Neural Networks for Estimating Net Ecosystem CO2 Exchange using Incomplete Eddy Covariance Data

H. Niska, N. Shurpali, P. Martikainen, M. Kolehmainen (Finland)


Modelling, imputation, missing data, and neural networks


Handling of missing eddy covariance (EC) data is necessary to construct daily and annual sums of net ecosystem CO2 exchange (NEE). This study aims at evaluating three different types of artificial neural network methods (ANNs), namely multi-layer perceptron (MLP), support vector regression (SVR) and self organizing map (SOM), for the estimation of NEE values in EC data. The performance of the methods is examined with realistic missing EC data patterns by means of several numerical performance indices. In general, the results obtained clearly show the high accuracy of ANNs, as they yielded the highest accuracies with SVR and MLP. The results also suggest that the NEE time series could be accurately estimated, even when frequent, large data gaps exist in the EC series.

Important Links:

Go Back