CO_2Flux Simulation Based On Adaptive Neuro-fuzzy Inference System

Adaptive neuro-fuzzy inference system(ANFIS)is used for CO2_flux simulation to obtain accurate simulation and provide research methods.Based on the data from 2015 to 2017 in Badaling station,feature importance is selected by Random forest(RF)and Gradient boosting decision tree(GBDT)which are both based on ensemble learning.Based on the feature selection results,CO2_flux is simulated by RF,GDBT and ANFIS,Support vector machine(SVM),Back-propagation neural network(BPNN)which are all based on different internal function.Optimal model of each method are selected and compared by using coefficient of determination(R2),index of agreement(IA)and other indicators to analyze the feasibility of ANFIS and the performance of different internal function.Combining feature selection results of RF and GDBT,R2 in training and test dataset of both RF and GDBT are equally 0.79 and 0.77 based on all feature input.5 important feature selected are photosynthetic active radiation,10cm soil temperature average at position 2,2m relative humidity average,2m air temperature average,soil moisture average at position 4.Combining the results of CO2_flux simulation,among the membership function of ANFIS,gaussian function is better than generalized S-shape function.ANFIS with 3 gaussian function is equivalent to ANFIS with 2 gaussian function.Optimal model is ANFIS with 2 gaussian function and its R2 in training and test dataset are 0.768 and 0.756,IA are 0.928 and 0.922.In ensemble learning,RF is better than GDBT and equivalent to the optimal model of ANFIS.Among the kernel function of SVM,gaussian function is better than polynomial function and standard S-shaped sigmoid function.Optimal model is SVM with gaussian function and its R2 in training and test dataset are 0.687 and 0.680,IA are 0.903 and 0.900.Among the activation function of BPNN,tanh function is better than standard S-shaped sigmoid function.Optimal model is BPNN with tanh function and equivalent to the optimal model of SVM.The simulation results of four optimal models on Badaling CO2_flux show that,(1)ANFIS is comparable to RF and superior to BPNN and SVM.ANFIS has no complicated parameter adjustment process.Simulation by ANFIS is feasible.(2)Among the internal function,gaussian function is superior to S-shaped function.Tanh function is superior to standard S-shaped sigmoid function.Generalized S-shape function is superior to standard S-shaped sigmoid function.(3)The performance of ANFIS is not positively correlated with the number of membership function.(4)Compared with simulation results of RF and GDBT based on 5 feature input and all feature input,the contribution rate of 5 feature is about 98%,which verifies the correctness of feature selection.