Towards global empirical upscaling of FLUXNET eddy covariance observations: validation of a model tree ensemble approach using a biosphere model

By Martin Jung, Markus Reichstein, A. Bondeau. 2009

700 citations recorded by the source corpus.

Abstract. Global, spatially and temporally explicit estimates of carbon and water fluxes derived from empirical up-scaling eddy covariance measurements would constitute a new and possibly powerful data stream to study the variability of the global terrestrial carbon and water cycle. This paper introduces and validates a machine learning approach dedicated to the upscaling of observations from the current global network of eddy covariance towers (FLUXNET). We present a new model TRee Induction ALgorithm (TRIAL) that performs hierarchical stratification of the data set into units where particular multiple regressions for a target variable hold. We propose an ensemble approach (Evolving tRees with RandOm gRowth, ERROR) where the base learning algorithm is perturbed in order to gain a diverse sequence of different model trees which evolves over time. We evaluate the efficiency of the model tree ensemble (MTE) approach using an artificial data set derived from the Lund-Potsdam-Jena managed Land (LPJmL) biosphere model. We aim at reproducing global monthly gross primary production as simulated by LPJmL from 1998–2005 using only locations and months where high quality FLUXNET data exist for the training of the model trees. The model trees are trained with the LPJmL land cover and meteorological input data, climate data, and the fraction of absorbed photosynthetic active radiation simulated by LPJmL. Given that we know the "true result" in the form of global LPJmL simulations we can effectively study the performance of the MTE upscaling and associated problems of extrapolation capacity. We show that MTE is able to explain 92% of the variability of the global LPJmL GPP simulations. The mean spatial pattern and the seasonal variability of GPP that constitute the largest sources of variance are very well reproduced (96% and 94% of variance explained respectively) while the monthly interannual anomalies which occupy much less variance are less well matched (41% of variance explained). We demonstrate the substantially improved accuracy of MTE over individual model trees in particular for the monthly anomalies and for situations of extrapolation. We estimate that roughly one fifth of the domain is subject to extrapolation while MTE is still able to reproduce 73% of the LPJmL GPP variability here. This paper presents for the first time a benchmark for a global FLUXNET upscaling approach that will be employed in future studies. Although the real world FLUXNET upscaling is more complicated than for a noise free and reduced complexity biosphere model as presented here, our results show that an empirical upscaling from the current FLUXNET network with MTE is feasible and able to extract global patterns of carbon flux variability.

Subfields: Atmospheric Science, Global and Planetary Change

View Paper: https://www.semanticscholar.org/paper/2b8c11b14da5dde6a4293ffa1a326114487c8bd5

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Atmospheric Science

In Global and Planetary Change

Citations in the corpus

The Canadian Earth System Model version 5 (CanESM5.0.3) (2019, background): “Figures 12 and 13 compare the geographical distribution and zonal averages of gross primary productivity (GPP) and latent and sensible heat fluxes over land with observationbased estimates from Jung et al. (2009).”
Disentangling climatic and anthropogenic controls on global terrestrial evapotranspiration trends (2015, background): “…Miralles et al (2011) CSIRO Modified Penman–Monteith 0.5°×0.5° SILO 1984–2005 Zhang et al (2010b) MPI Empirically derived fromFLUXNET 0.5°×0.5° GPCC 1982–2008 Jung et al (2009) NTSG Modified Penman–Monteith 0.5°×0.5° GPCC 1983–2006 Zhang et al (2010a) Diagnostic ET PRUNI (3 sets of…”
Predicting carbon dioxide and energy fluxes across global FLUXNET sites with regression algorithms (2016, methodology): “The splits are determined as described in Jung et al. (2009), but a certain fraction of data (default is one third) is randomly removed before”
Parameter Analysis and Estimates for the MODIS Evapotranspiration Algorithm and Multiscale Verification (2019, methodology): “We used a total of 175 flux sites from FLUXNET (Baldocchi et al., 2001; Jung et al., 2009) to identify the parameters to which theMOD16 algorithmwasmost sensitive so that we could optimize these parameters.”