Estimating gross carbon dioxide fluxes by eddy covariance net ecosystem exchange measurements and machine learning methods
Abstract
Eddy covariance (EC) technique is a powerful tool for assessing CO2 and energy exchanges between land ecosystems and atmosphere. In the last years, EC importance increased and now there are more than 700 monitoring sites globally distributed. EC provides direct measurements of the net ecosystem exchange (NEE) of CO2, that is the net balance between the CO2 uptake by the gross primary production (GPP) and CO2 release by ecosystem respiration (RECO). However, GPP and RECO are not directly measured but they are retrieved from NEE applying partitioning methods. The most popular and widely used in the FLUXNET community are the ones named as "Nighttime" (by Reichstein et al., 2005) and "Daytime" (by Lasslop et al., 2010). These methods apply non-linear relationships fitted on NEE measurements for estimating RECO (by Lloyd & Taylor equation) and GPP (by Michaelis & Menten equation). Few drivers are used (air temperature for Lloyd & Taylor, incoming radiation and vapor pressure deficit for Michaelis & Menten) and the functional relationships are prescribed. These approaches are based on strong knowledge assumption and allow a large applicability but have also some limitations in particular when additional drivers play an important effect on fluxes variability or when functional relationships deviate from the prescribed ones. In this experiment we investigated the feasibility to use machine learning (ML) algorithms for CO2 fluxes partitioning. ML experiment used a more comprehensive drivers dataset and no prescribed relationships. Two independent machine learning methods have been developed: 1) Random Forest have been trained for estimating RECO; b) Artificial Neural Network have been designed for directly estimate GPP and RECO starting from NEE data. In general, results show high consistency between partitioned fluxes using the standard methods and the output by ML in most of the study sites. However, differences among methods are emerging for water limited sites/years or in the diurnal cycle of gross CO2 fluxes and could be linked to interaction of drivers or to the inhibition of leaf respiration in daytime condition, differently accounted in the modeled output.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2018
- Bibcode:
- 2018AGUFM.B21D..05T
- Keywords:
-
- 0414 Biogeochemical cycles;
- processes;
- and modeling;
- BIOGEOSCIENCESDE: 0426 Biosphere/atmosphere interactions;
- BIOGEOSCIENCESDE: 0428 Carbon cycling;
- BIOGEOSCIENCESDE: 0438 Diel;
- seasonal;
- and annual cycles;
- BIOGEOSCIENCES