Prediction of the CO2 emission across grassland and cropland using tower-based eddy covariance flux measurements: a machine learning approach

In this research, the magnitude of the net ecosystem exchange (NEE) for both grasslands (GRA) and croplands (CRO) is estimated by different machine learning approaches (MLAs). There are two main goals including prediction/data gap filling of the NEE and developing a new MLA model. The variation of CO2 is affected by soil temperature and meteorological factors, including air temperature, latent heat, and sensible heat considered as inputs. Hourly data of three AmeriFlux sites have been collected for seven years. The normalized smoothed data are applied for modeling. Both artificial neural network (ANN) and genetic algorithm (GA) are the computational MLAs working by deep learning methods. In this study, a new GA-based model named integration of optimization with genetic algorithm and Fourier series (IOGAFS) was proposed for estimation of the NEE. The results show the IOGAFS and ANN methods have acceptable performance with 0.86 and 0.88 determination coefficient for CRO and 0.75 and 0.81 for GRA, respectively. Due to high performance of both methods, they can be used estimation of the NEE in similar ecosystems, mainly where there are no flux towers.