Gap Filling and Quality Assessment of CO2 and Water Vapour Fluxes above an Urban Area with Radial Basis Function Neural Networks

Abstract

Vertical turbulent fluxes of water vapour, carbon dioxide, and sensible heat were measured from 16 August to the 28 September 2006 near the city centre of Münster in northwest Germany. In comparison to results of measurements above homogeneous ecosystem sites, the CO 2 fluxes above the urban investigation area showed more peaks and higher variances during the course of a day, probably caused by traffic and other varying, anthropogenic sources. The main goal of this study is the introduction and establishment of a new gap filling procedure using radial basis function (RBF) neural networks, which is also applicable under complex environmental conditions. We applied adapted RBF neural networks within a combined modular expert system of neural networks as an innovative approach to fill data gaps in micrometeorological flux time series. We found that RBF networks are superior to multilayer perceptron (MLP) neural networks in the reproduction of the highly variable turbulent fluxes. In addition, we enhanced the methodology in the field of quality assessment for eddy covariance data. An RBF neural network mapping system was used to identify conditions of a turbulence regime that allows reliable quantification of turbulent fluxes through finding an acceptable minimum of the friction velocity. For the data analysed in this study, the minimum acceptable friction velocity was found to be 0.15 m s −1 . The obtained CO 2 fluxes, measured on a tower at 65 m a.g.l., reached average values of 12 µmol m −2 s −1 and fell to nighttime minimum values of 3 µmol m −2 s −1 . Mean daily CO 2 emissions of 21 g CO 2 m −2 d −1 were obtained during our 6-week experiment. Hence, the city centre of Münster appeared to be a significant source of CO 2 . The half-hourly average values of water vapour fluxes ranged between 0.062 and 0.989 mmol m −2 s −1 and showed lower variances than the simultaneously measured fluxes of CO 2 .

Testing the applicability of neural networks as a gap-filling method using CH<sub>4</sub> flux data from high latitude wetlands

“…Artificial neural networks that have previously been successfully implemented as a gap-filling method (Falge et al, 2001, Moffat et al, 2007 for carbon dioxide flux time series (Aubinet et al, 2000;Carrara et al, 2003;Valentini, 2003 andSchmidt et al, 2008) have been described as a robust, reliable and versatile tool. Nevertheless, their application is time consuming, particularly in finding the appropriate input variables, the appropriate number of hidden layers, and neurons/nodes within these layers, as well as the choice of training and test data sets (data rows).…”

“…The application of neural networks (Jain et al, 1996;Svozil et al, 1997;Elizondo and Góngora, 2005;Saxén and Pettersson, 2006) for data recovery and gap-filling (Aubinet et al, 2000;Gorban et al, 2002;Papale and Valentini, 2003;Ooba et al, 2006;Moffat et al, 2007 andSchmidt et al, 2008) has proven to be a very reliable tool in several scientific disciplines Dorling, 1998, 1999;Lek and Guégan, 1999;Lee and Jeng, 2002;Toptygin et al, 2005). In atmospheric sciences (Gardner and Dorling, 1998;Toptygin et al, 2005;Chattopadhyay G. and Chattopadhyay S., 2008), applying neural networks in forecasting has become a standard application.…”

“…Nelson et al (1999) showed better results for de-seasonalised time series, while others (Sharda and Patil, 1992;Franses and Draisma, 1997) found that neural networks are able to model seasonality directly and prior de-seasonalisation is not necessary. Regarding gap-filling of atmospheric trace gas fluxes (van Wijk and Bouten, 1999;Aubinet et al, 2000;Carrara et al, 2003;Papale and Valentini, 2003;Ryan et al, 2004;Ooba et al, 2006 andSchmidt et al, 2008), no de-seasonalisation of data was carried out prior to applying neural networks. As artificial neural networks have the ability to deal with complex data sets and as no de-seasonalisation of data was carried out in previous gap-filling studies (see above references) we decided to proceed in the same manner.…”

“…In order to add this seasonal and diurnal effect to their neural network, Papale and Valentini (2003) introduced several fuzzy data sets reflecting the diurnal and seasonal variation to reduce the linear cumulative numerical weight of time in relation to other variables. Adding this type of input to neural networks does not always increase the neural network performance, and has been shown by Schmidt et al (2008) to have sometimes little effect. Our analysis, however, showed an increase in network performance when including these fuzzy sets.…”

Estimativa De Fluxo De Calor Latente Em Reservatórios Através De Uma Rede Neural Artificial

“…( (Schmidt et al, 2008) e (Ooba et al, 2006)). Este trabalho avaliou a utilização do modelo de Redes Neurais Artificiais (RNA) backpropagation como alternativa para o preenchimento das séries temporais de fluxo de calor latente do lago de Furnas com dados artificiais.…”