ISSN No. (Prit): 097530 ISSN No. (Olie): 49-339 Forecastig of Tea Yield Based o Eergy Iputs usig Artificial Neural Networks (A case study: Guila provice of Ira) Farshad Soheili-Fard* ad Seyed Babak Salvatia** *Departmet of Biosystem Egieerig, Faculty of Agriculture, Uiversity of Tabriz, Tabriz, IRAN. **Tea Research Istitute, Lahija, Guila Provice, IRAN. (Correspodig author: Farshad Soheili-Fard) (Received 09 April, 05, Accepted May, 05) (Published by Research Tred, Website: www.researchtred.et) ABSTRACT: The objective of this study was the explorig relatio betwee eergy iputs ad tea yield usig artificial eural etwork (ANN) i the Guila provice of Ira. For this purpose, the eergy use patter was determied by collectio data from 30 tea farmers usig face-to-face questioaire method i the may village of studied regio. The results idicated the total eergy cosumptio ad yield of tea productio were 4644.04 ad 849.47 ha, respectively. The highest share of eergy cosumptio was beloged to itroge with 50.84%. I this study, the eergy idices coverig eergy use efficiecy, eergy productivity, specific eergy ad et eergy were calculated at 0.8, 0.3 MJ, 4.38 MJ ad -3774.57, respectively. Moreover, the share of eergy forms icludig direct, idirect, reewable ad o-reewable eergies was foud to be as 4.96%, 57.04%, 8.34% ad 7.66%, respectively. For forecastig of tea yield based o eergy iputs, ANN model developed by Back propagatio algorithm i this study. The results illustrated the ANN model with 733 architecture had the best coditio for predict of tea yield. With respect to ANN model, R, RMSE ad MAPE was computed as 0.968, 0.05 ad 0.006, respectively. I the last sectio of this study, sesitivity aalysis was applied by ANN for robustess of evaluated mode. The results disclosed the farmyard maure had the highest rate of sesitivity amog all iputs. Keywords: Artificial eural etworks, Eergy, Forecast, Model; Tea productio. INTRODUCTION Biological Forum A Iteratioal Joural 7(): 43438(05) The tea plat, Camellia siesis (L.) O. Kutze, family Theaceae, is a small evergree, pereial, crosspolliated plat ad grows aturally as tall as 5 m. However, uder cultivated coditios, a bush height of 6000 cm is maitaied for harvestig the teder leaves (Soheili -Fard et al., 04). Eergy has a ifluecig role i the developmet of key sectors of ecoomic importace such as idustry, trasport ad agriculture. This has motivated may researchers to focus their research o eergy maagemet. Eergy has bee a key iput of agriculture sice the age of subsistece agriculture. It is a established fact worldwide that agricultural productio is positively correlated with eergy iput. Agriculture is both a producer ad cosumer of eergy. It uses large quatities of locally available o-commercial eergy, such as seed, maure ad aimate eergy, as well as commercial eergies, directly ad idirectly, i the form of diesel, electricity, fertilizer, plat protectio, chemical, irrigatio water, machiery etc. Efficiet use of these eergies helps to achieve icreased productio ad productivity ad cotributes to the profitability ad competitiveess of agriculture sustaiability i rural livig. Eergy use i agriculture has bee icreasig i respose to icreasig populatio, limited supply of arable lad, ad a desire for higher stadards of livig. However, more itesive eergy use has brought some importat huma health ad eviromet problems so efficiet use of iputs has become importat i terms of sustaiable agricultural productio (Taheri-Garavad et al., 00). The artificial eural etwork (ANN) modelig is of competet soft computig techiques that edeavor mimickig the huma biological ervous system by itercoectig various artificial elemets, so called as euros. ANN has bee a dyamic iterest studyig filed with ever-icreasig applicatio for modelig i diversity of sciece ad egieerig cotexts.
The mai reaso for the acceptability ad applicability of ANN is that the methodology is comparable to statistical modelig ad ANNs could be dealt with as complemetary effort or a alterative approach to fittig o-liear data. ANN, which explores the iputoutput relatioships without ay give explicit iformatio o the processes, has bee extesively applied for describig complex o-liear relatioships withi various scietific disciplies (Tagh avifar ad Mardai, 05). I recet years may researcher used ANN method for modelig of eergy cosumptio i agricultural ad horticultural crop productio. Rahma ad Bala (00) predicted jute productio i Bagladesh usig ANNs. I aother study, Safa ad Samarasighe (0) developed a eural etwork model to predict eergy cosumptio of wheat productio i New Zealad. They also compared ANNs with the multiple liear regressio model (MLR) ad foud that ANNs ca predict eergy cosumptio better tha MLR. Khoshevisa et al. (03) ivestigated o ANN model of eergy use ad greehouse gas emissios of wheat productio i Esfaha provice, Ira. Soi et al. (03) examied CO emissios ad eergy use patters i rai-fed agricultural productio systems of ortheast Thailad. Nabavi-Pelesaraei et al. (03a) examied the ANN model for predictio of eggplat yield i orth of Ira. I aother study, the yield ad greehouse gas emissios of watermelo productio calculated by Nabavi-Pelesaraei (04a). Taghavifar ad Mardai (05) applied the ANN Items A. Iputs. Huma labor. Machiery 3. Diesel fuel 4. Chemical fertilizers (a) Nitroge (b) Phosphate (P O 5 ) 5. Farmyard maure 6. Biocides Soheili-Fard ad Salvatia 433 method for modelig of apple yield based o eergy iputs. With respect to above literature, the mai purposes of this study was to evaluate of eergy use patter of tea productio i Guila provice of Ira ad applyig this patter i modelig of tea yield usig ANN approach. MATERIALS AND METHODS A. Collectio data ad case study This study follows our previous study which was coducted o modelig ad relatioship betwee CO emitter iputs ad yield of tea i Lahija city of Guila Provice, Ira (Soheili-Fard et al., 04). Accordigly, data used i this study were obtaied from 30 tea farms from 5 villages i Guila provice of Ira i 03-04 crop years. B. Eergy cosumptio The amout of iputs used i agricultural productio practices (huma labor, machiery, diesel fuel, chemical fertilizers, farmyard maure, ad biocides) ad output (tea yield) were calculated per hectare ad the, these data were coverted to forms of eergy to evaluate the output-iput aalysis. I order to estimate output ad iput eergy, these iput data ad amout of output yield were multiplied with the coefficiet of eergy equivalet. Eergy equivalets of iputs ad output were coverted ito eergy o area uit. The previous studies (cited i Table ) were used to determie the eergy equivalets' coefficiets. Table : Eergy equivalet of iputs ad output i agricultural productio. Uit h h L Eergy equivalet (MJ uit ).96 6.7 56.3 66.4.44 0.3 0 Referece (Nabavi-Pelesaraei, 03b) (Nabavi-Pelesaraei, 03b) (Nabavi-Pelesaraei, 04b) (Mousavi-Avval, 0) (Rafiee et al., 00) (Khoshevisa et al., 03) (Hamedai et al., 0) B. Output. Tea 0.8 (Kitai, 999) For istace, diesel fuel eergy was estimated from the total fuel used i differet farm operatios for potato productio. Eergy cosumed was calculated usig a coversio factor ( liter of diesel fuel = 56.3 MJ) ad expressed i. Followig the calculatio of eergy iput ad output equivalets, to assess the eergy efficiecy of tea productio the idices of eergy cosumptio icludig eergy use efficiecy, eergy productivity, specific eergy (eergy itesity) ad et eergy were calculated as follow (Rafiee et al., 00):
Output eergy (MJ ha) Eergy use efficiecy = Iput eergy (MJ ha) Tea yield ( ha ) Eergy productivity = () Iput eergy (MJ ha ) Iput eergy (MJ ha ) Specific eergy = (3) Tea yield ( ha ) Net eergy = Output eergy (MJha )- Iput eergy(mjha ) (4) For the purpose of growth ad developmet eergy demad i agriculture is divided ito direct ad idirect eergies or reewable ad o-reewable eergies. Direct eergy (DE) covers huma labor ad diesel fuel, while idirect eergy (IDE) icludes eergy embodied i machiery, chemical fertilizers, farmyard maure ad biocides used i the tea farms. Reewable eergy (RE) cosists of huma labor ad farmyard maure, whereas o-reewable eergy (NRE) icludes machiery, diesel fuel, chemical fertilizer ad biocides. C. Artificial eural etworks (ANN) Iterest i usig artificial eural etworks (ANNs) for forecastig has led to a tremedous surge i research activities i the past two decades. They ca also be cofigured i various arragemets to perform a rage of tasks icludig classificatio, patter recogitio, data miig ad process modelig. ANNs are ispired to the huma brai fuctioality ad structure, which ca be represeted as a etwork of desely itercoected elemets called euros. They cosist of a great umber of processig elemets (euros) coected to each other ad the stregths of the coectios are called weights. For the modelig of physical systems, a feed forward back propagatio (BP) multilayered perceptro (MLP) structure is commoly used. The mai advatages of MLP structures are that, they are easy to use ad they require relatively little memory ad are geerally fast; also they have the ability to lear complex relatioships betwee iput ad output patters, which would be difficult to model with covetioal algorithmic methods (Pahlava et al., 0). A ANN structure usually cosists of a layer of iput euros, a layer of output euros ad oe or more hidde layers. I a MLP, there is o coectio betwee the euros i a give layer, so that the iformatio is trasferred from the (l )th layer to the lth layer. Exteral datasets eter the etwork through the iput odes ad through o-liear trasformatios; output values are geerated by the output odes. Hidde odes with appropriate o-liear trasfer Soheili-Fard ad Salvatia 434 fuctios are used to process the iformatio received by the iput odes (Omid et al., 009). Based o the above seteces, the eergy iputs icludig huma labor, diesel fuel, machiery, itroge, phosphate, farmyard maure ad were cosidered as iputs of ANN model; while, tea yield was chose as oly output of ANN model i this study. Several structures were evaluated usig the experimetal data to determie the best predictio model for the etwork. The iput weight matrixes are made up from all the liks betwee iput layers ad hidde layers ad the output weight matrix comprises all the liks betwee the hidde layers ad the output layers. Weight (w), which cotrols the propagatio value (x) ad the ou tput value (O) from each ode, is modified usig the value from the precedig layer accordig to Eq. (5) (Zhao et al., 009): ( i i ) O = f T + w x (5) where 'T' is a specific threshold (bias) value for each ode. 'f' is a o-liear sigmoid fuctio, which icreased uiformly The performace of the etwork ca be evaluated by comparig the error obtaied from the coverged/combied eural etwork rus ad the measured data. The error fuctio ca be writte as (Nabavi-Pelesaraei et al., 04a): E = ( t pk z pk ) (6) p p k where 'p' is the idex of the p traiig pairs of vectors, 'k' the idex of elemet i the output vector, 'z pk ' the kth elemet of the output vector whe patter p is preseted as iput to the etwork ad 't pk ' is the kth elemet of the pth desired patter vector. The mea square error (MSE) method is the most commoly used idicator of predictio error over all traiig vectors. MSE is very useful to compare differet models; it shows the etworks' ability to predict the correct output. The MSE ca be expressed as (Safa ad Samarasighe, 0): MSE = ( t i z i ) (7) i where 't i ' ad 'z i ' are the actual ad the predicted output for the ith traiig vector, ad 'N' is the total umber of traiig vectors. The coefficiet of determiatio (R) ad mea absolute error (MAE) betwee the predicted ad actual values were calculated usig the followig equatios (Pahlava et al., 0b):
Soheili-Fard ad Salvatia 435 ( ti zi ) = i= R () ti i= MAE = ( t i z i ) (3) t= where 't i ' ad 'z i ' are the predicted ad actual output for the i th orchardist, respectively. Basic iformatio o eergy iputs of tea productio was etered ito Excel 03 spreadsheets ad the Matlab (R04b) software package. RESULTS AND DISCUSSION A. Aalysis of iput-output eergy use i tea productio Table shows the quatities of iputs used i whole productio life ad establishmet of tea gree leaf ad their eergy equivaleces. Also Fig. shows the distributio percet of the eergy associated with the iputs. The results revealed that aroud 668 h of huma labor ad 4 h of machiery power per hectare were required to produce tea i the research area. The total eergy iput for various processes i the tea productio was calculated to be 4644.04. The highest average eergy cosumptio of iputs was for itroge (3458.7 ) which was accouted for about 5% of the total eergy iput (Fig. ), followed by huma labor (97.06, 8%). The irregular cosumptio of itroge i alfalfa productio is for misuderstadigs of farmers i the studied area. Really, we believed the high cosumptio of chemical fertilizer ca be icrease the yield per hectare. I other had, the low price of chemical fertilizers ad lack of expert supervisio ca be effective i the high rate of chemical fertilizer cosumptio i this regio. It should be oted, the average of eergy output (from tea yield) was calculated as 849.47 MJ per hectare with stadard deviatio of 356.4. Table : Amouts of iputs ad output of tea productio with their eergy equivalet. Items (uit) Quatity per uit area (ha) Total eergy equivalet ( ) Stadard deviatio A. Iputs. Huma labor (h). Machiery (h) 3. Diesel fuel (L) 4. Chemical fertilizers () (a) Nitroge (b) Phosphate (P O 5 ) 5. Farmyard maure () 6. Biocides () 667.89 4.44.7 354.68 66.35 350.90 3.3 97.06 53.4 6853.63 3458.7 85.34 05.7 397.6 389.08 667.3 3406.5 344.70 434.3 86.3.8 The total eergy iput - 4644.04 4897.5 B. Output. Tea () 054.34 849.47 356.4 The eergy use efficiecy, eergy productivity, specific eergy ad et eergy of tea productio i the Guila provice are listed i Table 3. The eergy use efficiecy i the productio of tea was foud to be 0.8. The eergy ratio is ofte used as a idex to examie the eergy efficiecy i crop productio. The eergy ratio for some crops are reported as 0.0 for basil, 9.03 for eggplat, 4.53 for peaut ad.9 for watermelo (Pahlava et al., 0; Nabavi Pelesaraei et al., 03a- 03b-04a). The eergy productivity of tea productio was calculated as 0.3 MJ. Obviously, the specific eergy was 4.38 i tea productio. Moreover, the et eergy of tea productio was foud to be -3774.57, which idicates that i this crop productio, the eergy loss because et eergy was egative. As ca be see i Table 3, the eergy forms of tea productio was icludig direct, idirect, reewable ad o-reewable eergies are demostrated. Accordigly, the average of eergy use was computed as 984.69, 639.35, 3076.33 ad 33067.7 for direct, idirect, reewable ad o-reewable eergies, respectively.
Soheili-Fard ad Salvatia 436 Fig.. The share of each iput i total eergy cosumptio of tea productio. Table 3: Amout of eergy idices ad eergy form categories of tea productio. Items Eergy use efficiecyficiecy Eergy productivity Specific eergy Net eergy gai Direct eergy Idirect eergy Reewable eergyergy No-reewable e eergy Total eergy iput The share of eergy iput as direct, idirect, reewable ad oreewable forms is illustrated i Fig.. With respect to above-metioed, the total cosumed eergy iput could be classified as direct eergy (4.96%), ad idirect eergy (57.04%) or reewable eergy (8.34%) ad o-reewable eergy (7.66%). Uit Value - 0.8 MJ 0.3 MJ 4.38-3774.57 984.69 639.35 3076.33 33067.7 4644.04 This idicates that tea productio depeds maily o o-reewable eergy (diesel fuel ad itroge) i the studied area. Therefore, it is clear that o-reewable eergy cosumptio was higher tha that of reewable i tea productio. Fig.. The cotributio of eergy forms each iput i tea productio of Guila provice, Ira.
B. Evaluatio ad aalysis of the ANN model Several MLP etworks were desiged, traied ad geeralized, usig the MATLAB R04b software package. Back propagatio algorithm was chose to build the predictio models. Soheili-Fard ad Salvatia 437 The best model was cosisted of a iput layer with seve iput variables, two hidde layers with thirtee euros i each layer, ad a output layer with oe output variable (7 33 structure). The results of ANN model are preseted i Table 4. Table 4: The best result of differet arragemet of models. Item Tea yield R 0.968 Based o results, the best topology has the highest coefficiet of determiatio (R) (with 0.968) ad the lowest values RMSE ad MAPE. Accordigly, the rate of RMSE ad MAPE was computed as 0.05 ad 0.006, respectively. So this model was selected as the best solutio for estimatig the tea productio yield o the basis of iput eergy i surveyed regio. Safa ad Samarasighe (0) reported o a eural etwork with two hidde layers that ca predict eergy cosumptio based o farm coditios (size of crop area), social factors (farmers' educatio level), ad eergy iputs (N ad P use, ad irrigatio frequecy). Pahlava et al. (0) predicted basil productio usig a ANN model icludig a iput layer ( with seve euros), two hidde layers (with 0 euros i each layer) ad a output layer (with oe euro). I aother study, a ANN model with a 0- structure was developed to model eergy cosumptio i watermelo productio (Nabavi-Pelesaraei et al., 04a). RMSE 0.05 MAPE 0.006 Table 5. Sesitivity aalysis results for iput eergies. C. Sesitivity aalysis Sesitivity aalysis via ANN raked ad selected the major ad iput variables through its aalysis usig partial differetial (Nabavi -Pelesaraei et al., 04a). Sesitivity aalysis was performed usig the best etwork selected, i order to assess the predictive ability ad validity of the developed models (Table 5). The share of each iput item of developed MLP model o desired output (tea yield) ca be see clearly. Sesitivity aalysis was used to test the robustess of the results of a model or system i the presece of ucertaity ad icreased the uderstadig of the relatioships betwee iput ad output variables i a system or model (Nabavi-Pelesaraei et al., 04a). The results idicated that the farmyard maure had the highest value of sesitivity aalysis with 0.058 o tea yield; followed by diesel fuel (0.050) ad biocides (0.04). I aother had, the lowest rate of sesitivity was beloged to itroge with 0.09 i tea productio of Guila provice, Ira. Iputs Sesitivity value o tea yield. Huma labor 0.033. Machiery 0.0 3. Diesel fuel 0.050 4. Nitroge 0.09 5. Phosphate (P O 5 ) 0.0 6. Farmyard maure 0.058 7. Biocides 0.04 CONCLUSION Total eergy cosumptio ad yield of tea productio were 4644.04 ad 849.47 ha, respectively. Nitroge had the highest share of eergy use amog all iputs, at 50.84%; followed by huma labor with 8.%. The average rates of eergy idices, icludig eergy use efficiecy, eergy productivity, specific eergy ad et eergy were calculated at 0.8, 0.3 MJ, 4.38 MJ ad -3774.57, respectively. The results of eergy forms aalysis idicated that the share of idirect ad o-reewable eergy was more tha direct ad reewable eergies, sigificatly. The best model for predictig of tea yield was a ANN model with a 733 structure. With respect to the best topology, R, RMSE ad MAPE was foud to be as 0.968, 0.05 ad 0.006, respectively. The results of the sesitivity aalysis showed that farmyard maure with 0.058 had the highest rate of sesitivity; followed by diesel fuel with 0.050. ACKNOWLEDGMENT The first author express his deep appreciatio to Mr. Saeed Soheili-Fard for fiacial support him i this study.
REFERENCES Hamedai, SR., Shabai, Z. & Rafiee, S. (0). Eergy iputs ad crop yield relatioship i potato productio i Hamada provice of Ira. Eergy. 36: 367-37. Khoshevisa, B., Rafiee, S., Omid, M., Yousefi, M. & Movahedi, M. (03). Modelig of eergy cosumptio ad GHG (greeh ouse gas) emissios i wheat productio i Esfaha provice of Ira usig artificial eural etworks. Eergy. 5: 333-338. Kitai, O. (999). Eergy ad biomass egieerig. I: CIGR hadbook of agricultural egieerig. St. Joseph, MI: ASAE. Mousavi-Avval, SH., Rafiee, S., Jafari, A. & Mohammadi, A. (0). Optimizatio of eergy cosumptio for soybea productio usig Data Evelopmet Aalysis (DEA) approach. Applied Eergy. 88: 3765-377. Nabavi-Pelesaraei, A., Shaker-Koohi, S. & Dehpour, MB. (03a). Modelig ad optimizatio of eergy iputs ad greehouse gas emissios for eggplat productio usig artificial eural etwork ad multi-objective geetic algorithm. Iteratioal Joural of Advaced Biological ad Biomedical Research. (): 478489. Nabavi-Pelesaraei, A., Abdi, R. & Rafiee, S. (03b). Eergy use patter ad sesitivity aalysis of eergy iputs ad ecoomical models for peaut productio i Ira. Iteratioal Joural of Agriculture ad Crop Scieces. 5(9): 93-0. Nabavi-Pelesaraei, A., Abdi, R. & Rafiee, S. (04a). Neural etwork modelig of eergy use ad greehouse gas emissios of watermelo productio systems. Joural of the Saudi Society of Agricultural Scieces. http://dx.doi.org/0.06/j.jssas.04.05.00. Nabavi-Pelesaraei, A., Abdi, R., Rafiee, S. & Bagheri, I. (04b). Determiatio of efficiet ad iefficiet uits for watermelo productio-a case study: Guila provice of Ira. Joural of the Saudi Society of Agricultural Scieces. http://dx.doi.org/0.06/j.jssas.04..00. Soheili-Fard ad Salvatia 438 Omid, M., Baharlooei, A. & Ahmadi, H. (009). Modelig dryig kietics of pistachio uts with multilayer feed-forward eural etwork. Dryig Techology. 7(0): 069077. Pahlava, R., Omid, M. & Akram, A. (0). Eergy iput-output aalysis ad applicatio of artificial eural etworks for predictig greehouse basil productio. Eergy. 37(): 776. Rafiee, S., Mousavi-Avval, SH. & Mohammadi, A. (00). Modelig ad sesitivity aalysis of eergy iputs for apple productio i Ira. Eergy. 35: 330-3306. Rahma, MM. & Bala, BK. (00). Modellig of jute productio usig artificial eural etworks. Biosystems Egieerig. 05(3): 350-356. Safa, M. & Samarasighe, S. (0). Determiatio ad modellig of eergy cosumptio i wheat productio usig eural etworks: "A case study i Caterbury provice, New Zealad". Eergy. 36(8): 540-547. Soheili-Fard, F., Ghassemzadeh, HR. & Salvatia, SB. (04). A ivestigatio of relatio betwee CO emissios ad yield of tea productio i Guila provice of Ira. Iteratioal Joural of Bioscieces. 4(): 7885. Soi, P., Taewichit, C. & Salokhe, V. (03). Eergy cosumptio ad CO emissios i raifed agricultural productio systems of Northeast Thailad. Agricultural Systems. 6: 5-36. Taheri-Garavad, A., Asakareh, A. & Haghai, K. (00). Eergy elevatio ad ecoomic aalysis of caola productio i Ira a case study: Mazadara provice. Iteratioal Joural of Evirometal Scieces. (): 36-4. Taghavifar, H. & Mardai, A. (05). Progosticatio of eergy cosumptio ad greehouse gas (GHG) emissios aalysis of apple productio i West Azarbayja of Ira usig Artificial Neural Network. Joural of Cleaer Productio. 87: 5967. Zhao, Z., Chow, TL., Rees, HW., Yag, Q., Xig, Z. & Meg, FR. (009). Predict soil texture distributios usig a artificial eural etwork model. Computers ad Electroics i Agriculture. 65(): 36-48.