gastropods used as fish feed, increasing

costs of pond maintenance, and dike erosion in the context of floods

or typhoons. The fluctuation of productive areas is not just affected

by seasonal changes. Other periodic factors such as epidemics, mar

ket price fluctuations, episodic conflict among stakeholders and the

unequal diffusion of farmingrelated innovations also impact on the

total surface area and distribution pattern of cropping systems. The

satellite image in Fig. 1 was obtained soon after the 1991 Pinatubo

volcanic eruption, revealing the parts of the delta affected by lahars.

After 1991, the decline in fish farming was the consequence both of

encroachment on the delta by volcanic debris and of rainfall anoma

lies (Mialhe, 2010). Despite a posteruptive upturn in aquacultural

activity, fish farming in 1993 had only regained its 1989 foot

print, indicating a pause in the previously ongoing growth spurt.

This is explained by the occurrence of secondary (raintriggered),

posteruptive lahars several years after the eruption. This caused

a further massive transfer of debris to the western municipalities

of the delta, clogging river channels, plugging drainage canals, and

disrupting much of the resource base. Increased production costs

in the wake of these sporadic events were felt as late as 2008.

During the 1960s and until 1976, fish ponds were clustered

in a belt no more that 10 km wide along the coastal delta front.

After 1976, however, aquacultural surfaces increased substantially.

By 1989, pond density had increased and spatial distribution had

expanded for the first time to more than 20 km inland. In 1991 and

1993, pond density in the outer delta belt decreased but expansion

into the hinterland reached a 25 km radius from the seashore.

Until the 2000s, alongside rice or fish monoculture the cropping

system in the Masantol area alternated between rice during the

dry season and fish farming during the rainy season. This pattern

reflected the urge to maintain some subsistence crops. However,

aquaculture has since taken over as a perennial activity, with mixed

farming being driven continuously further north and now subsist

ing only in the innermost confines of the Pampanga delta.

3. Model description

3.1. Fieldbased underpinnings

A livelihood framework was used to design surveys and ques

tionnaires submitted to the farmers during extensive field visits.

Use of such a framework allows a comprehensive analysis of house

hold assets, vulnerabilities, strategies and outcomes. An agronomic

survey was also conducted in order to characterize and evaluate

the agronomic performance at both plot and farm scales. Social

psychology theories of human behaviour (Jager et al., 2000) were

adapted to our local empirical findings about behaviour among

Pampanga farmers. In CHANOS, the main objective of farmers is

to attain satisfaction through the completion of several objectives.

All of these basic principles were implemented at the agent level.

The description of the model in the following sections adheres to

the ODD (overview, design concepts and details) protocol (Grimm

et al., 2006, 2010). Netlogo, a program developed at the Center for

F. Mialhe et al. / Agriculture, Ecosystems and Environment 161 (2012) 55– 69 57

Fig. 1. Location of study area.

Connected Learning (Wilensky, 1999), was used as the agentbased

modelling software.

3.2. Entities, state variables and scales

The model consists of entities and state variables operating

across a range of nested spatial scales. The elementary spatial unit

in Netlogo is the patch (Fig. 2). Here, patches were classified as

water body (WB), agricultural land (AL) or natural habitat (NH).

In the model, where salinity propagation through the delta is a

consequence of land subsidence, the main function of WB patches

is to diffuse dissolved salt to other patches. The main attributes

of NH patches are the land cover type (bare soil, open marsh

land, mangrove/nipa swamp) and their suitability for conversion

to aquaculture. The suitability level depends on the land cover and

the year. AL consists of patch aggregates (farms), each supporting

a unique cropping system (CS1: rice, CS2: seasonal rotation of rice

and aquaculture, or CS3: aquaculture). Each farm registers a salin

ity level. The best CS attribute in terms of economic performance

depends on the salinity level during the previous year. Finally, the

entire study area is divided into six rectangular sectors of equal sur

face area. These spatial units define the functional perimeter within

which the partial perception that agents have of their environment

can be modelled.

CHANOS works with two basic classes of agent, namely farm

ers and investors, and with an aggregate class defined as networks.

Investors are agents who acquire new land in favourable circum

stances. An investor’s main attributes are willingness to invest

and capacity to invest. Both of these attributes are related to the

conversion of NH patches. Furthermore, investors are capable of

perceiving an area composed of three of the six sectors previously

defined.

Every farmer owns one farm. Some of its attributes (household

size and behaviour type) do not change during the simulations,

whereas economic attributes (except the magnitude of expen

diture) change over time in relation to economic performance

and decisionmaking processes. Satisfaction and certainty are

the decisionmaking attributes. These attributes allow the farm

ers to decide which CS will be favoured at the next iteration.

Four cognitive strategies inspired by the consumat approach

(Jager et al., 2000) were implemented: repetition, social com

parison, imitation and deliberation. Networks aggregate agents

that have the same level of social capital. Social capital refers

to the number and quality of social connections. It is a proxy

for the quality of information required for choosing the best

CS, because we know from empirical observation during field

surveys that high social capital guarantees access to better infor

mation.

Three types of agent behaviour were implemented: rational

(type A), collectiveminded (type B), and boundedly rational (type

C) sensu (Jones, 1999). Differences are established on the basis of

objectives and cognitive strategies. The A and B decision models

represent opposite endmembers and are clearly less likely to be

representative of reality than the more hybrid variant C (Simon,

1997). However, models being ‘what if?’ experiments in their own

right, each of these behavioural models was useful for capturing

the fullest range of possible land use dynamics and exploring the

boundaries.

Atype agents are motivated by two economic objectives

(Table 1): one related to a shortterm strategy (immediate prof

its) and the other related to a mediumterm strategy (stability in

profits). The first objective is given a greater weight than the second

so as to better reflect the opportunistic strategies that we observed

during the field surveys. Btype agents have four objectives among

which three are collectivedriven, whereas Ctype agents have a

wider range of objectives. Fig. 3 shows the cognitive strategies that

are followed by the agents according to their type, their satisfaction

level and their uncertainty level.

Fig. 2. Class diagram of CHANOS.

Table 1

Farmers’ objectives entering into the computation of satisfaction levels.

Objective Computer instruction Atype Btype Ctype

1. To make profits Objective fulfilled when benefits are positive xxa x x

2. To adopt the same CS as their

neighbours

Make a list of the CS of the neighbours (Moore neighbourhood) and choose the

most frequent one. Objective fulfilled when the chosen CS is identical to the

current CS

3. To adopt the same CS as their

network members

Make a list of CS among network members and choose the most frequent one.

Objective fulfilled when the one chosen is the same as the current CS

4. To opt for the easier of two jobs Objective fulfilled when agent has adopted CS3 (aquaculture) x

5. To follow government guidelines Objective fulfilled when the agent’s CS is the same as the one recommended x x

6. To supply staple foods Objective fulfilled when agent is implementing CS1 x

7. To secure a stable income Objectives fulfilled when production of (year1) ≥ production of (year2) x x

a A double weight for the objective.

Fig. 3. Cognitive strategies of all agent types.

Lastly, the model introduces a class termed ‘external variables’

containing three subclasses. These include typhoons, (global) mar

kets and policybased (national) government recommendations.

During an ABM simulation these occur as discrete or continuous

events and are effective at different spatial scales (from individ

ual farmers to every farmer in the region). Typhoons are the only

discrete events, whereas government recommendations are con

tinuous but may change during the course of the simulations

(promoting, for example, CS3 at step 20 then CS1 at step 60). These

last two variables affect all the farmers.

The size of the basic spatial unit was set to 28.5 m. This resolu

tion corresponds to the spatial resolution of the Landsat 5 satellite

images used for model validation. Furthermore, this unit size was

selected because it allowed all the essential landscape elements to

be represented at a scale suited to the physical ‘grain’ of the actual

farmed landscape that prevails throughout the delta. Patch aggre

gates include farms and sectors. The combined administrative areas

of Masantol and Macabebe municipalities, which define the model

space, extend over 180 km2. Because LUC in the study area is mostly

agricultural, the most relevant unit time interval for the model is

the season (equivalent to six months). The simulations start in 1970

and last 40 years until 2010 (82 iterations).

3.3. Process overview and scheduling

Table 1 shows the different processes that occur during each

iteration. They are ordered according to their succession in time.

Physical processes always precede social responses. Each process

occurs at some observed frequency, typically every iteration or

every two iterations. Process order and recurrence are important

because the sequencing alters the successive system states. The

submodels are described in greater detail below.

3.4. Design concepts

Three important design concepts for understanding the model

are described here: adaptation, stochasticity and observation. Fur

ther details are provided in the Supplementary data available

online.

A state of adaptation (as opposed to inadaptation) refers to a

situation in which the outcome of an earlier decision is perceived

as positive and places the agent in a state of success or satisfaction.

Because decisionmaking is prompted by a change in the farmer’s

environment, adaptation here is of the reactive variety (and thus

distinct from anticipative adaptation: see Nelson et al., 2007). Adap

tation can be assessed at different scales, from the farmer to the

entire SES, and on different levels (social, economic, environmen

tal). Here we estimate the adaptation level of decisional outcomes

at farmstead scale based on criteria of social satisfaction, and at the

SES scale based on economic criteria (aggregated income). Adapta

tion level then refers in relative terms to how good the decisions

have been based on those aggregate criteria.

There are four sources of stochasticity in the model. Firstly, a ran

dom variable is implemented for calculating production, namely

±25% for aquaculture production and ±10% for rice production due

to inherent variations in yield recorded during the field surveys. For

scenarios relevant to environmental dynamic no. 4 (see Section 4),

this random variable reaches ±40% for aquaculture production after

the year 1990 in order to simulate high prawn mortality rates and

the variability of yields. Secondly, a random function was imple

mented for determining the best cropping system, i.e. the most

appropriate CS based on salinity levels, for farmers with low and

medium levels of social capital. These farmers obtain the good

information about the most appropriate CS in respectively 25% and

50% of cases, an outcome calibrated on the empirical observation

acquired from field surveys that higher social capital guarantees

access to better information. Thirdly, stochasticity was also intro

duced for the calculation of farm salinity when randomly selecting

the initial patch, pAL. This arises for computational reasons but is

appropriate given the relatively small size of the farms. Finally,

stochasticity was introduced in the selection of a convenient patch

by the investors. A convenient patch has a suitability level high

enough to comply with the willingness of the investors, i.e. the sum

of patch suitability (which varies between 0 and 10) and investor

willingness (also varies between 0 and 10) must equal or exceed

Spatial and statistical indicators were used to track the succes

sive state changes and process dynamics of the model (Table 2).

Land use (LU) and land use change (LUC) indicators were produced

Table 2

Processes involved in the model.

Process Description Affected entities (attribute) Ordering Recurrence

Physical

Subsidence Progressive sinking of the delta Patches (elevation) 1 (when activated) Annual (every two steps)

Hydrological surface

circulation

Variation of salinity front in delta

canals

Patches (salinity) 2 Seasonal (every step)

Social

Investment Investors (20) seek aggregated patches

suitable to farm

Investors (willingness/capacity)

Patch (suitability)

3 Annual

Satisfaction assessment Farmers assess their own satisfaction Farmers (objectives/satisfaction) 4 Seasonal

Certainty assessment Farmers assess their own certainty Farmers (objectives/uncertainty) 5 Seasonal

Farming system decision Farmers make a choice about the

cropping system

Farmers (CS) Farms (CS) 6 Seasonal

to inform about and map both land use patterns and dynam

ics. Six indicators (Table 3) were used to better understand the

decisionmaking process and are related to satisfaction, uncer

tainty, objectives and strategy.

3.5. Initialization

Two GIS layers were included in the model: the land use of 1970

from a topographic map (National Mapping and Resources Infor

mation Authority, or NAMRIA) (Fig. 4), and the elevation from a

digital elevation model (DEM) produced by interpolation of eleva

tion data points on another topographic map published in 2008

(NAMRIA). Because of active deltaic subsidence, we estimated the

elevation of 1970 by raising the present day elevation of patches

by 0.5 m. Initialization also encompasses salt diffusion processes

and the creation of spatial entities and agents. The salt diffusion

model is detailed below in the subsystem section. The farms were

created through an iterative clustering method that allows the vari

ance of both the size and the number of clusters to be controlled.

The attribute values of the farmers were both determined and dis

tributed according to the field survey results of Mialhe (2010).

3.6. Experimental and empirical data

The dates of the major typhoons that struck the study area were

provided by the Philippine Atmospheric, Geophysical and Astro

nomical Services Administration (PAGASA). Historical agricultural

farmgate prices, used to model the price increase of fishfarm

produce, were obtained from the Bureau of Agricultural Statistics

(BAS). Fig. 5 provides a chronology of the external variables that

have affected agent behaviour and LUC across the delta.

3.7. Submodels

3.7.1. The environmental submodels

Farm salinity was allocated at each iteration according to the

distance of the saltwater body closest to the farm. The subsidence

Table 3

Some indicators implemented in the model.

Variables Indicators

Land use Relative proportion of each land use category

Number of farmers conducting the most adapted farming

system

Paddytoaquaculture conversion

Land use maps

Satisfaction Average satisfaction

Number of satisfied farmers

Uncertainty Average uncertainty

Number of uncertain farmers

Objectives Proportion of farmers with every objective fulfilled

Strategies Proportion of farmers for each cognitive strategy

Economics Aggregated income

submodel simulates the annual subsidence of patches that are

higher than sealevel. We defined three subsidence intensities, with

annual subsidence rates ranging from 2.5 to 5 cm year−1. Modelling

of hydrological surface circulation relied on the fact that during

the dry season, low rainfall and low river discharge promote a

more intense penetration of seawater into the delta, whereas the

opposite happens during the rainy season. Fig. 6 details how the

algorithm works during the dry season.

The rainy season algorithm differs from its dry season counter

part in that it simulates the impact of rain and the diffusion of fresh

water.

3.7.2. The socioeconomic submodels

The investment submodel involves investors (n = 20) seeking

land to be converted to farming, under the double constraint that

the land must consist of continuous patches and that the area fits

the agent’s investment capacity.

The satisfaction and uncertainty submodels rely on the number

of completed and uncompleted objectives (see Table 1). Satisfaction

is the ratio of completed objectives over the total number of objec

tives, while certainty is the mean satisfaction level gained over the

last five iterations. At every step, the agent computes his profits, �:

k = P − C (1)

where P is the average daily income and C the daily household

(HH) consumption. The objective is fulfilled when � > 0. Informa

tion about income was gathered through fieldwork during the first

semester of 2008. The daily income averaged 432 Php (Philippino

pesos) for a fishpond owner, 363 Php for a farmer with a caretaker,

and 444 Php for a farmer without a caretaker (one US dollar was

equivalent to 42 Php in April 2008). The daily income (P) function

P = V ± (2)

where V is daily production and is the variation of the production,

which deviates by ±25% to ±40% from the daily average. Computa

tion of V relied on several observations: maximum daily agricultural

income is less than that for aquaculture, daily aquacultural income

is higher when salinity levels are high, and daily agricultural income

is higher when salinity levels are low. These observations were

combined with data obtained in the field concerning the economic

performance of farms (Mialhe, 2010): for each CS a scatter plot was

produced with salinity (in ppm) on the x axis and daily production

(in Php) on the y axis. Given the correlations between these two

variables, production levels in the model were therefore provided

as a function of salinity. The production functions were determined

from regression models for each CS (Table 4).

The function of daily HH consumption, C, is:

C = TR (3)

Fig. 4. Initial land use based on a 1970 topographic map (source: NAMRIA).

Table 4

Production functions for rice and aquaculture crops.

Salinity (ppm) Rice production Salinity (ppm) Aquacultural production

0–10 y = −25.27x + 345.29 0–35 y = −0.041x3 + 2.07x2 − 10.46x + 24

10–35 y = −2.48x + 83.49

Note: in the equations, y is the production (in Php) and x the salinity (in ppm).

Fig. 5. External forcing event chronology in the Pampanga delta between 1970 and 2010 (typhoons, government recommendations, prices, subsidence rates).

Fig. 6. Instructions flow chart for hydrological processes in the model during the dry season.

where T is the average consumption for every HH member and R is

the size of the HH. Both values were determined and distributed to

the farmers according to the field survey results.

Each completed objective equals 1 while any uncompleted

objective equals 0. Thus, the satisfaction function, S, is given by:

where O is the sum of objectives and N the number of objectives.

Certainty, M, is the mean of the satisfaction values summed over

the last five consecutive iterations:

where W is the sum of the last five satisfaction values. Certainty

is computed when agents have at least three satisfaction values

stored in their memory. Agents are satisfied when the satisfaction

value equals or exceeds 0.5 and are certain when the certainty value

likewise equals or exceeds 0.5.

Depending on their level of satisfaction and certainty, the farm

ers subsequently make a decision through the decision submodel.

Four cognitive strategies were implemented: repetition, imitation,

social comparison and deliberation. For an agent, repetition implies

conservation of the current CS; imitation is the reproduction of the

most common CS among a set of agents (the neighbours, the net

work members, or the satisfied agents); social comparison consists

in comparing one’s own CS with the CS of other groups of agents

(neighbours and network members) and adopting the majority CS

if one’s own represents less than a third of the total. Deliberation

consists in choosing the best CS based on the past year’s salinity

records. In the case of deliberation, the agents effectively choose

the best CS in 100%, 50%, and 25% of cases depending on their level

of social capital, respectively from high to low. Since this process

in reality mobilizes both cognitive capacities and time, the change

takes one season (6 months) to become effective.

4. Simulation experiments

The first experiment consisted of 30 runs (n = 30) of scenario

C4 (i.e. Ctype agent, dynamic no. 4) to assess the potential effect

of randomness on the modelling results. Scenario C4 was chosen

because it encompasses the complete set of agent strategies and

external events.

Subsequently, 12 scenarios were implemented by combining

agent behaviours with environmental dynamics (Table 5). The three

Table 5

Key to the 12 simulation scenarios.

Agent type Dynamic 1 Dynamic 2 Dynamic 3 Dynamic 4

A A1 A2 A3 A4

B B1 B2 B3 B4

C C1 C2 C3 C4

types of agent behaviour were implemented in separate scenarios.

Assessing the influence of environmental changes was achieved by

comparing the effects of four different environmental dynamics:

one involving no deltaic subsidence (dynamic 1), a second with a

constant subsidence rate during the entire simulation (dynamic 2),

a third with a higher subsidence rate after 1990 (dynamic 3), and

lastly a scenario that includes a higher subsidence rate punctuated

by external variables during the simulation (dynamic 4).

Finally, in order to know whether at least some of the model

scenarios produced realistic outcomes, which is a prerequisite for

any further integration of CHANOS in future resource management

frameworks, we tested the model outputs by comparing them with

land use maps generated from satellite images and topographic

maps. The landuse maps were computed from 1976 and 1989

Landsat images (MSS and Thematic Mapper, respectively) and from

a 2005 topographic map (Mialhe, 2010). For comparison purpose,

Landsat MSS images were resampled at the ground resolution of

Landsat 5. Confusion matrices were used to derive Cohen’s Kappa

coefficient, which measures the accuracy of a classification when

dealing with qualitative or categorical data (Foody, 2002). Matrix

rows tabulate the land use categories of the model outputs and the

columns define the land use provided by the reference land use

classification maps (Mialhe, 2010). The larger the Kappa coefficient

(it varies from 0 to 1), the better the match between the model land

use and the imageryderived land use maps.

5. Results

5.1. Randomness in the model

Table 6 shows the proportion of observations (n = 2460 for each

CS) after 30 runs that fell between ±1� or between ±2�. Overall

the results indicate a low impact of randomness in the model and

hence good replicability, highlighting the fidelity of the CHANOS

Some other sensitivity tests were carried out (see

Supplementary data). These tests show low sensitivity of the

model to stronger typhoons impacts (times 2 and times 3), to

investors with a lower financial capital, and to a higher attractivity

of natural habitats. On the other hand, the model is very sensitive

to accelerated land subsidence, to the introduction of a greater

proportion of agents with high social capital, and to agents with

larger households.

5.2. Assessment of the 12 scenarios

We first describe the outcomes of the farmers’ decisions in social

(satisfaction), economic (total earnings) and environmental (land

use) terms. We then combine social and economic outcomes to

Table 6

Results of statistical tests of randomness effects in the model.

Cropping system ±1� ±2�

CS1 (n = 2460) 68 96.9

CS2 (n = 2460) 69 96

CS3 (n = 2460) 68.5 97

All CS (n = 7380) 68.5 96.7

Note: n is the number of observations; � is the standard deviation.

periodize the simulations into time intervals for which the lev

els of adaptation have been estimated. This discretization helps

to explore the influence of SES and farmerlevel variables on the

decision patterns and finally on the LUC. Some additional figures

that show the detailed evolution of indicators are provided in the

Supplementary data accompanying this manuscript and are avail

able online.

5.2.1. Environmental outcomes

Three broad land use patterns are observed during the simu

lations: (i) CS1dominated land, (ii) CS3dominated land, and (iii)

equilibrium between CS1 and CS3. Patterns (i) and (ii) are further

divided into three subcategories: total domination by one CS, dom

ination with a few patches of other CS, domination with many

patches of other CS. Fig. 7 shows an example for three of these

categories.

Over the length of the simulation, three types of LUC pattern

were identified (Fig. 8): a progressive extension of paddy crops at

the expense of natural habitat (type 1: scenarios A1, B1, C1, B2,

B3 and B4), an extension of aquaculture at the expense of both

unfarmed land and paddy crops (type 2: scenarios A2, A3 and A4),

and an alternation of periods successively more favourable to CS1

or to CS3 (type 3: scenarios C2, C3 and C4). Several subtle differ

ences consist in a faster development of CS3 in scenario A4 than in

scenarios A2 and A3, the temporary emergence of CS3 in scenarios

B2, B3 and B4 accompanied by a proportional decrease of CS1, and

a stronger development of CS3 in scenario C4 than in scenarios C3

and C2 between steps 54 and 60.

Artificialization, i.e. replacement of NH by farms, and crop con

version are the major processes behind LUC. Artificialization occurs

consistently throughout any of the simulations and with a similar

intensity from one scenario to the other. This phenomenon explains

why aggregated income increases while averaged daily income

per capita decreases such as in scenario C1, in which aggregated

income gently rose throughout the simulation while daily income

per capita declined from 340 to 320 Php. In Atype agent scenarios,

conversions evolve progressively from the beginning until step 50

when the contingent of available land to be developed for rice crops

becomes limited. Its absence in scenarios involving Btype agents

is an indicator of the inertia of this agent category. It also indicates

that aquaculture development in Btype agent scenarios is only

due to investors. For Ctype agents, the total area of conversions

largely exceeds the areas under CS3, meaning that the conversions

are only temporary, whereas some are permanent in the case of A

type agents. Through these temporary conversions, Ctype agents

express here greater behavioural inertia than Atype agents in a

crop (paddy)dominated landscape.

5.2.2. Social and economic outcomes

Averaged satisfaction and aggregated income for all types of

agent and for dynamics 1, 2 and 4 are shown in Fig. 9. Because of the

similarity between scenarios arising from dynamics 2 and 3, only

the results concerning dynamic 2 are illustrated here. Fig. 9 reveals

that the dichotomous presence/absence of subsidence provokes a

major change in the evolution of these model outputs. Meanwhile,

the main differences between dynamics nos. 2 and 4 are shorter

lived except in the case of Atype agents, where they increased after

step 50. Oscillations are due to seasonal variations, while the dra

matic increase in income with Atype agents in dynamic no. 4 is due

to the higher prices of aquacultural products during that period.

5.2.3. Assessment of adaptation levels among farmers

We assessed the adaptation level of farmers’ decisions based on

the level and trend through time of social and economic outcomes.

Adaptation is greater when these outcomes have high value and

are on the increase; it is reduced when they have low value and

Fig. 7. Land use patterns at step 60 for scenarios A4, B4 and C4.

are decreasing (Fig. 10). Nine levels of adaptation were estimated

by combining the two outcome categories. The rationale behind the

adaptation ratings is based on the opposition between satisfied and

unsatisfied states, low and high satisfaction levels, and upper/lower

aggregated income brackets. We also considered the time trend as

an important longitudinal criterion.

In order to further assist in the analysis of model behaviour, sim

ulations were periodized into time intervals each characterized by

uniform adaptation levels (Fig. 11). The timing and sequencing of

real social and economic changes observed during the last 40 years

in the Pampanga delta is used as the reference frame against which

to assess the influence of model variables among model outcomes,

and therefore to appreciate model performance. These historical

periods were numbered in Fig. 11 for ease of reference during

subsequent analysis. Focus will be on transitions between more

uniform periods.

5.2.4. Influence of variables on cropping system decisions

5.2.4.1. Environmental variables. Subsidence and seasonal hydro

logical variations both modify the dissolved salt levels in water

bodies. Since production is affected by this change, it subsequently

modifies the fulfilment of objectives 1 and 7 and hence the decision

making process. With normal subsidence rates, mean salinity in all

water bodies increases rapidly at step 37 because salt water has

reached a flat portion of land and thus propagates swiftly. It then

reaches a plateau of 34 ppm around step 60. The overall dynamic is

a northward extension of brackish water followed by sea water.

The cropping systems are more productive when salinity levels

are maximal (CS3) or minimal (CS1). Thus, an extensive brackish

area is neither optimal in terms of production or income. However,

artificialization of land, which occurred throughout the simula

tion, slightly counterbalances this effect by sustaining a nearstable

or slightly increasing aggregated income (Period 2, Period 6–P2,

P6). When agents fail to adopt the most appropriate CS, any

extension of brackish water surfaces reduces both satisfaction and

income among agents (P13, P15, P20 and P22). Several periods of

rapid change among Atype agents are related to salinity changes.

Although salinity change explains P2 and P6, i.e. two scenarios that

are mainly characterized by a decreasing satisfaction, the strong

salinity change at step 37 induced some conversions (switch from

P2 to P3, noted P2/P3, and from P6 to P7, noted P6/P7). These

were followed by an immediate increase of both satisfaction and

aggregated income (P3, P7). Then, because some conversions were

not successful, some newly converted farmers became unsatisfied

(P3/P4, P7/P8) and encountered difficulties in adopting a better CS

(P4, P8). Finally, the extensive area of sea water allows a number

of conversions (P4/P5, P8/P9) that trigger adaptive states (P5, P9).

The stronger impact of salinity changes on Atype agent decision

making is due to the reduced set of objectives among Atype agents.

This places an important weight on objective 1 (2/3) by comparison

with B and Ctype agents (respectively 1/5 and 1/7). In substance,

deltaic subsidence negatively affects the farmers because it pro

duces environmental changes that alter the satisfaction state of the

farmers, because the changes are progressive and thus do not pro

voke radical shifts that would favour one particular CS, and also

because these changes are not anticipated.

Typhoons reduce profits among all types of agents, and also

ultimately reduce chances of fulfilling objective 1. But since their

impacts on profits are strong, they also affect the fulfilment of

objective 7. Atype agents are also more sensitive to this driver

because of their reduced set of objectives. Typhoons reduce profits

(objective 1) in such a strong way that objective 7 gets unfulfilled for

one whole year (2 iterations). However, objective 7 is fulfilled again

the following year because of the memory of the earlier losses. That

process increases the cognitive strategy changes among Atype

agents, but does not provoke significant changes in outcomes. The

impacts of typhoons on Btype agents are much more limited. Some

delayed impacts have been observed for Ctype agents following

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

iterations

Land u

0 20 40 60 80

Iterations

Land u

Natural habitatCS1 − rice

CS2 − agr i−aquacultureCS3 − aquaculture

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

0 20 40 60 80

Iterations

Land u

Fig. 8. Landuse dynamics of the study area for the 12 scenarios.

the two typhoons that appeared at steps 36 and 46. Such typhoons

altered cognitives strategies in a way that has favoured the adop

tion of both deliberation and social comparison strategies several

iterations later (P22/P23). These timedependent responses are due

to the existence of preferential successions in cognitive strategy

patterns. For example, deliberation typically occurs in situations

where repetition has been dominant among agents during the pre

ceding iteration. The impacts of typhoons diminish progressively

in all simulations because of the high, albeit changing, profitability

of CS3. Overall, typhoons play a minor role in the evolution of land

use when a CS is highly profitable.

5.3. Political and economic variables

Government recommendations only affect B and Ctype agents

because Atype agent decisions are not influenced by policy

changes. Many transitions occur when government recommen

dations change (P10/P11, P11/P12, P13/P14, P15/P16, P17/P18,

P18/P19, P20/P21 and P23/P24). With the exception of transi

tions P11/P12 and P18/P19, all the others provoke a shift to a

period marked by lower adaptation levels. Periods that followed

P11/P12 and P18/P19 register higher adaptation levels because

the governmentrecommended CS, namely aquaculture, is both

200 40 60 80

A−type agent

Iterations

atisfa

200 40 60 80

1000000

1500000

Iterations

Dynamic n°1 Dynamic n°2 Dynamic n°4

200 40 60 80

B−type agent

Iterations

atisfa

200 40 60 80

1000000

1500000

Iterations

200 40 60 80

C−type agent

Iterations

atisfa

200 40 60 80

1000000

1500000

Iterations

Fig. 9. Evolution of satisfaction and income indicators.

the most common among farmers and the best adapted given

the deltaic environment. In contrast, when the new government

recommended CS is neither the most common nor the most

adapted, transitions (P10/P11 and P17/P18) usher in lower adap

tation levels. When one of these two conditions (most common

and most adapted) is satisfied, the succeeding periods also exhibit

a diminished adaptation level (P13/P14, P15/P16, P20/P21, and

P23/P24). This variable influences the simulation path because it

instantaneously affects the satisfaction of all B and Ctype agents.

Positive outcomes can appear after some delay when the recom

mended CS is the most adapted and becomes increasingly popular

(P23). The main impact of the higher profitability of aquaculture

after 1990 is the increase in aggregated profits in scenario 10 (P9).

This effect is also strengthened by the extension of salt water areas.

As a (historically confirmed) consequence, Atype agents became

even more resilient to drivers such as typhoons. These two environ

mental drivers impact farmers’ decisions by altering the fulfilment

of objectives. However, the intensity with which these impacts

occur depends on whether the variables on which the objectives

rely are discrete (in the case of government recommendations) or

continuous (in the case of earnings). Impact intensity is strongest

in the first case.

5.4. Farmer behaviours

Adaptation levels among the successive periods experienced

during Atype agent scenarios either increased or stabilized. This

improvement is characteristic of adaptive behaviour. Average val

ues of adaptation indicators provide the same conclusion (Table 7):

overall, Atype agent scenarios exhibit higher mean indicators

Fig. 10. Adaptation levels from low (1black) to high (9white).

than others scenarios. An explanation for this is that cognitive

strategy adoption evolves in reaction to economic changes, which

themselves are driven by environmental changes such as salinity.

Adaptability, therefore, (i) results from the presence of objectives

that vary with environmental change or environmental quality,

(ii) also results from the reduced set of objectives that make

the agent more responsive, and (iii) can appear when agents

deploy a more limited range of cognitive strategies, particularly

when these contain deliberation. However, an Atype agent is also

made vulnerable by ignoring what other agents are doing and

200 40 60 80

scenario A1

Iterations

200 40 60 80

scenario A2

Iterations

P2 P3P4 P5

200 40 60 80

scenario A4

Iterations

P6 P7P8 P9P9

Adaptations levels (from low 1 to high 9)

200 40 60 80

scenario B1

Iterations

P10 P11 P12

200 40 60 80

scenario B2

Iterations

P13 P14

200 40 60 80

scenario B4

Iterations

P15 P16

200 40 60 80

scenario C1

Iterations

P17 P18 P19

200 40 60 80

scenario C2

Iterations

P20 P21

200 40 60 80

scenario C4

Iterations

P22 P23 P24

Fig. 11. Periodization of adaptation levels for nine scenarios.

Table 7

Adaptation indicators for the 12 scenarios.

Agent type Dynamic 1 Dynamic 2 Dynamic 3 Dynamic 4

Satisfaction (%) Aggregated

income

(×106 PhP)

income

(×106 PhP)

income

(×106 PhP)

income

(×106 PhP)

A 82 30.6 77 31 76 30.8 71 44.7

B 85 31.6 72 13.8 73 13.8 73 13.2

C 72 31.6 63 15.8 63 14.9 62 14.2

Table 8

Kappa coefficients obtained in the CHANOS model tests.

Year Dynamic 1 Dynamic 2 Dynamic 4

1976 1989 2005 1976 1989 2005 1976 1989 2005

Atype agent 0.62 0.4 0.23 0.63 0.46 0.55 0.64 0.47 0.57

Btype agent 0.62 0.36 0.19 0.62 0.37 0.19 0.61 0.36 0.19

Ctype agent 0.62 0.35 0.19 0.63 0.35 0.22 0.62 0.38 0.19

Note: Values express the quality of fit between the land use models produced by CHANOS and the historical land use mapped from Landsat imagery at three distinctive times

(1976, 1989, 2005). The models involve the implementation of environmental dynamics no. 1, no. 2, and no. 3.

by landscape homogenization. This vulnerability occurs because

the optimization process is only realized at the agent’s scale and

does not coevolve with an optimized landscape configuration. A

type agents could, for example, fall prey to the negative ecological

and economic feedbacks that are typical of uniform landscapes

(monoculture), which are likely to involve the erosion of ecosys

tem services and the rapid propagation of viral outbreaks (Kautsky

et al., 2000).

The main characteristic of Btype agents is their behavioural

inertia, which mainly results from a negative correlation between

satisfaction and economic performance (P14, P16, P2 and P24). B

type agents are not cognitively equipped to face up to a changing

environment. Constant high satisfaction works against behavioural

change. Such inertia is characterized by limited changes in spite of

a worsening situation for the farmers. Some events, such as deltaic

subsidence, reduce farmer satisfaction but not enough to modify

the main cognitive strategy, which among Btype agents is rep

etition. In these simulations, the investors are the only source of

diversity, which plays out through the introduction of CS3. Then

Btype agents are vulnerable because their actions are totally inde

pendent from economic rationality. Here, the absence of rationality

characterizes a total absence of link between decisionmaking and

environmental changes.

Simulations with Ctype agents exhibit more complex patterns.

Firstly, due to the increasing importance of areas under aquaculture

from step 20 in scenarios involving subsidence, the wide adop

tion of collective strategies such as social comparison or imitation

favours a number of conversions to aquaculture. This positive feed

back translates as a nonlinear evolution of CS3 between steps

20 and 60. Furthermore, shifts in government recommendations

strongly affect LUC, providing evidence that such agents are highly

sensitive to changes in just one objective when their satisfac

tion level approaches 0.5. The constantly low level of satisfaction

among Ctype agents is due to their wider set of objectives. The

low satisfaction indices promote an adaptation process through a

continuous change in cognitive strategies. In contrast, a high sat

isfaction level can inhibit strategic changes and the adoption of

a betterfitted farming system. This converges with the findings

obtained among Btype agents. The diverse set of cognitive strate

gies thus equips the farmer with a broader range of options. Ctype

agents are prone to introducing diversity and can also take advan

tage of an environmental change more swiftly than Btype agents.

The lack of adaptability among B and Ctype agents is also due

to the fact that the cognitive strategies most available to them

are rooted in criteria of social conformity (imitation, comparison)

rather than based on informed judgement about environmental

conditions and advantages determined by location. This limitation

feeds back on agricultural landscape structure and diversity.

5.5. Predictive performance of the models: a test against

historical landuse change in the Pampanga delta, 1976–2005

Table 8 lists the Kappa coefficient at patch level at three dates

for dynamics 1, 2 and 4. Without deltaic subsidence, all Kappa

coefficients are similar and decrease with time. The same trend

is observed with deltaic subsidence for B and Ctype agents, for

whom coefficients are of the same order. At any given date, coeffi

cients for Atype agents are the highest. However, with subsidence

they decrease between 1976 and 1989 and increase between 1989

and 2005. The higher value in 1976 is explained by the short period

between the beginning of the simulation (where Kappa equals

1) and the first measure. The lower index obtained in 1989 both

reflects the mislocation by the model of aquaculture farms, which

are shown to be situated farther north than in reality, and reveals

an overrepresentation of natural habitats (NH) in the model. The

last increase is the consequence of an almost complete disappear

ance of NH in the delta (total artificialization), both in the model

and in the field, so that the only differences in 2005 between model

and reality are due to some mismatches in the spatial distribution

of the cropping systems.

6. Conclusion

The goal of this study was to explore the influence of sev

eral variables on decisionmaking processes among farmers with

respect to their choice of cropping system, and to analyze the

resulting land use change. Analysis focused on farmer behavioural

patterns and how they are influenced by environmental, economi

cal and political variables. The modelled environment is located in

the Pampanga delta (Philippines). Three different agent behaviour

patterns (rational, collectiveminded and boundedly rational) and

four environmental dynamics (with or without deltaic subsidence,

with steady or accelerating subsidence, with or without additional

external influencing variables) were implemented. Twelve scenar

ios involving permutations of the behavioural and environmental

aspects were tested. Social (satisfaction), economic (income) and

landscaperelated (land use) decisions were analyzed based on

their outcomes and compared using various spatial and statistical

indicators. These outcomes further helped to assess the rela

tive influence of the various variables. In the absence of deltaic

subsidence, LUC turned out to be similar from one agent to another,

whereas in the presence of subsidence, all LUC patterns differed

from one another. Simulations with (Atype) rational agents exhib

ited a distinctive pattern involving a pervasive development of

aquaculture at the expense of both natural habitat and rice crops.

Collectiveminded (Btype) agents showed a low tendency to adapt

to environmental changes because few of their objectives take

account of such changes and because their behaviours rely exclu

sively on collective strategies. Boundedly rational (Ctype) agents

saw their satisfaction levels affected by the variation of external

forcing variables. These variations produce spatially fragmented

LUC (i.e. heterogeneous landscape mosaics) and generate alter

nating periods of time during which contrasting land use choices

are promoted. The impacts of factors such as deltaic subsidence,

government policy, natural hazards and market prices were also

assessed through the variations they generated among the differ

ent categories of outcomes. The CHANOS ABM makes it possible

to test hypotheses about how decision making affects land use in

a range of historical, social and environmental settings potentially

much wider and diverse than the worked example presented in this

prototype study. CHANOS can also be used as a decisionsupporting

tool in participatory approaches by observing, understanding,

comparing, and discussing outcomes involving different variables

that relate to behavioural changes, policy options, environmen

tal changes and market options. Furthermore, the comparison of

model land use maps with observed land use maps computed from

remotely sensed imagery is a procedure that can measure how

realistically or accurately the modelling process can reproduce his

torical evidence – and thus how reliably it can eventually be used

as a predictive tool.

Acknowledgments

We are grateful to the journal editor and to two anonymous

referees for their high quality reviews of the manuscript. The con

tribution to data acquisition through interviews and field surveys

from farmers and other stakeholders in the Pampanga delta has

been priceless. FM benefited from a PhD grant allocated by the

French government.

Appendix A. Supplementary data

Supplementary data associated with this article can be found, in

the online version, at http://dx.doi.org/10.1016/j.agee.2012.07.016.

References

AcostaMichlik, L., Espaldon, V., 2008. Assessing vulnerability of selected farmingcommunities in the Philippines based on a behavioural model of agent’s adaptation to global environmental change. Global Environ. Change 18 (4), 554–563.

Bakker, M.M., van Doorn, A.M., 2009. Farmerspecific relationships between landuse change and landscape factors: introducing agents in empirical land usemodelling. Land Use Policy 26 (3), 809–817.

Bandini, S., Manzoni, S., Vizzari, G., 2009. Agentbased modelling and simulation: an informatics perspective. J. Artif. Soc. Soc. Simul. 12 (4), 4,http://jasss.soc.surrey.ac.uk/12/4/4.html.

Berkes, F., Colding, J., Folke, C., 2003. Navigating Socialecological Systems: BuildingResilience for Complexity and Change. Cambridge University Press, Cambridge.

Foody, G.M., 2002. Status of land cover classification accuracy assessment. RemoteSens. Environ. 80 (1), 185–201.

Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., GossCustard, J.,Grand, T., Heinz, S.K., Huse, G., Huth, A., Jepsen, J.U., Jorgensen, C., Mooij, W.M.,Muller, B., Pe’er, G., Piou, C., Railsback, S.F., Robbins, A.M., Robbins, M.M., Rossmanith, E., Ruger, N., Strand, E., Souissi, S., Stillman, R.A., Vabo, R., Visser, U.,deAngelis, D.L., 2006. A standard protocol for describing individualbased andagentbased models. Ecol. Modell. 198, 115–126.

Grimm, V., Berger, U., DeAngelis, D.L., Polhill, J.G., Giske, J., Railsback, S.F., 2010. TheODD protocol: a review and first update. Ecol. Modell. 22, 2760–2768.

Huigen, M.G.A., 2004. First principles of the MameLuke multiactor modelling framework for land use change, illustrated with a Philippine case study. J. Environ.Manage. 72 (1–2), 5–21.

Jager, W., Janssen, M.A., De Vries, H.J.M., De Greef, J., Vlek, C.A.J., 2000. Behaviour incommons dilemmas: Homo economicus and Homo psychologicus in an ecologicaleconomic model. Ecol. Econ. 35 (3), 357–379.

Janssen, M.A., Ostrom, E., 2006. Empirically based, agentbased models. Ecol. Soc.11 (2), 13, http://www.ecologyandsociety.org/vol11/iss2/art37/ (art. no. 37,online).

Jones, B.D., 1999. Bounded rationality. Annu. Rev. Polit. Sci. 2, 297–321.Kautsky, N., Rönnbäck, P., Tedengren, M., Troell, M., 2000. Ecosystem perspectives

on management of disease in shrimp pond farming. Aquaculture 191 (1–3),145–161.

Lambin, E.F., Turner, B.L., Geist, H.J., Agbola, S.B., Angelsen, A., Bruce, J.W., Coomes,O.T., Dirzo, R., Fisher, G., Folke, C., George, P.S., Homewood, K., Imbernon, J.,Leemans, R., Li, X., Moran, E.F., Mortimore, M., Ramakrishnan, P.S., Richards, J.F.,Skanes, H., Steffen, W., Stone, G.D., Svedin, U., Veldkamp, T.A., Vogel, C., Xu, J.,2001. The causes of land use and landcover change: moving beyond the myths.Global Environ. Change 11, 261–269.

Lambin, E.F., Meyfroidt, P., 2011. Global land use change, economic globalization,and the looming land scarcity. Proc. Natl. Acad. Sci. U. S. A. 108 (9), 3465–3472.

Macal, C.M., North, M.J., 2010. Toward teaching agentbased simulation. In: PaperPresented at the Proceedings of the 2010 Winter Simulation Conference, Baltimore.

Matthews, R.B., Gilbert, N.G., Roach, A., Polhill, J.G., Gotts, N.M., 2007. Agentbased land use models: a review of applications. Landscape Ecol. 22 (10),1447–1459.

Meyfroidt, P., Rudel, T.K., Lambin, E.F., 2010. Forest transitions, trade, and the globaldisplacement of land use. Proc. Natl. Acad. Sci. U. S. A. 107 (49), 20917–20922.

Mialhe, F., 2010. Le développement de l’aquaculture saumâtre dans l’Aire Pacifique.Evolution des paysages, dynamiques socioéconomique et impacts environnementaux dans deux territoires au Pérou et aux Philippines. Thèse de Doctorat.Université Paris 7 Diderot, Paris.

Nelson, D.R., Adger, W.N., Brown, K., 2007. Adaptation to environmental change:contributions of a resilience framework. Annu. Rev. Environ. Resour. 32,395–419.

Polhill, J.G., Sutherland, L.A., Gotts, N.M., 2010. Using qualitative evidence to enhancean agentbased modelling system for studying land use change. J. Artif. Soc. Soc.Simul. 13 (2), 10, http://jasss.soc.surrey.ac.uk/13/2/10.html (online).

Rodolpho, K.S., Siringan, F.P., 2006. Global sealevel rise is recognised, but floodingfrom anthropogenic land subsidence is ignored around northern Manila Bay,Philippines. Disasters 30 (1), 118–139.

Simon, H.A., 1997. Models of Bounded Rationality. Empirically Grounded EconomicReason. MIT Press, 479 pp.

Smajgl, A., Brown, D.G., Valbuena, D., Huigen, M.G.A., 2011. Empirical characterisation of agent behaviours in socioecological systems. Environ. Modell. Soft. 26(7), 837–844.

Soria, J.L.A., Siringan, F.P., Rodolpho, K., 2005. Compaction rates and paleosea levelsalong the delta complex north of Manila Bay, Luzon Island, Philippines. Sci. Dilim.17 (2), 39–45.

Turner II, B.L., 2010. Vulnerability and resilience: coalescing or parallelingapproaches for sustainability science? Global Environ. Change 20 (4), 570–576.

Veldkamp, A., Verburg, P.H., 2004. Modelling land use change and environmentalimpact. J. Environ. Manage. 72 (1–2), 1–3.

Verburg, P.H., Schot, P., Dijst, M., Veldkamp, A., 2004. Land use change modelling:current practice and research priorities. GeoJournal 61 (4), 309–324.

Wilensky, U., 1999. NetLogo. Center for Connected Learning and ComputerbasedModeling, Evanston, IL.

Supplementary Data

The first section is dedicated to the description of some design concepts behind the model. It

is then followed by some technical points concerning the submodels. Thereafter, several

pseudo-codes are included with comments. The results of sensibility tests are then presented

in a table. Finally, eight figures detailed the evolution of indicators during the simulations.

Design concepts

Emergence. The land use pattern observed after every iteration emerges from the combination

of individual agent decisions.

Objectives. A farmer’s goal is to attain satisfaction and certainty (confidence in making the

right decision). Those two attributes are computed on the basis of whether or not the

objectives have been realized. Depending on behavioural category (rational, collective-

minded or boundedly rational), the number of objectives ranges from two to seven. Model

farmers calculate their satisfaction and certainty at every iteration. The goal of investors is to

become farmers by converting natural habitat (NH). Such actions require that their

willingness and capacity to invest are compatible with the suitability level of NH.

Learning. A learning process occurs if the farmer chooses to imitate the dominant cropping

system (CS) of a group of agents and that this dominant CS is identical to the farmer’s current

CS. This process is assimilated to an intensification that accrues from the sharing of

experience with others and results in an increase in earnings by 30 Php (Phlippine pesos)

every time a farmer encounters such a situation. However, since the situation can occur

several times during a simulation, learning-related gains are limited to 150 Php.

Sensing. Farmers do not directly perceive salinity on their farms. They receive information

about it from their network. Every investor is able to perceive the suitability level of patches

located in three sectors only, randomly selected out of the six existing ones.

Interaction. Farmers interact with their neighbours and with the members of their network by

asking them information. The exchanged information is only about CS but can be considered

as a proxy for beliefs, practices, knowledge, or preferences.

Collectives. Farmers belong to networks within which information is shared. Farmer

membership is based on the level of social capital, with four networks characterized by low,

two by medium and one by high social capital, respectively. Those seven networks have equal

sizes.

Some details about submodels

Farm salinity allocation. Farm salinity is calculated as follows: choice of a random patch

(pAL) on the farm, computation of the salinity of the pAL’s nearest WB patch, and assignment

of that value to the entire aggregate AL patch. Rainfall reduces the salinity of all patches

containing water. The diffusion of freshwater was simulated first by iteratively reducing the

salinity of patches located within the 5-patches radius buffer zone around zero-salinity

patches, and then by dividing the salinity of patches located in 2- and 3-patch radius buffer

zones of the zero salinity patches respectively by 2 and by 1.5. This operation was

implemented in order to reduce the step effect between fresh and saline patches.

Investment submodel. The investment submodel is a two-step process. At first, investors

randomly choose a convenient patch, then this patch iteratively aggregates other convenient

neighbouring patches (Moore neighbourhood). The process stops when the size of the

aggregate reaches or exceeds the investment capacity of the investor (expressed in patch

units). If the aggregate cannot grow any more due to a lack of suitable patches and if the size

does not reach the investment capacity, then the farmer randomly chooses another convenient

patch. If there is no such patch, the investor is eliminated and the next investor takes his turn.

References cited

Jager, W., Janssen, M.A., De Vries, H.J.M., De Greef, J., Vlek, C.A.J., 2000. Behaviour in

commons dilemmas: Homo economicus and Homo psychologicus in an ecological-economic

model. Ecol. Econ. 35 (3), 357–379.

Some examples of pseudo-code

The pseudo-codes written for the creation of farms, the diffusion of freshwater and for

investment processes are shown here.

1. Creation of farms

To attribute-cluster

ask patches of rice [

set cluster variable randomly from 0 to 5] ;; give a random number to every patches of

aggregate ;;call the function aggregate

find-clusters ;;call the function find-clusters

To aggregate ;; this function gives to the central patch the modal cluster number in a specific

neighborhood; we can defined the number of repetition; then adjustment of the size of clusters

can be done through the variation of either/both number of repetitions or size of radius

repeat 10 times [ ;; procedure repeated n times

ask patches of rice [

let a [cluster] of patches of rice in-radius 3

let b modes a

let main_cluster one-of b

set cluster main_cluster

To find-clusters

let counter 0

ask patches [with CS1][

set exploitation nobody]

loop [

let seed one-of patches with [CS1 and exploitation = nobody]

if seed = nobody

[ stop ]

ask seed[

set exploitation counter + 1

set counter counter + 1

grow-cluster

To grow-cluster

ask neighbors with [

exploitation = nobody and cluster = [cluster] of myself][

set exploitation [exploitation] of myself

grow-cluster

2. Diffusion of freshwater during the rainy season

to diffuse-freshwater

let a 0.8 ;; coefficient of salinity reduction to simulate the rain

ask patches with [OS = 2 or OS = 3 and elev > 1.25][ ;;water bodies patches at the exception

of the sea and with an elevation superior to 1.25 m

set salt a * salt]

repeat 8 [

ask patches with [OS = 3 or OS = 2 and salt = 0][

ask patches in-radius 5 with [OS = 2 or OS = 3 and salt > 0 and elev > 1.25][

set salt salt - 2

if salt < 0 [set salt 0]

ask patches with [OS = 3 or OS = 2 and salt = 0][

set salt salt / 1.5

set salt salt / 2

3. Investment algorithm

to invest

set counter min [who] of investors

investment

ask investors [die]

to investment

set counter-cluster 1

ask patches [ set groupe 0]

if counter <= max [who] of investors[

ask investors with [who = counter][

let a [willingness_invest] of one-of investors with [who = counter]

ask patches with [(OS = 4 or OS = 5 or OS = 6 or OS = 7) and (attractivity >= (10 - a))

and (visible_area= [visible_area_1] of myself or visible_area= [visible_area_2] of myself)][

set invest counter]]

set limit_area [willingness_capacity] of one-of investisseurs with [who = counter]

to seed

let a one-of patches with [invest = counter and group = 0]

ifelse a != nobody [

ask a[

set group counter-cluster]

growth][

investment]

to growth

ask patches with [invest = counter and group = 0][

set neighbors-cluster count neighbors with [invest = counter and group = counter-cluster]]

let a one-of patches with [invest = counter and group = 0 and neighbors-cluster > 0]

ifelse a != nobody [

ask patches with [invest = counter and group = 0 and neighbors-cluster > 0][

set group counter-cluster]

set size-cluster count patches with [invest = counter and group = counter-cluster]

if size-cluster >= limit_area [

choose

investment]

growth

set counter-cluster counter-cluster + 1

to choose

let number_group max [group] of patches with [invest = counter]

let number_farms max [farm_number] of patches

ask patches with [invest = counter and groupe = number_groupe][

set exploitation numero_exploitation + 1]

ask one-of patches with [invest = counter and group = numero_groupe][

sprout-exploitants 1

Sensitivity tests

Table 1 - Sensitivity test results

Variable Instance variable

in the model

Instance variable

tested

% of observations that

fell between

± 1 � of the reference

simulation (C4) values

% of observations that

fell between

± 2 � of the reference

simulation (C4) values

Attraction Between 3 and 7

Minus 2 50 81

Plus 2 58 93

Household

spending

35-50-70

Minus 20 31 52

Plus 20 41 72

Household

Between 3 and 5

Minus 2 39 74

Plus 2 32 48

Investors 20

10 36 63

40 34 56

Investor

capital

Between 12 and

Times 2 44 79

Divide by 2 68 92

Learning

Absent 42 66

Social Capital 1/6 of the agents 3/6 31 48

Subsidence

Times 2 29 43

Times 3 17 34

Typhoon Minus 100 Php

Minus 200 61 91

Minus 300 62 92

Figure 1 – Proportion of rice and aquaculture farmers with profits

0 20 40 60 80

scenario 1

Iterations

fits(%

0 20 40 60 80

scenario 4

Iterations

fits(%

0 20 40 60 80

scenario 10

Iterations

fits(%

0 20 40 60 80

scenario 2

Iterations

fits(%

0 20 40 60 80

scenario 5

Iterations

fits(%

0 20 40 60 80

scenario 11

Iterations

fits(%

0 20 40 60 80

scenario 3

Iterations

fits(%

0 20 40 60 80

scenario 6

Iterations

fits(%

0 20 40 60 80

scenario 12

Iterations

fits(%

Rice farmers Aquaculture farmers

Figure 2 – Proportion of farmers with fulfilled objectives

0 20 40 60 80

scenario 1

Iterations

0 20 40 60 80

scenario 4

Iterations

0 20 40 60 80

scenario 10

Iterations

0 20 40 60 80

scenario 2

Iterations

0 20 40 60 80

scenario 5

Iterations

0 20 40 60 80

scenario 11

Iterations

0 20 40 60 80

scenario 3

Iterations

0 20 40 60 80

scenario 6

Iterations

0 20 40 60 80

scenario 12

Iterations

Objective 1

Objective 2

Objective 3

Objective 4

Objective 5

Objective 6

Objective 7

Figure 3 – Proportion of strategies adopted by farmers

0 20 40 60 80

scenario 1

Iterations

0 20 40 60 80

scenario 4

Iterations

0 20 40 60 80

scenario 10

Iterations

Repetition

Deliberation

0 20 40 60 80

scenario 2

Iterations

0 20 40 60 80

scenario 5

Iterations

0 20 40 60 80

scenario 11

Iterations

RepetitionLocal Imitation

Global Imitation

0 20 40 60 80

scenario 3

Iterations

0 20 40 60 80

scenario 6

Iterations

0 20 40 60 80

scenario 12

Iterations

Repetition

Imitation

DeliberationSocial comparison

* Difference between local and global imitations is related to the scale of the reference set of agents (from neighbors –local- to network –global)

Figure 4 – Average daily income according to cropping systems

0 20 40 60 80

scenario 1

Iterations

0 20 40 60 80

scenario 4

Iterations

0 20 40 60 80

scenario 10

Iterations

0 20 40 60 80

scenario 2

IterationsA

0 20 40 60 80

scenario 5

Iterations

0 20 40 60 80

scenario 11

Iterations

0 20 40 60 80

scenario 3

Iterations

0 20 40 60 80

scenario 6

Iterations

0 20 40 60 80

scenario 12

Iterations

Aquaculture farmers Rice farmers Agri-aqua farmers

Figure 5 - New areas under aquaculture and converted areas

0 20 40 60 80

1200 scenario 1

Iterations

0 20 40 60 80

1200 scenario 4

Iterations

0 20 40 60 80

1200 scenario 10

Iterations

Aquaculture new areas Converted areas

0 20 40 60 80

1200 scenario 2

Iterations

0 20 40 60 80

1200 scenario 5

Iterations

0 20 40 60 80

1200 scenario 11

Iterations

0 20 40 60 80

1200 scenario 3

Iterations

0 20 40 60 80

1200 scenario 6

Iterations

0 20 40 60 80

1200 scenario 12

Iterations

*converted areas are related to the CS1 (rice) lands that have been converted to CS3 (aquaculture) lands while

aquaculture new areas are related to natural habitats converted to CS3

Figure 6 – Cumulated converted areas compared with CS3 areas

0 20 40 60 80

scenario 1

Iterations

0 20 40 60 80

scenario 4

Iterations

0 20 40 60 80

scenario 10

Iterations

Cumulated converted areas CS3 - aquaculture

0 20 40 60 80

scenario 2

Iterations

0 20 40 60 80

scenario 5

Iterations

0 20 40 60 80

scenario 11

Iterations

0 20 40 60 80

scenario 3

Iterations

0 20 40 60 80

scenario 6

Iterations

0 20 40 60 80

scenario 12

Iterations

Figure 7 – Evolution of global mean salinity of water bodies with and without

subsidence

0 20 40 60 80

Without subsidence

iterations

linity(p

0 20 40 60 80

With subsidence

iterations

linity(p

Figure 8 – Proportion of water body categories with and without subsidence

0 20 40 60 80

Without subsidence

iterations

0 20 40 60 80

With subsidence

iterations

brackish water sea water fresh water

An agentbased model for analyzing land use dynamics in response to farmer behaviour and...

Documents

Transcript of An agentbased model for analyzing land use dynamics in response to farmer behaviour and...

FARMER-TO-FARMER PROGRAM (PCAC), SIUNA

Analyzing meaning - Refubium

Analyzing Toxicity Through Electrophilicity

Abstract delta modeling

Determinants of food insecurity among smallholder farmer ...

Analytical Abstracts on Farmer Participatory Research - - ODI

farmer - Chicken Farmers of Canada

farmer field schools in Uganda - WUR eDepot

Analyzing Project Cash Flows

Changing Course to Secure Farmer Livelihoods Worldwide

2006-DEC_VOL-132_NO-1.pdf - Delta Tau Delta Archive

Human Resources Conference - National Council of Farmer ...

Module 9: Farmer Organisational Development

Analyzing political speeches

Delta dunarii ppt

Untitled - The Farmer Myanmar

Delta Nailing System

Delta Sensorless Vector Control Compact ... - Delta Electronics

6 Issue - 11 November - the Indian farmer

Sustainable Solutions on Farms - Farmer Time