Matemáticas aplicadas, estadística y computación

 

Artificial neural networks for sustainable agribusiness: a case study of five energetic crops

 

Redes neurales artificiales para agronegocios sostenibles: un caso de estudio de cinco cultivos energéticos

 

Mircea Untaru^* , Vasile Rotarescu, Liliana Dorneanu

 

loan Slavici University, 144, Dr. A. Paunescu-Podeanu Street, 300587, Timisoara, Romania. *Author for correspondence. (titusslavici@yahoo.com).

 

Received: January, 2011.
Approved: July, 2012.

 

Abstract

The growing agricultural economic environment referred to as agribusiness requires continuous balanced cost-benefit solutions. The use of artificial intelligence in this area provides complex solutions that are easily applicable. The objective of this study was to elaborate on innovative instruments from the field of artificial intelligence for the decision-making process related to energetic crops. The field of expertise of this paper is strongly related to current issues of sustainable development. The methodology used is artificial neural networks (ANN) and compares it with other tools. The targeted results regard the optimization of decision-making processes and the forecast of financial results in an agricultural economy. A case study of forecasting profits of five energetic crops is included.

Key words: energetic crops, neural networks, agribusiness, resource saving.

 

Resumen

El crecimiento del entorno agrícola-económico conocido como agronegocios requiere continuas soluciones equilibradas costo-beneficio. El uso de la inteligencia artificial en esta área ofrece soluciones complejas de fácil aplicación. El objetivo de este estudio fue elaborar instrumentos innovadores del campo de la inteligencia artificial para el proceso de toma de decisiones en cultivos energéticos. El campo de especialidad de este artículo está muy relacionado con los problemas actuales del desarrollo sostenible. La metodología usada es redes neuronales artificiales (ANN) y la compara con otras herramientas. Los resultados previstos consideran la optimización de los procesos de toma de decisiones y pronostican resultados financieros en una economía agrícola. Se incluye un estudio de caso de pronosticar utilidades de cinco cultivos energéticos.

Palabras clave: cultivos energéticos, redes neurales, agronegocios, ahorro de recursos.

 

INTRODUCTION

The objective of this research was to contribute through practical solutions to optimize a balanced cost-benefit use of natural resources in agribusiness. Thus, the article suggests the use of energetic crops by firms. The research ensures an accessible artificial neural network (ANN) for helping such firms in profit estimation and decision making regarding the cultivation of those crops.

Environmental sustainability is a debated and supported matter in Europe and the world; energy savings is an urgent matter in agriculture. According to Hernandez-Santoyo and Sánchez-Cifuentes (2003), there is a need for new processes in the area of energy savings and their solution involves the use of a trigeneration system. Within this framework created by regulations and international agreements for the preservation and protection of the environment, a scientific approach to the agricultural economy has become increasingly important since the late 1990s. Hu and Kao (2007) point out that between 2001 and 2006, global sites where energy-saving ratios significantly improved include China, Hong Kong, Philippines, USA (which has the highest energetic efficiency level) and Chile-México-Taiwan. In Eastern Europe and Asia, Fischer et al. (2005) studied agro-ecological zones and supported the elaboration of global, regional and national assessments of agricultural potentials; the study provide a standardized framework for identifying climate soil and soil conditions relevant to crops. The models in this paper use environmental contexts as independent variables in forecasting profitable crops in agricultural businesses.

From the traditional features of the primary sector, the trend over the past decades has been toward the establishment of a business environment. Agricultural business (or agribusiness) represents an area ofparticular concern.The EU created a Common Agricultural Policy for European countries, later followed by the establishment of structural funds. This adequate specific legal European framework increased the regional development in weak rural economic areas. Bonanno (2003) describes three main directions concerning agriculture in Eastern Europe in the late 1990s: a wide variety in the primary sector, the role of the market in the reorganization of Eastern European agriculture and changing social stratification in rural regions. Previous guidelines have been developed within the context of agricultural transformation from an on-demand system to a market-oriented one. The use of energetic plants provides sustainable solutions for heating, but such crops can also grow in swamp areas, capitalizing on lands categorized as useless. The utilization of such plants in agribusiness is highly significant. However, in less developed countries, managers of agribusiness are hesitant to use such crops as a source of profit.

Different approaches are used to forecast profitable crops in agricultural businesses. Some studies combine interdisciplinary areas such as non-linear and dynamic agricultural systems. These require advanced techniques and technologies for sustainable solutions in the primary sector. Moreover, the promotion of these technologies supports cost decreases and catalyzes sustainable development in agriculture (Murase, 2000).

Artificial neural networks are an effective tool for dealing with the complexity of agricultural data (high-variance, noise-affected, qualitative data). The use of neural networks in agribusiness has become frequent since the mid-1990s, and most of the research has addressed yield prediction, spatial modeling (irrigation management and spectral vegetation indices) and spatial-temporal forecasting (time series of the yield in a specific region).

One of the first uses of neural networks in agriculture is found in the study of Joerding et al. (1994), for estimating the yield of Bermuda grass (Cynodon dactylon) depending on the quantities of various fertilizers used on the crop. A study by Heinzow and Tol (2003) included seven of the most common crops in Germany and the authors constructed regional time series of crop yield under climate change, recording time series of temperature, precipitation and crop yields as input variables. According to the authors, the study was suited for the German regions considered and also for the adjacent European regions. Savin et al. (2007) constructed a neural network that predicted the crop yield for winter wheat in some regions of southern Russia with an average accuracy of approximately 75 %. Recent studies have proposed neural networks for forecasting yields for rice (Oryza sativa) in Northern India (Kumar, 2011), onion (Allium cepa) (Stastny et al., 2011), tomato (Solanum lycopersicum) (Qaddoum et al., 2011) and sunflower (Helianthus annuus) in saline fields (Dai et al., 2011), taking into account the crop yield from an economical point of view. All of these studies require a level of knowledge more advanced than that of an investor without agricultural background and whose main objective is to establish a profitable agricultural business.

Our study considers five types of energetic crops in the moderate temperate continental climate specific to Western Romania. For each crops profit per ha was estimated for a three-year period. Six factors that influence the profit were identified, including natural disasters. The paper presents a model created through ANN to determine the profit level of agricultural firms regarding the use of energetic crops.

 

MATERIALS AND METHODS

Theoretical aspects of artificial neural networks and their application to the case studies

In this section, a short summary of the main theoretical aspects of ANN and their use in this study is provided. Neural networks are considered the most accurate tools for economic decision-making (Slavici, 2006). Its learning capacity is most likely the fundamental feature of a neural network, and modern algorithms aim to emulate this characteristic by using computers (Russell and Norvig, 2002).

The fundamental features of ANN can be classified into two categories:

1) The ANN architecture which defines the structure and determines the number of neurons and their connections; its other features include the input/output (I/O), the synapse intensity, deviations and activations.

2) The ANN properties that consist of: the learning method, the synapse reactivation, and the continuous associations and comparisons of new information with existing one to assess how the new information is categorized.

Our ANN is trained to store a wide database regarding energetic plants and all of their parameters and usage features with various types of crops and geographical areas in Eastern Europe. The database has been subsequently used for economic forecasting and decision-making. The economic results of the crops have been forecasted, and decisions were made regarding the cultivation of the most profitable energetic crops.

General description of the database used in the application

A database was constructed using data collected by Ioan Slavici University by administering yearly questionnaires from 2008 to 2010 in Eastern Europe (three countries) on a sample of 120 agricultural firms. Additional data was collected during a series of experiments on the energetic willow (Salix viminalis) at the University's field laboratories. The collected data have been statistically processed using Statgraphics Centurion; the mean and dispersion were computed and the noise of the data (irrelevant values) was eliminated. The hypotheses were derived from the interpretation of the questionnaire. The University has agreed to the publication of the studies and results concerning the database that was used.

The average results of the costs and generated income are presented in Tables 1 and 2.

The energetic crops for this study were energetic willow, miscanthus (Miscanthus giganteus), reed (Phragmites australis), sorghum (Sorghum saccharatum) and rapeseed (Brassica rapa oleifera). However, the users can introduce data for any type of energetic crop.

The data presented in Tables 1 and 2 represent average values of the costs and corresponding incomes. The database was constructed over three years by collecting multiple data, including plant type, soil, climate and specific conditions (groundwater and surface water gleying, salinity/sodicity, topsoil texture, soil pollution, slope, landslides, groundwater depth, liability to inundation, total porosity, calcium carbonate levels, soil reaction, physiologically useful volume, humus reserve, excess moisture, etc.) and economic criteria (land concession price, grouping of areas for insuring a minimum of 150 ha, the wages of the local workforce, etc.). Next, the incomes over multiple years were determined, and the final step was the determination of the profits.

The profit was calculated as the difference between income and total cost. In addition, the income has been calculated as a product of production and price per unit.

Sustainable decisions using artificial neural networks in cultivating energetic plants

For simulation purposes, we used MATLAB and its neural network toolbox to try predicting the financial gain of harvesting five energetic crops(energetic willow, miscanthus, reed, sorghum saccharatum and rapeseed) under various quantifiable conditions.

The list of conditions fed to the system is extracted from the following six one-dimensional column vectors Xi, with the value of i ranging from 1 to 6. Thus, X1 represents "Plant type", X2 -"Year of observation relative to planting year", X3 - "Rainfall/ year", X4 - "State of the economy", X5 - "Natural disasters" and X6 - "Demand/supply market elasticity".

Because the study examined five plants, "Plant type" varies from 1 to 5 with the following interpretations: energetic willow plant, miscanthus plant, reed, sorghum saccharatum and rapeseed.

Even though the simulation takes into account the data collected from 2008 to 2010 through surveys, the actual plantations were conducted 10 years before the data were collected. Thus, the "Year of observation" vector contains relative values ranging from 1 to 10.

The production quantity is expressed in t/ha and rainfall is expressed in average mm/year. The "State of the economy" parameter can take only one of two values: "1," meaning that the economy was in a normal state when the data were collected, or "0," which is interpreted as indicating "bad economical times" or "economic crisis."

The "Natural disasters" vector reflects values of the amount of damage sustained by the crop in that year when being hit by a natural calamity (wildfire, drought, hail, etc.). Its values can be interpreted as follows: "0" indicates there were few or no natural problems associated with the crop and a 100 to 75 % harvesting record was obtained, "0.33" indicates that the crop was partially affected by nature and a 75 to 50 % harvesting record was obtained, "0.66" indicates that the crop sustained massive natural damage and a 50 to 25 % estimated harvesting potential was recovered, and "1" indicates that the crop was almost completely destroyed by natural events and only 25 to 0 % of it was recovered. The last input parameter of the system was the "Demand/supply" coefficient: a value less than "1" indicate that there was a demand deficit, a value of "1" indicate that the market was at an equilibrium and a value exceeding "1" indicate a supply deficit. By merging these vectors, the training matrix P = [X1 X2 X3 X4 X5 X6] was obtained. Any line (row) in this newly formed matrix is considered a training sample for the neural network.

In the construction and use of the network (Figure 1), three steps were taken:

Step 1: Training of the network

The training of the network used a set of 90 records, randomly chosen, representing 75 % of the input data.

Step 2: Network validation

After training the network, the process continued with the actual evaluation of its performance (validation of ANN according to the latest version of the MATLAB program). The validation set consisted of 25 inputs, representing 20.83 % of the input data. This result was achieved by acquiring further input samples, applying them to the network and examining and comparing its output. It is from this output that the error rate per sample is calculated and further summed.

Step 3: Testing the network

The ANN, which has been trained in Step 1 and validated in Step 2, was tested using 5 records, representing 4.17 % of the input data.

 

RESULTS AND DISCUSSION

Artificial neural network performance

For the training phase there was a strong correlation between targets and outputs, the correlation coefficient was 1 and the error 2.5e-007 (Figure 2A). For the validation phase the correlation coefficient between the targets and outputs was 0.999, showing that the outputs were properly estimated (Figure 2B). For the test phase, the correlation coefficient between the outputs and the targets was 0.997, an indicator of the high accuracy of the network (Figure 2C). Thus, the overall correlation coefficient between outputs and targets was 0.999 (Figure 2D). This represents an indicator of the high performance achieved by the network we propose.

The validation phase is summarized in Table 3. The quantities X1 to X6 represent the inputs of the network, as depicted above. Oo the obtained output (the profit predicted by the network), Co the correctly measured output (the profit actually obtained while harvesting a particular crop) and Err (%) the signed percentage error, given by

The absolute average error value observed from these evaluation data was 5.69 %. In the end, we obtained the means of accurately predicting the gain of harvesting five energetic crops, taking into account various non-quantifiable parameters such as rainfall, market climate and/or demand elasticity.

The target group of the present application is composed of the economic agents in the agricultural economic environment. Except for the classic economic benefit obtained by using such methods, social benefits can be obtained. These consist of quality improvement ofthe lands (sanitation ofwatery lands, cleared forests), saved energy (or consumption avoided), reduction of greenhouse gases that affect the environment and new jobs for an unskilled labor force during the break from the crop when the leaves fall. Additionally, it must be considered that during autumn agricultural machinery is used only for the crops. According to the European Commission, all of these benefits must be quantified (the revenue from the energy sale at appropriate shadow prices). The latter can be approximated, wherever possible, by estimating the willingness to pay for energy (the quantification of marginal costs incurred to acquire energy when installing and using private generators) (DG Regional Policy, 2008).

Once constructed, because of its ease of use, ANN represents an innovative tool for business management that is available even for medium-skilled agricultural managers. Such an instrument represents an investment for small enterprises and will result in new innovative technologies being included in the agricultural sector, which Eastern Europe especially lacks. Directions for future research include the development of a quantitative analysis of the social benefits regarding this case.

Artificial neural networks optimization to increase the forecast accuracy

In this section we summarize the optimization of the neural network presented previously. Hence, it necessary to stress the use of several statistical tools regarding the optimization of the neural networks used. In this case, the statistical tools are not for the economic analysis but for the network optimization (Slavici, 2006). From the user perspective, an ANN functions on the black box principle. In contrast, the computer scientist must create the most effective black box for users. Thus, the parameters of the network must be optimized.

The optimization function is the forecast accuracy and the input variables are the network topology, the learning method, the non-linearity function established, the number of hidden layers, the neuron number of each layer and the number of training epochs. The optimization was performed in two phases. First, a quasi-empirical optimization of the ANN was made using different values for the network's parameters. Second, the results obtained were statistically analyzed and a relationship between the input variables and the forecast accuracy was obtained. This phase was achieved according to the following methodology:

1) The most relevant variables were chosen by using the Fischer test, consisting of the number of training epochs (e) and the number of layers (s).

2) Scientific planning of the experiment was made (factorial model).

3) In the end, a second-degree model for the forecast accuracy was obtained:

Its graphic representation is as follows (Figure 3):

Following the optimization process described, the forecast accuracy increased. According to Figure 3, the forecast accuracy varied from an average value of 64 % with classical methods to over 95 % when the number of layers is close to 4 and the number of training epochs is close to 350. It can also be observed that the forecast accuracy is more affected by the number of training epochs than the number of layers, that is, for a given number of training epochs, the forecast accuracy is nearly constant when the number of layers increases. This result is useful for identifying the optimum value for the network parameters that ensures the maximum forecast accuracy.

 

CONCLUSIONS

In conclusion, out study provides useful information for the economic agents investing in agribusiness, in Eastern Europe or similar geo-climatic regions: a database containing information about previous experience in similar conditions and an optimized neural network which could accurately predict the outcome of various energetic crops. Thus, this paper expands the actual knowledge with information corresponding to a previously not investigated context: the outcome of several types of energetic plants in a specific geo-climatic area.

 

LITERATURE CITED

Bonanno, A. 2003. Some reflections on Eastern European agriculture. An introductory essay to the special issue on agriculture in Eastern Europe. Agric. Human Values 10(1):2-10.         [ Links ]

Dai, X., Z. Huo, and H. Wang. 2011. Simulation for response of crop yield to soil moisture and salinity with artificial neural network. Field Crops Res. 121: 441-449.         [ Links ]

DG Regional Policy. 2008. Guide to cost-benefit analysis of investment projects. http://ec.europa.eu/regional_policy/sources/docgener/guides/cost/guide2008_en.pdf (Accessed: November 2010).         [ Links ]

Fischer, G., S. Prieler, and H. Van Velthuizen. 2005. Biomass potentials of miscanthus, willow and poplar: results and policy implications for Eastern Europe in Northern and Central Asia Proceedings of the joint IEA Bioenergy Task 30 and Task 31 Workshop Sustainable Bioenergy Production Systems: Environmental, Operational and Social Implications 28(2): 119-132.         [ Links ]

Heinzow, T., and R. S. J Tol. 2003. Prediction of crop yields across four climate zones in Germany: an artificial neural network approach, No FNU-34, Working Papers, Research unit Sustainability and Global Change, Hamburg University, http://EconPapers.repec.org/RePEc:sgc:wpaper:34. (Accessed: May, 2012).         [ Links ]

Hernandez-Santoyo, J., and A. Sanchez-Cifuentes. 2003. Trigeneration: an alternative for energy savings. Appl. Energy 76(1-3): 373-382.         [ Links ]

Hu, J., and C. Kao. 2007. Efficient energy-saving targets for APEC economies. Energy Policy 35(1): 372-382.         [ Links ]

Joerding, W., Y. Li, and D. Young. 1994. Feedforward neural network estimation of a crop yield response function. J. Agric. Appl. Econ. 26(I): 252-263.         [ Links ]

Kumar, P. 2011. Crop yield forecasting by adaptive neuro fuzzy inference system. Mathematical Theory and Modeling 1 (3): 1-7.         [ Links ]

Murase, H. 2000. Artificial intelligence in agriculture Computers and Electronics Agric. 29(1-2): 1-2.         [ Links ]

Qaddoum, K., E. Hines, and D. Illiescu. 2011. Adaptive neuro-fuzzy modeling for crop yield prediction. Proceedings of the 10th WSEAS International Conference on Artificial Intelligence, Knowledge Engineering and Data Bases (AIKED'11): 199-204.         [ Links ]

Russell, S., and P. Norvig. 2002. Artificial Intelligence: A Modern Approach. 2^nd Edition. Prentice Hall Series in Artificial Intelligence. Fort Collins, CO. 1132 p.         [ Links ]

Savin, I., D. Stathakis, T. Negre, and V. A. Isaev. 2007. Prediction of crop yields with the use of neural networks. Russian Agric. Sci. 33(6): 361-363.         [ Links ]

Slavici, T. 2006. Artificial Intelligence. Eurostampa Publishing House. Timisoara, Romania. 208 p.         [ Links ]

Stastny, J., V. Konecny, and O. Trenz. 2011. Agricultural data prediction by means of neural network, Agric. Econ.—Czech. 57(7): 356-361.         [ Links ]