Estimation of terrestrial ecosystem NEP coupled with remote sensing and ground data

Net ecosystem productivity (NEP) is a key indicator that reflects the net absorption or net emissions of carbon in terrestrial ecosystems to measure the carbon sink capacity of vegetation. This paper builds machine learning models such as Random Forest Regression model (RFR), Support Vector Regression machine model (SVR) and Back Propagation Neural Network model (BPNN), coupled with Fluxnet2015 data set and NDVI data to build machine learning training data set. Based on different machine learning models, centrally learning the characteristics of the training data set and mining the implicit relationship information between each impact factor and NEP. The estimation results of each machine learning model are evaluated through $\mathrm{R}^{2}$, RMSE and other accuracy indicators and analyze the importance of the characteristics that affect NEP prediction. Provide scientific basis and technical support for ecological environmental monitoring and protection. The research results show that the Random Forest Regression model (RFR) and the Support Vector Regression machine model (SVR) have good prediction results for NEP data, and $\mathrm{R}^{2}$ is greater than 0.7. NDVI is an important influence parameter for the overall prediction of NEP of vegetation types. When predicting NEP data of different vegetation types, the prediction results of the Random Forest Regression model (RFR) and the Support Vector Regression machine model (SVR) are generally good. Except for permanent wetland, all other types of $\mathrm{R}^{2}$ are greater than 0.6, and the Back Propagation Neural Network model’s NEP prediction results for permanent wetland are better than the Random Forest Regression model and the Support Vector Regression machine model. These three machine models can well predict the NEP of mixed forests, and the accuracy of the prediction results of permanent wetland NEP is poor compared with other types. The accuracy of the prediction results of the Support Vector Regression machine model for woody savannas is better than that of the other two machine learning models. NDVI and latent heat flux have an important impact on the NEP prediction of various vegetation types. Characteristics such as soil water content, sensible heat flux and net radiation are one of the key influencing factors in the NEP prediction of specific vegetation types. Research shows that NEP estimation based on machine learning and multisource data can provide remote sensing monitoring results of carbon revenue and expenditure in terrestrial ecosystems similar to real observation on the ground, and can provide scientific basis and technical support for the formulation of ecological and environmental protection policies and ecological environmental monitoring and protection.