Phuong Nguyen, Milton Halem

The advent of direct high-resolution global surface measurements of CO2 from the recently launched NASA Orbiting Carbon Observatory (OCO-2) satellite offers an opportunity to improve the estimate of Net Ecosystem Exchange (NEE) over land. Long-term measurements of CO2flux obtained from eddy covariance instruments on Flux towers show large annual differences with NEE calculated by inverse methods using Land Surface Photosynthetic models. This suggests consideration of alternative approaches for calculating seasonal to annual global CO2 flux over land. Recent advances in deep machine learning models, including recurrent neural nets, have been successfully applied to many inverse measurement problems in the Earth and space sciences. We present evaluations of two deep machine learning models for estimating CO2 flux or NEE using station tower data acquired from the DOE Atmospheric Radiation Measurement (ARM), AmeriFlux and Fluxnet2015 station datasets. Our results indicate that deep learning models employing Recurrent Neural Networks (RNN) with the Long Short Term Memory (LSTM) provide significantly more accurate predictions of CO2 flux (~22% -28% improvements) than Feed Forward Neural Nets (FFNN) in terms of Root Mean Square Errors, correlation coefficients and anomaly correlations with observations. It was found that using heat flux as input variables also produce more accurate CO2 flux or NEE predictions. A non-intuitive machine learning metaphysical result was observed by the omission of CO2 concentrations as an input variable. Neural net models, in most cases, produce comparable accuracies of CO2 flux or NEE inferences, when trained with and without CO2 for the same station data.