Elaboração de redes neuronais para descrever a epidemia da ferrugem do cafeeiro
Arquivos
Data
2001
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Resumo
O presente trabalho teve por objetivo avaliar o potencial das redes neuronais como método alternativo aos sistemas fundamentais para descrever a epidemia de ferrugem, uma das principais causas de perdas na cultura do café. As redes foram elaboradas com dados de incidência da ferrugem do cafeeiro coletados em Lavras-MG, no período de 13/02/98 a 20/04/2001, e variáveis climáticas. Para construção das redes, as variáveis climáticas foram selecionadas pela análise de regressão "stepwise" ou pelo sistema Braincel TM , o qual foi utilizado no desenvolvimento das redes. Séries temporais também foram empregadas na elaboração de redes. Foram testadas 59 redes e 26 modelos de regressão. A escolha dos melhores modelos foi baseada nos menores valores do quadrado médio do desvio (QMD) e erro médio de previsão (EMP). Para os modelos ajustados pela regressão, também foi considerado o maior valor do coeficiente de determinação (R 2 ). O melhor modelo de rede neuronal apresentou QMD = 4,36 e EMP = 2,43% e incluiu as variáveis: temperatura média, dias com precipitação, umidade relativa do ar e insolação, defasadas 30, 45, 60 e 60 dias, respectivamente. O melhor modelo ajustado pela análise de regressão foi desenvolvido com 29 variáveis climáticas selecionadas na elaboração de rede e apresentou
EMP = 6,58%, QMD = 4,36 e R 2 = 0,80. As redes neuronais elaboradas a partir de séries temporais também foram adequadas para descrever a epidemia. As observações da incidência das cinco semanas anteriores resultaram em um modelo com EMP = 7,01% e QMD = 3,95.
The present work had for objective to evaluate the potential of the neural networks (NN) as alternative method to the basic systems to describe the rust epidemic, one of the main causes of loss in the coffee crop. The NN had been elaborated from data of coffee rust collected in Lavras - MG, between 13/02/98 a 20/04/2001 and climatic variable. The NN had been built with climatic variable selected by the "stepwise" regression analysis or by the Braincel Tm system, which was used in the development of the networks. Time series had been also used in the elaboration of networks. 59 networks and 26 models of regression had been tested. The choice of the best models was based on the lesser values of the mean square deviation (MSD) and mean prediction error (MPE). For the regression models also the biggest value of the determination coefficient was considered (R 2 ). The best model of neural network presented MSD = 4,36 and MPE =2.43% and included the variable: mean temperature, days with precipitation, relative humidity of air and insolation, on the 30, 45, 60 and 60 days before the occurrence, respectively. The best regression model was developed with 29 selected climatic variable in the network elaboration and presented MPE =6.58%, MSE=4,36 and R 2 =0,80. The elaborated neural networks from time series had been also adjusted to describe the epidemic. The dates of the incidence of the four previous fortnight resulted in a model with MPE =7,01% and MSD =3,95.
The present work had for objective to evaluate the potential of the neural networks (NN) as alternative method to the basic systems to describe the rust epidemic, one of the main causes of loss in the coffee crop. The NN had been elaborated from data of coffee rust collected in Lavras - MG, between 13/02/98 a 20/04/2001 and climatic variable. The NN had been built with climatic variable selected by the "stepwise" regression analysis or by the Braincel Tm system, which was used in the development of the networks. Time series had been also used in the elaboration of networks. 59 networks and 26 models of regression had been tested. The choice of the best models was based on the lesser values of the mean square deviation (MSD) and mean prediction error (MPE). For the regression models also the biggest value of the determination coefficient was considered (R 2 ). The best model of neural network presented MSD = 4,36 and MPE =2.43% and included the variable: mean temperature, days with precipitation, relative humidity of air and insolation, on the 30, 45, 60 and 60 days before the occurrence, respectively. The best regression model was developed with 29 selected climatic variable in the network elaboration and presented MPE =6.58%, MSE=4,36 and R 2 =0,80. The elaborated neural networks from time series had been also adjusted to describe the epidemic. The dates of the incidence of the four previous fortnight resulted in a model with MPE =7,01% and MSD =3,95.
Descrição
Trabalho apresentado no Simpósio de Pesquisa dos Cafés do Brasil (2. : 2001 : Vitória, ES). Resumos. Brasília, D.F. : Embrapa Café, 2001. 181p. : il.
Palavras-chave
Ferrugem do cafeeiro Redes neuronais, Artificial intelligence Epidemiology Coffea arabica Neural network.
Citação
Pinto, A. C. S.; Pozza, E. A. Elaboração de redes neuronais para descrever a epidemia da ferrugem do cafeeiro. In: Simpósio Brasileiro de Pesquisa dos Cafés do Brasil (2. : 2001 : Vitória, ES). Anais. Brasília, D.F. : Embrapa Café, 2001. (CD-ROM), p. 1141-1150