A Machine Learning Method for Carbon Flux Partitioning Using Solar-induced Fluorescence Measurements

Abstract:

The measured net ecosystem exchange (NEE) is the net balance between gross primary production (GPP) and ecosystem respiration (Reco). Partitioning NEE into its components is critical for understanding the terrestrial carbon cycle. The standard partitioning methods are based on assumed relationships (e.g. Lloyd & Taylor equation), which however, may be biased due to unaccounted processes (e.g. light inhibition of respiration) and neglected important predictors (e.g. soil moisture). Here, we apply neural networks (NNs) to partition NEE using comprehensive environmental predictors combined with solar-induced fluorescence (SIF) which is directly linked to plant photosynthesis. We test the robustness of our approach across different vegetation types using both field measurements and synthetic data generated by the coupled fluorescence-photosynthesis model (SCOPE). The NNs show good performance in emulating SCOPE model in predicting carbon fluxes (R²>0.93) and SIF (R²=0.94). The partitioned fluxes agree well with the output of the standard methods on daily and weekly scales, while showing small systematic difference on the (half-)hourly scale that could be attributed to the light inhibition of leaf respiration. More importantly, NNs can accurately retrieve the SIF-GPP relationship at leaf and canopy scales and simulate the linearization of SIF-GPP relationship when scaling from leaf to canopy level. It can also reveal the response of the SIF-GPP relations to environmental factors. And we find that for all tested sites, relative humidity and the ratio of diffuse radiation affect the relationship between SIF and GPP. Therefore, the NN model is an effective way to partition NEE and to improve the understanding of the ecophysiological response of SIF-GPP relationship.