A Neural Network Model for Estimating Carbon Fluxes in Forest Ecosystems from Remote Sensing Data

Forests are among the main places on Earth where carbon is collected and accumulated. However, instrumental assessment of carbon fluxes is possible only for small areas. When solving the scaling problem, machine learning methods are used, which allow transforming the Earth’s surface reflectance intensities in different spectral ranges into ground-based in situ observations. We suggest a regression neural network model of the multilayer perceptron type for assessment of carbon fluxes. The model is trained on FLUXNET network data for a station located in a boreal coniferous forest (56.4615° N, 32.9221° E). Using the vegetation indices NDVI and EVI measured by the MODIS spectroradiometer onboard the Aqua satellite, the air temperature at an altitude of 2 m, and total precipitation as input data, the model estimates the gross primary production (GPP), net ecosystem exchange (NEE), ecosystem respiration (TER), and some other parameters which characterize water and energy fluxes. The statistical assessments for the test dataset show high correlation coefficients ( R ) and Nash–Sutcliffe coefficients (NSE): R > 0.9 and NSE ≥ 0.87 for GPP and TER; R = 0.4 and NSE = 0.15 for NEE.