RECORD

An Integrated Platform for Agro-ecosystems Study

Ronan Tr

epos, H

ene Raynal and Gauthier Quesnel

INRA, UR875 Biom

etrie et Intelligence Artiﬁcielle, F-31326 Castanet-Tolosan, France

Keywords:

Agro-ecosystems, Modeling and Simulation Platform.

Abstract:

The complexity of models developped in order to improve the agricultural systems and the need to have

efﬁcient tools to build them, simulate and analyse them have motivated the conception and the development of

the new modelling and simulating platform RECORD. This paper describes how this initiative of the French

National Institute for Agricultural Research has been conducted. A pragmatic strategy consisting in integrating

heterogeneous tools into a same framework has been used. The example of the integration of the 3 different

tools: VLE, R, Python illustrates how the users’ requirements have been fullﬁlled. The platform is currently

used in various projects, and on the basis of the ﬁrst experiences, we conclude on the interest of this strategy.

We underline that this strategy must be accompanied by efforts on developping user’s training and coaching on

these powerful tools. Finally software design should facilitate collaborative developments, which will motivate

our future works.

1 INTRODUCTION

Agriculture is currently driven by continuously evolv-

ing challenges (e.g. climate change, environmental

concerns, market globalisation, food safety and qual-

ity) and requires the development of innovative sus-

tainable production systems. Modelling and simu-

lation approaches are usefull in identifying the agri-

cultural systems that respond to current social, eco-

nomic, political and environmental concerns. How-

ever, building, testing, evaluating and using models

(i.e. simulating different combinations of manage-

ment, soil and climate conditions and policy and envi-

ronmental constraints) is far from being a straightfor-

ward task. Numerous models that can address ques-

tions about agro-ecosystem functioning already ex-

ist. The major issue now is how to combine or cou-

ple these models and use them at different spatial

and temporal scales rather than develop new models.

To overcome the problems which arise when build-

ing, simulating and reusing models, generic com-

puting platforms have been created. Most of these

platforms capitalise on advances in computer sci-

ence (e.g. object-oriented, modular and generic pro-

gramming) and propose model repositories to facili-

tate their use and re-use (e.g. CropSyst (Van Evert

and Campbell, 1994) or ICASA (Bouma and Jones,

2001)).

To take advantage of these developments the

French cropping-system research community tries to

deﬁne a strategy to developp an integrated modelling

platform, named RECORD (REnovation and COOR-

Dination of agro-ecosystems modelling), to gather,

link and build models and companion tools to answer

new agricultural questions (Bergez et al., 2012). We

present in this paper the pragmatic approach we have

used to developp the platform. It is essentially based

on dedicated software development and reuse of ex-

isting tools. The ﬁrst section gives an overview of

the functional requirements of the platform. Section

2 explains how the RECORD platform was designed

in order to fulﬁll the initial requirements.

2 REQUIREMENTS FOR THE

PLATFORM

The functional requirements for the platform were de-

ﬁned following a survey targeting modellers from the

Environment and Agronomy department of INRA and

seminar in 2007.

Scope and Users. The platform must be able to

simulate cropping systems and interactions with the

surrounding territory. Four types of users were identi-

ﬁed: i) researchers working on and designing crop-

186

Trépos R., Raynal H. and Quesnel G..

RECORD - An Integrated Platform for Agro-ecosystems Study.

DOI: 10.5220/0004058801860189

In Proceedings of the 2nd International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH-2012),

pages 186-189

ISBN: 978-989-8565-20-4

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: The coupling of a decision model, an assessment model called Bilan, and the corn crop model 2CV designed into

VLE. Up right, the hierarchical structure of the model is given into a tree structure. The model embeds coupled models (e.g.

2CV, CropFull) and atomic models (e.g. Decision, SoilSWCB).

ping systems ii) researchers using cropping system

models iii) extension service personnel who wants to

test new cropping-systems proposals and iv) graduate

students being trained in modelling.

Accepting Various Time and Spatial Scales.

Hourly simulations up to several decades for sustain-

ability studies have to be permitted. Spatial aggrega-

tions include a plot or a set of plots representing one

or more farms, including, if necessary, interstitial ar-

eas and larger areas such as a territory when dealing

with natural resources (e.g. a catchment). Transfers

between different spatial units must also be modelled

(e.g. water exchange, pests, genes, spores).

Accepting Various Formalisms. Most current

models are dynamic with discrete time steps. How-

ever, the platform also should be able to incorpo-

rate static and stochastic models and different for-

malisms (e.g. difference equations, differential equa-

tions, Markov chains, state charts, cellular automata).

Focusing on Management. It is particularly im-

portant that the platform be able to handle manage-

ment models for cropping systems. Modelling fac-

tors such as technical operations sequences, compe-

tition among agricultural tasks, spatial distribution of

agricultural practices and choice of crop rotations in a

ﬁeld must be possible.

Using Existing Codes. Existing crop models cover

a wide variety of crops, crop management options

and spatial and temporal scales. The architecture of

the platform should enable integration of these dif-

ferent types of models, either by recoding or by en-

capsulating existing code. The design of the platform

should also allow model developpers to share easily

their models and generic developments through a col-

laborative development strategy.

Using Models. The platform should facilitate i) im-

plementation of optimisation methods by simulation,

ii) implementation of methods of multicriteria choice,

iii) comparison of cropping-systems models and iv)

use of data-mining techniques to exploit simulation

results. Deployment of models on web sites should

be facilitated by generic web development.

3 METHODS AND TOOLS

Applications in agriculture or environment requires

frequently the integration of several models, the pos-

sibility to share them and to explore them with nu-

merical methods. For modelisation and simulation,

the platform relies on DEVS theory and VLE soft-

ware. The use of R software is favoured for statisti-

cals works on models (parameterization, evaluation).

Finally, python language and its libraries offer a good

solution for web development.

DEVS Theory. DEVS (Discrete Event System

Speciﬁcation) is a theoretical formalism for mod-

elling and simulating dynamic systems with discrete

events. It provides a robust solution for the coupling

and integration of heterogeneous models and the im-

plementation of complex hierarchical systems as it

was proved by formal mathematical proofs (Zeigler

RECORD-AnIntegratedPlatformforAgro-ecosystemsStudy

187

et al., 2000). There are two types of models in DEVS,

atomic and coupled models, which exchange infor-

mation in the form of events. Based on DEVS theory,

the abstract simulator DSDE (Barros, 1998), is able

to modify model structure during a simulation with

an atomic model called executive.

VLE Software. VLE (Quesnel et al., 2009) is a

free and open-source software developed in C++

which provides simulators, modelling tools, libraries,

an IDE (Integrated Development Environment) and

a SDK (Software Development Kit). It is a

generic modelling, simulation and analysis environ-

ment based on the DEVS formalism, and implements

a DSDE simulator. VLE offers a library of extensions

which correspond to commonly-used mathematical

formalisms: cellular automata, statecharts, ordinary

differential equations (including classical numerical

integration schemes such as Runge Kutta and QSS

(Cellier et al., 2008)), difference equations and a de-

cision making extension based on planiﬁcation of ac-

tivities. We identiﬁed the three last cited extensions

as a basis for modeling agro-ecosystems. Solutions

for complex models that involve large territories are

provided. For example modelling hundreds of plots

is achieved by using “executive” models that gener-

ate the plots and their connections from GIS (Geo-

graphical Information System) data. A GUI (Graphi-

cal User Interface) is used to edit models, to set initial

conditions and to perfom simple simulations (Figure

1). For multi-simulation, one can express experiment

plans as combinations of initial conditions and par-

allelisation of simulations on different cores can be

performed.

Statistical and Mathematical Works with Models.

Statistical methods for parameter estimation, valida-

tion and sensitivity analysis are necessary for imple-

menting and analysing agro-ecosystem models (Wal-

lach et al., 2006). Platform RECORD relies on a

package rvle that links VLE to the statistical language

R (R Development Core Team, 2011) in order to per-

form models calibration and exploration. For exam-

ple, Table 1 gives indices resulting from a sensitivity

analysis performed with fast99 (Saltelli et al., 1999)

on the crop model 2CV. It concerns 5 decision vari-

ables for irrigation and targets a net margin output.

Based on this analysis, one can search for the couple

of values of water supply and stock (ie. the variables

that impact the most the output), that lead to a maxi-

mal expected net margin. The stochastic behavior of

the model comes from climatic data produced by a

a stochastic generator of climatic series available into

the platform. Results of a simulated annealing optimi-

sation, performed with R function optim, suggest the

couple (16 mm, 121 mm) which leads to an expected

output of 898 euros per hectare.

Concerning optimisation of agro-ecosystems

models, challenges such as multi-objective (e.g.

economic and ecologic criterion) and optimisation

under uncertainty (when e.g. models depend on

climatic data) arise rapidly; see e.g. (Crespo et al.,

2010). Current works also concern embedding opti-

misation into simulation. For example, optimising the

behavior of an agent that reacts to a system such as

a biophysical model, by mean of e.g. reinforcement

learning (Garcia, 1999) requires that the optimisation

is ﬁnely coupled to simulation.

Table 1: Indices (ordered by total effect) of 5 decision vari-

ables resulting from a sensitivity analysis.

variable

main

interactions total

effect

water stock 0.54 0.29 0.83

water supply 0.37 0.39 0.76

irrigation

0.20 0.20 0.40

delay

threshold on

0.002 0.042 0.044

soil water

threshold on

0.003 0.040 0.043

available water

Web Applications. The models in the RECORD plat-

form could be of real interest as educational tools to

improve understanding of agro-ecosystems or as tools

for decision makers to integrate agronomic, social,

economic and environmental issues. However, di-

rect use of these models by non-scientists or by users

with limited computed skills could be problematic be-

cause i) they often require signiﬁcant parameterisa-

tion to adapt them to individual contexts, ii) model

validity is context-dependent and iii) working directly

on the platform requires a minimum level of train-

ing. Therefore, we proposed a web-oriented tool, de-

signed to develop web applications around the mod-

els implemented on the platform (Figure 2). This

tool is based on the development of a speciﬁc pack-

age: PyVLE that enables the use of VLE from Python

applications. This package provides a set of func-

tions that enable the management of VLE simulations

from Python scripts. Therefore, VLE simulators can

be parameterized, launched from Python and results

are available for further uses such as graphical repre-

sentations. A framework (such as Pylons, Django or

Zope) to help in developing the application is used.

The database holds the information used to specify

the various simulated contexts (e.g. climate, soil, va-

rietal characteristics, crop distributions). There is one

model per context, and several contexts are simulated

at the same time. The results are aggregated into in-

SIMULTECH2012-2ndInternationalConferenceonSimulationandModelingMethodologies,Technologiesand

Applications

188

dicators and graphs which are easily understandable

by professionals. Some applications are already on-

line and used by scientists, students and agricultural

advisors.

User browser

Service Web

PyVLE

VLE

Framework Pylon

Python

User database

Server

Figure 2: The main elements of the RECORD web archi-

tecture.

4 CONCLUSIONS AND

DISCUSSION

Since 2007, choices have been made for the develop-

ment of the platform RECORD. First, it has been de-

cided that all features that were required should not be

part of one unique tool, in light of the large range of

disciplines that are involved. Nevertheless the techni-

cal solutions should not restrain the different commu-

nities to a conﬁned set of methods and tools. Techni-

cal solutions for model development, model analysis

and web applications are the following:

• The software VLE provides an IDE for designing

and coupling models. Coupling properties are en-

sured by the DEVS theory which is the basis of

VLE. This IDE can be used to perform the ﬁrst

simulations.

• R and its libraries are used for analysing and ex-

ploring models. The platform relies on a speciﬁc

R package that links R to VLE.

• Python and its libraries are the basis of web ap-

plications development for RECORD projects,

thanks to a link between Python and VLE.

A large set of methods and tools are thus possibly

carried out into the platform. Due to this integration

strategy, there are few assumptions that are made re-

garding the methods that are involved. For example,

parameter estimation of a model using R can be per-

formed in different ways. The challenge of RECORD

is to identify well suited methods with the help of ex-

perts and to build training courses which are provided

through an E-learning web site and 3-days practice

sessions (2 per year).

We ﬁnally argue that integrating tools is the best

solution to develop a collaborative platform. An-

other important task that the platform team addresses

is to facilitate the collaborative work. The solutions

proposed and which are currently under construction

comprise the package management system into VLE,

attachment of meta-data to packages and models for

documentation and conﬁguration.

REFERENCES

Barros, F. J. (1998). Abstract simulators for the dsde for-

malism. In Proceedings of the 30th conference on

Winter simulation, WSC ’98, pages 407–412, Los

Alamitos, CA, USA. IEEE Computer Society Press.

Bergez, J.-E., Chabrier, P., Gary, C., Jeuffroy, M.,

Makowski, D., Quesnel, G., Ramat, E., Raynal, H.,

Rousse, N., Wallach, D., Debaeke, P., Durand, P.,

Duru, M., Dury, J., Faverdin, P., Gascuel-Odoux, C.,

and Garcia, F. (2012). An open platform to build, eval-

uate and simulate integrated models of farming and

agro-ecosystems. Environmental Modelling And Soft-

ware. In press.

Bouma, J. and Jones, J. W. (2001). An international collab-

orative network for agricultural systems applications

(icasa). Agricultural Systems, 70:355–368.

Cellier, F., Ernesto Kofman, E., Migoni, G., and Bortolotto,

M. (2008). Quantized state system simulation. In Pro-

ceedings of Summer Simulation Multiconference 08.

Crespo, O., Bergez, J., and Garcia, F. (2010). Multiobjec-

tive optimization subject to uncertainty: Application

to irrigation strategy management. Computers and

Electronics in Agriculture, 74(1):145–154.

Garcia, F. (1999). Use of reinforcement learning and simu-

lation to optimise wheat crop technical management.

In Proceedings of the International Congress on Mod-

elling and Simulation (MODSIM 99), pages 801–806,

Hamilton, New-Zealand.

Quesnel, G., Duboz, R., and Ramat, E. (2009). The Virtual

Laboratory Environment – An operational framework

for multi-modelling, simulation and analysis of com-

plex dynamical systems. Simulation Modelling Prac-

tice and Theory, 17:641–653.

R Development Core Team (2011). R: A Language and

Environment for Statistical Computing. R Foundation

for Statistical Computing, Vienna, Austria.

Saltelli, A., Tarantola, S., and Chan, K. P.-S. (1999).

A quantitative model-independent method for global

sensitivity analysis of model output. Technometrics,

41:39–56.

Van Evert, F. K. and Campbell, G. S. (1994). Cropsyst -

a collection of object-oriented simulation-models of

agricultural systems. Agronomy Journal, 86:325–331.

Wallach, D., Makowski, D., and Jones, J. (2006). Work-

ing with dynamic crop models: evaluation, analysis,

parameterization, and applications. Elsevier.

Zeigler, B. P., Kim, D., and Praehofer, H. (2000). Theory of

modeling and simulation: Integrating Discrete Event

and Continuous Complex Dynamic Systems. Aca-

demic Press.

RECORD-AnIntegratedPlatformforAgro-ecosystemsStudy

189