ANALYSING FOREST TRANSPIRATION MODEL ERRORS WITH ARTIFICIAL NEURAL NETWORKS

Abstract

A Single Big Leaf (SBL) forest transpiration model was calibrated on half-hourly eddy correlation measurements. The SBL model is based on the Penman-Monteith equation with a canopy conductance controlled by environmental variables. The model has eight calibration parameters, which determine the shape of the response functions. After calibration, residuals between measurements and model results exhibit complex patterns and contain random and systematic errors. Artificial Neural Networks (ANNs) were used to analyse these residuals for any systematic relations with environmental variables that may improve the SBL model. Different sub-sets of data were used to calibrate and validate the ANNs. Both wind direction and wind speed turned out to improve the model results. ANNs were able to find the source area of the fluxes of the Douglas fir stand within a larger heterogeneous forest without using a priori knowledge of the forest structure. With ANNs, improvements were also found in the shape and parameterisation of the response functions. Systematic errors in the original SBL model, caused by interdependencies between environmental variables, were not found anymore with the new parameterisation. After the ANNs analyses, about 80% of the residuals can be attributed to random errors of eddy correlation measurements. It is finally concluded that ANNs are able to find systematic trends even in very noisy residuals if applied properly.