Application Of Artificial Neural Network In Modeling Carbon Flux

In terrestrial ecosystems,data deficiency is a common problem while Eddy covariance technology is used to monitor the flux data.Finding a better way to develop effective methods for simulating the CO2 flux data is one of the hottest issues in the researches of global climate change.As part of data quality improvement and analysis,exploring the high-accuracy model settles the foundation of simulating the data accurately.Aiming at this problem,this paper deeply analyzed the ecological characteristics and sequential characteristics of CO2 flux data collected from Beijing Olympic Forest Park.The major results of this thesis can be summarized as follows:(1)In growing season,the values of CO2 flux were simulated by XGBoost and artificial neural network(ANN)model based on gradient descent rule.The results showed that the importance ranking of factors was carried out by XGBoost model,and characterization of ecosystem responses to climatic controls was calculated by artificial neural network model efficiently.(2)After analyzing the different time scales of CO2 flux,this paper stated the working principle of time series and the change trend in different months.By calculating the autocorrelation coefficient,the results showed that three points before the target had significant impact on the target time.(3)A kind of integration model named ANN-LSTM composed of ecological domain features and sequential features was proposed and the accuracy of models were evaluated using the coefficient of determination(R2),the mean squared root error(RMSE),the mean absolute error(MAE)and the index of agreement(IA).Compared with other models including Back Propagation neural network(BP-ANN),BP-ANN improved by Moffat(Moffat-ANN),Long short-term memory neural network(LSTM)with single-factor and multi-factors,the ANN-LSTM model got the best performance yielding a R2 of 0.792,RMSE of 3.012 ?mol/(m2·s),MAE of 2.000 ?mol/(m2·s)and IA of 0.934 on train dataset,and R2 of 0.811,RMSE of 3.081 ?mol/(m2·s),MAE of 2.057 ?mol/(m2·s)and IA of 0.937 on test dataset.Compared with other models,this method could improve the accuracy of simulation effectively.