Modeling seasonal variations of long-term soil CO2 emissions in an orchard plantation in a semiarid area, SE Turkey

Emissions of soil CO2 under different management systems have a significant effect on the carbon balance in the atmosphere. Soil CO2 emissions were measured from an apricot orchard at two different locations: under the crown of trees (CO2-UC) and between tree rows (CO2-BR). For comparison, one other measurement was performed on bare soil (CO2-BS) located next to the orchard field. Analytical data were obtained weekly during 8 years from April 2008 to December 2016. Various environmental parameters such as air temperature, soil temperature at different depths, soil moisture, rainfall, and relative humidity were used for modeling and estimating the long-term seasonal variations in soil CO2 emissions using two different methods: generalized linear model (GLM) and artificial neural network (ANN). Before modeling, data were randomly split into two parts, one for calibration and the second for validation, with a varying number of samples in each part. Performances of the models were compared and evaluated using means absolute of estimations (MAE), square root of mean of prediction (RMSEP), and coefficient of determination (R2) values. CO2-UC, CO2-BR, and CO2-BS values ranged from 11 to 3985, from 9 to 2365, and from 8 to 1722 kg ha−1 week−1, respectively. Soil CO2 emissions were significantly correlated (p < 0.05) with some environmental variables. The results showed that GLM and ANN models provided similar accuracies in modeling and estimating soil CO2 emissions, as the number of samples in the validation data set increased. The ANN was more advantageous than GLM models by providing a better fit between actual observations and predictions and lower RMSEP and MAE values. The results suggested that the success of environmental variables for estimations of CO2 emissions using the two methods was moderate.