Modeling photosynthesis of the tropical tree Hymenaea courbaril L. using artificial neural networks
Madeleine
Barriga
Puente de la Vega, Escola Politécnica da USP - Laboratório de Automação Agrícola, madelein@usp.br
(Presenting)
Antonio Mauro
Saraiva, Escola Politécnica da USP - Laboratório de Automação Agrícola, amsaraiv@usp.br
Hernán
Prieto
Schmidt, Escola Politécnica da USP - Departamento de Energia e Automação Eléctricas, hernan@pea.usp.br
Marcos
Silveira
Buckeridge, Instituto de Botânica de São Paulo - Seção de Fisiologia e Bioquímica de Plantas, msbuck@usp.br
The study of current climatic changes caused by the emission of greenhouse gases such as carbon dioxide is extremely complex and involve the integration of several scientific fields. An important aspect of the problem is the evaluation of the carbon exchange between plants and the atmosphere. Due to the non-linear characteristics of the processes involved in this phenomenon, the development of tools to forecast carbon assimilation has been quite difficult. However, the development of modern artificial intelligence techniques such as artificial neural networks (ANNs) has greatly improved the forecast potential of complex non-linear processes. In the present work, a large set of data concerning CO2 assimilation by young plants of Hymenaea courbaril has been used to train ANNs. The multilayer perceptron with backpropagation learning algorithm was used. Leaves at three levels of six-months-old plants (at the beginning of the experiment) were monitored for CO2 assimilation during 1 year, twice a week between 6am and 6pm. Sixty five percent of all data collected was used for ANN training and the remaining 35% was used for model testing and validation. Training was performed with different combinations of variables in order to determine the more appropriate sets that give best results (lowest error). In a first phase of training, data corresponding to each season was used and in a second phase the whole data set was used. Modelling through ANNs showed to be very efficient, reaching an average error level of 8% only. These results suggest that this approach could be useful for forecasts of carbon assimilation and sequestration by tropical trees.
Submetido por Madeleine Lita Barriga Puente de la Vega em 25-MAR-2004
Tema Científico do LBA: CD (Armazenamento e Trocas de Carbono)