Revista Ingenierías Universidad de Medellín ISSN: Universidad de Medellín Colombia

Transcription

1 Revsta Ingenerías Unversdad de Medellín ISSN: Unversdad de Medellín Colomba Corrales, Davd Camlo; Peña Q., Andrés J.; León, Carlos; Fgueroa, Apolnar; Corrales, Juan Carlos Early warnng system for coffee rust dsease based on error correctng output codes: a proposal Revsta Ingenerías Unversdad de Medellín, vol. 13, núm. 25, julo-dcembre, 2014, pp Unversdad de Medellín Medellín, Colomba Avalable n: How to cte Complete ssue More nformaton about ths artcle Journal's homepage n redalyc.org Scentfc Informaton System Network of Scentfc Journals from Latn Amerca, the Carbbean, Span and Portugal Non-proft academc project, developed under the open access ntatve

2 Revsta Ingenerías Unversdad de Medellín Early warnng system for coffee rust dsease based on error correctng output codes: a proposal Davd Camlo Corrales * Andrés J. Peña Q. ** Carlos León *** Apolnar Fgueroa **** Juan Carlos Corrales ***** Recbdo: 25/10/2013 Aceptado: 27/06/2014 Abstract Colomban coffee producers have had to face the severe consequences of the coffee rust dsease snce t was frst reported n the country n Recently, machne learnng researchers have tred to predct nfecton through classfers such as decson trees, regresson Support Vector Machnes (SVM), non-determnstc classfers and Bayesan Networks, but t has been theoretcally and emprcally demonstrated that combnng multple classfers can substantally mprove the classfcaton performance of the consttuent members. An Early Warnng System (EWS) for coffee rust dsease was therefore proposed based on Error Correctng Output Codes (ECOC) and SVM to compute the bnary functons of Plant Densty, Shadow Level, Sol Acdty, Last Nghttme Ranfall Intensty and Last Days Relatve Humdty. Key words: Coffee Rust Dsease, Early Warnng System, ECOC, SVM, Codeword. * M.Sc. n Telematcs Engneerng, and Researcher of Telematcs Engneerng Group and Envronmental Study Group at Unversty of Cauca, Colomba. Address. Carrera 2 # 1A-25. Popayán, Colomba. Tel. +57 (8) Ext E-mal. dcorrales@uncauca.edu.co ** M.Sc. n Meteorology, and Researcher at Natonal Coffee Research Center (Cencafé), Colomba. Address: Chnchná Km 4 va. Manzales, Colomba. Tel. +57 (6) E-mal. andres.pena@cafedecolomba.com *** M.Sc. n Electrcal Engneerng, and CEO of CreaTIC Corporaton - Parquesoft, Colomba. Address: Calle 17 N # 6 53, Popayán, Colomba. Tel E-mal. cleon65@corporaconcreatc.com **** Doctor of Bologcal Scences, and Full Professor and Leader of the Envronmental Study Group at Unversty of Cauca, Colomba. Address: Carrera 2 # 1A-25. Popayán, Colomba. Tel. +57 (8) Ext E-mal. apolnar@uncauca.edu.co ***** Doctor of Phlosophy n Scences, Specalty Computer Scence, and Full Professor and Leader of the Telematcs Engneerng Group at Unversty of Cauca, Colomba. Address: Calle 5 No Popayán, Colomba. Tel. +57 (8) Ext E-mal. jcorral@uncauca.edu.co Revsta Ingenerías Unversdad de Medellín, vol. 13, No. 25 ISSN julo-dcembre de 2014/248 p. Medellín, Colomba

3 58 Davd Camlo Corrales Andrés J. Peña Q. Carlos León Apolnar Fgueroa Juan Carlos Corrales Sstema de alerta temprana para la roya en el café basado en códgos de salda de correccón de error: una propuesta Resumen Los productores de café colombanos han sufrdo severas consecuencas por la Roya desde que fue reportada por prmera vez en el país en el año Recentemente, nvestgadores de aprendzaje automátco han ntentado predecr la roya a través de clasfcadores como: arboles de decsón, máqunas de vector de soporte, clasfcadores no determnístcos y redes bayesanas, pero se ha demostrado teórca y empírcamente que la combnacón de múltples clasfcadores puede mejorar sustancalmente el rendmento en la clasfcacón. En este sentdo es propuesto un sstema de alerta temprana para la roya en el café, basado en códgos de salda de correccón de error y máqunas de vector de soporte para calcular las funcones bnaras de la densdad de planta, el nvel de sombra, la acdez del suelo, la ntensdad de lluva en la últma noche, y en últmos días, con humedad relatva. Palabras clave: roya, sstema de alerta temprana, ECOC, SVM, Codeword. Revsta Ingenerías Unversdad de Medellín

4 Early warnng system for coffee rust dsease based on error correctng output codes: a proposal 59 INTRODUCTION Coffee producton s the man agrcultural actvty n Colomba. More than famles depend on coffee harvest for ther sole ncome. As such, dseases, pests and even low prces mpact terrbly on the economc and socal aspects of the man coffee-growng regons. Coffee rust, frst reported n 1983 [1], s the most mportant and severe dsease currently affectng the producton of Colomban coffee. Resstant varetes have been developed through mprovement wth genes of Tmor Hybrd (a plant that features natural resstance to the dsease) as a soluton to the rust problem [2], yet more than 50 percent of the country s coffee crop s stll susceptble n the productve phase. Studes on coffee rust have concluded that the spores carryng the nfecton are spread by clmatc elements such as wnd and ranfall [3], wnd beng the vector for long dstance spore transport, whle precptaton droplets are responsble for vertcal propagaton from nfected leaves or sol [3]. Once spores make contact wth a susceptble leaf, the nfecton process s mproved by hgh shadow ndex, hgh humdty (atmosphere and leaf), sol acdty, hgh coffee tree densty and low sol fertlty. The warnng system proposed heren seeks to detect each of the favorable condtons that coffee rust requres to nfect the crop and, by takng prophylactc measures (bologcal, chemcal and cultural control), thus allow preventon of the onset of the dsease. Havng dentfed the partcular condtons favorable to coffee rust, the system s desgned to alert growers to the settled presence of condtons for nfecton, sendng out a graded warnng (None, Very Low, Low, Medum, Hgh, Very Hgh) to all coffee growers n the detecton area. Machne learnng researchers have n recent tmes attempted to predct nfectons through such classfers as: decson trees [4], regresson Support Vector Machnes [5], non-determnstc classfers [6] and Bayesan Networks [7]. However, t has been demonstrated theoretcally and emprcally that usng combnatons of multple classfers can substantally mprove upon the performance of consttuent members [8]. It was therefore proposed usng an ECOC to alert about coffee rust dsease. Ths method s consdered excellent for applyng bnary learnng algorthms to multclass learnng problems [9]. The rest of the paper s organzed as follows: the SVM classfer and ECOC method are ntroduced n Secton and Secton respectvely. The fnal proposal of ECOC n alertng to the rsk of coffee rust dsease s presented n Secton 61, whle Secton 63 reports a number of conclusons. 1. SUPPORT VECTOR MACHINES FOR BINARY CLASSIFICATION Support vector machnes (SVM) consttute an effectve bnary data classfcaton method. The key dea of SVMs s the use of a mappng functon that projects the gven nput feature space nto a hgh dmensonal feature space to fnd an optmal hyperplane, Revsta Ingenerías Unversdad de Medellín, vol. 13, No. 25 ISSN julo-dcembre de 2014/248 p. Medellín, Colomba

5 60 Davd Camlo Corrales Andrés J. Peña Q. Carlos León Apolnar Fgueroa Juan Carlos Corrales havng the largest margn of separaton between dfferent classes wth mnmum error rate. SVMs use a porton of the data to tran and fnd several support vectors that represent the tranng data. These support vectors are then formed nto a model by the SVM, representng each category. For a lnearly separable bnary classfcaton wth N an n -dmensonal vector x and the label of the class that vector y,.e., {( x, y) } = 1 and y = { + 1, 1 }, the SVM separates the two classes of ponts usng the classfcaton decson functon fwb, = sgn( w x + b), where w s an nput vector, x s an adaptve weght vector, and b s a bas. SVM fnds the parameters w and b for the optmal hyperplane to maxmze the geometrc margn [10], 2 T ww, y w x + b w, subject to mn ( ) (1) 2. ERROR CORRECTING OUTPUT CODES (ECOC) ECOC s a powerful approach to dealng wth mult-class problems based on the combnaton of bnary classfers. The ECOC approach works n two steps [10]: 1. The codng step. In ths step, a set of B dfferent bparttons of the class label set { C,, 1 Cl } are constructed, and subsequently B bnary classfers h,, 1 hb are traned over the parttons. 2. The decodng step. In ths step, gven an nstance x, a codeword s generated by usng the outputs of B bnary classfers. The codeword s then compared to the base codeword of each class, and the nstance s assgned to the class wth the most smlar codeword. The central task of decodng s to fnd the base codeword w (correspondng to class c ) whch s the closest to the codeword v of the gven test nstance. The bnary decodng scheme most used s presented next: Hammng decoder (HD): Ths scheme s based on the assumpton that the learnng task can be modeled as an error-correctng communcaton problem [11]. The measure s gven by: ( ( )) ( ) j j 1 sgn v w j HD v, w = ( 2) 2 Typcally, the parttons of the class set are specfed by a codng matrx M, whch can appear n two forms, bnary form and ternary form. In ths paper use of the bnary lxb form s proposed, where M { 1, + 1 }. Fg. 1 provdes an example of a bnary codng matrx, whch transforms a four-class problem nto fve bnary classfcaton problems. In the fgure, the regons coded by +1 are consdered as a class, whle regons coded by -1 are consdered as another class [10]. Consequently, the bnary classfers Revsta Ingenerías Unversdad de Medellín

6 Early warnng system for coffee rust dsease based on error correctng output codes: a proposal 61 selected for ths paper are Support Vector Machne, consdered the most accurate for classfcaton problems [12]. Ths classfer s presented n secton 2. Fg. 1. Bnary ECOC for a 4-class problem. An nstance x s classfed to class c 3 usng the Hammng (HD) (Source:[10]) 3. EARLY WARNING SYSTEM FOR COFFEE RUST DISEASE (EWSCRD) The EWSCRD approach detects all favorable condtons that coffee rust requres to nfect a crop. These favorable condtons are dentfed through nput data from coffee growers or from agro-meteorologcal staton data. The archtecture proposed for EWSCRD conssts of fve components: Crop envronment, Storage data, Data set framework, ECOC framework, and Vew layer. These are shown n Fg. 2 (a), and are explaned n the followng: Vew layer E-mal Moble Access SMS Voce ECOC Framework Decoder Support Vector Machne Bnary Framework Inputs Data set framework Storage Data Crop Envronment Data tranng processng Crop DB Crop Manager Attrbutes Manager Weather DB Weather Informaton (a) EWSCRD archtecture (b) Bnary ECOC for a 6-class Fg. 2. (a) Archtecture; (b) Bnary ECOC for a 6-class, ncludes 64 examples. Revsta Ingenerías Unversdad de Medellín, vol. 13, No. 25 ISSN julo-dcembre de 2014/248 p. Medellín, Colomba

7 62 Davd Camlo Corrales Andrés J. Peña Q. Carlos León Apolnar Fgueroa Juan Carlos Corrales Crop envroment (CE). Conssts of two sub-components: the weather nformaton sub-component, whch accesses weather data (temperature, ranfall, etc.) from the agro-meteorologcal staton network and farmer observaton, and the crop manager subcomponent, whch comples detaled crop nformaton (sowng spacng between plants, plant shadows, sol ph) through the data nserted by farmer. Storage data (SD). Once nformaton s captured n the weather and crop subcomponets of CE, the data are stored n the weather database (Weather DB) and crop database (Crop DB) respectvely. Data set framework (DSF). Ths component creates the tranng data from two subcomponents: the attrbutes manager, where an expert user defnes the attrbutes and classes for buldng the data tranng, and the data tranng processng subcomponent whch transforms the data from the Weather DB, Crop DB and the attrbutes manager subcomponent to the data tranng set. ECOC framework (ECOCF): The ECOCF receves the data tranng bult n DFS n order to tran the SVM bnary framework (SVMBF) subcomponent. In ths sense, each functon fpd, fsl, fsa, fpmr, fpnri, frhpd s consdered a bnary classfer (Fg. 2 (b)) bult through SVMBF, whch solves the followng problems: Plant densty ( f PD ) : related to the number of coffee trees per unt area. Less than 6,500 trees per hectare s consdered low densty (0). More than 6,500 trees s consdered hgh (1). Shadow level ( f SL ): related to the presence of trees (larger than the coffee crop) n the unt area. If the shade provded by these s greater than 25% t s consdered a hgh level of shade (1). Below 25% s consdered a low level of shade (0). Sol acdty ( f SA ): related to low sol ph. Sol ph values below 5.5 are consdered acdc (1). Values greater than 5.5 are consdered non-acdc (0). Prevous month ranfall ( f PMR ): related to leaf mosture n the prevous month. When t has raned more than 120 mllmeters n the prevous month, the ranfall s consdered hgh (1); less than ths threshold s consdered low (0). Prevous nghttme ranfall ntensty ( f PNRI ): related to the extent that the leaf and the canopy of the plantaton was wet durng the prevous week. If there were more than three days n the week wth ran at nght, t s consdered hgh (1), whle f the number of days wth ran at nght was less than three, t s consdered low (0). Revsta Ingenerías Unversdad de Medellín

8 Early warnng system for coffee rust dsease based on error correctng output codes: a proposal 63 Relatve humdty n the prevous days ( f RHPD ): related to the atmosphere and ts ablty to dry the leaves. Ten days wth relatve humdty over 85% promotes the dsease (1); less than 85% does not promote the dsease (0). The combnaton the outputs for each functon (0 or 1) generates a codeword x (Fg. 2 (b)) n the decoder subcomponent. The central task of ths subcomponent s to fnd the codeword correspondng to one of the classes (None, Very Low, Low, Medum, Hgh, Very Hgh) that s the closest to the codeword x. Thus, the process for choosng the most smlar codeword s based on computng the Hammng dstance. Vew Layer: shows the results obtaned n ECOCF through varous devces such as moble phone (SMS), e-mal, voce, etc. 4. CONCLUSIONS Whle coffee rust s the most mportant coffee crop dsease n Colomba, local growers have no access to early warnng systems (EWS) for management of the dsease. Studes have been done relatng to the desgn of just such systems, but n most cases proposals have been based on lnear approxmatons (unvarate) that connect some envronmental factor (e.g. temperature, relatve humdty) to the dsease [13]. As a consequence, the models are not very accurate and when the farmer comes to use the system t s possble that he does not have the nput data relatng to the relevant envronmental factor. Ideally, the farmer should need only to have access to a system of alert. Wth ths n mnd, ECOC s consdered excellent for applyng bnary learnng algorthms to multclass learnng problems [9] and SVM s the most accurate algorthm for classfcaton problems [12]. As such, ECOC + SVM may be a powerful tool to solve the partcular problem of coffee rust nfecton. The proposed EWSCRD gves the farmer ths alert opton. More than ths, EWSCRD s easy to use n any coffee growng scenaro, whether meteorologcal data s avalable or not ths latter beng the realty n Colomba for the most part. In other words, wth EWSCRD the absence of models relatng clmate, management, and sol wth producton, pests and dseases e.g. for coffee rust n Colomba, can be solved n that the alert s related to the presence (or absence) of dentfed events (clmate, sol or management). The mportance s seen n avodance of coffee rust dsease and a consequent vtal enhancement n coffee growng ncome. 5. ACKNOWLEDGEMENTS Ths paper was presented n the 1st CSIRO Workshop on Semantc Machne Learnng and Lnked Open Data (SML2OD2013) for Agrcultural and Envronmental Informatcs was held n conjuncton wth the 12th Internatonal Semantc Web Conference Revsta Ingenerías Unversdad de Medellín, vol. 13, No. 25 ISSN julo-dcembre de 2014/248 p. Medellín, Colomba

9 64 Davd Camlo Corrales Andrés J. Peña Q. Carlos León Apolnar Fgueroa Juan Carlos Corrales (ISWC 2013), on 22nd October 2013, n Sydney, Australa. The authors are grateful to the Envronmental Study Group (GEA), the Telematcs Engneerng Group (GIT) of the Unversty of Cauca, and AgroCloud project of the RICCLISA Program for techncal and scentfc support. REFERENCES [1] E. Sheber and G. A. Zentmyer, Coffee rust n the western hemsphere n Plant dsease, ed, [2] J. C. Zapata, G. M. Ruíz, F. N. d. C. d. Colomba, and C. N. d. I. d. Café, La varedad Colomba: seleccón de un cultvar compuesto resstente a la roya del cafeto : Premo Naconal de Cencas Fundacón Alejandro Angel Escobar, 1986: Cencafé, [3] S. Becker, n La propagacón de la roya del cafeto, ed: Socedad alemana de cooperacón técnca ltda. (GTZ), 1979, p. 70. [4] M. E. Cntra, C. A. A. Mera, M. C. Monard, H. A. Camargo, and L. H. A. Rodrgues, The use of fuzzy decson trees for coffee rust warnng n Brazlan crops, n Intellgent Systems Desgn and Applcatons (ISDA), th Internatonal Conference on, 2011, pp [5] O. Luaces, L. H. A. Rodrgues, C. A. A. Mera, Jos, #233, R. Quevedo, et al., Vablty of an alarm predctor for coffee rust dsease usng nterval regresson, presented at the Proceedngs of the 23rd nternatonal conference on Industral engneerng and other applcatons of appled ntellgent systems - Volume Part II, Cordoba, Span, [6] O. Luaces, L. H. A. Rodrgues, C. A. Alves Mera, and A. Bahamonde, Usng nondetermnstc learners to alert on coffee rust dsease, Expert Systems wth Applcatons, vol. 38, pp , [7] C. B. Pérez-Arza, A. E. Ncholson, and M. J. Flores, Predcton of Coffee Rust Dsease Usng Bayesan Networks, presented at the The Sxth European Workshop on Probablstc Graphcal Models, Granada (Span), [8] L. L, B. Zou, Q. Hu, X. Wu, and D. Yu, Dynamc classfer ensemble usng classfcaton confdence, Neurocomputng, vol. 99, pp , [9] T. G. Detterch and G. Bakr, Error-correctng output codes: a general method for mprovng multclass nductve learnng programs, presented at the Proceedngs of the nnth Natonal conference on Artfcal ntellgence - Volume 2, Anahem, Calforna, [10] Z. H. Zhou, Ensemble Methods: Foundatons and Algorthms: Chapman and Hall/CRC;, [11] N. J. Nlsson, Learnng machnes: foundatons of tranable pattern-classfyng systems. New York: McGraw-Hll, [12] S. B. Kotsants, Supervsed Machne Learnng: A Revew of Classfcaton Technques, presented at the Proceedngs of the 2007 conference on Emergng Artfcal Intellgence Applcatons n Computer Engneerng: Real Word AI Systems wth Applcatons n ehealth, HCI, Informaton Retreval and Pervasve Technologes, [13] G. C. Gómez, Dsemnacón de Hemlea vastatrx Berk y Br., n Taller sobre Roya del cafeto, ed Manzales: Federacón Naconal de Cafeteros - Centro Naconal de Investgacones de Café, 1982, pp Revsta Ingenerías Unversdad de Medellín