Footprint Aware Carbon Dioxide Flux Gap-Filling for Agricultural Eddy Covariance Systems.

Missing observational data for eddy covariance or any field instrumentation is ubiquitous. While there are well tested and rigorous methods such as marginal distribution sampling (MDS) to gap-fill observations, there are many other machine learning methods such as random forests, gradient boosting and deep learning. This project investigated several deep learning architectures to incorporate footprint variability as well as quantifying the epistemic and aleatoric uncertainty of the imputing model. A generative adversarial network, using flux footprint information, had improved epistemic uncertainty and was found to be competitive with MDS. This type of deep learning imputation architecture can be a valuable tool for eddy covariance gap filling.