| The eddy covariance technique for measuring the exchange of mass and energy between the land surface and atmosphere yields data records with substantial gaps, reported to be as long as 30 to 40% of a time series annually (at a half-hourly time step). The application of these data sets in modeling studies as well as on varying time scales and under non-ideal conditions, requires some interpolation method to infer values for the missing data. This study will consider a neural network regression model for a flux record from a semi-arid riparian site and examine the model's responsiveness to variability in the data available for training. The neural network sensitivity to flux data used for training is evaluated. Model response worsened under reduced training data availability and was dependent on the characteristics of the data. |
