Assessing neural networks with wavelet denoising and regression models in predicting diel dynamics of eddy covariance-measured latent and sensible heat fluxes and evapotranspiration

Abstract

Eddy covariance (EC)-measured data were used to develop multiple nonlinear regression (MNLR) models of latent (LE) and sensible heat (H s) fluxes, and micrometeorological station-measured actual evapotranspiration (ET). Discrete wavelet transform (DWT) with symmlets (sym10), coiflets (coif10), and daubechies (db10) was used to decompose time series signals of LE, H s, and ET into frequency components in order to feed denoised output data into 26 artificial neural networks (ANNs) with different learning algorithms, based on independent validation-derived values of coefficient of determination (r 2), root mean square error (RMSE), mean absolute error (MAE), wavelet neural networks (WNNs) with coif10-1 and db10-1 outperformed ANNs, and MNLR models. The best ones out of 26 WNNs appeared to be multilayer perceptrons (MLPs) for LE and H s, and time-delay network (TDNN) for ET, while the best ones out of 26 ANNs were determined as TDNN for LE, MLP for H s, and generalized feedforward network (GFF) for ET. The combination of batch mode and Levenberg–Marquardt algorithm was adopted in the ANNs and WNNs more frequently and generated better accuracy metrics than the combinations of online mode and Momentum algorithm, and batch mode and Momentum algorithm.