Comparisons of gap-filling methods for carbon flux dataset: A combination of a genetic algorithm and an artificial neural network

In many cases of field measurements, missing data of net ecosystem CO2 exchange (NEE) constitute a quarter to almost half of all data. Those omissions result from accidents of instruments, unfavorable weather, and quality-control screening. Accurate gap-filling methods are needed for interpolating these missing data. This study evaluated the performance of interpolation methods for gap-filling of NEE data: a conventional method with empirical equations of bioreaction whose parameters were estimated using nonlinear regression (NR) methods, an artificial neural network (ANN) method, and an automated ANN method (genetic neural network, GNN). In the GNN method, parameters of an ANN model, such as initial weight matrixes and input-data-selections, were determined automatically using a genetic algorithm (GA).A tested dataset was observed in a Japanese larch plantation in northern Japan for 5 months from May to September 2002. The available dataset was divided into two subsets to train and validate the models. Averaged coefficients of determination (r2) of ANN models between observed and estimated values of NEE were almost identical to that of the NR method (r2 = 0.86). The performance of the GNN method (r2 = 0.88, averaged value) was somewhat better than those of the two methods. Using the GNN method, the mean daily NEE was estimated at −0.29 mol m−2 d−1 for the 5-month period using ANNs, thereby showing better performance. The mean daily NEE, which was reported previously, was compatible with that from the NR method (−0.23 mol m−2 d−1).From those results, it was concluded that the proposed GNN method offers better performance for gap-filling and high availability because of the obviated need for specialization of ecological or physiological mechanisms.