Estimation of CO2 fluxes across different biomes using machine learning approaches

Quantifying CO2 fluxes over terrestrial land is crucial to better understand the global carbon cycle and the contribution of ecosystems to climate change. In addition, ecosystems such as croplands and forests have the potential to sequester carbon in the soil and vegetation. Robust tools to simulate CO2 fluxes with high accuracy are needed to identify best practices and management for carbon sequestration.
In this study, the Net Ecosystem Exchange (NEE) from different networks (ICOS, NEON) is used to develop machine learning (ML) approaches to simulate daily CO2 fluxes. These biome specific approaches use as input high spatial and temporal resolution optical remote sensing products combined with meteorological data. The biomes considered are cropland, deciduous forest, evergreen forest and grassland. Different ML models were tested and the ExtraTreesRegression model seems to be better suited for all biomes except grassland where an SVR model was more appropriate. The features identified as most important among the remote sensing products are NDVI and NDMI while among meteorological variables, global radiation, air temperature and fraction of diffuse radiation appears as more relevant.
The predicted results show good agreement with daily observations, with R2 of 0.82 over cropland. The performance of the model in simulating CO2 fluxes over forests is more contrasted with good accuracy over deciduous forests (R2 of 0.72) but low confidence over evergreen forests (R2 of 0.29). Finally the model was also applied to grassland, but the small size of the dataset combined with the high heterogeneity of soil and climatic conditions of grassland sites led to low correlation with observations (R2 of 0.44).
This work demonstrates the potential of a machine learning-based method to assess CO2 fluxes across different biomes, and should be further explored due to its ease of use and application.