NEURAL NETWORKS FOR ESTIMATING NET ECOSYSTEM CO_2 EXCHANGE USING INCOMPLETE EDDY COVARIANCE DATA

摘要

Handling of missing eddy covariance (EC) data isrnnecessary to construct daily and annual sums of netrnecosystem CO2 exchange (NEE). This study aims atrnevaluating three different types of artificial neuralrnnetwork methods (ANNs), namely multi-layer perceptronrn(MLP), support vector regression (SVR) and selforganizingrnmap (SOM), for the estimation of NEE valuesrnin EC data. The performance of the methods is examinedrnwith realistic missing EC data patterns by means ofrnseveral numerical performance indices. In general, thernresults obtained clearly show the high accuracy of ANNs,rnas they yielded the highest accuracies with SVR andrnMLP. The results also suggest that the NEE time seriesrncould be accurately estimated, even when frequent, largerndata gaps exist in the EC series.