Aplicação de redes neurais artificiais na análise de dados de molhamento foliar por orvalho
Carregando...
Data
2006-12-11
Autores
Mathias, Ivo Mário [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O trabalho descrito nesta tese apresenta o desenvolvimento de um sistema computacional denominado PMNeural, baseado em Redes Neurais Artificiais (RNAs). A finalidade do sistema é o tratamento de dados climáticos e de molhamento foliar por orvalho, visando reconhecer padrões de comportamento de variáveis meteorológicas em relação ao molhamento foliar por orvalho. Para determinar as melhores arquiteturas e algoritmos de treinamento de RNAs, bem como, definir quais as variáveis climáticas que influenciam significativamente na ocorrência do molhamento foliar, foram utilizados dois simuladores: o simulador SNNS (Stuttgart Neural Network Simulator) versão 4.2, que utiliza plataforma operacional Linux e o simulador JavaNNS - Java Neural Network Simulator 1.1, com ambiente de execução Windows, o qual é baseado no SNNS. Foram utilizados dados climáticos de três estudos de caso, dois destes referentes à cultura do trigo, oriundos de locais e datas diferentes. Base de Dados 1 - Fazenda Capão do Cipó, em Castro - PR, safra de inverno de 2003. Base de Dados 2 - Campo Demonstrativo e Experimental da Fundação ABC - Fazenda Palmeirinha, em Piraí do Sul - PR., safra de inverno de 2005. Base de Dados 3 - Posto Agrometeorológico ESALQ/USP em Piracicaba - SP, período entre julho e setembro de 2005. Um quarto estudo de caso foi elaborado a partir dos arquivos dos estudos de casos 1, 2 e 3, utilizando-se as variáveis climáticas comuns, juntamente com seus respectivos índices de molhamento. Dentre os algoritmos de treinamento testados nos simuladores, o Resilient 2 Propagation (Rprop) foi o que apresentou as menores taxas de erro em relação aos outros...
The work described in this thesis presents the development of a computational system named PMNeural based on Artificial Neural Networks (ANNs). The system has for purpose the handle of climatic and leaf wetness data, aiming to recognize patterns of behavior of meteorological variables in relation to the wetness from dew. Two simulators were used in order to determine the best architecture and ANNs training algorithms, as well as, to define which the climatic variables that influence significantly in the leaf wetness occurrence: the SNNS (Stuttgart Neural Network Simulator) version 4.2 for Linux platform, and the JavaNNS - Java Neural Network Simulator 1.1, for Windows platform, which is based on the SNNS. Climatic data of three case studies were used, two related to wheat culture, obtained from different places and dates. Dataset 1 - Capão do Cipó Farm, in Castro - PR, 2003. Dataset 2 - Palmeirinha Farm in Piraí do Sul - PR, 2005 winter crop. Dataset 3 - Meteorological Station of ESALQ/USP in Piracicaba - SP, from July to September, 2005. A fourth case study was elaborated from datasets of the case studies 1, 2 and 3, using the common climatic variables together with their respective wetness indexes. After testing the training algorithms in the simulators, the Resilient Propagation (Rprop) presented lower training errors than the others evaluated methods: Backpropagation Standard, Backpropagation for batch training, Backpropagation with momentum term, Backpropagation with chunkwise update, Backpropagation with Weight Decay and Quickprop. It was verified 4 also that, among the climatic variables used for classification of leaf wetness from dew, the inclusion of the schedule had influenced in the obtaining better ANNs results... (Complete abstract, click electronic access below)
The work described in this thesis presents the development of a computational system named PMNeural based on Artificial Neural Networks (ANNs). The system has for purpose the handle of climatic and leaf wetness data, aiming to recognize patterns of behavior of meteorological variables in relation to the wetness from dew. Two simulators were used in order to determine the best architecture and ANNs training algorithms, as well as, to define which the climatic variables that influence significantly in the leaf wetness occurrence: the SNNS (Stuttgart Neural Network Simulator) version 4.2 for Linux platform, and the JavaNNS - Java Neural Network Simulator 1.1, for Windows platform, which is based on the SNNS. Climatic data of three case studies were used, two related to wheat culture, obtained from different places and dates. Dataset 1 - Capão do Cipó Farm, in Castro - PR, 2003. Dataset 2 - Palmeirinha Farm in Piraí do Sul - PR, 2005 winter crop. Dataset 3 - Meteorological Station of ESALQ/USP in Piracicaba - SP, from July to September, 2005. A fourth case study was elaborated from datasets of the case studies 1, 2 and 3, using the common climatic variables together with their respective wetness indexes. After testing the training algorithms in the simulators, the Resilient Propagation (Rprop) presented lower training errors than the others evaluated methods: Backpropagation Standard, Backpropagation for batch training, Backpropagation with momentum term, Backpropagation with chunkwise update, Backpropagation with Weight Decay and Quickprop. It was verified 4 also that, among the climatic variables used for classification of leaf wetness from dew, the inclusion of the schedule had influenced in the obtaining better ANNs results... (Complete abstract, click electronic access below)
Descrição
Palavras-chave
Inteligencia artificial, Agricultura, Umidade, Redes neurais, Plantas - Efeito da umidade, Artificial intelligence, Agriculture, Wetness index, Wetness duration
Como citar
MATHIAS, Ivo Mário. Aplicação de redes neurais artificiais na análise de dados de molhamento foliar por orvalho. 2006. xi, 120 f. Tese (doutorado) - Universidade Estadual Paulista, Faculdade de Ciências Agronômicas, 2006.