Evaluation of an Artificial Neural Network based method for surface heat flux estimates over heterogeneous landscapes

A new Artificial Neural Network (ANN)-based method for measuring surface-layer fluxes at low cost without compromising accuracy is presented. It is based on a one-year dataset of surface sensible and latent heat fluxes from eddy-covariance stations, along with classical meteorological variables from automatic weather stations, acquired over a meadow and a maize crop field. For each of these landscapes, ANNs are trained to estimate observed heat fluxes based on relevant variables that basically describe radiative and atmospheric forcing. The costly eddy-covariance station is required only for a short measurement period to collect sufficient reference data to train ANNs. Then, the estimation of heat fluxes relies on measurements from the affordable automatic weather station. The final objective is to deploy this method during a one-year field campaign to monitor different land-surface covers. Three questions are investigated here: the experimental scenario that allows to constitute the best training dataset, the accuracy of the ANNs estimates across different timescales, and the ANNs performance according to the land-surface cover. From a general perspective, the ANNs estimates have quite similar performance for all training datasets, irrespective of the vegetation cover. However, for conditions that were not well sampled in the training dataset, higher discrepancies were observed between ANNs estimated flux and those measured using eddy-covariance station. The comparison between the two vegetation covers indicated a degradation in ANNs performance for the maize crop compared to the meadow, with an increase of the bias up to 40 % and a decrease of the Pearson correlation coefficient up to 8 % .