HR: 16:00h
AN: B42D-01 INVITED     [PDF]
TI: Carbon Fluxes Assessment Using Artificial Neural Networks
AU: * Papale, D
EM: darpap@unitus.it
AF: DISAFRI University of Tuscia, via C. de Lellis, Viterbo, 01100 Italy
AU: Valentini, R
EM: rik@unitus.it
AF: DISAFRI University of Tuscia, via C. de Lellis, Viterbo, 01100 Italy
AB: The feed-forward backpropagation artificial neural networks (ANNs), used in this application, are non-linear models able to reproduce non-linear functions. Unlike process based and multiple constrain models, to use ANNs it is not necessary to know the relationship between input and output variables, because this is found directly by the neural network. To do this and also to set the values of the parameters, the ANNs have to be trained using examples with the input values and also the correct output (supervised approach). We used eddy covariance carbon flux measurements to train artificial neural networks to simulate carbon fluxes at different scales. ANNs showed good performances when applied to gap-fill and correct the flux datasets at site level. The input variables used are meteorological data (air and soil temperatures, photosynthetic photon flux density, air humidity etc.) and four fuzzy values for the seasons, while the output is the carbon dioxide flux. An intercomparison between different gap-filling techniques showed that ANNs are able to reconstruct the annual trend of fluxes better than the others classic methods (regressions, look-up tables, mean daily values.). Another ANNs possible application is the NEE spatialization at regional and continental scales. After a first methodological experiment where NEE data from the Euroflux sites were used to obtain NEE maps of Europe with weekly time resolution and 1 km of spatial resolution (Papale and Valentini 2003), other analysis showed the potentiality of this approach to generalize from site to region. The GLOFLUX initiative (http://gaia.agraria.unitus.it/gloflux.html), lunched one year ago, has the aim to combine the available datasets into an artificial neural networks system to spatialize carbon fluxes at global scale.
UR: http://gaia.agraria.unitus.it/gloflux.html
DE: 0400 Biogeosciences
DE: 1615 Biogeochemical processes (4805)
DE: 1694 Instruments and techniques
SC: Biogeosciences [B]
MN: 2003 Fall Meeting