Quantifying biosphere–atmosphere exchange of CO2 using eddy covariance, wavelet denoising, neural networks, and multiple regression models

Abstract

Net ecosystem exchange of CO2 (NEE) over a temperate peatland in northwestern Turkey was directly measured using the eddy covariance (EC) method for 590 days. Both the response variables of diurnal and nocturnal NEE (NEEd and Reco-n) and the explanatory variables of latent heat (LE), relative humidity (RH), and atmospheric CO2 and H2O concentrations (AtmCO2 and AtmH2O) were denoised with discrete wavelet transform (DWT) using coiflet (coif10-6). Denoised NEE fluxes and their temporal components were modeled using multiple linear regression (MLR), polynomial regression (PR) and artificial neural network (ANN) models as a function of LE, RH, AtmCO2, AtmH2O, air temperature (Tair), day of year (DOY), and local time. Peak NEEd flux, and peak Reco-n efflux were −0.37mgCO2m−2s−1 in late July and 0.27mgCO2m−2s−1 in mid-August. Mean annual NEE was estimated at −1157gCO2m−2 which is in agreement with previous results of peatland studies. The use of DWT-augmented ANN, MLR and PR models significantly increased predictive power and reduced uncertainties in predicting the temporal dynamics of the biosphere–atmosphere CO2 exchange, relative to the models without DWT denoising. Out of 28 DWT-augmented ANNs, multilayer perceptron (MLP) and recurrent network (RN) models were the best diurnal and nocturnal ones, respectively, based on accuracy metrics derived from training, cross-validation and independent validation. Among the DWT-based ANN, MLR and PR models, diurnal MLP and nocturnal MLR outperformed the others. Wavelet-augmented ANN and MLR models appear to be a promising tool to quantify diurnal and nocturnal NEE dynamics, respectively.