Estimating CO2 Flux at 30 Meter Resolution Using Machine Learning

We present an approach to using a machine learning based regression model to estimate CO2 fluxes at 30 meter spatial resolution. The method uses eddy covariance measurements of CO2 obtained from in situ stations (FLUXNET) as primary reference data. Multispectral satellite observations collected by Landsat are combined with meteorological information to form feature vectors that are used as predictor variables. The XGBoost machine learning algorithm is used to train the regression models on a per-land cover basis. The resulting models can be used to estimate CO2 fluxes wherever Landsat satellite imagery is available. Moreover, the approach provides a framework that is extensible to other satellite imagery types and will improve in accuracy as more primary reference data becomes available. We present results of the method as applied to examples in the agricultural sector.