=Paper=
{{Paper
|id=Vol-2030/HAICTA_2017_paper18
|storemode=property
|title=Predicting Operational and Environmental Efficiency of Primary Sectors of EU Countries, by Implementing Artificial Neural Networks
|pdfUrl=https://ceur-ws.org/Vol-2030/HAICTA_2017_paper18.pdf
|volume=Vol-2030
|authors=George Vlontzos
|dblpUrl=https://dblp.org/rec/conf/haicta/Vlontzos17
}}
==Predicting Operational and Environmental Efficiency of Primary Sectors of EU Countries, by Implementing Artificial Neural Networks==
https://ceur-ws.org/Vol-2030/HAICTA_2017_paper18.pdf
Predicting operational and environmental efficiency of
primary sectors of EU countries, by implementing
Artificial Neural Networks
George Vlontzos1
1
Dep. of Agriculture, Crop Production and Rural Development, School of Agricultural
Sciences, University of Thessaly, e-mail: gvlontzos@agr.uth.gr
Abstract. One of the most important policy reforms for the European Union
(EU) agriculture was the implementation of the Agenda 2000, which
establishes a new framework for subsidies management, decoupled from both
crop and animal production for the vast majority of products. One of the main
goals of this new policy framework is the improvement of its environmental
impact. Additionally, there is a need for the implementation of new efficiency
assessment and prognostication tools for the evaluation of EU farming,
because the influence of market forces has been increased substantially.
Regarding prognostication of crop and animal output, as well as Green House
Gas (GHG) emissions, the application of Artificial Neural Networks (ANNs)
is being proposed, succeeding satisfactory quality characteristics for the
models being proposed for operational and environmental predictions in EU
agriculture.
Keywords: Artificial Neural Networks, Agriculture, Efficiency, Common
Agricultural Policy, Environment, Energy
1 Introduction
It is a continuous goal of the European Union (EU) Common Agricultural Policy
(CAP) to improve both operational and environmental efficiency of agricultural
holdings, aiming by this way to increase the competitiveness of EU primary sectors
as a whole in a globalized production and trading framework. The quantification of
this approach is being expressed by the 20-20-20 strategy which focuses on
increasing the energy efficiency by 20%, reducing the CO2 emissions by 20% and
produce 20% of overall energy consumed by renewable energy resources (European
Commission, 2011). One of the most important policy reforms for the EU agriculture
was the implementation of the Agenda 2000, with the establishment of a totally new
framework for subsidies management, decoupled from both crop and animal
production. Since the year 2005 the new subsidy scheme has come into force,
providing by this way the ability to the EU to fully comply with the last World Trade
Organization (WTO) agreement of the Uruguay Round (European Commission,
2013). Under this new framework, the subsidy scheme has a pure supportive role on
the producers’ income, increasing by this way the impact of their managerial
172
decisions on the improvement of efficiency of their holdings.
Quite important is the ability of policy makers to assess the level of success of
policies being planned, before their implementation. Throughout the years it has been
proven that this is not an easy task, due to the fact that this level of success is heavily
depended on the assumptions being made as well as the suitability of the models
being used for such estimations. Regarding agriculture, there are models focusing on
the impact of policies on agricultural trade and development, as well as other tasks
like biophysical and environmental ones. All these widely recognized models are
based on various mathematical methodologies, providing useful information for
significant issues of agriculture, like land management, agricultural trade and
agricultural income. There are though other prognostication methodologies, being
already used in many scientific fields with significant success, like the Artificial
Neural Networks (ANNs). In this paper ANNs are being used as a tool to estimate
future performance of EU countries primary sectors in both operational and
environmental terms.
2 Background
A very important and promising methodology for both performance assessment and
prognostication is the Artificial Neural Network (ANN). ANNs have been used for
predicting purposes for various economic activities. In agriculture quite important is
to establish models for predicting yields, turnovers, and recently undesirable outputs
like GHG emissions. A recent study presented an ANN model for predicting wheat
yield and GHG emissions having a 11-3-2 structure, with R2 0.99 and 0.998 for yield
and GHG emissions respectively (Khoshevisan et al, 2013). On the same trend,
ANNs were used to build models for prognostication of environmental parameters in
potato production. ANN model having 11-10-6 structure achieved the best
performance for this purpose (Khoshevisan et al, 2013). Another case study of ANNs
for predicting greenhouse basil production determined satisfactory results, having a
7-20-20-1 structure and R2 of 0.976 (Pahlavan et al, 2012). A more policy oriented
use of ANNs was applied for cropland change in Romania. This application allows
land-change scientists to identify the spatial determinants based on the observed
changes and to manage complex factional relationships coexisting in agricultural
production process (Lakes et al, 2009).
Especially for CAP, prognostication of the impact of various reforms being
planned was and is a continuous goal. For this reason the Global Trade Analysis
Project (GTAP) applied general equilibrium model has been used taking into
consideration the GTAP global data base (Hertel, 1997; McDougall, 1998). In the
case of the EU enlargement towards central and eastern European countries the
model implementation projected increased agricultural production for these countries
and significant financial transfers from EU taxpayers to the central and eastern
European farmers. The macroeconomic costs for the EU were found to be limited
(Bach et al, 2000). Focusing on agricultural land management issues, the combined
implementation of GTAP and the biophysical (IMAGE) model for the EU after the
173
enforcement of the 2003 CAP reform showed that there will be no drastic decrease
on agricultural land use will occur in the next 30 years due to increased demand for
food globally. On the contrary, significant changes on land use are expected in
developing areas like Africa (Meijl et al, 2006). Another modelling approach is the
International Model for Policy Analysis of Agricultural Commodities and Trade
(IMPACT) model. It has been developed by the international food policy research
institute focusing on connecting food supply chain and water supply and demand
(Rosegrant et al, 2008). More recent attempts to simulate farm operation in regards
of policy recommendations lead to the Farm System Simulator (FSSIM). This is a
bio-economic farm model linking micro and macro analysis of farming systems in
specific regions (Louhichi et al, 2010).
A new forecasting methodology is being proposed to predict outputs on both
operational and environmental level, by the implementation of ANNs. It will be
applied separately on an operational and environmental basis, attempting by this way
to identify which methodological approach is simpler, regarding ANN structure, and
which model performs better for prognostication purposes.
3 Methodological approach
For everyone who participates in a direct or indirect way in agricultural production
process, it is very important to have the ability to foresee the impact of the
implementation of specific policies and interventions in general in the near future. As
it was presented in the literature review section, up to now there have developed
models for this purpose, taking into consideration parameters affected by the CAP.
The recent radical changes though towards a liberalized and more market oriented
policy approach, provide the framework for the implementation of reliable models
for prognostication purposes, being used for a long period of time for both economic
and engineering activities (Chinchuluumet al, 2008; Zopounidis and Pardalos, 2010).
Such models are the ANNs. The implementation of ANNs is being used to predict
crop and animal production, as well as GHGs emissions, by using available data sets,
in order to examine the suitability of both of them for prediction purposes (Table
1,2). ANNs time series problem definition requires the arrangement of input vectors
and target vectors as well. The type series problem being used aims to predict future
values of a time series y(t) based on past values of that time series and from past
values of a second time series x(t). This prediction form is called Nonlinear
Autoregressive with Exogenous input (NARX), with the formula describing it to be
the following:
y(t)=f(y(t-1),…,y(t-d),x(t-1),…, (t-d)) (1)
The input and target vectors are randomly divided into three sets, the training,
validation and generalization ones. The ration among them is 70%, 15% and 15%
respectively. NARX is a two-layer feedforward network consisted of a sigmoid
function in the hidden layer and a linear transfer function in the output layer. The
output is fed back to the input of the network through delays. The Levenberg-
174
Marquardt algorithm is used for training the network. The comparison of the two
networks will verify which approach, the operational or the environmental one, is the
most appropriate for prediction purposes of the impact of CAP on efficiency
improvement of agriculture of EU countries.
Table 1: Basic statistics of Inputs
Agricultural Fixed capital
Land Chemicals Energy Fertilizers consumption Labour
Medium 6,660.1 358.8 770.2 516.0 1,791.3 1,199.8
Standard
Deviation 11,236.78 596.26 1049.89 1180.28 1800.68 1859.82
Max 35,177.8 3,021.5 4,502.7 4,604.5 12,377.4 7,307.4
Min 9.7 0.5 5.4 1.0 3.8 3.2
Source: Eurostat
Table 2: Basic statistics of Outputs
Animal Output Crop Output GHG Emissions
Medium 5,032.9 6,624.2 17.8
Standard Deviation 8,658.58 5,695.11 28.65
Max 25,987.7 44,407.2 100.5
Min 67.3 43.2 0.1
Source: Eurostat
It is obvious that there is quite significant variation for every input and output being
used. Such differences in such cases are expectable due to the considerably different
sizes of primary sectors of EU countries.
4 Solutions and recommendations
For the implementation of ANNs two data sets were used, following the same
approach with the DEA Window models. All models were trained, tested and
validated by using the MATLAB® 2015b software. The first ANN aims to
prognosticate crop and animal output, by using as inputs only the non-energy
dependent ones, which are agricultural land, labour and capital. The second ANN
aims to prognosticate not only the operational outputs, based on the operational
inputs, but the GHGs too, by adding as inputs all the relevant energy depended ones,
like fertilizers, agrochemicals and fuels.
The best performance of the first model was achieved by applying 12 hidden
neurons and 3 delays. For this structure the Mean Square Error (MSE) is 0.052629 at
the 5th epoch, with 11 epochs being tried. This score is significantly low and can be
175
considered as acceptable. The network was created and trained in an open loop form,
in order to have the ability to get correct past outputs during the training period and
produce the correct current outputs. The R2 for validation was 0.9744 which is
acceptable too, with R2 for all three stages of the model to be 0.93524.
Figure 1. MSE scores operational ANN
The second ANN succeeds the best performance is a different structure. It consists of
11 hidden neurons and 4 delays. The following figure presents the MSE for training,
test and validation of the model, achieving the best MSE for validation 0.11325 at
epoch 5 after implementing 11 epochs. The R2 for validation is 0.97365 and the R2
for all three stages of the model to be 0.98344. Comparing the two models it is
obvious that although the qualitative characteristics of both of them are quite
satisfactory, the ANN using the energy dependent inputs and undesirable output
performs better, because it requires a simpler structure and the R2 overall score is
higher too. These findings provide considerable hints that using energy dependent
data for efficiency estimations in agriculture is more safe, compared with the use of
pure operational one, signifying at the same time that using market oriented data sets
lead to more reliable forecasting results. It remains to be seen in the near future
though, when there will be available data from non-energy pure operational inputs
not affected by policy interventions, if this qualitative difference between ANNs will
still remain or not.
176
Figure 2. MSE scores environmental
5 Conclusion
Efficiency in agriculture, especially after the recent and radical reforms of
CAP towards more liberalized subsidy management practices, is a top
priority issue for farmers, policy makers and taxpayers. It is proven that when
farming managerial practices are driven by market forces, there is an
improved efficiency outcome, verifying that CAP reforms are heading
towards the right direction, having as precondition a globalized trading
environment for agricultural products. Implementation of ANNs propose a
new methodological approach for ex ante policy evaluation, utilizing
knowhow from other activities, like engineering and economics, which are
more market oriented, compared with the majority of agricultural products
being produced in the EU. The widely accepted advantages of this
methodology are expected to provide safer prognoses regarding operational
activities and environmental safety, increasing by this way the level of
177
success of CAP, improving at the same the utility of financial transfers from
taxpayers to farmers.
References
1. Bach, C.F. Frandsen, S.E., Jensen, H.G. (2000). Agricultural and
Economy-Wide Effects of European Enlargement: Modelling the
Common Agricultural Policy. Journal of Agricultural Economics,
51(2), 162-180.
2. Chinchuluun, R., Lee, W.S., Bhorania, J., Pardalos, P.M. (2009).
Clustering and Classification Algorithms in Food and Agricultural
Applications:A Survey. P.J. Papajorgji, P.M. Pardalos (eds.),
Advances in Modeling Agricultural Systems, Springer Science +
Business Media, LLC
3. European Commission (2013). Overview of CAP reform 2014-2020.
Retrieved May 19, 2016, fromhttp://ec.europa.eu/agriculture/policy-
perspectives/index_en.htm
4. Ghasemi Mobtaker, H., Akram, A., Keyhani, A., Mohammadi, A.
(2012). Optimization of energy required for alfalfa production using
data envelopment analysis approach. Energy for Sustainable
Development, 16, 242-248.
5. Heidari, M.D., & Omid, M. (2011). Energy use patterns and
econometric models of major greenhouse vegetable productions in
Iran. Energy, 36, 220-225.
6. Hertel, T. (1997). Global Trade Analysis: Modeling and
Applications. Published by Cambridge University Press.
7. Khoshnevisan, B., Rafiee, S., Omid, M., Mousazadeh, H. (2013).
Applying data envelopment analysis approach to improve energy
efficiency and reduce GHG (greenhouse gas) emission of wheat
production. Energy, 58, 588-593.
8. Khoshnevisan, B., Rafiee, S., Omid, M., Mousazadeh, H. (2013).
Reduction of CO2 emission by improving energy use efficiency of
greenhouse cucumber production using DEA approach. Energy, 55,
676-682.
9. Lakes, T., Muller, D., Kruger, C. (2009). Cropland change in
southern Romania: a comparison of logistic regressions and artificial
neural networks. Landscape Ecology, 24, 1195-1206.
10. McDougall, R.A., A. Elbehri, T.P. Truong (1998). Global Trade
Assistance and Protection: The GTAP 4 Data Base, Center for
Global Trade Analysis, Purdue University.
11. Meijl, H.V., Rheenen, T.V., Tabeau, A., Eickhout, B. (2006). The
178
impact of different policy environments on agricultural land use in
Europe. Agriculture Ecosystems and Environment, 114, 21-38.
12. Pahlavan, R., Omid, M.,Rafiee, S.,Mousavi-Avval, S.H. (2012).
Optimization of energy consumption for rose production in Iran.
Energy for Sustainable Development, 16, 236-241.
13. Seiford, L.M. and Zhu, J. (2002). Modeling undesirable factors in
efficiency evaluation. European Journal of Operational Research,
142, 16-20.
179