Subscribe Now Subscribe Today
Abstract
Fulltext PDF
References
Research Article
 
Neural Network for Farm Household Output Prediction



Muhd. Khairulzaman Abdul Kadir, Arjunan Subramanian, Nadaraj Miniappan and Mohd. Zaki Ayob
 
ABSTRACT

In ensuring that food security is at an acceptable level, all of its range of indicators needs to be monitored and maintained. In this study, farm household behaviour such as the mode of labour, type of fertilizer being used; cost of each component for the farm and so forth can be used to predict the farm household output of crops. The crop output is one indicator of food security for household level. The dataset used is based on the Village Level study (VLS) by the International Crops Research Institute for Semi Arid Tropics (ICRISAT). This dataset consists of 37 features and 29 samples of households in 1975. For the prediction model, the Optimum Weight and Threshold Neural Network (OWTNN) is proposed on 37 inputs and one output of crops for each household and compared with the Artificial Neural Network (ANN) performance. The result of the proposed model shows that OWTNN gives a better performance than ANN, which is 99% compared to 86%. Hence, it proves that the proposed model can be offered as one of the best predictors for farm household output. The model also shows that farm household behaviour can affect farm crop output.

Services
Related Articles in ASCI
Similar Articles in this Journal
Search in Google Scholar
View Citation
Report Citation

 
  How to cite this article:

Muhd. Khairulzaman Abdul Kadir, Arjunan Subramanian, Nadaraj Miniappan and Mohd. Zaki Ayob, 2016. Neural Network for Farm Household Output Prediction. Journal of Applied Sciences, 16: 594-599.

DOI: 10.3923/jas.2016.594.599

URL: https://scialert.net/abstract/?doi=jas.2016.594.599
 
Received: April 11, 2016; Accepted: June 29, 2016; Published: November 15, 2016

INTRODUCTION

In 2010 UK Department of Food and Rural Affairs (DEFRA) report1, one of the key features of food security was based on household food security where everyone in all places should be able to access and afford a healthy food1. The headline indicator for the food security theme is the proportion of income that a household spends on food. India, in the case of villages such as Aurepalle, the village considered here, households both plant their own crops and work in other farms to provide food. In some instances, the farms are owned and operated for consumed food and profitable sales where some farm owner sold the crops for its profits. Thus, the food supply in this case was based mostly on crops from the farm or other farm household outputs within the same village2.

The International Crops Research Institute for the Semi-Arid Tropics (ICRISAT) is one organization which studies the socioeconomics, agro-biological, institutional constraints to agricultural development in Semi-Arid Tropical (SAT) areas and they also help in testing and modifying the new technologies generated by themselves for the benefits of people in Indian villages2. For our case study, one dataset for the year 1975 from Aurepalle village in India was taken based on the ICRISAT study data to predict the farm household output of crops. The ‘farm household’ term being used rather than ‘farm’ because of the categories on the dataset. This dataset was based on the research of small farmer, medium farmer and large farmer2. Previous studies have shown that many factors can reflect farm household crop outputs such as the types of labour being used, the type of fertilizer being used, the cost of each component for the farm and so on3-8.

Here, based on the village dataset, the farm household crops output will be referred to as grains, vegetables or fruits, these are also for fodder and seedlings use as well as human consumption2.

In summary, this research studies farm household crop output predictions, making use of the ICRISAT data for Aurepalle village in the year of 1975. The model utilised is an Optimized Weight and Threshold Neural Network (OWTNN) the performance of which will be compared with the traditional Artificial Neural Network (ANN).

LIMITATIONS IN THE STATE OF THE ART

The ANN is one of the best and well known tools for prediction because of its fast and effectiveness in solving non-linear problems9,10. Generally, ANNs have 3 layers, the input layer (I), the hidden layer (can be more than one layer, J and K) and the output layer (L) as shown in Fig. 1. The interconnection between each layer consists of weights (WIJ, WJK and WKL) and at each neuron there are also thresholds connection (BI, BK and BL). An ANN can be used with either multiple inputs to single output or multiple inputs to multiple outputs11.

Although, an ANN offers a good tool in prediction it does travel to local extrema and the convergence is slow9. Moreover, when the ANN is not able to generalize, over-fitting and under-fitting can occur especially when the dataset is divided into training, validation and testing parts12,13. Most of these problems happen because the final interconnecting weight and threshold of the network are constantly and rapidly changing in an uncontrolled manner during the training phase14.

Fig. 1:Basic ANN architecture

In solving these problems, a Genetic Algorithm (GA) is used to optimize the weight and the threshold of ANN and the process will be explained in later sections.

MATERIALS AND METHODS

The GA is a well-known technique for its capability in optimizing certain datasets or any network architecture. Furthermore, it offers an effective search technique based on the principle of genetics and natural selection11,15. The advantages of adding a GA to an ANN are that it can avoid local minima, which it finds rapidly and discards. It also can search every region simultaneously with great efficiency through the GA operations and parameters of selection, crossover probability and mutation probability9,10,15,16.

In this study, ANN had multiple inputs and one output with one hidden layer with 30 neurons. The number of hidden neurons was selected based on multiple ANN training and was taken from the best regression value. The data was divided to 70% for training, 15% for validation and 15% for testing. Then, the GA will optimize the weights and the thresholds of the ANN in Fig. 1 by its aforementioned bio-inspired evolutionary operators. The process of the OWTNN is shown in the flow diagram of Fig. 2.

This model was simulated in the MATLABTM 2010 environment where each parameter was selected as Fig. 2. The inputs and the output of the ANN were as in Table 1, consisting of 37 features with 29 samples. As explain previously, the dataset was based on Aurepalle village in India for the year 1975.

RESULTS AND DISCUSSION

From Fig. 2, GA will randomly initialize all the weight and threshold values for the ANN based on the size of chromosomes for each population.

Fig. 2:OWTNN process using GA

Chromosomes consist of a number of genes which is an array of variable values to be optimized11,15. The chromosomes sizes will depend on the total interconnections between the number of inputs, number of hidden neurons and number of output neurons in ANN. After that, the ANN will be trained iteratively by the GA based on the population size for each chromosome until it achieves the best Mean Square Error (MSE) as Eq. 1 for the fitness function.

In Eq. 1, NNout is the ANN output and T is the actual output based on the dataset. The result of the optimization process is shown in Fig. 3 where the plot are based on MSE (fitness value) and the average distance between individual chromosomes at each generation:

(1)

where, Ntotal is the total number of outputs.

When the GA optimization is performed, the best weight and threshold values generate by the GA is applied to the ANN and it will calculate the farm household output prediction as in Fig. 4. This shows that OWTNN gives better performance than ANN (Fig. 5), R = 0.99 compared to 0.87. The performance comparison of ANN is not good because of the over-fitting and under-fitting problems for validation and testing, as shown in Fig. 4, which have been pointed out.

Table 1:Inputs and output variable

Fig. 3(a-b): GA optimization result, (a) Best: 0.029793, Mean: 0.27315 fitness values and (b) Average distance between individuals

Fig. 4(a-d): ANN results, (a) Training, (b) Validation, (c) Test and (d) All

Fig. 5(a-d): OWTNN results, (a) Training, (b) Validation, (c) Test and (d) All

If compare with Fig. 5, all data division part (training, validation and testing) show considerable generalization ability.

CONCLUSION

In this study, OWTNN has been successfully applied to predict the farm household crop output for each household in Aurepalle, India with a 14% improvement in the regression value compares to original ANN methods (to 0.99). It also shows that each activity or component being used and the value of each component on the farm can affect the crops output of a household.

In addition, the proposed model shows how to achieve good generalizations of the ANN network by optimizing the weight and the threshold for each neuron. The model offers the prospect of an excellent prediction tool for farm household crop output. This method offers the option of timely analysis of household food security in order to inform future modifications of farming activity patterns to achieve better crops outputs.

ACKNOWLEDGMENT

We would like to thank the School of Engineering, University of Warwick for their support. Thank you also to Arjunan Subramanian for gaining permission to use the data and to the ICRISAT in making the data available.

REFERENCES
Betechuoh, B.L., T. Marwala and T. Tettey, 2007. Using inverse neural networks for HIV adaptive control. Int. J. Comput. Intell. Res., 3: 11-15.
Direct Link  |  

Bhende, M.J. and J.V. Venkataram, 1994. Impact of diversification on household income and risk: A whole-farm modelling approach. Agric. Syst., 44: 301-312.
CrossRef  |  Direct Link  |  

DEFRA., 2010. UK food security assessment: Detailed analysis. Department for Environment Food and Rural Affairs (DEFRA), January 2010.

Demuth, H.B. and M.H. Beale, 2004. Neural Network Toolbox for Use with MATLAB: User Guide. The Math Works Inc., USA., Pages: 474.

Goldberg, D.E., 1997. Genetic Algorithm in Search, Optimization and Machine Learning. Addison-Wesley, USA.

Haupt, R.L. and S.E. Haupt, 2004. Practical Genetic Algorithms. 2nd Edn., John Wiley and Sons, New York, USA., ISBN-13: 9780471455653, Pages: 253.

Jha, D. and R. Sarin, 1981. An analysis of levels, patterns and determinants of fertilizer use on farms in selected regions of semi-arid tropical India. Progress Report No. 25, International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), Andhra Pradesh, India, October 1981.

Karimi, H., F. Yousefi and M.R. Rahimi, 2011. Correlation of viscosity in nanofluids using Genetic Algorithm-Neural Network (GA-NN). Heat Mass Transfer, 47: 1417-1425.
CrossRef  |  Direct Link  |  

Negnevitsky, M., 2005. Artificial Intelligence: A Guide to Intelligent Systems. 2nd Edn., Addison-Wesley, Boston, MA USA., ISBN: 9780321204660, Pages: 415.

Rajendra, M., P.C. Jena and H. Raheman, 2009. Prediction of optimized pretreatment process parameters for biodiesel production using ANN and GA. Fuel, 88: 868-875.
CrossRef  |  Direct Link  |  

Singh, R.P., B.L. Jain and V.B. Rao, 1984. Feature of traditional farming systems in selected villages of Madhya Pradesh. Working Paper, Economics Progress Report No. 61, International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), India.

Singh, R.P., H.P. Binswanger and N.S. Jodha, 1985. Manual of instructions for economic investigators in ICRISAT's village level studies (revised). International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), Andhra Pradesh, India, May 1985.

Skoufias, E., 1993. Seasonal labor utilization in agriculture: Theory and evidence from agrarian households in India. Am. J. Agric. Econ., 75: 20-32.
CrossRef  |  Direct Link  |  

Skoufias, E., 1994. Risk and seasonality in empirical model of the farm household. J. Econ. Dev., 19: 93-116.

Walker, T.S. and K.V. Subba Rao, 1982. Yield and net return distributions in common village cropping systems in the semi-arid tropics of India. International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), Andhra Pradesh, India, August 1982.

Zhang, Q. and C. Wang, 2008. Using genetic algorithm to optimize artificial neural network: A case study on earthquake prediction. Proceedings of the 2nd International Conference on Genetic and Evolutionary Computing, September 25-28, 2008, Hubei, pp: 128-131.

©  2018 Science Alert. All Rights Reserved
Fulltext PDF References Abstract