Prediction And Analysis Of Ecosystem Net Carbon Exchange Based On LightGBM And RF

Carbon cycle has become one of the most important issues of international human concern.Net Ecosystem Exchange(NEE)helps to measure the balance of carbon cycle.The ecosystem provides a reference for the feedback mechanism of global climate change and the prediction of regional climate change,and is of great significance for understanding the relationship between ecosystems and environmental factors.In order to improve the accuracy of NEE prediction and obtain the objective relationship behind the flux and meteorological data more accurately,study the application of Light GBM in NEE prediction and explore new methods;and analyze the dynamic changes of important environmental factors in the growing season and non-growing season Characteristic responses,and analysis of the impact on NEE dynamics on diurnal and annual scales.In this study,using the daily 30-minute meteorological and flux data sets of Changbai Mountain Temperate Broad-leaved Forest Observatory for five consecutive years,the data sets were divided into three types,namely growing season data sets,non-growing season data sets and Do not split the data set;use 24 environmental variable factor characteristics and 5consecutive years of data set to generate the reference frame of the Pearson correlation coefficient environmental variable,and calculate the important score of each environmental factor variable in the reference frame through the Gini coefficient of the RF model,After sorting,the top 5 environmental factors were selected to analyze the importance of environmental factors on a daily scale and an annual scale.At the same time,the 24 sorted environmental factors were divided into 5 groups,and the combination of environmental factors that affected the accuracy of the model’s prediction of NEE was screened out.The Light GBM regression model was used to train and predict NEE,and the model parameters were optimized.The accuracy of the model was verified by calculating the coefficient of determination(R~2),mean absolute error(MAE),and root mean square error(RMSE).Validate and compare with different machine learning models,and then perform correlation analysis and daily-scale and annual-scale analysis of NEE in the growing season and non-growing season.The results show:(1)During the growing season,NEE was significantly correlated with net radiation(NR),ecosystem latent heat flux(LE),solar radiation(SR)and ecosystem sensible heat flux(Hs),and was significantly correlated with water vapor pressure above the canopy(WVP2)Not significant;NEE in the non-growing season has a significant correlation with photosynthetically active radiation(PAR)and net radiation(NR),and has a significant correlation with second-layer soil temperature(Ts2),near-surface air temperature(Ta1),and first-layer soil volumetric water content(SWC1)Not significant;the growth season acts as a carbon sink,the non-growth season acts as a carbon source,and the overall carbon sink,and NR is the main environmental factor affecting the change of NEE in both the growing season and the non-growing season at the daily and annual scales;(2)The non-growth data set has an impact on the training and prediction of the model,but has the least impact on the Light GBM regression model;the Light GBM regression model in the test set has an R~2of 0.802,an MAE of 0.078μmol/(m~2·s),and an RMSE of 0.13μmol/(m~2·s),compared with the other three regression models(RF,SVR and KNN),the Light GBM regression model has higher prediction accuracy.