Modeling Net Ecosystem Carbon Dioxide Exchange Using Temporal Neural Networks after Wavelet Denoising
MetadataShow full item record
Eddy covariance (EC) time-series data obtained from flux towers are noisy due to both stochastic atmospheric turbulence and deterministic processes, and no standard data-denoising protocols exist for them. The potential of six temporal artificial neural networks (ANNs) augmented with and without three orthogonal wavelet functions was tested for predicting net ecosystem exchange of carbon dioxide (CO2) based on a long-term EC data set for a temperate peatland. Multiple comparisons were made of (1) temporal ANNs with and without discrete wavelet transform (DWT) denoising; (2) denoising with the orthogonal wavelet families of Daubechies, Coiflet, and Symmlet; (3) different decomposition levels; (4) time-delay neural network, time-lag recurrent network, and recurrent neural network; (5) online learning versus batch learning algorithms; and (6) diel, diurnal, and nocturnal periods. The coefficient of determination, root mean square error, and mean absolute error performance metrics were used for multiple comparisons based on training, cross-validation, and independent validation of the temporal ANNs as a function of 24 explanatory variables contained in an EC data set. Integration of the temporal ANNs and DWT denoising provided more accurate and precise estimates of net ecosystem CO2 exchange.