Net Ecosystem Exchange (NEE) simulation in maize using artificial neural networks

Abstract

The rate of the net ecosystem exchange of CO2 (NEE) between the atmosphere and vegetated surface plays important role to achieve a better understanding regarding the global carbon budget. The eddy covariance method is known as an accurate and direct approach to measure the CO2 flux in small scale from a hundred meter to several kilometers. Some of resources and practical limitations prevent the utilization of this technique. Thus, simulation methods can be applied to estimate the NEE in the scale of larger than eddy covariance measurement capability. One class of simulation models, Artificial Neural Networks (ANNs), are powerful tools to identify the complicated non-linear relations among the input and output vectors. They used a set of input variables to simulate the measured NEE values by investigation of the relations (hidden rules) among them.In this study, multi-layer perceptron neural network trained by steepest descent Back-Propagation (BP) algorithm was tested to simulate the NEE above two maize sites (rain-fed &irrigated) near Mead, Nebraska. In addition, the sensitivity of NEE with respect to each input was analyzed over the growth stages.Network training and testing was fulfilled using on two different subsets of the hourly data which were selected from days of year (DOY) 169 to 222 for 2001, 2003, 2005, 2007, and 2009.The results showed a high correlation between actual and estimated NEE values (R2 > 0.9620). The RMSE for NEE reached 0.0681 and 0.0642 (mg m−2 s −1) for irrigated and rain-fed sites respectively. Thus, the results indicate the reliability and efficiency of this technique to simulate the NEE. The sensitivity analysis indicated that the most effective inputs on the NEE were identified Rn and LAI for irrigated site and Rn, and VPD for rain-fed site. Furthermore, to achieve the minimal set of input in order to speed up the analysis procedures; the impact intensity of each input on NEE was recognized by deactivation of each input vector in network training.