Research On Carbon Flux Prediction And Response Of Multiple Terrestrial Ecosystems Based On Deep Learning

Global carbon cycle change plays an important role in global climate warming.It is of great research value to explore the atmospheric carbon exchange and the relationship between carbon flux and various ecological factors.Currently,mainstream modeling is based on ecological mechanism or ecological long-term continuous observation data,the application of deep learning theory and technology is still in its infancy.First,this paper takes the long-term continuous observation flux data of7 kinds of terrestrial ecosystems in FLUXNET,including evergreen broad-leaved forest,evergreen coniferous forest,deciduous broad-leaved forest,mixed forest,grassland,farmland and wetland as the research object,using deep learning model attention,encoder-decoder,LSTM and machine learning model ANN,SVM,ELM and GRNN for carbon flux simulation,comparing simulation results of different models on different sites.Secondly,Random Forest is used to analyze the different responses to ecological indicators of each site's carbon flux to ecological factors.Finally,the time series feature theory is used to study the reasons for the difference in the simulation results of carbon flux data.The thesis research conclusions are:(1)The attention model on sites US-GLE,US-WCr,GF-Guy,CH-Lae,US-SRG,US-Myb,CH-Oe2,the root mean square errors are 1.48,3.10,4.93,5.32,1.20,3.74,1.76?mol/(m~2·s),the average absolute errors are 0.94,1.88,3.23,3.54,0.67,2.37,1.26?mol/(m~2·s),and the determination coefficients are 0.91,0.93,0.89,0.84,0.89,0.90,0.96,and the consistency coefficients are 0.95,0.96,0.94,0.90,0.94,0.95,0.97,respectively,which have advantages over traditional machine learning models in the accuracy of carbon flux data prediction.(2)The random forest model was used to calculate the importance scores of carbon flux from January to December for light,temperature,and precipitation.The results showed that forest ecosystems were mainly affected by PPFD and TS;grassland ecosystems were mainly affected by TS,PPFD,and SWC.Impact;The main factors affecting farmland ecosystems are TS,SWC and VPD;wetland ecosystems are most affected by TS,PPFD and VPD.(3)Calculate the time series characteristics such as binned entropy and approximate entropy of each terrestrial ecosystem site.The results show that for GF-Guy site with similar deep learning and machine learning effects,the carbon flux data has high complexity,and the measured value distribution is randomly fluctuating around the mean.There is no certain trend,and it is difficult to predict.